CN110543929B - Wind speed interval prediction method and system based on Lorenz system - Google Patents

Abstract

The invention discloses a wind speed interval prediction method and system based on the Lorenz system. The prediction method includes: obtaining the original wind speed sequence; performing variational mode decomposition (VMD) on the wind speed sequence to obtain the denoised sequence and noise remainder; establishing a long-short-term neural network prediction model (LSTM), and denoising the Preliminary predictions are made in sequence to obtain preliminary prediction results; by defining wind speed ramp events (WSR) and wind speed ramp rates, the preliminary prediction results are revised to obtain revised wind speed prediction results; the influence of the atmospheric dynamic system on wind speed is described by the Lorenz equation , and obtain the Lorenz disturbance sequence (LDS); fit the LDS through B-spline interpolation method, and fix the confidence interval of the fitting result to obtain the upper and lower limits of the wind speed disturbance interval; combine the corrected wind speed prediction results and the wind speed disturbance interval The sum is performed to obtain the interval prediction result of wind speed; the wind speed interval prediction method or system of the present invention significantly improves the accuracy and reliability of the prediction model, and can obtain high-precision prediction results.

Description

A wind speed interval prediction method and system based on Lorenz system

Technical field

The invention relates to the field of wind speed prediction, and in particular to a wind speed interval prediction method and system based on the Lorenz system.

Background technique

In recent years, the global energy situation has become increasingly severe and energy demand has continued to increase. As a clean and renewable energy, wind energy has attracted much attention in new energy applications around the world, and wind power grid connection technology has become a hot spot in international research. The International Energy Agency (IEA) reports that wind power generation is expected to increase by 325 million kilowatts in 2019. However, wind fluctuations often lead to instability in the power system after wind power integration. Therefore, effective wind speed prediction can promote the intelligent development of the wind power industry, and accurate wind speed prediction is an important prerequisite for large-scale development and utilization of wind power.

Contents of the invention

The purpose of the present invention is to provide a wind speed interval prediction method and system based on the Lorenz system to obtain high-precision prediction results while improving the reliability of the prediction model.

In order to achieve the above objects, the present invention provides the following solutions:

A wind speed interval prediction method based on Lorenz system, the method includes:

Obtain the original wind speed sequence; perform variational mode decomposition (VMD) on the wind speed sequence to obtain the denoised sequence and noise remainder;

Establish a long-short-term neural network prediction model (LSTM) to make preliminary predictions on the denoised sequence and obtain preliminary prediction results;

By defining the wind speed ramp event (WSR) and wind speed ramp rate, the preliminary forecast results are revised to obtain the revised wind speed forecast results;

The Lorenz equation is used to describe the impact of the atmospheric dynamic system on wind speed, and the Lorenz disturbance sequence (LDS) is obtained;

Fit the LDS through the B-spline interpolation method, and fix the confidence interval for the fitting result to obtain the upper and lower limits of the wind speed disturbance interval;

The corrected wind speed prediction results and wind speed disturbance intervals are summed to obtain the wind speed interval prediction results.

Optionally, obtaining the denoised sequence and noise remainder specifically includes:

Get raw wind speed data;

Set the number of decompositions; iteratively calculate the set containing all modes and their center frequencies under the determined number of decompositions; obtain the intrinsic mode function components and noise remainder of the wind speed signal;

Reconstruct the intrinsic mode function components of the wind speed signal to obtain the denoised wind speed sequence.

Optionally, establishing a long-term and short-term neural network prediction model, making preliminary predictions on the denoised sequence, and obtaining preliminary prediction results. The specific process includes:

Divide the denoised wind speed sequence into a training set and a test set according to the ratio of 9:1;

Set the network structure of the long-term and short-term neural network, including the number of input layer neurons, the number of hidden layer neurons and the number of output layer neurons;

Training is performed by inputting the training set into the long-short-term neural network;

Input the test set into the trained long-short-term neural network to obtain preliminary prediction results of wind speed.

