CN117175595B - Power grid regulation and control method and system based on multi-level data - Google Patents

Abstract

The invention provides a power grid regulation and control method and system based on multistage data, which relate to the technical field of power grid regulation and control and comprise the following steps: acquiring initial power grid data, inputting the initial power grid data into a preset first load prediction model, preprocessing the initial power grid data through the first load prediction model, determining first load input, and combining the first load prediction model according to the first load input to obtain second load input; according to the second load input, combining a preset second load prediction model, transmitting the second load input into the second load prediction model in the sequence from the root node to the leaf node to obtain second load output, and according to the second load output, combining a preset result optimization algorithm to determine a load prediction result; according to the load prediction result, combining with the initial data of the power grid, constructing an energy action space, and according to the energy action space, determining a regulation and control optimization strategy through a preset strategy optimization algorithm.

Description

A power grid control method and system based on multi-level data

Technical field

The present invention relates to the technical field of power grid control, and in particular to a power grid control method and system based on multi-level data.

Background technique

Grid regulation is the basic data required for power grid operation monitoring, analysis and calculation by software application systems in the field of power system dispatching and operation. However, with the increasing demand for electricity, the grid load is also increasing, so it is necessary to conduct real-time implementation of grid regulation strategies. Updated and optimized.

In the existing technology, CN114548653A, a data collection method, system and electronic equipment for a power grid load control platform provides a data collection method, system and electronic equipment for a power grid load control platform. Through the mechanism and means of integrating data, the data is When comprehensively utilized, it can provide good support for decision-making analysis, improve the system's return on investment, effectively integrate data from various business systems, centralize and unify management, transform it into valuable information, and provide unified information for related business applications. Data support truly realizes one data, one entrance, unified export, and multi-level applications, automatically obtains relevant data from multiple heterogeneous data sources, and conducts validity verification based on the attribute relationships of the data to achieve automatic calculation, statistics, summary, and Automated operation mode meets different business scenario applications and provides applications in different business fields.

CN107451188A, a method and system for releasing multi-level node combinations of power grid control model data, discloses a method and system for releasing multi-level node combinations of power grid control model data, and sets the range of multi-level node combinations of power grid control model; in the Within the scope of the multi-level node combination of the power grid control model, a configuration file of the multi-level nodes involved in the container object is established; the configuration file of the multi-level nodes involved in the container object is accessed through a single attribute node, and the multi-level nodes included in the container object are published. Data, the technical solution provided by the present invention combines the object data hierarchical definition under a specific container, and provides a method for overall access to the full data of the power grid control model in the equipment container by setting exclusive characteristic identifiers. This method does not change the definition of the OPCUA object access service, but only adds hierarchical orchestration of objects and attributes in the device container on the service implementation side, so that the application can quickly obtain the required power grid control model data.

To sum up, although the existing technology can read the data of existing equipment in the power grid and adjust the power grid control strategy according to the preset mode, it cannot make adaptive adjustments to the power grid based on real-time load, nor does it take into account time and historical power consumption. The impact of the strategy on the power grid control strategy, therefore, this solution can at least solve some of the existing problems.

Contents of the invention

Embodiments of the present invention provide a power grid control method and system based on multi-level data, which are used to dynamically adjust the power grid control strategy according to the real-time status of the power grid.

The first aspect of the embodiment of the present invention, a power grid control method based on multi-level data, includes:

Obtain the initial data of the power grid, input the initial data of the power grid into a preset first load prediction model, preprocess the initial data of the power grid through the first load prediction model, determine the first load input, and according to the first load prediction model A load input, combined with the first load prediction model, to obtain a second load input;

According to the second load input, combined with the preset second load prediction model, the second load input is transferred in the second load prediction model in the order of root node to leaf node to obtain the second load output. , according to the second load output, combined with the preset result optimization algorithm, determine the load prediction result, wherein the second load model is a tree model constructed based on the extreme gradient boosting algorithm;

According to the load prediction results, combined with the initial data of the power grid, an energy action space is constructed. According to the energy action space, a control optimization strategy is determined through a preset strategy optimization algorithm, wherein the strategy optimization algorithm is based on an improved near-end Built using strategy optimization algorithms.

In an alternative implementation,

The initial data of the power grid is input into a preset first load prediction model, and the initial data of the power grid is preprocessed through the first load prediction model to determine the first load input. According to the first load input , combined with the first load prediction model, the second load input includes:

Acquire historical power consumption data through the database, obtain power grid status data according to smart sensors installed in power grid nodes, add the historical power consumption data and the power grid status data to the same set, and record them as the power grid initial data;

Perform data cleaning and feature selection operations on the initial data of the power grid through the first load prediction model to obtain the first load input;

According to the first load input, determine the forward hidden state and the backward hidden state through the bidirectional state network in the first load prediction model;

According to the forward hidden state and the backward hidden state, combined with the preset residual module and self-attention mechanism, the input data of a specific layer in the bidirectional state network is directly added to the output of the bidirectional state network data, determine the second load input.

