CN114225662B - Hysteresis model-based flue gas desulfurization and denitrification optimal control method - Google Patents

Abstract

The invention relates to a flue gas desulfurization and denitrification optimization control method based on a hysteresis model, which comprises the steps of establishing a simulation model of a desulfurization and denitrification system of a power plant through dynamic simulation software according to a modularized modeling and control strategy; establishing a first lag time prediction model for the time delay of determining the existence of the PH value of the absorption tower in the desulfurization system, and establishing a second lag time prediction model for the time delay of determining the existence of the concentration of the nitrogen oxides in the flue gas at the inlet of the denitration system through a flue gas online monitoring device CEMS; predicting the sulfur dioxide concentration at a desulfurization outlet and the nitrogen oxide concentration at a denitration inlet in a support vector machine model; controlling the slurry spraying amount according to the sulfur dioxide concentration predicted value and the first lag time predicted model, and controlling the ammonia spraying amount according to the nitrogen oxide concentration predicted value and the second lag time predicted model; and issuing the control parameters to a simulation model of the desulfurization and denitrification system for intelligent diagnosis.

Description

An optimization control method for flue gas desulfurization and denitrification based on hysteresis model

Technical field

The invention belongs to the technical field of desulfurization and denitrification, and specifically relates to an optimization control method for flue gas desulfurization and denitrification based on a hysteresis model.

Background technique

With the rapid development of various industries, the resulting air pollution has become more and more serious. The emissions of sulfur dioxide and nitrogen oxides have remained high. The resulting problems of air pollution and acid rain are also very serious, and thermal power plants are It is one of the main sources of sulfur dioxide and nitrogen oxide emissions, so it is urgent to control the emissions of sulfur dioxide and nitrogen oxides from power plants. The State Council has also proposed a plan to strengthen pollutant emission reduction and continue to promote desulfurization and denitrification in the power industry. It requires newly-built coal-fired units to fully implement desulfurization and denitrification to achieve emission standards. Active coal-fired units that have not yet installed desulfurization and denitrification facilities must have supporting flue gas facilities. Desulfurization and denitrification facilities, units that cannot stably meet emission standards must be modified.

The common process of desulfurization system is limestone-gypsum wet flue gas desulfurization. The whole process mainly includes absorption tower system, limestone slurry preparation system and gypsum dehydration treatment system. The absorption tower is equipped with two outlets, one outlet is the gypsum slurry outlet, and the detection is real-time. The PH value is to detect the PH value of the gypsum slurry outlet; the other outlet is the clean flue gas outlet to detect the concentration of the dioxide building, that is, to detect the concentration of sulfur dioxide at the clean flue gas outlet; however, the absorption tower flue gas reaction is a big lag and slow. The desulfurization system is a dynamic process, and the desulfurization system is a complex control system. The conventional PID control strategy sets the pH value or adjusts the slurry spray volume based on experience. It is difficult to accurately control the limestone slurry spray volume, and the slurry pH value is difficult to effectively control and Make sure the value is within the valid range.

A common process in the denitrification system is SCR denitrification. The main influencing factor is the amount of ammonia. Too little ammonia will lead to incomplete reaction, causing the outlet nitrogen oxide concentration to exceed the standard. If the amount of ammonia is too much, unreacted ammonia will be discharged from the system with the flue gas. , causing air pollution and blockage of downstream equipment; the efficient and stable operation of the denitrification system is the key to achieving the standard of nitrogen oxide emission concentration in the flue gas of power plants. Due to the nonlinear and hysteresis characteristics of the denitrification system, traditional PID control technology is difficult to maintain the outlet of the denitrification system. The concentration of nitrogen oxides is stable, resulting in excessive fluctuations in the concentration of nitrogen oxides at the outlet. On the one hand, the concentration of nitrogen oxides at the outlet frequently exceeds the standard. On the other hand, in order to ensure the compliance rate of the concentration of nitrogen oxides at the outlet, it is necessary to increase the amount of ammonia sprayed. The cost of ammonia injection in the denitrification system increases accordingly.

How to solve the control problems caused by the hysteresis characteristics of the desulfurization and denitrification system of power plants, ensure effective control of the nitrogen oxide concentration at the desulfurization outlet, reduce the fluctuation of nitrogen oxide concentration at the denitrification outlet, accurately control the slurry spray volume, and reduce the cost of ammonia spraying in the denitrification system is an important issue for power plants An important direction for the operation and regulation of desulfurization and denitrification systems.

Based on the above technical problems, it is necessary to design a new optimization control method for flue gas desulfurization and denitrification based on the hysteresis model.

Contents of the invention

The technical problem to be solved by the present invention is to provide an optimized control method for flue gas desulfurization and denitrification based on a hysteresis model, which can truly simulate the actual working scene of desulfurization and denitrification of a power plant, and can intuitively display and learn the actual operation status of desulfurization and denitrification of a power plant and carry out systematic Adjustment of parameters.

The technical solution adopted by the present invention is:

An optimized control method for flue gas desulfurization and denitrification based on a hysteresis model, which includes the following steps: Step 1: Construct a model of each component of the desulfurization and denitrification of the power plant, build a control system based on the on-site control strategy, and establish a complete simulation model of the desulfurization and denitrification system of the power plant ; Step 2: Establish a first lag time prediction model for the delay in measuring the pH value of the absorption tower in the desulfurization system of the power plant, and determine the delay in measuring the concentration of flue gas nitrogen oxides through the flue gas online monitoring device CEMS at the entrance of the denitrification system of the power plant. Establish a second lag time prediction model; Step 3: Predict the sulfur dioxide in the flue gas at the outlet of the desulfurization system of the power plant and the flue gas nitrogen oxide concentration at the inlet of the denitrification system; Step 4: Predict the first lag time based on the predicted value of sulfur dioxide concentration The model controls the slurry spray amount of the desulfurization system, and controls the ammonia spray amount of the denitrification system based on the nitrogen oxide concentration prediction value combined with the second lag time prediction model; Step 5: Combine the slurry spray amount and the ammonia spray amount The control parameters are sent to the power plant desulfurization and denitrification system simulation model for intelligent diagnosis.

Further, in step 1, after the dynamic simulation software is used to model each component of the power plant desulfurization and denitrification based on the modular modeling method, the corresponding control system is built based on the on-site control strategy to establish a complete power plant desulfurization and denitrification system simulation model. Specifically include:

The power plant desulfurization system selects a limestone-gypsum wet desulfurization system, which at least includes a flue gas system, an absorption tower system, a limestone slurry preparation system, a gypsum slurry dehydration system, a wastewater treatment system and an electrical system; the power plant denitrification system selects the SCR method Flue gas denitration system, which at least includes flue gas system, SCR reactor system, sonic soot blowing system, and liquid ammonia storage and supply system;

The dynamic simulation software is based on the mass conservation, momentum conservation and energy conservation equations. During the modeling process, according to the process flow of the limestone-gypsum wet desulfurization system and the SCR flue gas denitrification system, the corresponding component modules are selected from the model library and connected. , input the initial data and complete the model construction of the power plant desulfurization and denitrification system;

Based on the on-site control strategy, we build an analog control system, a sequence control system and a logic control system, and use basic algorithm modules for configuration to achieve the same functions as the actual control system, and establish a complete simulation model of the power plant desulfurization and denitrification system.

Furthermore, the simulation model of the power plant desulfurization and denitrification system also includes:

During the model development and debugging process, the physical data collected by the desulfurization and denitrification system of the actual power plant are compared with the virtual data obtained based on the desulfurization and denitrification simulation model of the power plant to determine whether the error exceeds the threshold. If it exceeds the threshold, cluster learning will be used to detect errors with larger errors. Classify the virtual data, combine the corresponding historical data as input, perform error learning through the neural network, output the correction coefficient to correct the error data of the virtual data, and perform virtual and real fusion of the corrected virtual data and physical data to generate a verified power plant Desulfurization and denitrification simulation model.