Optionally, modify the preliminary prediction results for the defined wind speed ramp event and wind speed ramp rate to obtain the revised wind speed prediction results, which specifically includes:

Calculate the wind speed gradient; introduce the definition of wind speed ramp according to the wind speed gradient;

When the absolute value of the gradient is greater than threshold 1, and the positive gradient at the current moment increases beyond the positive gradient threshold 2 or the negative gradient decreases beyond the negative gradient threshold 3, use the error at the current moment to correct the wind speed prediction value at the next moment;

For threshold 1, threshold 2 and threshold 3, it is converted into a multi-objective optimization problem, with the goal of minimizing the root mean square error, and the particle swarm algorithm (PSO) is used to solve the optimization problem.

Optionally, describe the influence of the atmospheric dynamic system on wind speed through the Lorenz equation, and obtain the LDS, which specifically includes:

Given the initial conditions (0, 1, 1), solve the Lorenz equation and obtain the three-dimensional LDS;

According to the Chebyshev distance, the three-dimensional LDS is converted into a one-dimensional perturbation sequence.

Optionally, the LDS is fitted through the B-spline interpolation method, and the confidence interval is fixed for the fitting result to obtain the upper and lower limits of the wind speed disturbance interval, which specifically includes:

Fit the distribution of LDS through B-spline difference and obtain the B-spline difference fitting function;

Fix the confidence intervals at 90% and 98% respectively, and calculate the upper and lower quantile points of the fitting function;

The upper quantile is set as the upper limit of the interval prediction, and the lower quantile is set as the lower limit of the interval prediction.

Optionally, the modified wind speed prediction result and the wind speed disturbance interval are summed to obtain the wind speed interval prediction result, which specifically includes:

By adding and subtracting the upper and lower quantiles under the 90% confidence interval to the corrected wind speed prediction result, the wind speed interval prediction result under the 90% confidence interval is obtained;

By adding and subtracting the upper and lower quantiles under the 98% confidence interval to the corrected wind speed prediction result, the wind speed interval prediction result under the 98% confidence interval is obtained.

The invention also provides a wind speed interval prediction system based on the Lorenz system. The prediction system includes:

The wind speed data acquisition and denoising module is used to obtain the original wind speed sequence; perform variational mode decomposition on the wind speed sequence to obtain the denoised sequence and noise remainder;

The wind speed data preliminary prediction module is used to establish a long- and short-term neural network prediction model, perform preliminary predictions on the denoised sequence, and obtain preliminary prediction results;

The preliminary prediction result correction module based on wind speed climbing is used to modify the preliminary prediction results by defining wind speed climbing events and wind speed climbing rates, and obtain revised wind speed prediction results;

Obtain the LDS module, which is used to describe the impact of the atmospheric dynamic system on wind speed through the Lorenz equation, and obtain the LDS;

Obtain the upper and lower limits of the disturbance module, which is used to fit the LDS through the B-spline interpolation method, fix the confidence interval for the fitting results, and obtain the upper and lower limits of the wind speed disturbance interval;

The prediction module is used to sum the corrected wind speed prediction results and wind speed disturbance intervals to obtain the wind speed interval prediction results.

Optionally, the wind speed data acquisition and denoising module specifically includes:

Obtain the eigenmodal component unit, which is used to obtain the original wind speed data; set the number of decompositions; iteratively calculate the set containing all modes and their center frequencies under the determined number of decompositions; obtain the eigenmodes of the wind speed signal State function components and noise remainders;

Obtain the denoised wind speed sequence unit, which is used to reconstruct the intrinsic mode function components of the wind speed signal and obtain the denoised wind speed sequence.

Optionally, the preliminary wind speed data prediction module specifically includes:

Divide and divide the wind speed sequence unit, which is used to divide the denoised wind speed sequence into a training set and a test set according to a ratio of 9:1;

Set the network structural unit of the long-term and short-term neural network, which is used to include the number of input layer neurons, the number of hidden layer neurons and the number of output layer neurons;

The training model unit is used for training by inputting the training set into the long and short-term neural network;

The prediction unit is used to input the test set into the trained long and short-term neural network to obtain preliminary prediction results of wind speed.

Optionally, the preliminary prediction result correction module based on wind speed climbing specifically includes:

Define the wind speed ramping unit, which is used to calculate the wind speed gradient; introduce the definition of wind speed ramping based on the wind speed gradient;

Define the correction method unit, which is used to correct the next moment with the error of the current moment when the absolute value of the gradient is greater than the threshold A and the positive gradient at the current moment increases beyond the positive gradient threshold B or the negative gradient decreases beyond the negative gradient threshold C. wind speed prediction value;

Solving the threshold unit is used to convert the thresholds A, B and C into a multi-objective optimization problem, with the goal of minimizing the root mean square error, and use PSO to solve the optimization problem.