In an alternative implementation,

According to the second load input, combined with the preset second load prediction model, the second load input is transferred in the second load prediction model in the order of root node to leaf node to obtain the second load prediction model. Load output, based on the second load output, combined with the preset result optimization algorithm, determines the load prediction result, wherein the second load model is a tree model constructed based on the extreme gradient boosting algorithm, including:

Obtain the second load input, add the second load input to the second load prediction model, and pass the second load input downward from the root node of the second load prediction model, at each leaf At the node, the transmission direction is determined according to the preset node splitting conditions. If the preset node splitting conditions are met, the transmission is to the left. If not, the transmission is to the right until the transmission is completed, and the second load output is obtained. ;

According to the second load output, combined with the result optimization algorithm, the abnormal values in the second load are removed and corrected according to the data trend to determine the load prediction result.

In an optional implementation, the method further includes training the second load prediction model:

Obtain the second load input and convert the second load input into a feature vector. For continuous feature values, divide the continuous feature values into discrete intervals through the equal-frequency bucketing strategy and perform discretization to obtain the feature value range. According to the feature value range, a histogram is constructed through a lightweight gradient boosting algorithm;

Based on the leaf node priority growth strategy, the first gain value of the current leaf node is determined through the node gain function, the node with the maximum gain among the current leaf nodes is selected for splitting, the maximum gain value is obtained, and a loss function is constructed based on the maximum gain value. According to The loss function calculates the gradient and second-order derivative of the current leaf node, and constructs a second-order Taylor expansion of the objective function;

For each leaf node, traverse the eigenvalues of all leaf nodes, combine the gradient and second derivative of the current leaf node, determine the leaf node with the smallest loss function value after splitting, record it as a split leaf node, and split the split leaf node into Two leaf nodes, and calculate the gradient and second-order derivative of the leaf node after splitting, and repeat the operation until the stopping condition is met.

In an alternative implementation,

Determine the first gain value of the current leaf node through the node gain function, select the node with the maximum gain among the current leaf nodes for splitting, obtain the maximum gain value, and construct a loss function based on the maximum gain value as shown in the following formula:

;

Among them, D represents the data set of the current node, Loss(D) represents the loss function, y _i represents the real label of sample i , y _avg represents the mean value of the current node, y _split represents the node mean value after splitting, max _split () represents the Choose the splitting method that maximizes the loss function among all possible methods.

In an alternative implementation,

The energy action space is constructed based on the load prediction results and the initial data of the power grid. According to the energy action space, the control optimization strategy is determined through a preset strategy optimization algorithm, including:

Obtain the load prediction result and the power grid initial data, define a state space according to the load prediction result and the power grid initial data, determine a first adjustment range and a second adjustment range according to the state space, and determine the first adjustment range and the second adjustment range according to the state space. The first adjustment range and the second adjustment range determine the energy action space in combination with the pre-acquired power unit characteristics;

According to the energy action space, combined with system losses and grid stability, a reward function is determined. According to the reward function, a grid action is selected according to the policy optimization algorithm at each time step. According to the grid action, the state is iterated space and the energy action space. According to the iterated state space and energy action space, combined with the preset strategy optimization algorithm, the control optimization strategy is determined.

In an alternative implementation,

According to the energy action space, combined with system loss and grid stability, the reward function is determined as follows:

;

Among them, R(s,a) represents the reward value, s represents the system state, a represents the grid action, Loss(s,a) represents the system loss, α represents the weight coefficient, and VoltageStability(s,a) represents the voltage stability.

A second aspect of the embodiment of the present invention provides a power grid control system based on multi-level data, including:

The first unit is used to obtain the initial data of the power grid, input the initial data of the power grid into a preset first load prediction model, preprocess the initial data of the power grid through the first load prediction model, and determine the first load input, According to the first load input, combined with the first load prediction model, a second load input is obtained;

The second unit is configured to transmit the second load input in the second load prediction model in the order of root node to leaf node according to the second load input and combined with the preset second load prediction model. , obtain the second load output, and determine the load prediction result according to the second load output in combination with the preset result optimization algorithm, where the second load model is a tree model constructed based on the extreme gradient boosting algorithm;

The third unit is used to construct an energy action space based on the load prediction results and combined with the initial data of the power grid. According to the energy action space, determine the control optimization strategy through a preset strategy optimization algorithm, wherein the strategy optimization algorithm It is built based on the improved proximal strategy optimization algorithm.

A third aspect of the embodiment of the present invention,

An electronic device is provided, including:

processor;

Memory used to store instructions executable by the processor;

Wherein, the processor is configured to call instructions stored in the memory to execute the aforementioned method.