Further, in step 2, using change point detection, time window sliding, correlation analysis and machine learning models to establish the first lag time prediction model for the delay in measuring the pH value of the absorption tower in the power plant desulfurization system includes: using change point Detection, time window sliding and correlation analysis methods are used to establish the slurry pH value response lag time identification algorithm process and a machine learning model is used to establish the first lag time prediction model;

The slurry pH value response lag time identification algorithm process includes:

Select the working condition in which the sulfur dioxide concentration value at the outlet of the absorption tower changes after the PH value of the absorption tower slurry is adjusted as the identification object;

Divide the time window Δt into two equally spaced time windows Δt _i1 and Δt _i2 , gradually slide forward on the time axis, and calculate the average difference in sulfur dioxide concentration within the two time windows. If it exceeds the set threshold, then is the working condition change point t _i . If it is less than the set threshold, continue to slide the time window forward until the working condition change point is detected or the time window slides to the cut-off time point;

Based on the working condition change point t _i and the time window Δt, the slurry pH value time series and sulfur dioxide concentration time series from the beginning of the working condition change to the end of the time window are obtained respectively;

Step by step move the sulfur dioxide concentration time series forward, set the maximum number of moving steps k, obtain a new sulfur dioxide concentration series through the forward movement, and construct the sulfur dioxide concentration time lag matrix V;

Calculate the Pearson correlation coefficient r of the slurry pH value time series and each column in the matrix V. The delay time corresponding to the maximum correlation coefficient is the PH value response lag time t ₁ under the working condition;

Using machine learning models to establish the first lag time prediction model includes:

After collecting the original data characteristics in the power plant desulfurization system and performing preprocessing, they are substituted into the slurry pH value response lag time identification algorithm process to perform pH value delay identification, and the relationship between the delay time and different operating data characteristics is obtained; the original data characteristics At least include: the load of the boiler, the air supply volume of the boiler, the flow rate of the limestone slurry, the input amount of the limestone slurry, the sulfur dioxide content, the calcium carbonate content in the limestone, and the distance data characteristics from the absorption tower to the PH measurement point;

The identified operating data features that can cause changes in pH value are converted into features with more working condition characteristics through feature transformation, reducing the correlation between data features. By normalizing the transformed data features, eliminating The impact of dimensions;

Use the correlation analysis method to perform correlation analysis on the original operating data features to obtain the correlation coefficient between each operating data feature and the PH value response lag time. The higher the correlation coefficient, the more relevant the data features and lag time are;

The feature fusion method is used to fuse the operating data features according to the level of the correlation coefficient to form a new fusion feature. The original operating data features and the new fused features are used as sample data, and the training set in the sample data is input to the machine according to the preset ratio. A first lag time prediction model under different operating data changing conditions is established in the learning model; through the first lag time prediction model, the pH value response lag time can be calculated according to different operating data characteristics.

Further, in step 2, change point detection, time window sliding, correlation analysis and machine learning models are used to establish the time delay for measuring the concentration of flue gas nitrogen oxides through the flue gas online monitoring device CEMS at the entrance of the power plant denitration system. The second lag time prediction model includes: using change point detection, time window sliding and correlation analysis methods to establish a CEMS measurement lag time identification algorithm process and using a machine learning model to establish a second lag time prediction model;

The CEMS measurement lag time identification algorithm process includes:

After the inlet nitrogen oxide concentration changes, the working condition in which the CEMS measurement value changes is selected as the identification object; the nitrogen oxide concentration is measured through the heating duct and the analysis cabinet, the flow of flue gas in the heating duct and the concentration measurement in the analysis cabinet. There is a certain time lag;

Divide the time window Δt′ into two equally spaced time windows Δt _i1 ′ and Δt _i2 ′, gradually slide forward on the time axis, and calculate the average difference in CEMS measurement values within the two time windows. If it exceeds the set threshold , then this moment is the working condition change point t _i ′. If it is less than the set threshold, continue to slide the time window forward until the working condition change point is detected or the time window slides to the cut-off time point;

Based on the working condition change point t _i ′ and the time window Δt′, the nitrogen oxide concentration value time series and the CEMS measurement value time series from the beginning of the working condition change to the end of the time window are obtained respectively;

Gradually move the CEMS measurement value time series forward, set the maximum number of moving steps k, obtain a new CEMS measurement value sequence through the forward movement, and construct the CEMS measurement value time lag matrix V′;

Calculate the time series of nitrogen oxide concentration values and the Pearson correlation coefficient r′ of each column in the matrix V′. The delay time corresponding to the maximum correlation coefficient is the nitrogen oxide concentration measurement lag time t ₂ under the working condition;

Using machine learning models to establish a second lag time prediction model includes:

After collecting the original data characteristics in the power plant denitrification system and preprocessing, they are substituted into the CEMS measurement lag time identification algorithm process for delay identification, and the relationship between the delay time and different operating data characteristics is learned; the original data characteristics at least include: boiler load, coal burning Type, coal supply amount, combustion temperature, air volume and flue gas volume;

The identified operating data features that can cause changes in nitrogen oxide concentration are converted into features with more operating condition characteristics through feature transformation, reducing the correlation between data features, and normalizing the transformed data features. , eliminate the influence of dimensions;

The correlation analysis method is used to perform correlation analysis on the original operating data characteristics, and the correlation coefficient between each operating data characteristic and the nitrogen oxide concentration value measurement lag time is obtained. The higher the correlation coefficient, the more relevant the data characteristics and the lag time are;

The feature fusion method is used to fuse the operating data features according to the level of the correlation coefficient to form a new fusion feature. The original operating data features and the new fused features are used as sample data, and the training set in the sample data is input to the machine according to the preset ratio. In the learning model, a second lag time prediction model under different operating data changing conditions is established; through the second lag time prediction model, the CEMS measurement lag time can be calculated based on different operating data characteristics.

Further, the machine learning model selects the XGBoost model, which is an ensemble learning algorithm using the boosting method. The base learner selects the CART decision tree, and k CART functions {f ₁ , f ₂ ,..., f _k } are added to form an ensemble. Tree model; the objective function of the model consists of a loss function and a regular term. The loss function is approximated by a second-order Taylor expansion; and the key parameters are tuned to improve the accuracy of model prediction. The key parameters include the maximum depth of the tree, sub-tree Samples, proportion of randomly sampled columns for each tree, minimum leaf node sample weight and learning rate;

Among them, the construction of the model starts from the root node, sorts the training set data according to each data feature, uses the greedy method to calculate the profit of each feature, selects the feature with the largest profit as the split feature, and maps the training set data to the corresponding Leaf nodes, the generated leaf nodes are recursed until the restriction conditions are reached, the decision tree generation process ends, and then the weights of the decision tree leaf nodes are calculated from the first-order and second-order derivatives of the loss function, which are used as the fitting target of the next tree. , repeat the recursive execution until the conditions are met, and the model establishment ends.

Further, in step 3, after collecting the historical operating parameters of the power plant desulfurization system and denitrification system, select operating parameters that are strongly related to the power plant desulfurization and input them into the first support vector machine model constructed to analyze the flue gas at the outlet of the power plant desulfurization system. Prediction of sulfur dioxide concentration, including:

Use the collected historical operating parameters of the power plant desulfurization system as sample data, and conduct correlation analysis on the sample data to remove sample data whose correlation with the sulfur dioxide concentration in the flue gas at the outlet of the limestone-gypsum wet desulfurization system is less than the preset value. The remaining sample data is used as operating data that is strongly related to the desulfurization system; the historical operating parameters of the desulfurization system at least include the sulfur dioxide concentration at the inlet, the nitrogen oxide concentration, the unit load, the limestone slurry circulation pump current, the slurry supply volume, and the flue gas at the absorption tower outlet. Sulfur dioxide concentration and slurry pH;

Perform data preprocessing on the operating data that is strongly related to the desulfurization system, and use the preprocessed data to construct the first support vector machine model;

Collect real-time operating data related to power plant desulfurization and input it into the constructed first support vector machine model to obtain the predicted value of sulfur dioxide concentration in the flue gas at the outlet of the power plant desulfurization system;

Wherein, the data preprocessing includes: filling in missing values and outliers and normalizing the operating data that is strongly related to the desulfurization system to obtain a preprocessed desulfurization data sequence, recorded as F=[f ₁ , f ₂ , f ₃ ,…, f _n ], f _i is the desulfurization data at the i-th time point in the processed desulfurization data sequence;

The desulfurization data sequence F is subjected to wavelet threshold denoising processing, and the noisy data with noise is decomposed by wavelet to obtain the real data information, which is recorded as P = [p ₁ , p ₂ , p ₃ ,..., p _m ], p _i is the desulfurization data at the i-th time point in the real desulfurization data sequence.

Further, in step 3, operating parameters that are strongly related to the denitrification of the power plant are selected and input into the second support vector machine model constructed to predict the concentration of flue gas nitrogen oxides at the entrance of the denitrification system of the power plant, specifically including:

The collected historical operating parameters of the power plant denitrification system are used as sample data, and the Pearson correlation coefficient is used to calculate the correlation between the sample data and the flue gas nitrogen oxide concentration at the entrance of the power plant denitrification system. Data combinations with high correlation are selected based on the correlation. As operating data strongly related to the denitration system; the historical operation data of the denitration system at least includes ammonia injection mass flow, boiler load, SCR inlet flue gas temperature, SCR inlet flue gas oxygen content, SCR inlet nitrogen oxide concentration, and SCR denitration efficiency;

Perform data preprocessing on the operating data that is strongly related to the denitrification system, and use the preprocessed data to build a second support vector machine model;

Collect real-time operating data related to power plant denitration and input it into the constructed second support vector machine model to obtain the predicted value of flue gas nitrogen oxide concentration at the entrance of the power plant denitration system;

Among them, the data with strong correlation and extremely strong correlation are selected as the data with high correlation with the denitrification system. The calculation formula is:

X is the input sample data feature, Y is the concentration of nitrogen oxides at the inlet, cov(X,Y) represents the covariance of _X and _Y ; σ The correlation coefficient between When 0.2≤ρ<0.4, it is called weak correlation; when 0.0≤ρ<0.2, it is called extremely weak correlation or no correlation.