Optionally, the obtaining LDS module specifically includes:

Obtain the three-dimensional LDS unit, which is used to give the initial conditions (0, 1, 1), solve the Lorenz equation, and obtain the three-dimensional LDS;

Obtain the one-dimensional LDS unit, which is used to convert the three-dimensional LDS into a one-dimensional LDS based on the Chebyshev distance.

Optionally, the module for obtaining the upper and lower limits of disturbance specifically includes:

The fitting unit is used to fit the distribution of LDS through B-spline difference and obtain the B-spline difference fitting function;

Quantile point calculation unit, used to fix the confidence intervals at 90% and 98% respectively, and calculate the upper and lower quantile points of the fitting function;

The unit for determining the upper and lower limits of the interval prediction is used to set the upper quantile point as the upper limit of the interval prediction and the lower quantile point as the lower limit of the interval prediction.

Optionally, the prediction module specifically includes:

The interval prediction result unit under the 90% confidence interval is used to obtain the wind speed interval prediction result under the 90% confidence interval by adding and subtracting the upper and lower quantiles under the 90% confidence interval to the corrected wind speed prediction result;

The 98% confidence interval lower interval prediction result unit is used to obtain the wind speed interval prediction result under the 98% confidence interval by adding and subtracting the upper and lower quantiles under the 98% confidence interval to the corrected wind speed prediction result.

Compared with the existing technology, the present invention has the following technical effects:

The wind speed prediction method and system of the present invention are wind speed interval prediction processes based on the Lorenz system. First, the signal decomposition technology VMD is used to perform the noise reduction process. Secondly, the long-short-term neural network is used to predict the noise-reduced data. Then, the wind speed gradient is introduced to correct the preliminary prediction results and obtain the revised wind speed prediction results. On this basis, the Lorenz perturbation theory is introduced to describe the atmospheric dynamic system, the Lorenz equation is solved to obtain the LDS, and B-spline interpolation is used to fit the distribution of the LDS. By fixing different confidence intervals of the fitting function, the upper limit and sum of the interval predictions are obtained. lower limit. This invention considers the influence of the atmospheric dynamic system on wind speed, and proposes a wind speed interval prediction method that effectively improves the accuracy of the short-term wind speed prediction model and the reliability of the prediction results.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the drawings of the present invention. Embodiments, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without exerting creative efforts.

Figure 1 is a flow chart of the wind speed interval prediction method based on the Lorenz system in Embodiment 1 of the present invention;

Figure 2 is a flow chart of the prediction method provided by Embodiment 2 of the present invention;

Figure 3 is a structural block diagram of a prediction system provided by Embodiment 3 of the present invention;

Figure 4 is a schematic diagram of the prediction results of six single models of the present invention on data set 1;

Figure 5 is a schematic diagram of the prediction results of six single models of the present invention on data set 2;

Figure 6 is a schematic diagram of the modified prediction results of the present invention on data set 1;

Figure 7 is a schematic diagram of the modified prediction results of the present invention on data set 2;

Figure 8 is a schematic diagram of the interval prediction results of the present invention on data set 1 when the confidence interval is 90%;

Figure 9 is a schematic diagram of the interval prediction results of the present invention on data set 1 when the confidence interval is 98%;

Figure 10 is a schematic diagram of the interval prediction results of the present invention on data set 2 when the confidence interval is 90%;

Figure 11 is a schematic diagram of the interval prediction results of the present invention on data set 2 when the confidence interval is 98%.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all the 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.

In order to make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

Figure 1 is a schematic flow chart of the wind speed interval prediction method of the present invention. As shown in Figure 1, the wind speed interval prediction method based on the Lorenz system includes the following steps:

Step 11: Obtain the original wind speed sequence, perform variational mode decomposition on the wind speed sequence, and obtain the denoised sequence and noise remainder.