The fourth aspect of the embodiment of the present invention,

A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

According to the embodiment of the present invention, the solution can accurately predict the load of the power grid through the preset first and second load prediction models. This helps the system better understand future power demand and formulate more effective regulation strategies. The use of a tree model built based on the extreme gradient boosting algorithm can better capture the complex relationships between loads. This model can provide more accurate secondary load prediction results and help improve the understanding of the grid status. By combining the load prediction results with the initial grid data, an energy action space can be constructed. This space reflects the energy regulation options available in the system and provides a basis for subsequent strategy optimization. In summary, this solution can better adapt to the complexity of power grid operation and improve the regulation of the power grid through load forecasting and strategy optimization algorithms. ability.

Description of the drawings

Figure 1 is a schematic flow chart of a power grid control method based on multi-level data according to an embodiment of the present invention;

Figure 2 is a schematic structural diagram of a power grid control system based on multi-level data according to an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments These are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

The technical solution of the present invention will be described in detail below with specific examples. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

Figure 1 is a schematic flow chart of a power grid control method based on multi-level data according to an embodiment of the present invention. As shown in Figure 1, the method includes:

S1. Obtain the initial data of the power grid, input the initial data of the power grid into the preset first load prediction model, preprocess the initial data of the power grid through the first load prediction model, determine the first load input, and determine the first load input according to the preset first load prediction model. The first load input is combined with the first load prediction model to obtain a second load input;

The initial data of the power grid specifically refers to historical power consumption data, distributed energy data and power grid status data. The historical power consumption data refers to the power consumption data per unit time in the past, and the distributed energy data refers to photovoltaic power generation and wind power generation. Real-time or historical power generation data of distributed energy sources such as hydropower and hydropower. Grid status data refers to information such as voltage, frequency and equipment operating status of each node in the grid.

The first load input refers to the result of inputting the initial data of the power grid into the first load prediction model for processing. The preprocessing specifically includes using the first load prediction model to perform data cleaning and feature selection on the initial data of the power grid, This can include operations such as handling missing values, outliers, and selecting the most relevant features.

In an alternative implementation,

The initial data of the power grid is input into a preset first load prediction model. The initial data of the power grid is preprocessed through the first load prediction model to determine the first load input. According to the first load input, combined with The first load prediction model obtains the second load input including:

Acquire historical power consumption data through the database, obtain power grid status data according to smart sensors installed in power grid nodes, add the historical power consumption data and the power grid status data to the same set, and record them as initial power grid data;

Perform data cleaning and feature selection operations on the initial data of the power grid through the first load prediction model to obtain the first load input;

According to the first load input, determine the forward hidden state and the backward hidden state through the bidirectional state network in the first load prediction model;

According to the forward hidden state and the backward hidden state, combined with the preset residual module and self-attention mechanism, the input data of a specific layer in the bidirectional state network is directly added to the output of the bidirectional state network data, determine the second load input.

Obtain data from the historical power consumption database, including timestamps and corresponding power consumption information. Obtain power grid status data through smart sensors set in power grid nodes, including current, voltage, frequency and other information. Combine historical power consumption data with the power grid. State data is merged into a single dataset where timestamps are used to align the data.

Use the first load prediction model for data cleaning and feature selection, including processing missing values, outliers, selecting the most relevant features, etc. Based on the cleaned and feature selected data, generate the first load input and input the first load As the input value of the first load forecasting model;

For example, suppose you have a power grid and want to predict the power load for the next day. Historical power consumption data and grid status data related to smart sensors in grid nodes are collected. For missing time points, the average value of before and after data is used to fill in the data. Through a threshold-based method, data beyond the reasonable range are regarded as anomalies. value and process it, calculate the correlation between each feature and the target (load), select the feature with high correlation, and obtain the first load input through the above operations.

The reasonable range of data can be determined based on historical power consumption data, or can be adjusted accordingly based on the date and time. For example, the use of air conditioners in summer will increase the grid load, so the reasonable range in the above scheme can be based on the historical summer average. Load is determined.

Using a deep learning framework, import the required layers, optimizers and other tools, load the pre-trained first load prediction model, which contains a bidirectional state network, obtain the first load input data, and ensure the format and structure Matching the model inputs, the first load input data is added to the loaded first load prediction model, and the forward and backward hidden states are obtained through the output of the model.

The forward hidden state is the information retained by the model from the beginning to the end of the sequence when processing the input sequence. The forward hidden state contains all the information before the current time step in the sequence and is what the model has learned before the current time step. The backward hidden state is the information retained by the model from the end to the beginning of the sequence when processing the input sequence. The backward hidden state contains all the information after the current time step in the sequence and is the model's current time. The representation learned after the step.