Further, constructing the first support vector machine model and the second support vector machine model includes: using the cuckoo optimization method to determine the optimal support vector machine parameters: initializing the parameters of the cuckoo optimization algorithm, and based on the parameters of the cuckoo optimization method , Levi’s flight searches for the nest position with step size adaptive and dynamic adjustment: i=1,2,...,n; where, x _i ^(t+1) is the nest position of the i-th bird's nest in the t-th generation; a is the step control amount, used to control the search range of the step, which obeys Normal distribution; L(λ) is the Levy random walk path; the step size adaptive dynamic adjustment strategy is:

step _i = step _min + (step _max - step _min )d _i

Among them, step _i is the current search step, step _max is the maximum value of the step, step _min is the minimum value of the step, n _i is the position of the i-th bird's nest, and n _best is the bird's nest corresponding to the current minimum fitness. Position, d _max is the maximum value of the distance between the nest and other nests corresponding to the current minimum fitness;

Use the training set in the preprocessed data to train the support vector machine model, calculate the fitness of each bird's nest position, and retain the bird's nest corresponding to the minimum fitness to the next iteration;

Determine whether the minimum fitness meets the preset termination conditions. If it does, the nest position of the nest corresponding to the minimum fitness is the determined optimal support vector machine parameter. If not, remove the nests with the highest fitness and start again. Adjust the position of the nest;

According to the determined optimal support vector machine parameters, train the support vector machine model: establish a support vector machine training program based on the kernel function, input variables and output variables form a mapping relationship through the support vector machine model, and use the support vector machine training program to train the training samples The data is learned and trained, and N support vectors _Xi ^* are obtained after learning and training, i=0,1,...,N, thus forming a support vector machine model:

_Among ^them _{, _} is the input preprocessed desulfurization data or denitrification data, Y(X) represents the predicted value of sulfur dioxide concentration of the support vector of the power plant desulfurization system or the predicted value of nitrogen oxide concentration of the support vector of the denitrification system, K(·) represents the predicted value of the support vector machine Kernel function, the kernel function selects one of Gaussian function, polynomial function, linear function and radial basis function.

Further, step 5 includes in detail: after inputting the slurry spray volume control parameters, the ammonia spray volume control parameters and the relevant configuration parameters of the power plant desulfurization and denitrification system operation into the power plant desulfurization and denitrification system simulation model, through the set The expert diagnosis module compares the obtained real-time operating parameters of the power plant's desulfurization and denitrification system with the simulation result data of the simulation model, obtains the deviation, and implements a pre-alarm based on whether the deviation exceeds the preset threshold;

Among them, the expert diagnosis module is equipped with an intelligent diagnosis strategy, which uses preset logic to determine the relevant operating status, data deviation, and pre-alarm information, comprehensively outputs preliminary diagnosis result information, and then calls the expert database knowledge information for comparison to analyze the intelligent diagnosis Whether the conclusion information drawn by the strategy is associated with or consistent with the knowledge information of the expert database, and outputs diagnostic analysis results, operation guidance or task orders; the knowledge information of the expert database includes stored preset knowledge and information about abnormal faults that have occurred; The pre-alarm includes the parameter exceeding the preset threshold, the time when the parameter exceeds the preset threshold, and abnormal fault information.

The positive effects of the present invention are:

(1) The present invention uses dynamic simulation software to build a model of each desulfurization and denitrification component of the power plant based on the modular modeling method, and builds a corresponding control system based on the on-site control strategy to establish a complete power plant desulfurization and denitrification system simulation model, which can truly simulate the power plant. Actual working scenarios of desulfurization and denitrification, and can visually display and learn the actual operation of desulfurization and denitrification of power plants and adjust system parameters;

(2) In the present invention, during the model development and debugging process, the physical data collected by the desulfurization and denitrification system of the actual power plant are compared with the virtual data obtained based on the desulfurization and denitrification simulation model of the power plant to determine whether the error exceeds the threshold. Class learning classifies virtual data with large errors, combines the corresponding historical data as input, performs error learning through the neural network, outputs correction coefficients to correct the error data of the virtual data, and compares the corrected virtual data and physical data between virtual and real data. Fusion generates a verified power plant desulfurization and denitrification simulation model. After comparison and analysis of virtual and real data, the neural network can be used to correct errors, improve the precision and accuracy of the power plant desulfurization and denitrification simulation model, and provide predictive control for subsequent desulfurization and denitrification systems. Do the basics well;

(3) The present invention uses change point detection, time window sliding, correlation analysis and machine learning models to respectively establish a first lag time prediction model for the time delay in measuring the pH value of the absorption tower in the power plant desulfurization system, and to predict the entrance of the power plant denitrification system. The second lag time prediction model is established based on the time delay in measuring the concentration of nitrogen oxides in the flue gas through the flue gas online monitoring device CEMS. It can analyze, calculate and establish a prediction model for the lag time existing in the desulfurization and denitrification system of the power plant, and quickly and effectively obtain the corresponding information. The lag affects data characteristics and lag time;

(4) The present invention collects the historical operating parameters of the power plant desulfurization system and denitrification system, selects the operating parameters that are strongly related to the power plant desulfurization and inputs them into the constructed first support vector machine model to analyze the sulfur dioxide in the flue gas at the outlet of the power plant desulfurization system. Predict the concentration of flue gas nitrogen oxides at the entrance of the denitration system of the power plant, and select operating parameters that are strongly related to the denitrification of the power plant and input them into the second support vector machine model to predict the concentration of flue gas nitrogen oxides at the entrance of the denitrification system of the power plant. Predict the sulfur dioxide concentration value in the flue gas and predict the nitrogen oxide concentration value at the entrance of the denitrification system to improve the accuracy of the prediction value;

(5) The present invention controls the slurry spray volume of the desulfurization system based on the predicted value of sulfur dioxide concentration combined with the first lag time prediction model, and controls the slurry spray volume of the denitrification system based on the predicted value of nitrogen oxide concentration combined with the second lag time prediction model. Controlling the amount of ammonia sprayed can effectively control the concentration of nitrogen oxides at the desulfurization outlet in combination with the first lag time, accurately control the slurry spray volume, control the slurry pH value within the effective range, and at the same time reduce the nitrogen at the denitrification outlet in combination with the second lag time. The concentration of oxides fluctuates, and the amount of ammonia sprayed can be precisely controlled to reduce the cost of ammonia spraying in the denitrification system;

(6) The present invention performs intelligent diagnosis by sending the control parameters of the slurry spray amount and the ammonia spray amount to the simulation model of the power plant desulfurization and denitrification system, and sets expert database knowledge information and intelligent diagnosis strategies in the expert diagnosis module. Compare the system's real-time operating parameters and simulation data to achieve alarms and diagnosis, and output diagnostic analysis results, operation guidance or task orders, realizing effective processing and diagnostic analysis of power plant desulfurization and denitrification system data.

Description of the drawings

Figure 1 is a flow chart of the method of the present invention;

Figure 2 is a process flow diagram of the limestone-gypsum wet desulfurization system of the present invention;

Figure 3 is a schematic diagram of the inlet CEMS device in the SCR flue gas denitration system of the present invention.

Detailed ways

As shown in Figure 1, this embodiment 1 provides a flue gas desulfurization and denitrification optimization control method based on a hysteresis model. The flue gas desulfurization and denitrification optimization control method includes:

After using dynamic simulation software to build a model of each component of the power plant desulfurization and denitrification based on the modular modeling method, a corresponding control system was built based on the on-site control strategy to establish a complete power plant desulfurization and denitrification system simulation model; the power plant desulfurization system uses limestone-gypsum Wet desulfurization system, the power plant denitrification system adopts SCR flue gas denitrification system;

Change point detection, time window sliding, correlation analysis and machine learning models were used to establish a first lag time prediction model for the time delay in measuring the pH value of the absorption tower in the desulfurization system of the power plant, and online monitoring of flue gas at the entrance of the denitrification system of the power plant. The device CEMS detects the delay in the concentration of nitrogen oxides in the flue gas and establishes a second lag time prediction model;

After collecting the historical operating parameters of the power plant desulfurization system and denitrification system, select the operating parameters that are strongly related to the power plant desulfurization and input them into the first support vector machine model to predict the sulfur dioxide concentration in the flue gas at the outlet of the power plant desulfurization system, and select The operating parameters that are strongly related to the denitrification of the power plant are input into the second support vector machine model constructed to predict the concentration of flue gas nitrogen oxides at the entrance of the denitrification system of the power plant;

The slurry spray volume of the desulfurization system is controlled based on the predicted sulfur dioxide concentration value combined with the first lag time prediction model, and the spray amount of the denitrification system is controlled based on the predicted value of nitrogen oxide concentration combined with the second lag time prediction model. The amount of ammonia is controlled;

The control parameters of the slurry spray amount and the ammonia spray amount are sent to the simulation model of the power plant desulfurization and denitrification system for model verification and intelligent diagnosis.