Currently, VMD can filter out sequences with frequency characteristics, leaving a cluttered noise part. In the present invention, the signal processing method variational mode decomposition method is used to decompose and reconstruct the wind speed, and perform denoising to obtain the denoising sequence and the noise remainder. Specific steps are as follows:

Step 111: The denoising sequence constructs a constrained variational model under the current modal number: first, for each modal component (A _K (t) and/> Represent the instantaneous amplitude and instantaneous phase of u _k (t) respectively,/> is a non-decreasing function, t is the sampling time), and its analytical signal is obtained through Hilbert transform; then, the center frequency of each analytical signal is estimated, and the spectrum of each analytical signal is transformed to the base frequency band; finally, Gaussian smoothing of the frequency-shifted signal is used The index estimates the bandwidth of each modal component and constructs a constrained variational model as shown in Equation (1).

In the formula, is the center frequency of the component signal u _k (t),/> It means finding the partial derivative of t for the function, j ² =-1, and x(t) is the denoising sequence.

Step 112: Introduce the quadratic penalty factor α and the Lagrange multiplication operator λ(t) to change the constrained problem into a non-constrained problem. The extended Lagrange expression is as follows:

Step 113: Obtain equation (3) through Fourier isometric transform and other processes to achieve adaptive decomposition of the signal.

in, They are/> λ(t), Fourier transform of x(t). The center frequency is updated by the following formula:

Step 114: Iteratively update until convergence meets the following conditions:

Step 12: Establish a long-term and short-term neural network prediction model, conduct preliminary predictions on the denoised sequence, and obtain preliminary prediction results.

With the continuous development of deep learning technology, the concept of time series is applied in the structural design of recurrent neural network (RNN). Therefore, RNN has good time series analysis capabilities. As an improved RNN, the long-short-term neural network inherits RNN's ability to analyze time series data and makes up for RNN's shortcomings in long-term memory. Since the long-short-term neural network model can effectively maintain a long-term memory, it has also achieved certain results in the field of wind speed prediction.

Step 121: Divide the denoised wind speed sequence into a training set and a test set according to 9:1, with the first 90% of the data as the training set and the last 10% of the data as the test set;

Step 122: Determine the structure of the long-term and short-term neural network. The number of input layer neurons is 3, the number of hidden layer neurons is 5, and the number of output layer neurons is 1;

Step 123: Use the data of the training set and the test set to construct the input matrix according to formula (6),

v ₁ , v ₂ , v ₃ ... represent the 1st, 2nd, 3rd... wind speed values;

Step 124: Input the constructed training set input matrix into the long and short-term neural network by rows, train the long and short-term neural network, and then input the constructed test set input matrix into the trained long and short-term neural network by rows. , to obtain preliminary prediction results of wind speed.

Step 13: Modify the preliminary forecast results by defining wind speed ramp events and wind speed ramp rates to obtain revised wind speed forecast results, including:

Step 131: Define the wind speed ramp event according to Equation 7,

|v(t ₀ +Δ _t )-v(t ₀ )|/Δ _t ＞v _val (7)

v(t ₀ ) is the wind speed value at time t ₀ , Δ _t is the time interval, and v _val is the threshold value of wind speed ramp;

Step 132: Define the wind speed gradient according to Equation 8,

k(i)＝(v(i)-v(i-1))/imterval (8)

v(i) is the i-th wind speed value, interval is the time interval;

Step 133: Determine whether the i-th or i-1 wind speed value is greater than the wind speed climbing threshold WRR _val , that is, determine whether Formula 9 or Formula 10 is true.

|k(i-1)|＞WRR _val (9)

|k(i)|＞WRR _val (10)

If established, the wind speed climb correction is transformed into a multi-objective optimization problem. When the forward slope increases to a certain extent and exceeds the forward threshold WRR _up , that is, Formula 11,

Or the negative climb is reduced to a certain extent and exceeds the negative threshold WRR _down , that is, Formula 12,

Correct the wind speed prediction value at the next moment according to Equation 13,

Represents the wind speed prediction value before wind speed ramp correction,/> Represents the wind speed prediction value after wind speed ramp correction;

Step 134: If neither Formula 9 nor Formula 10 holds true, the wind speed prediction value at this time will not be revised, that is, it will remain unchanged before and after the correction.