Obtain the forward hidden state and backward hidden state of a specific layer in the bidirectional state network, and perform residual connection between the forward and backward hidden states with the original input by adding the corresponding elements. For a specific layer, you may need to choose different residuals Connection strategy, such as choosing direct additive connection or weighted additive connection, using the self-attention mechanism to calculate attention weights, and applying the attention weights to the weighted combination of hidden states, adjusting the bidirectional hidden states, and determining the residual module and the hidden state processed by the self-attention mechanism, and use this hidden state as the second load input;

For example, assuming that there is a task to predict the load condition of the power grid and define a prediction model containing multiple calculation layers. Each calculation layer has a corresponding forward hidden state and backward hidden state. We set the residual module Residual connections are performed in specific layers. In each calculation layer, the corresponding forward hidden state and backward hidden state are extracted, and the weights of the extracted forward hidden state and backward hidden state are dynamically adjusted according to the self-attention mechanism. , adjusting the bidirectional state network through attention weights to determine the second load input.

In this step, by integrating historical power consumption data and grid status data, the initial grid data forms a comprehensive data set containing rich information, which improves the comprehensive understanding of the overall status of the grid and helps to establish a more comprehensive and accurate Load forecasting model; the first load forecasting model performs data cleaning and feature selection on the initial data of the power grid, extracts key load forecasting features, and improves the model's ability to understand and express the input data by removing noise and selecting important features, and enhances It improves the generalization ability of the model; the bidirectional state network is used to extract the forward and backward hidden states of each time step, which can capture the relevant information in the time series, better reflect the dynamic changes of the load, and improve the model The ability to model sequences; the introduction of residual modules and self-attention mechanisms in specific layers strengthens the model's ability to process sequence information. Through residual connections, the model better retains the original input information; self-attention mechanism It enhances the model's attention to different time steps in the sequence and improves its ability to capture key information.

In summary, this embodiment effectively improves the ability of power grid load prediction by integrating multi-source data and introducing complex model structures and processing mechanisms, and provides more intelligent support for power grid regulation.

S2. According to the second load input, combined with the preset second load prediction model, the second load input is transferred in the order of root node to leaf node in the second load prediction model to obtain the second load prediction model. Load output, determine the load prediction result according to the second load output and combined with the preset result optimization algorithm, wherein the second load model is a tree model constructed based on the extreme gradient boosting algorithm;

The tree model is a type of machine learning model based on a tree structure. It constructs a tree structure for prediction or classification through recursive binary division of input data; the root node is a decision tree or tree model. The starting node is located at the top of the tree. It is the entrance of the tree. Starting from the root node, each node of the tree is connected to the node of the next layer through edges. The task of the root node is to select a feature and then divide the input data according to the feature. The leaf nodes is the end node of the tree, it has no child nodes. At the leaf node, the model gives the final output or prediction result. The leaf node contains an output value. This value is determined by the leaf node that the input data finally reaches after passing the judgment condition on each branch of the tree. The extreme gradient boosting is a gradient boosting algorithm, which is a type of ensemble learning.

In an alternative implementation,

According to the second load input, combined with the preset second load prediction model, the second load input is transferred in the second load prediction model in the order of root node to leaf node to obtain the second load output. , according to the second load output, combined with the preset result optimization algorithm, the load prediction result is determined, wherein the second load model is a tree model constructed based on the extreme gradient boosting algorithm and includes:

Obtain the second load input, add the second load input to the second load prediction model, and pass the second load input downward from the root node of the second load prediction model, at each leaf At the node, the transmission direction is determined according to the preset node splitting conditions. If the preset node splitting conditions are met, the transmission is to the left. If not, the transmission is to the right until the transmission is completed, and the second load output is obtained. ;

According to the second load output, combined with the result optimization algorithm, the abnormal values in the second load are removed and corrected according to the data trend to determine the load prediction result.

Starting from the root node of the tree model, recursively transfer the second load input to each node of the tree according to the preset node splitting conditions. For each node, determine the transfer direction: if the node splitting conditions are met, transfer to the left , otherwise pass to the right, repeat this process until reaching the leaf node, at the leaf node, obtain the output value of the node, that is, the predicted value of the second load, this output value will be used as the final second load output;

The node splitting condition is set according to the specific node. For example, there is now a feature vector (x, y). The splitting condition of the first node is that x is greater than 1. If it is met, enter the left subtree. If If not, enter the right subtree.

Use the result optimization algorithm to optimize the second load output, use the outlier detection algorithm to identify and remove outliers in the second load output; analyze the trend of the second load output, you can use sliding window, exponential smoothing and other methods, based on past Observed values are used to correct possible outliers to make them consistent with the overall trend of the data; the final load forecast result is determined by combining the second load output after removing outliers and trend correction.

For example, assume that there is a time series containing historical load data, in which each time point has a corresponding second load value, and statistical methods such as Z-score or IQR are used to detect outliers. Assume that at a certain time point The second load value exceeds a certain threshold and is considered an outlier. The data points detected as outliers are removed from the second load output. This can be done by deleting the outliers or using reasonable replacement values (such as the average of the preceding and following moments). value) to fill in, use sliding window, exponential smoothing and other methods to analyze the trend of the second load output. If the second load value at a certain point in time obviously deviates from the trend, it may be an abnormal point and needs to be corrected; use sliding window averaging Or sliding window median and other methods, calculate the average or median of adjacent time points at each time point, compare the second load value at the current time point with the calculated value, and make corrections. The corrected second load output is combined with the previous result of removing outliers to obtain the final load forecast result.