It should be noted that the slurry spray volume of the desulfurization system is controlled based on the sulfur dioxide concentration prediction value combined with the first lag time prediction model: the set value is the output target value, which in the actual system is the sulfur dioxide in the flue gas at the outlet. Concentration; the operable variable is the spray volume of the slurry. Its size can be adjusted by the controller so that it acts on the target object so that the output reaches the desired value. During the control process, the lag effect of the pH value measurement is considered, and the pH needs to be carried out a certain time in advance. value detection in order to open the slurry control valve in advance to control the slurry spray volume. And control the ammonia injection amount of the denitrification system according to the predicted nitrogen oxide concentration value combined with the second lag time prediction model: calculate the first ammonia injection amount under the current operating conditions based on the inlet nitrogen oxide concentration predicted value, and send it to The denitrification system measures the actual measured value of the outlet nitrogen oxide concentration, deviates the actual measured value from the outlet nitrogen oxide concentration set value and then inputs it into the controller to obtain the second ammonia injection amount, and sends it to the denitrification system. The denitrification system determines the deviation according to the first The ammonia injection amount and the second ammonia injection amount control the ammonia injection amount of the denitrification reactor. During the control process, the hysteresis effect of the nitrogen oxide concentration value measurement needs to be considered. The nitrogen oxide concentration value needs to be measured a certain time in advance in order to control the injection in advance. Ammonia amount.

In this embodiment, after the dynamic simulation software is used to model each component of the desulfurization and denitrification of the power plant according to the modular modeling method, the corresponding control system is built according to the on-site control strategy, and a complete simulation model of the desulfurization and denitrification system of the power plant is established, including :

The power plant desulfurization system selects a limestone-gypsum wet desulfurization system, which at least includes a flue gas system, an absorption tower system, a limestone slurry preparation system, a gypsum slurry dehydration system, a wastewater treatment system and an electrical system; the power plant denitrification system selects the SCR method Flue gas denitration system, which at least includes flue gas system, SCR reactor system, sonic soot blowing system, and liquid ammonia storage and supply system;

The dynamic simulation software is based on the mass conservation, momentum conservation and energy conservation equations. During the modeling process, according to the process flow of the limestone-gypsum wet desulfurization system and the SCR flue gas denitrification system, the corresponding component modules are selected from the model library and connected. , input the initial data and complete the model construction of the power plant desulfurization and denitrification system;

Based on the on-site control strategy, we build an analog control system, a sequence control system and a logic control system, and use basic algorithm modules for configuration to achieve the same functions as the actual control system, and establish a complete simulation model of the power plant desulfurization and denitrification system.

The present invention uses dynamic simulation software to model each component of the desulfurization and denitrification of the power plant based on the modular modeling method, and then builds a corresponding control system based on the on-site control strategy to establish a complete simulation model of the desulfurization and denitrification system of the power plant, which can truly simulate the desulfurization and denitrification of the power plant. Actual working scenarios, and can intuitively display and learn the actual operation conditions of desulfurization and denitrification of power plants and adjust system parameters.

As shown in Figure 2, it should be noted that the main process sequence of desulfurization and denitrification is denitrification-dust removal-desulfurization. The process flow of the limestone-gypsum wet desulfurization system used in this patent includes: after the raw flue gas from the tail of the boiler passes through the denitrification system and dry electrostatic precipitator, it enters the absorption tower under the action of the induced draft fan. The entire absorption tower absorbs Integrated with oxidation, the upper part is the absorption area and the lower part is the oxidation area. The oxidation fan located at the bottom of the tower continuously blows in oxidized air, while the slurry circulation pump continuously pumps the limestone slurry from the bottom of the absorption tower to the upper spray layer. The flue gas entering the absorption tower contacts the limestone slurry sprayed from the upper part in reverse direction. After the fully reacted clean flue gas passes through the demister at the top of the absorption tower to remove the mist droplets carried in the gas, it is directly discharged from the top of the absorption tower into the internal space of the Heller air cooling system, and finally along with the water vapor in the cooling tower Emitted to the atmosphere. As the reaction proceeds, the density of the gypsum slurry in the absorption tower slurry continues to increase. When the density reaches a certain value, the gypsum discharge pump is started and the gypsum slurry is sent to the gypsum dehydration system. After dehydration, the by-product gypsum is formed, and the remaining slurry is returned to the system. Medium recycling thus improves the utilization rate of desulfurization absorbent.

In the limestone-gypsum wet flue gas desulfurization system process, the main processes are: chemical reactions such as dissolution and oxidation of sulfur dioxide and limestone slurry, and the process of obtaining the by-product gypsum. The above process is the most important process of the entire process. .

The SCR method flue gas denitrification process includes: the liquid ammonia tanker transports the liquid ammonia into the liquid ammonia storage tank through the unloading compressor. The liquid ammonia in the storage tank enters the liquid ammonia evaporator through its own pressure, and is heated and evaporated by the water bath It is ammonia gas, and then enters the ammonia gas buffer tank to stabilize the pressure and is sent to the SCR reaction zone. Before entering the SCR reactor, the air and ammonia sent by the dilution fan are uniformly mixed, and then introduced into the SCR reactor to participate in the chemical reaction.

In this embodiment, the simulation model of the power plant desulfurization and denitrification system also includes:

During the model development and debugging process, the physical data collected by the desulfurization and denitrification system of the actual power plant are compared with the virtual data obtained based on the desulfurization and denitrification simulation model of the power plant to determine whether the error exceeds the threshold. If it exceeds the threshold, cluster learning will be used to detect errors with larger errors. Classify the virtual data, combine the corresponding historical data as input, perform error learning through the neural network, output the correction coefficient to correct the error data of the virtual data, and perform virtual and real fusion of the corrected virtual data and physical data to generate a verified power plant Desulfurization and denitrification simulation model.

During the model development and debugging process, the present invention can use neural networks to correct errors after comparing and analyzing virtual and real data, thereby improving the precision and accuracy of the desulfurization and denitrification simulation model of the power plant, and providing a basis for predictive control of the subsequent desulfurization and denitrification system. Good foundation.

In this embodiment, the use of change point detection, time window sliding, correlation analysis and machine learning models to establish the first lag time prediction model for the time delay in measuring the pH value of the absorption tower in the power plant desulfurization system includes: using change points Detection, time window sliding and correlation analysis methods are used to establish the slurry pH value response lag time identification algorithm process and a machine learning model is used to establish the first lag time prediction model;

The slurry pH value response lag time identification algorithm process includes:

Select the working condition in which the sulfur dioxide concentration value at the outlet of the absorption tower changes after the PH value of the absorption tower slurry is adjusted as the identification object;

Divide the time window Δt into two equally spaced time windows Δt _i1 and Δt _i2 , gradually slide forward on the time axis, and calculate the average difference in sulfur dioxide concentration within the two time windows. If it exceeds the set threshold, then is the working condition change point t _i . If it is less than the set threshold, continue to slide the time window forward until the working condition change point is detected or the time window slides to the cut-off time point;

Based on the working condition change point t _i and the time window Δt, the slurry pH value time series and sulfur dioxide concentration time series from the beginning of the working condition change to the end of the time window are obtained respectively;

Step by step move the sulfur dioxide concentration time series forward, set the maximum number of moving steps k, obtain a new sulfur dioxide concentration series through the forward movement, and construct the sulfur dioxide concentration time lag matrix V;

Calculate the Pearson correlation coefficient r of the slurry pH value time series and each column in the matrix V. The delay time corresponding to the maximum correlation coefficient is the PH value response lag time t ₁ under the working condition;

Using machine learning models to establish the first lag time prediction model includes:

After collecting the original data characteristics in the power plant desulfurization system and performing preprocessing, they are substituted into the slurry pH value response lag time identification algorithm process to perform pH value delay identification, and the relationship between the delay time and different operating data characteristics is obtained; the original data characteristics At least include: the load of the boiler, the air supply volume of the boiler, the flow rate of the limestone slurry, the input amount of the limestone slurry, the sulfur dioxide content, the calcium carbonate content in the limestone, and the distance data characteristics from the absorption tower to the PH measurement point;

The identified operating data features that can cause changes in pH value are converted into features with more working condition characteristics through feature transformation, reducing the correlation between data features. By normalizing the transformed data features, eliminating The impact of dimensions;

Use the correlation analysis method to perform correlation analysis on the original operating data features to obtain the correlation coefficient between each operating data feature and the PH value response lag time. The higher the correlation coefficient, the more relevant the data features and lag time are;

The feature fusion method is used to fuse the operating data features according to the level of the correlation coefficient to form a new fusion feature. The original operating data features and the new fused features are used as sample data, and the training set in the sample data is input to the machine according to the preset ratio. A first lag time prediction model under different operating data changing conditions is established in the learning model; through the first lag time prediction model, the pH value response lag time can be calculated according to different operating data characteristics.

It should be noted that due to the installation position of the pH measuring element in the desulfurization system, the time required for detection using a special pH detector and the reaction time between limestone slurry and sulfur dioxide will cause a pure lag in the pH value detection. This pure lag makes the measurement The signal cannot reflect changes in the pH value of the absorption liquid in the absorption tower in time, resulting in a time delay in the pH value measured by the electrode of the pH detector.