Step 135: For the determination of the three thresholds, use the particle swarm algorithm (PSO);

PSO was first proposed by Eberhart and Kennedy in 1995. Its basic concept originated from the study of bird foraging behavior. Use a particle to simulate the above-mentioned individual bird. Each particle can be regarded as a search individual in the N-dimensional search space. The current position of the particle is a candidate solution to the corresponding optimization problem, and the flight process of the particle is the individual. The search process. The flight speed of particles can be dynamically adjusted according to the historical optimal position of the particle and the historical optimal position of the population. The particle has only two attributes: speed and position. Speed represents the speed of movement, and position represents the direction of movement. The optimal solution that each particle searches for individually is called the individual extreme value, and the best individual extreme value in the particle swarm is regarded as the current global optimal solution. Continuously iterate, updating speed and position. Finally, the optimal solution that satisfies the termination condition is obtained.

Step 1351: Initialization, first set the maximum number of iterations to 5, the number of independent variables of the objective function to 3 WRR _up , WRR _down , and WRR _val , the maximum speed of the particles to 1.5, the learning factor to 1, and the position information to the entire In the search space, we randomly initialize the speed and position in the speed interval and search space, set the particle swarm size to 10, and randomly initialize a flying speed for each particle.

Step 1352: Solve the individual extreme value and the global optimal solution. Define the objective function as the root mean square error between the predicted wind speed and the real wind speed. The individual extreme value is the optimal solution (pbest) found for each particle. From these optimal The solution finds a global value (gbest), which is called the global optimal solution. Compare with the historical global optimal and update.

Step 1353: Update velocity and position according to Equations 14 and 15,

v _i ＝v _i +c ₁ ×rand()×(pbest _i -x _i )+c ₂ ×rand()×(gbest _i -x _i ) (14)

x _i =v _i +x _i (15)

i=1, 2,...N, N is the total number of particles, _vi is the speed of the particle, rand() is a random number between (0, 1), x _i is the current position of the particle , c ₁ , c ₂ are learning factors;

Step 1354: The termination condition is that the set number of iterations or the difference between the algebras meets the minimum limit 1e10-8;

Step 14: Use the Lorenz equation to describe the impact of the atmospheric dynamic system on the wind speed and obtain the LDS.

In 1963, meteorologist Edward Lorenz calculated aperiodic phenomena from a deterministic equation (later known as the Lorenz equation). The Lorentz system is the earliest chaotic motion dissipative system discovered in numerical experiments. Its equation of state (Lorentz equation) is a simplified model of weather forecasting. Specific steps are as follows:

Step 141: Establish the Lorenz equation as shown in Equation 14, set the parameters σ = 10, b = 8/3, r = 28, the initial value is (0, 1, 1), and obtain the three-dimensional LDS;

Step 142: According to the Chebyshev distance formula, perform dimensionality reduction processing on the three-dimensional LDS and convert it into a one-dimensional LDS;

d(C _n -C ₀ )=max(|x _n -x ₀ |, |y _n -y ₀ |, |z _n -z ₀ |) (17)

Step 15: Fit the LDS through B-spline interpolation method, fix the confidence interval for the fitting result, and obtain the upper and lower limits of the wind speed disturbance interval.

Step 151: Draw a frequency histogram for LDS, and set the number of intervals in the frequency histogram to 25;

Step 152: Use B-spline interpolation method to fit the frequency histogram and obtain the fitting function;

Step 153: Set the confidence interval, and obtain the upper and lower confidence limits based on the fitting function. The upper confidence limit takes a positive value and the lower confidence limit takes a negative value to obtain the upper and lower limits of the wind speed disturbance interval;

Step 16: Sum the corrected wind speed prediction results and wind speed disturbance intervals to obtain the wind speed interval prediction results.

Step 161: Add the revised wind speed prediction result to the upper limit of the disturbance interval to obtain the upper limit of the wind speed interval prediction;

Step 162: Subtract the lower limit of the disturbance interval from the revised wind speed prediction result to obtain the lower limit of the wind speed interval prediction.