In this embodiment, by transferring the second load input to the leaf nodes of the tree model layer by layer, the model can be finely divided according to different characteristics and conditions, thereby improving the accuracy of load prediction; through the preset node splitting conditions, the model can Adaptively perform different divisions based on the characteristics of the input data, making the model more adaptable to complex power grid load changes; the resulting optimization algorithm, combined with outlier detection and correction strategies, helps remove outliers in the second load output. This improves the robustness of load forecasting, allowing the model to better adapt and predict when faced with abnormal situations. By analyzing the trend of the secondary load output and using methods such as sliding windows and exponential smoothing, it can better capture and predict Correct data trends. This helps make the load forecast results smoother and more realistic.

In an optional implementation, the method further includes training a second load prediction model:

Obtain the second load input and convert the second load input into a feature vector. For continuous feature values, divide the continuous feature values into discrete intervals through the equal-frequency bucketing strategy and perform discretization to obtain the feature value range. According to the feature value range, a histogram is constructed through a lightweight gradient boosting algorithm;

Based on the leaf node priority growth strategy, the first gain value of the current leaf node is determined through the node gain function, the node with the maximum gain among the current leaf nodes is selected for splitting, the maximum gain value is obtained, and a loss function is constructed based on the maximum gain value. According to The loss function calculates the gradient and second-order derivative of the current leaf node, and constructs a second-order Taylor expansion of the objective function;

For each leaf node, traverse the eigenvalues of all leaf nodes, combine the gradient and second-order derivative of the current leaf node, determine the leaf node with the smallest loss function value after splitting, record this leaf node as a split leaf node, and record the The split leaf node is split into two leaf nodes, and the gradient and second derivative of the split leaf node are calculated, and the operation is repeated until the stopping condition is met.

Obtain the historical data of the second load, including various characteristics such as temperature, season, time, etc., and convert the second load input into a feature vector. For continuous features, the equal-frequency bucketing strategy is used for discretization to obtain the discrete values of the features. Each sample will be converted into a feature vector containing each feature, a tree model is initialized, and hyperparameters such as the maximum depth of the tree and learning rate are set. After each feature is discretized, a histogram is constructed based on a lightweight gradient boosting algorithm. .

Starting from the root node, initialize the entire tree, traverse each leaf node, and determine the first gain value of the current leaf node through the node gain function. For example, starting from the root node, the entire tree starts to be empty. For each leaf node, calculate its first gain value, which is determined by the node gain function. The node gain function is usually an indicator representing the quality of model fitting, such as square loss, absolute loss, etc. It indicates the degree of performance improvement of the model on the current node. In the case of square loss, the gain can be the mean square error reduction before and after the node split. For each leaf node, calculate the node gain value of its left child node and right child node after the split.

Select the leaf node with the maximum gain for splitting, construct a loss function based on the maximum gain value, usually using square loss or absolute loss, calculate the gradient and second-order derivative of the current leaf node according to the loss function, and construct the second-order Taylor expansion of the objective function .

Traverse the eigenvalues of the current leaf node, perform a split test on each eigenvalue, and combine the gradient and second derivative of the current leaf node to determine the leaf node with the smallest loss function value after splitting. Use the eigenvalues that meet the conditions as the split point, split the current leaf node into two leaf nodes, calculate the gradient and second-order derivative of the split leaf node, and repeat the operation until the depth of the tree reaches the set value or the number of samples in the node is less than a certain threshold.

The equal-frequency bucketing strategy is a method of dividing continuous feature values into discrete intervals to ensure that each interval contains the same number of samples. The equal-frequency bucketing strategy simplifies complexity by segmenting data while reducing Sensitivity to outliers and improve model robustness. The specific strategy of equal frequency classification is to sort the feature values in ascending order and determine the number of buckets as needed. For example, if you want to divide the features into 10 buckets, then divide the data into 10 equal parts. Quantity calculates the corresponding quantile, and divides the feature values into corresponding buckets based on the calculated quantile. Each bucket contains the same number of samples, and each bucket is discretized with a label or value that represents the bucket.

In this embodiment, by using the lightweight gradient boosting algorithm and the node priority growth strategy, the tree model can be constructed more efficiently, reducing the computational complexity, improving the training speed of the model, and performing feature discretization through the equal-frequency bucketing strategy. Helps improve the robustness of the model, reduce sensitivity to outliers, and enable the model to better adapt to different data distributions. Through node splitting and optimization operations, the tree model can better capture the nonlinear relationship of the data and improve the prediction performance of the model.

In an optional implementation, the first gain value of the current leaf node is determined through the node gain function, the node with the maximum gain among the current leaf nodes is selected for splitting, the maximum gain value is obtained, and the loss is constructed based on the maximum gain value The function is shown in the following formula:

;

Among them, D represents the data set of the current node, Loss(D) represents the loss function, y _i represents the real label of sample i , y _avg represents the mean value of the current node, y _split represents the node mean value after splitting, max _split () represents the Choose the splitting method that maximizes the loss function among all possible methods.