In this embodiment, the change point detection, time window sliding, correlation analysis and machine learning model are used to establish the time delay for measuring the concentration of flue gas nitrogen oxides through the flue gas online monitoring device CEMS at the entrance of the power plant denitration system. The second lag time prediction model includes: using change point detection, time window sliding and correlation analysis methods to establish a CEMS measurement lag time identification algorithm process and using a machine learning model to establish a second lag time prediction model;

The CEMS measurement lag time identification algorithm process includes:

After the inlet nitrogen oxide concentration changes, the working condition in which the CEMS measurement value changes is selected as the identification object; the measurement of the nitrogen oxide concentration passes through the heating duct and the analysis cabinet, the flow of the flue gas in the heating duct and the concentration in the analysis cabinet Measurement, there is a certain time lag;

Divide the time window Δt′ into two equally spaced time windows Δt _i1 ′ and Δt _i2 ′, gradually slide forward on the time axis, and calculate the average difference in CEMS measurement values within the two time windows. If it exceeds the set threshold , then this moment is the working condition change point t _i ′. If it is less than the set threshold, continue to slide the time window forward until the working condition change point is detected or the time window slides to the cut-off time point;

Based on the working condition change point t _i ′ and the time window Δt′, the nitrogen oxide concentration value time series and the CEMS measurement value time series from the beginning of the working condition change to the end of the time window are obtained respectively;

Gradually move the CEMS measurement value time series forward, set the maximum number of moving steps k, obtain a new CEMS measurement value sequence through the forward movement, and construct the CEMS measurement value time lag matrix V′;

Calculate the time series of nitrogen oxide concentration values and the Pearson correlation coefficient r′ of each column in the matrix V′. The delay time corresponding to the maximum correlation coefficient is the nitrogen oxide concentration measurement lag time t ₂ under the working condition;

Using machine learning models to establish a second lag time prediction model includes:

After collecting the original data characteristics in the denitrification system of the power plant and performing preprocessing, they are substituted into the CEMS measurement lag time identification algorithm process to perform delay identification, and the relationship between the delay time and different operating data characteristics is learned; the original data characteristics at least include: boiler Load, coal type, coal supply amount, combustion temperature, air volume and flue gas volume;

The identified operating data features that can cause changes in nitrogen oxide concentration are converted into features with more operating condition characteristics through feature transformation, reducing the correlation between data features, and normalizing the transformed data features. , eliminate the influence of dimensions;

The correlation analysis method is used to perform correlation analysis on the original operating data characteristics, and the correlation coefficient between each operating data characteristic and the nitrogen oxide concentration value measurement lag time is obtained. The higher the correlation coefficient, the more relevant the data characteristics and the lag time are;

The feature fusion method is used to fuse the operating data features according to the level of the correlation coefficient to form a new fusion feature. The original operating data features and the new fused features are used as sample data, and the training set in the sample data is input to the machine according to the preset ratio. A second lag time prediction model under different operating data changing conditions is established in the learning model; through the second lag time prediction model, the CEMS measurement lag time can be calculated based on different operating data characteristics.

The present invention uses change point detection, time window sliding, correlation analysis and machine learning models to respectively establish a first lag time prediction model for the time delay in measuring the pH value of the absorption tower in the desulfurization system of the power plant, and to predict the smoke passing through the entrance of the denitrification system of the power plant. The gas online monitoring device CEMS establishes a second lag time prediction model for measuring the concentration of nitrogen oxides in the flue gas. It can analyze, calculate and establish a prediction model for the lag time existing in the desulfurization and denitrification system of the power plant, and quickly and effectively learn the corresponding lag effects. Data characteristics and lag times.

As shown in Figure 3, it should be noted that the nitrogen oxide concentration in the denitrification system is measured using a flue gas online monitoring system (CEMS). There will be a certain measurement lag during the CEMS measurement process. The CEMS flue gas online monitoring system extracts gas from the flue through heat pipe extraction and sampling. The gas goes through dust removal, heating, insulation and other links, and is guided to the pretreatment system to remove particulate matter, H ₂ O and corrosive gases, and finally transported To the flue gas analyzer, during this process, it can be seen that the measurement of the concentration of nitrogen oxides in the flue gas must pass through the heating duct and the analysis cabinet. The flow of the flue gas through the heating duct and the measurement of the concentration in the analysis cabinet require a certain amount of time. time, resulting in a certain delay in the CEMS measurement. This measurement delay will have an impact on the subsequent control of the ammonia injection amount of the denitrification system, making it impossible to respond in time during the ammonia injection amount control process, and increasing the ammonia injection amount control. The difficulty is that the concentration of nitrogen oxides at the outlet of the denitrification system will fluctuate more. When the concentration of nitrogen oxides at the inlet changes more rapidly, the measurement lag error will become larger.

It should be noted that data with strong and weak correlations in the original data can be found through data correlation analysis and calculation. However, a lower correlation coefficient does not mean that there is no relationship between the data characteristics and the output volume. It can only indicate that the The degree of linear correlation between data features and output volume is not high, but there may be some nonlinear correlations. The fusion features related to output volume can be studied through data fusion.

The essence of the fusion method based on linear regression and multi-layer perceptron algorithm is to build a basic prediction model y=f(x), and use the learned As a new fusion feature, if the mapping function f(·) is a linear (such as linear regression) function, the fusion feature obtained will also be a linear combination of the original features; if the mapping function f(·) is nonlinear (such as a multi-layer perceptron) function, the obtained fused features will be a nonlinear combination of the original features.

In addition to linear fusion of features, multi-layer perceptrons can also be used to extract nonlinear relationships in original features, and the feature data can be input into the model to fit the nitrogen oxide concentration, thereby obtaining the nonlinear feature fusion feature F _MLP . The nodes of the hidden layer are set to 100, the number of iterations is set to 200 generations, and the activation function uses the relu function. The essence of combining features through exponential fusion is different from the above two methods based on mapping fitting.

The method based on exponential fusion is to combine the original features in the form of exponential power to form a new fusion feature, such as Among them, f ₁ , f ₂ ,...f _k are the filtered original features, and a, b,..., m are the parameters that need to be determined for each feature. By searching different parameter combinations in a parameter space, the fusion feature space is obtained. The correlation coefficient between each index fusion feature and the delay time is calculated in the fusion feature space, and the feature with the largest correlation coefficient is used as the final fusion feature.

In this embodiment, the machine learning model selects the XGBoost model, which is an integrated learning algorithm using the boosting method. The base learner selects the CART decision tree and applies k CART functions {f ₁ , f ₂ ,..., f _k }. Add to form an integrated tree model; the objective function of the model consists of a loss function and a regular term, and the loss function is approximated by a second-order Taylor expansion; and the key parameters are tuned to improve the accuracy of model prediction, and the key parameters include the maximum tree Depth, subsamples, proportion of randomly sampled columns for each tree, minimum leaf node sample weight and learning rate;

Among them, the construction of the model starts from the root node, sorts the training set data according to each data feature, uses the greedy method to calculate the profit of each feature, selects the feature with the largest profit as the split feature, and maps the training set data to the corresponding Leaf nodes, the generated leaf nodes are recursed until the restriction conditions are reached, the decision tree generation process ends, and then the weights of the decision tree leaf nodes are calculated from the first-order and second-order derivatives of the loss function, which are used as the fitting target of the next tree. , repeat the recursive execution until the conditions are met, and the model establishment ends.

In this embodiment, after collecting the historical operating parameters of the desulfurization system and denitrification system of the power plant, the operating parameters that are strongly related to the desulfurization of the power plant are selected and input into the first support vector machine model constructed to analyze the flue gas at the outlet of the desulfurization system of the power plant. Prediction of sulfur dioxide concentration, including:

Use the collected historical operating parameters of the power plant desulfurization system as sample data, and conduct correlation analysis on the sample data to remove samples whose correlation with the sulfur dioxide concentration in the flue gas at the outlet of the limestone-gypsum wet desulfurization system is less than the preset value data, and the remaining sample data is used as operating data strongly related to the desulfurization system; the historical operating parameters of the desulfurization system at least include the sulfur dioxide concentration at the inlet, the nitrogen oxide concentration, unit load, limestone slurry circulation pump current, slurry supply volume, absorption The sulfur dioxide concentration of the flue gas at the tower outlet and the pH value of the slurry;

Perform data preprocessing on the operating data that is strongly related to the desulfurization system, and use the preprocessed data to build the first support vector machine model;

Collect real-time operating data related to power plant desulfurization and input it into the constructed first support vector machine model to obtain the predicted value of sulfur dioxide concentration in the flue gas at the outlet of the power plant desulfurization system.

Wherein, the data preprocessing includes: filling in missing values and outliers and normalizing the operating data that is strongly related to the desulfurization system to obtain a preprocessed desulfurization data sequence, recorded as F=[f ₁ , f ₂ , f ₃ ,…, f _n ], f _i is the desulfurization data at the i-th time point in the processed desulfurization data sequence;

The desulfurization data sequence F is subjected to wavelet threshold denoising processing, and the noisy data with noise is decomposed by wavelet to obtain the real data information, which is recorded as P = [p ₁ , p ₂ , p ₃ ,..., p _m ], p _i is the desulfurization data at the i-th time point in the real desulfurization data sequence.