In order to verify that the method of the present invention has good prediction performance for wind speed data of actual wind farms, the wind speed of the Sotavento wind farm in Galicia, Spain, was used to conduct data simulation experiments. Figure 2 is a flow chart of the prediction method provided in this embodiment. As shown in Figure 2, the specific process includes:

Step 21: Obtain the original wind speed sequence

In order to better verify the prediction ability of the method proposed in the present invention, this embodiment selects two data sets from a seaside wind farm in Galicia, Spain to establish a prediction model. The time of data set 1 is from 00:00 on September 1, 2018 to 22:40 on September 7, 2018. The time of the second data set is from 08:00 on March 9, 2019 to 06:30 on March 16, 2019. They represent wind speeds in different seasons with time intervals of 10 minutes, with 1000 wind speeds in Dataset 1 and 996 wind speeds in Dataset 2. Since there are four scattered vacancies in Dataset 2, we use linear interpolation as shown in Equation (18) to complete them.

x _t represents the t-th moment, y _t represents the wind speed value corresponding to that moment, x represents any moment between x _t and x _t+i , and y is its corresponding unknown wind speed value.

For each data set, the first 900 data (90% of the entire data) are used as the training set, and the last 100 data are used as the test set. In other words, the forecast period is 16 hours and 30 minutes.

Step 22: Define error metrics

For deterministic prediction, this article selects the most commonly used mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE) for evaluation. Their calculation formula is shown below.

y _t represents the true wind speed at time t, Indicates the predicted wind speed at time t.

For interval prediction, both interval coverage (ratio of the number of intervals containing actual wind speed) and average interval diameter should be considered. Therefore, formula (23) and formula (24) are defined to describe the interval coverage and average interval diameter respectively.

N _t indicates whether the t-th interval covers the t-th true wind speed value. If N _t =1, it means that the t-th true wind speed is covered. If N _t =0, it means that the t-th true wind speed is not covered. R _cover represents the interval coverage, and d _average represents the average interval diameter. Obviously, the higher the coverage rate, the smaller the average diameter, and the better the interval prediction effect, which shows that the prediction accuracy is still high when the prediction interval is small.

Step 23: Deterministic prediction results and discussion

Step 231: First, compare the autoregressive moving average model (ARMA), support vector machine (SVM), gradient boosting decision tree (GBDT), extreme gradient boosting tree (XGBoost), backpropagation neural network (BP) and LSTM. The prediction results of six single wind speed prediction models and the errors on the two data sets were calculated, as shown in Tables 1 and 2.

Step 232: Determine the parameters of each model. The autocorrelation coefficient parameter p and partial autocorrelation coefficient parameter q of ARMA are determined according to the minimum information criterion (AIC). The most commonly used radial basis function is selected as the core function of SVM. The learning rate and number of training sets of GBDT and XGBOOST were obtained by minimizing MAE respectively, and the number of boosting trees for both models was set to 100. The structure of BP is three layers, and the number of nodes is 3-30-1. LSTM is also set to three layers, and the number of nodes is 3-5-1. This article sets the number of hidden layer nodes in LSTM to be smaller than BP, because too many nodes will increase the calculation time, especially for LSTM (in our programming environment, when using six nodes, it will consume more than one minute) , which will greatly increase the calculation time of the entire model), and even if LSTM has fewer hidden layer nodes, its prediction performance is still better than BP.

Table 1 Single model prediction error on data set 1

Table 2 Single model prediction error on data set 2

Step 233: Analysis of table calculation results. As can be seen from Table 1 and Table 2, all error indicators of the LSTM neural network are the smallest on any data set. As can be seen from Table 1, GBDT has the largest error. Compared with GBDT, the MSE of LSTM is reduced by 69.83%, MAE is reduced by 46.78%, RMSE is reduced by 45.08%, and MAPE is reduced by 53.85%. In other words, LSTM has the highest prediction accuracy among commonly used single prediction models. It can be seen from Table 2 that the support vector machine has the largest error. Compared with SVM, the MSE of LSTM is reduced by 84.63%, MAE is reduced by 61.09%, RMSE is reduced by 60.79%, and MAPE is reduced by 67.47%. That is, LSTM has the highest prediction accuracy compared to the baseline prediction model.

Step 234: The prediction results of the six models on the two data sets are shown in Figure 4 and Figure 5 respectively. As can be clearly seen from Figures 4 and 5, there is an obvious one-step lag in the prediction results of LSTM, so we denoised the data and modified the prediction results based on WSR, thus improving the LSTM. The prediction errors of the two data sets are shown in Table 3 and Table 4 respectively.