This function calculates the loss function and selects the node with the largest gain among the current leaf nodes for splitting. This means that when selecting split nodes, the model focuses more on improving the fitting effect of the overall model, making the split child nodes more likely to contain more information and better capture the differences between samples. Through the loss function By maximizing the selection splitting method, the model pays more attention to improving the fitting effect of the training set, which is expected to improve the prediction performance of the model. This is particularly important in power grid regulation, because accurate load forecasting is crucial for reasonable dispatch of the power grid. At the same time, because the loss function takes into account the mean square error of the nodes after splitting, the model is more suitable for data with different distributions. In summary, this paper The function can make the lightweight gradient boosting algorithm build a tree model more effectively, improve the performance and robustness of the model, and achieve more accurate and reliable load forecasting in power grid regulation.

S3. According to the load prediction results and combined with the initial data of the power grid, an energy action space is constructed. According to the energy action space, a control optimization strategy is determined through a preset strategy optimization algorithm. The strategy optimization algorithm is based on an improved Constructed by proximal strategy optimization algorithm.

The energy action space refers to the collection of energy dispatching plans that can be selected by the system in power grid regulation. It describes various energy dispatch decisions that the system can take given certain initial data of the power grid and load forecast results; the strategy optimization algorithm is a proximal method for optimizing strategies, by Multiple policy update steps are performed to ensure the conservativeness of the update by minimizing the relative probability ratio of the policy update.

In an alternative implementation,

According to the load prediction results, combined with the initial data of the power grid, an energy action space is constructed. According to the energy action space, a control optimization strategy is determined through a preset strategy optimization algorithm, wherein the strategy optimization algorithm is based on an improved near-end The strategy optimization algorithm constructed includes:

Obtain the load prediction result and the power grid initial data, define a state space according to the load prediction result and the power grid initial data, determine a first adjustment range and a second adjustment range according to the state space, and determine the first adjustment range and the second adjustment range according to the state space. The first adjustment range and the second adjustment range determine the energy action space in combination with the pre-acquired power unit characteristics;

According to the energy action space, combined with system losses and grid stability, a reward function is determined. According to the reward function, a grid action is selected according to the policy optimization algorithm at each time step. According to the grid action, the state is iterated space and the energy action space. According to the iterated state space and energy action space, combined with the preset strategy optimization algorithm, the control optimization strategy is determined.

Obtain the load prediction results and the initial data of the power grid from the database or other data sources, and combine the load prediction results and the initial data of the power grid to define the state space. The state space may include grid node status, load demand, energy supply status, etc. According to the defined state space, combined with the characteristics of the power supply unit obtained in advance, including the start-stop characteristics of the generator and the charging and discharging characteristics of the energy storage system, the energy action space is defined. Including active power adjustment, reactive power adjustment, power market transactions, etc.

For example, the maximum active power of the generator: 1000 kW, the minimum active power output (Pmin): 100 kW, the active power adjustment granularity: 50 kW, the maximum reactive power output (Qmax): 300 kVAR, the minimum reactive power output ( Qmin): -200 kVAR, then the active power range of the generator's energy action space: 100 kW - 1000 kW, with 50 kW as the adjustment granularity, and the reactive power range: -200 kVAR - 300 kVAR, with 50 kVAR as the adjustment granularity .

According to the power grid status and the output of the power supply unit, calculate the system loss, consider voltage stability, frequency stability and other factors, define the grid stability index, combine the system loss and the grid stability index, and construct a reward function. For example, the reward function can be the system The negative value of the loss, because the smaller the system loss, the better, taking into account the impact of grid stability, and using the state space and energy action space defined in the previous steps as the search space of the optimization algorithm.

At each time step, a strategy optimization algorithm is used to search for optimal actions in the defined state space and energy action space. The optimization algorithm generates new grid actions, updates the output of the power supply unit, and simulates the response of the grid. According to the current grid state, Calculate the reward of the current time step based on the output of the power unit and the defined reward function, use the reward information to update the parameters or model in the policy optimization algorithm, and repeat the above steps until the maximum number of iterations or the preset stop condition is reached, such as 50 iterations. , based on the final strategy model, determine the control optimization strategy, that is, the power unit actions that should be taken under a given grid state.

The first adjustment range specifically refers to the adjustment range of reactive power, the second adjustment range specifically refers to the adjustment range of active power, and the state space is a set that describes the status of the power grid and the system, including but not limited to grid node voltage, Grid status such as frequency, system load demand, the energy action space specifically refers to the adjustable action range of the power supply unit, including the adjustment range of active power and reactive power, the start and stop status of the generator, the charging and discharging status of the energy storage system, and others Actions of the power supply unit, the characteristics of the power supply unit refer to energy devices such as generators and energy storage systems in the power supply system.