In this embodiment, the selected operating parameters that are strongly related to the denitrification of the power plant are input into the second support vector machine model constructed to predict the concentration of flue gas nitrogen oxides at the entrance of the denitrification system of the power plant, specifically including:

The collected historical operating parameters of the power plant denitrification system are used as sample data, and the Pearson correlation coefficient is used to calculate the correlation between the sample data and the flue gas nitrogen oxide concentration at the entrance of the power plant denitrification system. Based on the correlation, the highly relevant parameters are selected. The data combination is used as operating data that is strongly related to the denitrification system; the historical operating data of the denitrification system at least includes ammonia injection mass flow, boiler load, SCR inlet flue gas temperature, SCR inlet flue gas oxygen content, SCR inlet nitrogen oxide concentration, SCR denitrification efficiency;

Perform data preprocessing on the operating data that is strongly related to the denitrification system, and use the preprocessed data to construct a second support vector machine model;

Collect real-time operating data related to power plant denitration and input it into the constructed second support vector machine model to obtain the predicted value of flue gas nitrogen oxide concentration at the entrance of the power plant denitration system;

Among them, the data with strong correlation and extremely strong correlation are selected as the data with high correlation with the denitrification system. The calculation formula is:

X is the input sample data feature, Y is the concentration of nitrogen oxides at the inlet, cov(X,Y) represents the covariance of _X and _Y ; σ The correlation coefficient between When 0.2≤ρ<0.4, it is called weak correlation; when 0.0≤ρ<0.2, it is called extremely weak correlation or no correlation.

It should be noted that data preprocessing for the operating data that is strongly related to the denitrification system includes: removing duplicate data, time completion and resampling, missing value filling and outlier replacement processing, etc.

In this embodiment, the construction of the first support vector machine model and the second support vector machine model includes:

Use the cuckoo optimization method to determine the optimal support vector machine parameters: initialize the parameters of the cuckoo optimization algorithm, and search the nest position with the Levi's flight adaptively dynamically adjusted step size according to the parameters of the cuckoo optimization method: i=1,2,...,n; where, x _i ^(t+1) is the nest position of the i-th bird's nest in the t-th generation; a is the step control amount, used to control the search range of the step, which obeys Normal distribution; L(λ) is the Levy random walk path; the step size adaptive dynamic adjustment strategy is:

step _i = step _min + (step _max - step _min )d _i

Among them, step _i is the current search step, step _max is the maximum value of the step, step _min is the minimum value of the step, n _i is the position of the i-th bird's nest, and n _best is the bird's nest corresponding to the current minimum fitness. Position, d _max is the maximum value of the distance between the nest and other nests corresponding to the current minimum fitness;

Use the training set in the preprocessed data to train the support vector machine model, calculate the fitness of each bird's nest position, and retain the bird's nest corresponding to the minimum fitness to the next iteration;

Determine whether the minimum fitness meets the preset termination conditions. If it does, the nest position of the nest corresponding to the minimum fitness is the determined optimal support vector machine parameter. If not, remove the nests with the highest fitness and start again. Adjust the position of the nest;

According to the determined optimal support vector machine parameters, train the support vector machine model: establish a support vector machine training program based on the kernel function, input variables and output variables form a mapping relationship through the support vector machine model, and use the support vector machine training program to train the training samples The data is learned and trained, and N support vectors _Xi ^* are obtained after learning and training, i=0,1,...,N, thus forming a support vector machine model:

_Among ^them _{, _} is the input preprocessed desulfurization data or denitrification data, Y(X) represents the predicted value of sulfur dioxide concentration of the support vector of the power plant desulfurization system or the predicted value of nitrogen oxide concentration of the support vector of the denitrification system, K(·) represents the predicted value of the support vector machine Kernel function, the kernel function selects one of Gaussian function, polynomial function, linear function and radial basis function.

This invention predicts the sulfur dioxide concentration in the flue gas at the outlet of the power plant desulfurization system by collecting historical operating parameters of the power plant desulfurization system and denitrification system, and then selecting operating parameters that are strongly related to the power plant desulfurization and inputting them into the constructed first support vector machine model. , and select the operating parameters that are strongly related to the denitrification of the power plant and input them into the second support vector machine model to predict the nitrogen oxide concentration of the flue gas at the entrance of the denitrification system of the power plant. The support vector machine model can be used to predict the flue gas at the outlet of the desulfurization system. Predict the sulfur dioxide concentration value in the system and predict the nitrogen oxide concentration value at the inlet of the denitrification system to improve the accuracy of the prediction value.

In practical applications, the correctness of the predicted sulfur dioxide concentration in the flue gas at the outlet and the nitrogen oxide concentration in the flue gas at the inlet is evaluated through the root mean square error RMSE and the average absolute percentage error MAPE. The calculation formula is:

Among them, y _i is the actual value of the sulfur dioxide concentration at the outlet or the flue gas nitrogen oxide concentration at the inlet, is the predicted value of the sulfur dioxide concentration at the outlet or the predicted value of the flue gas nitrogen oxide concentration at the inlet, n is the number of predicted samples, the smaller the RMSE and MAPE values are, the predicted value of the sulfur dioxide concentration at the outlet or the flue gas nitrogen oxide concentration at the inlet The closer the predicted value of substance concentration is to reality, the higher the accuracy.

In this embodiment, the control parameters of the slurry spray amount and the ammonia spray amount are sent to the power plant desulfurization and denitrification system simulation model for model verification and intelligent diagnosis, including:

After inputting the slurry spray volume control parameters, the ammonia spray volume control parameters and the relevant configuration parameters for the operation of the power plant desulfurization and denitrification system into the power plant desulfurization and denitrification system simulation model, the obtained power plant desulfurization information is obtained through the set expert diagnosis module. The real-time operating parameters of the denitrification system are compared with the simulation result data of the simulation model to obtain the deviation, and a pre-alarm is implemented based on whether the deviation exceeds the preset threshold;

Among them, the expert diagnosis module is equipped with an intelligent diagnosis strategy, which determines the relevant operating status, data deviation, and pre-alarm information through preset logic, comprehensively outputs preliminary diagnosis result information, and then calls the expert database knowledge information for comparison and analysis. Whether the conclusion information drawn by the intelligent diagnosis strategy is associated with or consistent with the knowledge information of the expert database, and output diagnostic analysis results, operation guidance or task orders; the knowledge information of the expert database includes stored preset knowledge and abnormal faults that have occurred Information; the pre-alarm includes parameters exceeding the preset threshold, the time when the parameter exceeds the preset threshold, and abnormal fault information.

The present invention controls the slurry spray amount of the desulfurization system based on the predicted value of sulfur dioxide concentration combined with the first lag time prediction model, and controls the ammonia spray amount of the denitrification system based on the predicted value of nitrogen oxide concentration combined with the second lag time prediction model. Control can be combined with the first lag time to effectively control the concentration of nitrogen oxides at the desulfurization outlet, accurately control the slurry spray volume, control the pH value of the slurry within the effective range, and at the same time reduce the concentration of nitrogen oxides at the denitration outlet in combination with the second lag time. Fluctuate, accurately control the amount of ammonia sprayed, and reduce the cost of ammonia spraying in the denitrification system.

The present invention performs intelligent diagnosis by sending the control parameters of the slurry spray amount and the ammonia spray amount to the power plant desulfurization and denitrification system simulation model, and sets expert database knowledge information and intelligent diagnosis strategies in the expert diagnosis module to perform real-time monitoring of the system. The operating parameters and simulation data are compared to achieve alarm and diagnosis, and diagnostic analysis results, operation guidance or task orders are output, realizing effective processing and diagnostic analysis of power plant desulfurization and denitrification system data.

In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system embodiments described above are only illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible implementation architecture, functions and computer program products of systems, methods and computer program products according to multiple embodiments of the present invention. operate. In this regard, each block in the flowchart or block diagram may represent a module, segment, or portion of code that contains one or more executable functions for implementing the specified logical function instruction. It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two consecutive blocks may actually execute substantially in parallel, or they may sometimes execute in the reverse order, depending on the functionality involved. It will also be noted that each block of the block diagram and/or flowchart illustration, and combinations of blocks in the block diagram and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts. , or can be implemented using a combination of specialized hardware and computer instructions.

In addition, each functional module in various embodiments of the present invention can be integrated together to form an independent part, each module can exist alone, or two or more modules can be integrated to form an independent part.

If functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention essentially or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code. .

Taking the above-mentioned ideal embodiments of the present invention as inspiration and through the above description, relevant workers can make various changes and modifications without departing from the scope of the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the description, and must be determined based on the scope of the claims.