In this embodiment, by defining the state space and determining the first adjustment range and the second adjustment range, the system can consider the power grid status and adjustment requirements at a higher level and in a more comprehensive dimension. This helps to improve the flexibility and accuracy of regulation. Taking into account the characteristics of the power supply unit, such as the start and stop characteristics of the generator, the charge and discharge efficiency of the energy storage system, etc., it helps to build a more realistic energy action space. Designing reward functions and selecting strategic optimization algorithms suitable for power grid control problems can make the system make control decisions more intelligently. Control decisions based on energy action space not only consider the optimization of system losses, but also comprehensively consider the stability of the power grid. This has Help maintain the safe operation of the power grid and improve the economic efficiency of electric energy utilization.

In summary, this embodiment comprehensively considers the system status, energy characteristics and control requirements, and performs real-time decision-making and iterative optimization in an intelligent manner, thereby comprehensively improving the efficiency of power grid control and making the power grid more secure, stable and efficient.

In an optional implementation, according to the energy action space, combined with system loss and grid stability, the reward function is determined as follows:

;

Among them, R(s,a) represents the reward value, s represents the system state, a represents the grid action, Loss(s,a) represents the system loss, α represents the weight coefficient, and VoltageStability(s,a) represents the voltage stability.

By using this function and minimizing system losses, the control system tends to improve the effective utilization of electric energy and reduce energy losses, thereby promoting the economy of the power grid. By introducing the voltage stability term, the reward function makes system control pay more attention to the power grid. stability. This helps prevent voltage instability caused by regulation decisions and improves the reliability of the power grid. The reward function takes into account two important factors: system loss and voltage stability. By adjusting the weight coefficient, the trade-off between economy and stability can be balanced. Such a comprehensive optimization goal helps to achieve multi-objective optimization of power grid regulation decisions. In summary, by using this function, the two aspects of economy and stability are fully considered. By weighing the relationship between the two, the power grid regulation system can achieve multi-objective optimization. With the support of advanced data, control decisions can be made more intelligently and comprehensively to further improve the control efficiency of the power grid.

Figure 2 is a schematic structural diagram of a power grid control system based on multi-level data according to an embodiment of the present invention. As shown in Figure 2, the system includes:

The first unit is used to obtain the initial data of the power grid, input the initial data of the power grid into a preset first load prediction model, preprocess the initial data of the power grid through the first load prediction model, and determine the first load input, According to the first load input, combined with the first load prediction model, a second load input is obtained;

The second unit is configured to transmit the second load input in the second load prediction model in the order of root node to leaf node according to the second load input and combined with the preset second load prediction model. , obtain the second load output, and determine the load prediction result according to the second load output in combination with the preset result optimization algorithm, where the second load model is a tree model constructed based on the extreme gradient boosting algorithm;

The third unit is used to construct an energy action space based on the load prediction results and combined with the initial data of the power grid. According to the energy action space, determine the control optimization strategy through a preset strategy optimization algorithm, wherein the strategy optimization algorithm It is built based on the improved proximal strategy optimization algorithm.

The invention may be a method, apparatus, system and/or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions thereon for performing various aspects of the invention.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, but not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: The technical solutions described in the foregoing embodiments can still be modified, or some or all of the technical features can be equivalently replaced; and these modifications or substitutions do not deviate from the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention. scope.

Claims (7)