Claims (4)

1. The flue gas desulfurization and denitrification optimization control method based on the hysteresis model is characterized by comprising the following steps of:

step 1: model construction is carried out on all parts of the desulfurization and denitrification of the power plant, a control system is built according to a field control strategy, and a complete simulation model of the desulfurization and denitrification system of the power plant is built;

step 2: a first lag time prediction model is established for the time delay of determining the existence of the PH value of the absorption tower in the power plant desulfurization system, and a second lag time prediction model is established for the time delay of determining the existence of the nitrogen oxide concentration of the flue gas at the inlet of the power plant denitration system through a flue gas online monitoring device CEMS;

step 3: predicting the concentration of sulfur dioxide in flue gas at the outlet of a desulfurization system of a power plant and the concentration of nitrogen oxides in flue gas at the inlet of a denitration system;

step 4: controlling the slurry spraying amount of the desulfurization system according to the sulfur dioxide concentration predicted value and the first lag time predicted model, and controlling the ammonia spraying amount of the denitration system according to the nitrogen oxide concentration predicted value and the second lag time predicted model;

step 5: the control parameters of the slurry spraying quantity and the ammonia spraying quantity are issued to a simulation model of a desulfurization and denitrification system of the power plant for intelligent diagnosis;

in step 2, establishing a first lag time prediction model for determining a time delay existing in the PH value of the absorption tower in the power plant desulfurization system by using a variable point detection, time window sliding, correlation analysis and machine learning model comprises: establishing a slurry PH value response lag time identification algorithm flow by adopting a variable point detection, time window sliding and correlation analysis method and establishing a first lag time prediction model by adopting a machine learning model;

The slurry PH response lag time identification algorithm flow comprises:

after the PH value of the slurry in the absorption tower is adjusted, the working condition that the concentration value of sulfur dioxide at the outlet of the absorption tower is changed is selected as an identification object; equally dividing the time window Δt into two equally spaced time windows Δt _i1 And Deltat _i2 Sliding forward on the time axis gradually, calculating the average difference value of the sulfur dioxide concentration in two time windows, and if the average difference value exceeds a set threshold value, taking the moment as a working condition change point t _i If the time window is smaller than the set threshold value, continuing to slide the time window forwards until the working condition change point is detected or the time window slides to the cut-off time point; based on the working condition change point t _i And a time window delta t, respectively obtaining a slurry PH value time sequence and a sulfur dioxide concentration time sequence from the beginning of the working condition change to the end of the time window;

gradually advancing the sulfur dioxide concentration time sequence, setting a maximum movement step number k, obtaining a new sulfur dioxide concentration sequence through advancing, and constructing a sulfur dioxide concentration time lag matrix V;

calculating the Pelson correlation coefficient r of each column in the slurry PH value time sequence and matrix V, wherein the delay time corresponding to the maximum correlation coefficient is PH value response lag time t under the working condition ₁ ；

The building of the first lag time prediction model using the machine learning model includes:

After original data features in a power plant desulfurization system are collected and preprocessed, substituting the original data features into the slurry PH value response lag time identification algorithm flow to carry out PH value delay identification, and acquiring the relation between delay time and different operation data features; the raw data features include at least: the load of the boiler, the air supply quantity of the boiler, the flow rate of limestone slurry, the input quantity of the limestone slurry, the sulfur dioxide content, the calcium carbonate content in the limestone and the distance data characteristic from the absorption tower to the PH measuring point;

converting the operation data characteristics which are obtained through identification and can cause PH value change into characteristics with more working condition characteristics through a characteristic conversion mode, reducing the correlation among the data characteristics, and eliminating the influence caused by dimension through carrying out normalization processing on the converted data characteristics;

carrying out correlation analysis on the original operation data characteristics by adopting a correlation analysis method to obtain correlation coefficients of the operation data characteristics and PH response lag time, wherein the higher the correlation coefficients are, the most correlation between the data characteristics and the lag time is shown;

adopts a characteristic fusion method to fuse the operation data characteristics according to the height of the correlation coefficient to form new fusion characteristics,

Taking the original operation data characteristics and the new fusion characteristics as sample data, and inputting a training set in the sample data into a machine learning model according to a preset proportion to establish a first lag time prediction model under different operation data change working conditions; the PH value response lag time can be calculated according to different operation data characteristics through the first lag time prediction model;

in step 2, establishing a second lag time prediction model for the time delay of determining the existence of the concentration of the nitrogen oxides in the flue gas at the inlet of the power plant denitration system through the flue gas online monitoring device CEMS by adopting a variable point detection, time window sliding, correlation analysis and machine learning model comprises the following steps: establishing a CEMS measurement lag time identification algorithm flow by adopting a variable point detection, time window sliding and correlation analysis method and establishing a second lag time prediction model by adopting a machine learning model;

the CEMS measurement lag time identification algorithm flow comprises:

after the concentration of the nitrogen oxide at the inlet is selected to change, the working condition that the CEMS measured value changes is selected as an identification object; the measurement of the concentration of the nitrogen oxides is carried out through the heat tracing pipe and the analysis cabinet, and the smoke flows in the heat tracing pipe and the concentration in the analysis cabinet is measured, so that a certain time lag exists;

Equally dividing the time window Δt' into two equally spaced time windows Δt _i1 ' and Deltat _i2 ' gradually sliding forward on a time axis, calculating the average difference value of CEMS measured values in two time windows, and if the average difference value exceeds a set threshold value, taking the moment as a working condition change point t _i If the threshold value is smaller than the set threshold value, continuing to slide the time window forwards until the working condition change point is detected or the time window slides to the cut-off time point;

based on the working condition change point t _i 'and time window Deltat', respectively obtaining a nitrogen oxide concentration value time sequence from the beginning of working condition change to the end of the time window and a CEMS measurement value time sequence;

gradually advancing the CEMS measurement time sequence, setting a maximum movement step number k, obtaining a new CEMS measurement sequence through advancing, and constructing a CEMS measurement time lag matrix V';

calculating the pearson correlation coefficient r 'of each column in the time sequence and matrix V' of the nitrogen oxide concentration value, wherein the delay time corresponding to the maximum correlation coefficient is the nitrogen oxide concentration measurement lag time t under the working condition ₂ ；

The building of the second lag time prediction model using the machine learning model includes:

after the original data characteristics in the power plant denitration system are collected and preprocessed, the original data characteristics are substituted into a CEMS measurement lag time identification algorithm flow to carry out delay identification, and the relationship between the delay time and different operation data characteristics is obtained;

The raw data features include at least: boiler load, coal type, coal supply, combustion temperature, air quantity and smoke quantity;

converting the operation data characteristics which are obtained through identification and can cause the concentration change of the nitrogen oxides into characteristics with more working condition characteristics through a characteristic conversion mode, reducing the correlation among the data characteristics, and eliminating the influence caused by dimension through carrying out normalization processing on the converted data characteristics;

carrying out correlation analysis on the original operation data characteristics by adopting a correlation analysis method to obtain correlation coefficients of each operation data characteristic and the nitrogen oxide concentration value measurement lag time, wherein the higher the correlation coefficient is, the most correlation between the data characteristic and the lag time is shown;

the method comprises the steps of adopting a feature fusion method to fuse the running data features according to the height of a correlation coefficient to form new fusion features, taking the original running data features and the new fusion features as sample data, inputting a training set in the sample data into a machine learning model according to a preset proportion, and establishing a second lag time prediction model under different running data change working conditions; calculating CEMS measurement lag time according to different operation data characteristics through a second lag time prediction model;

The machine learning model selects an XGBoost model, is an integrated learning algorithm adopting a boosting method, and a base learner selects a CART decision tree and applies k CART functions { f } ₁ ,f ₂ ,…f _k Adding to form an integrated tree model; the target function of the model consists of a loss function and a regular term, and the loss function approximates by adopting second-order Taylor expansion; performing optimization operation on key parameters to improve the accuracy of model prediction, wherein the key parameters comprise the maximum depth of a tree, subsamples, the number of randomly sampled columns of each tree, the minimum leaf node sample weight and the learning rate;

the construction of the model starts from a root node, training set data are ordered according to each data feature, a greedy method is adopted to calculate the benefit of each feature, the feature with the largest benefit is selected as a splitting feature, the training set data are mapped to corresponding leaf nodes, the generated leaf nodes are recursively subjected to constraint condition until the constraint condition is reached, the decision tree generation process is finished, then the first-order derivative and the second-order derivative of a loss function are used for calculating the weight of the decision tree leaf node, the weight is used as a fitting target of the next tree, the recursion is repeated until the condition is met, and the model establishment is finished;