1. A power grid control method based on multi-level data, characterized by including: Obtain the initial data of the power grid, input the initial data of the power grid into a preset first load prediction model, preprocess the initial data of the power grid through the first load prediction model, determine the first load input, and according to the first load prediction model A load input, combined with the first load prediction model, to obtain a second load input; The initial data of the power grid is input into a preset first load prediction model, and the initial data of the power grid is preprocessed by the first load prediction model to determine the first load input. According to the first load input , combined with the first load prediction model, the second load input includes: Acquire historical power consumption data through the database, obtain power grid status data according to smart sensors installed in power grid nodes, add the historical power consumption data and the power grid status data to the same set, and record them as the power grid initial data; Perform data cleaning and feature selection operations on the initial data of the power grid through the first load prediction model to obtain the first load input; According to the first load input, determine the forward hidden state and the backward hidden state through the bidirectional state network in the first load prediction model; According to the forward hidden state and the backward hidden state, combined with the preset residual module and self-attention mechanism, the input data of a specific layer in the bidirectional state network is directly added to the output of the bidirectional state network In the data, determine the second load input; According to the second load input, combined with the preset second load prediction model, the second load input is transferred in the second load prediction model in the order of root node to leaf node to obtain the second load output. , according to the second load output, combined with the preset result optimization algorithm, determine the load prediction result, wherein the second load model is a tree model constructed based on the extreme gradient boosting algorithm; According to the second load input, combined with the preset second load prediction model, the second load input is transferred in the second load prediction model in the order of root node to leaf node to obtain the second load prediction model. Load output, based on the second load output, combined with the preset result optimization algorithm, determines the load prediction result, wherein the second load model is a tree model constructed based on the extreme gradient boosting algorithm, including: Obtain the second load input, add the second load input to the second load prediction model, and pass the second load input downward from the root node of the second load prediction model, at each leaf At the node, the transmission direction is determined according to the preset node splitting conditions. If the preset node splitting conditions are met, the transmission is to the left. If not, the transmission is to the right until the transmission is completed, and the second load output is obtained. ; According to the second load output, combined with the result optimization algorithm, the abnormal values in the second load output are removed and corrected according to the data trend to determine the load prediction result; according to the load prediction result, combined with the power grid initialization data, construct an energy action space, and determine the control and optimization strategy through a preset strategy optimization algorithm based on the energy action space, wherein the strategy optimization algorithm is constructed based on an improved proximal strategy optimization algorithm; The energy action space is constructed based on the load prediction results and the initial data of the power grid. According to the energy action space, the control optimization strategy is determined through a preset strategy optimization algorithm, including: Obtain the load prediction result and the power grid initial data, define a state space according to the load prediction result and the power grid initial data, determine a first adjustment range and a second adjustment range according to the state space, and determine the first adjustment range and the second adjustment range according to the state space. An adjustment range and the second adjustment range are combined with the pre-obtained power supply unit characteristics to determine the energy action space; According to the energy action space, combined with system losses and grid stability, a reward function is determined. According to the reward function, a grid action is selected according to the policy optimization algorithm at each time step. According to the grid action, the state is iterated space and the energy action space. According to the iterated state space and energy action space, combined with the preset strategy optimization algorithm, the control optimization strategy is determined. 2. The method of claim 1, further comprising training the second load prediction model: Obtain the second load input and convert the second load input into a feature vector. For continuous feature values, divide the continuous feature values into discrete intervals through the equal-frequency bucketing strategy and perform discretization to obtain the feature value range. According to the feature value range, a histogram is constructed through a lightweight gradient boosting algorithm; Based on the leaf node priority growth strategy, the first gain value of the current leaf node is determined through the node gain function, the node with the maximum gain among the current leaf nodes is selected for splitting, the maximum gain value is obtained, and a loss function is constructed based on the maximum gain value. According to The loss function calculates the gradient and second-order derivative of the current leaf node, and constructs a second-order Taylor expansion of the objective function; For each leaf node, traverse the eigenvalues of all leaf nodes, combine the gradient and second-order derivative of the current leaf node, determine the leaf node with the smallest loss function value after splitting, record it as a split leaf node, and split the split leaf node into Two leaf nodes, and calculate the gradient and second-order derivative of the leaf node after splitting, and repeat the operation until the stopping condition is met. 3. The method according to claim 2, characterized in that the first gain value of the current leaf node is determined through the node gain function, and the node with the maximum gain among the current leaf nodes is selected for splitting to obtain the maximum gain value and based on The maximum gain value constructs the loss function as shown in the following formula: ; Among them, D represents the data set of the current node, Loss(D) represents the loss function, y _i represents the real label of sample i , y _avg represents the mean value of the current node, y _split represents the node mean value after splitting, max _split () represents the Choose the splitting method that maximizes the loss function among all possible methods. 4. The method according to claim 1, characterized in that the reward function is determined according to the energy action space, combined with system loss and power grid stability, as shown in the following formula: ; Among them, R(s,a) represents the reward value, s represents the system state, a represents the grid action, Loss(s,a) represents the system loss, α represents the weight coefficient, and VoltageStability(s,a) represents the voltage stability. 5. A power grid control system based on multi-level data, used to implement the multi-level data-based power grid control method according to any one of the preceding claims 1-4, characterized in that it includes: The first unit is used to obtain the initial data of the power grid, input the initial data of the power grid into a preset first load prediction model, preprocess the initial data of the power grid through the first load prediction model, and determine the first load input, According to the first load input, combined with the first load prediction model, a second load input is obtained; The second unit is configured to transmit the second load input in the second load prediction model in the order of root node to leaf node according to the second load input and combined with the preset second load prediction model. , obtain the second load output, and determine the load prediction result according to the second load output in combination with the preset result optimization algorithm, where the second load model is a tree model constructed based on the extreme gradient boosting algorithm; The third unit is used to construct an energy action space based on the load prediction results and combined with the initial data of the power grid. According to the energy action space, determine the control optimization strategy through a preset strategy optimization algorithm, wherein the strategy optimization algorithm It is built based on the improved proximal strategy optimization algorithm. 6. An electronic device, characterized in that it includes: processor; Memory used to store instructions executable by the processor; Wherein, the processor is configured to call instructions stored in the memory to execute the method according to any one of claims 1 to 4. 7. A computer-readable storage medium with computer program instructions stored thereon, characterized in that when the computer program instructions are executed by a processor, the method of any one of claims 1 to 4 is implemented.