In step 3, after collecting historical operation parameters of a power plant desulfurization system and a denitration system, selecting operation parameters which are strongly related to power plant desulfurization, inputting the operation parameters into a constructed first support vector machine model to predict the concentration of sulfur dioxide in flue gas at an outlet of the power plant desulfurization system, and specifically comprising the following steps:

taking the collected historical operation parameters of the power plant desulfurization system as sample data, carrying out correlation analysis on the sample data, removing the sample data with the correlation of the sulfur dioxide concentration in the flue gas at the outlet of the limestone-gypsum wet desulfurization system being smaller than a preset value, and taking the rest sample data as operation data which is strongly correlated with the desulfurization system; the historical operating parameters of the desulfurization system at least comprise sulfur dioxide concentration at an inlet, nitrogen oxide concentration, unit load, limestone slurry circulating pump current, slurry supply quantity, flue gas sulfur dioxide concentration at an outlet of an absorption tower and slurry PH value;

performing data preprocessing on the operation data which is strongly related to the desulfurization system, and constructing a first support vector machine model by utilizing the preprocessed data;

collecting real-time operation data related to power plant desulfurization, and inputting the real-time operation data into a constructed first support vector machine model to obtain a predicted value of the concentration of sulfur dioxide in flue gas at the outlet of a power plant desulfurization system;

Wherein the data preprocessing comprises: filling up missing values and abnormal values and normalizing the operation data which are strongly related to the desulfurization system to obtain a preprocessed desulfurization data sequence, wherein the preprocessed desulfurization data sequence is marked as F= [ F ] ₁ ,f ₂ ,f ₃ ,…,f _n ]，f _i Desulfurizing data of the ith moment point in the processed desulfurizing data sequence;

wavelet threshold denoising is carried out on the desulfurization data sequence F, wavelet decomposition is carried out on noisy data with noise, real data information is obtained, and the real data information is recorded as P= [ P ] ₁ ,p ₂ ,p ₃ ,…,p _m ]，p _i The desulfurization data of the ith moment point in the real desulfurization data sequence;

in step 3, selecting an operation parameter related to the denitration intensity of the power plant, inputting the operation parameter to a constructed second support vector machine model, and predicting the concentration of the nitrogen oxides in the flue gas at the inlet of the denitration system of the power plant, wherein the method specifically comprises the following steps:

taking the collected historical operation parameters of the power plant denitration system as sample data, calculating the correlation between the sample data and the concentration of the nitrogen oxides in the flue gas at the inlet of the power plant denitration system by adopting a Pearson correlation coefficient, and selecting a data combination with high correlation as operation data which is strongly correlated with the denitration system according to the correlation; the denitration system historical operation data at least comprises ammonia injection mass flow, boiler load, SCR inlet flue gas temperature, SCR inlet flue gas oxygen content, SCR inlet nitrogen oxide concentration and SCR denitration efficiency;

Performing data preprocessing on operation data which are strongly related to the denitration system, and constructing a second support vector machine model by utilizing the preprocessed data;

acquiring real-time operation data related to power plant denitration and inputting the real-time operation data into a constructed second support vector machine model to acquire a flue gas nitrogen oxide concentration predicted value at an inlet of a power plant denitration system;

the data with strong correlation and extremely strong correlation are selected as the data with high correlation with the denitration system, and the calculation formula is as follows:

x is the characteristic of input sample data, Y is the concentration of nitrogen oxides at the inlet, cov (X, Y) represents the covariance of X, Y; sigma (sigma) _X Sum sigma _Y The standard deviation of X and Y respectively, ρ represents the correlation coefficient between two variables, and the value range is [ -1,1]The method comprises the steps of carrying out a first treatment on the surface of the When ρ is 0.8 or less<At 1, it is called extremely strong correlation; when ρ is 0.6 or less<At 0.8, it is called strong correlation; when ρ is 0.4 or less<0.6, referred to as moderate correlation; when 0.2 is less than or equal to ρ<0.4, called weak correlation; when 0.0 is less than or equal to ρ<0.2, is referred to as very weakly correlated or uncorrelated; constructing a first support vector machine model and a second support vector machine model, comprising: adopting a cuckoo optimization method to determine optimal support vector machine parameters: initializing parameters of a cuckoo optimization algorithm, and searching bird nest positions by step-length self-adaptive dynamic adjustment of Lewy flight according to the parameters of the cuckoo optimization method: x is x _i ^(t+1) ＝x _i ^(t) +a⊕L(λ)，i＝1,2,…,n；

Wherein x is _i ^(t+1) A bird nest position of the ith bird nest in the t generation; a is a step control quantity, which is used for controlling the searching range of the step and obeys the front distribution; l (lambda) is a Lewy random walk; the step length self-adaptive dynamic adjustment strategy is as follows:

step _i ＝step _min +(step _max -step _min )d _i ；

wherein step _i Step for the current search step _max Step is the maximum value of step length _min Is the minimum value of the step length, n _i Is the position of the ith nest, n _best D, the bird nest position corresponding to the bird nest is the current minimum fitness _max The current minimum fitness corresponds to the maximum value of the distance between the bird nest and other bird nests;

training a support vector machine model by adopting a training set in the preprocessed data, calculating the fitness of each bird nest position, and reserving the bird nest corresponding to the minimum fitness to the next iteration;

judging whether the minimum fitness meets a preset termination condition, if so, determining the bird nest position of the bird nest corresponding to the minimum fitness as the determined optimal support vector machine parameter, and if not, removing a plurality of bird nests with the highest fitness, and readjusting the bird nest position;

training a support vector machine model according to the determined optimal support vector machine parameters: establishing a support vector machine training program based on a kernel function, forming a mapping relation between an input variable and an output variable through a support vector machine model, performing learning training on training sample data by using the support vector machine training program, and obtaining N support vectors X through learning and training _i ^* I=0, 1, …, N, forming a support vector machine model:

wherein X is _i ^* Support vector representing desulfurization or denitrification system of power plant, Y _i Sulfur dioxide concentration representing power plant desulfurization system support vector or nitrogen oxide concentration representing denitration system support vector, alpha _i The coefficient representing the ith support vector, X is the input preprocessed desulfurization data or denitration data, Y (X) represents the sulfur dioxide concentration predicted value of the support vector of the desulfurization system of the power plant or the nitrogen oxide concentration predicted value of the support vector of the denitration system, K (·) represents the kernel function of the support vector machine, and the kernel function selects one of a Gaussian function, a polynomial function, a linear function and a radial basis function.

2. The method for optimizing control over desulfurization and denitrification of flue gas based on a hysteresis model according to claim 1, wherein in step 1, after the model construction is carried out on each component of desulfurization and denitrification of a power plant according to a modularized modeling method by dynamic simulation software, a corresponding control system is built according to a field control strategy, and a complete simulation model of the desulfurization and denitrification system of the power plant is built, which comprises the following steps: the power plant desulfurization system is a limestone-gypsum wet desulfurization system, and at least comprises a flue gas system, an absorption tower system, a limestone slurry preparation system, a gypsum slurry dehydration system, a wastewater treatment system and an electrical system; the power plant denitration system selects an SCR method flue gas denitration system and at least comprises a flue gas system, an SCR reactor system, a sound wave soot blowing system and a liquid ammonia storage and supply system; the dynamic simulation software selects and connects corresponding component modules from a model library according to mass conservation, momentum conservation and energy conservation equations and the technological process of a limestone-gypsum wet desulfurization system and an SCR flue gas denitration system in the modeling process, and inputs initial data to complete the model construction of the power plant desulfurization and denitration system; and constructing an analog quantity control system, a sequence control system and a logic control system according to a field control strategy, configuring by adopting a basic algorithm module, realizing the same function as an actual control system, and establishing a complete simulation model of the desulfurization and denitrification system of the power plant.

3. The hysteresis model-based flue gas desulfurization and denitrification optimization control method according to claim 2, wherein the simulation model of the power plant desulfurization and denitrification system further comprises: in the process of model development and debugging, physical data acquired by an actual power plant desulfurization and denitrification system and virtual data acquired based on a power plant desulfurization and denitrification simulation model are compared, whether errors exceed a threshold value is judged, if so, virtual data with larger errors are classified through cluster learning, corresponding historical data are combined as input, error learning is carried out through a neural network, correction coefficients are output to correct the error data of the virtual data, and virtual-real fusion is carried out on the corrected virtual data and the physical data to generate a verified power plant desulfurization and denitrification simulation model.

4. The method for optimizing control over flue gas desulfurization and denitrification based on hysteresis model according to claim 1, wherein the step 5 comprises the following steps: after the slurry spraying quantity control parameter, the ammonia spraying quantity control parameter and the relevant configuration parameters of the operation of the power plant desulfurization and denitrification system are input into the power plant desulfurization and denitrification system simulation model, comparing the acquired real-time operation parameters of the power plant desulfurization and denitrification system with simulation result data of the simulation model through a set expert diagnosis module to obtain deviation, and realizing pre-alarm through whether the deviation exceeds a preset threshold value; the expert diagnosis module is internally provided with an intelligent diagnosis strategy, and is used for judging the related running state, data deviation and pre-alarm information conditions through preset logic, comprehensively outputting diagnosis preliminary result information, then calling expert database knowledge information for comparison, analyzing whether conclusion information obtained by the intelligent diagnosis strategy is related to or consistent with the expert database knowledge information, and outputting diagnosis analysis results, running instructions or task sheets; the knowledge information of the expert database comprises stored preset knowledge and information of abnormal faults; the pre-alarm comprises a parameter exceeding a preset threshold value, and time when the parameter exceeds the preset threshold value and abnormal fault information.