Table 3 Comparison of corrected prediction errors on data set 1

Table 4 Comparison of corrected prediction errors on data set 1

In Tables 4 and 5, we underline the minimum error achieved by WD-LSTM for different number of decomposition levels (lev) (lev is from 2 to 5 in both tables). For different numbers of IMFs (the range is 4 to 8 in Table 4 and the range is 3 to 7 in Table 5), we bold the minimum error achieved by VMD-LSTM. It can be seen from these two tables that when the number of WD decomposition layers is 2, the error of WD-LSTM reaches the minimum. Moreover, before WSR correction, most of the VMD-LSTM errors are smaller than the minimum error of WD-LSTM. This fully proves that the denoising effect of VMD is much better than that of WD. On these two data sets, after correcting the prediction results through WSR, the errors of each model were significantly reduced. The reduction ratio is between 5% and 30%. The minimum error on data set 1 is obtained by VMD-LSTM-PSOR (IMF=6), and the minimum error on data set 2 is obtained by VMD-LSTM-PSOR (IMF=4).

Step 235: Analysis of revised prediction results. Figures 6 and 7 are the prediction results of LSTM, VMD-LSTM and VMD-LSTM-PSOR for data set 1 when IMF=6 and data set 2 when IMF=4 respectively. As can be seen from these two figures, the solid line representing the LSTM prediction results has an obvious one-step lag. After VMD preprocessing, the fluctuations of the wind speed sequence are weakened and the wind speed curve becomes stable. Finally, WSR with particle swarm optimization is used to correct the prediction results so that the wind field is closer to the actual wind speed curve. The line representing the VMD-LSTM-PSOR prediction results is closest to the line representing the actual wind speed.

Step 24: Interval prediction results and discussion

Step 241: Fit LDS using KDE and B-spline interpolation respectively. Based on the deterministic prediction results, interval predictions of wind speed were obtained according to different confidence intervals. The coverage, average diameter and uncovered points are shown in Table 5.

Table 5 Analysis of interval prediction results

Step 242: Analysis of interval prediction results. Table 5 calculates the interval prediction results obtained by different fitting methods under 90% confidence interval and 98% confidence interval respectively. For Dataset 1, on the one hand, the mean diameters of KDE and B-spline interpolation are almost the same for the 90% confidence interval (1.5009m/s and 1.5075m/s respectively), but the coverage of B-splines is 2 higher than that of KDE %. B-spline interval prediction covers the 2nd and 74th points, but KDE does not. The situation is similar for the 98% confidence interval. The average diameter is almost the same, but the coverage of B-spline is 3% higher than KDE. B-spline interval prediction includes three points, namely points 19, 81 and 95, but KDE does not. The results show that when the confidence intervals are the same and the average diameters are basically the same, the fitting effect of KDE on LDS is not as good as that of B-spline interpolation. The confidence interval of B-spline interpolation can cover more real wind speed points. On the other hand, for the same fitting method, the larger the confidence interval, the larger the average diameter, which is consistent with statistical principles but cannot lead to higher coverage. Because a higher confidence interval means that the interval prediction may have a higher upper limit and a higher lower limit, the coverage will also change (for the KDE-fitted LDS, the 19th, 81st, and 85th are covered at the 90% confidence interval point and the 98% confidence interval does not. Conversely, the 98% confidence interval covers point 84 but the 90% confidence interval does not. This situation is similar to B-spline interpolation fitting LDS). This means that fitting LDS can achieve higher coverage and better interval prediction results without using high confidence intervals.

Step 243: We plotted the results of interval prediction on the two data sets, as shown in Figure 8, Figure 9, Figure 10, and Figure 11 respectively. Figures 8 and 10 plot the interval prediction results with a 90% confidence interval, and Figures 9 and 11 plot the interval prediction results with a 98% confidence interval. The interval prediction results cover most of the actual wind speed points, indicating that the wind speed interval prediction based on LDS is effective. The prediction results obtained by fitting LDS with B-spline interpolation are better than the LDS fitted by KDE. The latter is a more effective method. The method is consistent with that shown in Table 5.

Figure 3 is a schematic structural diagram of the wind speed prediction system of the present invention. As shown in Figure 3, the prediction system includes:

The wind speed data acquisition and denoising module 31 is used to obtain the original wind speed sequence; perform variational mode decomposition on the wind speed sequence to obtain the denoised sequence and noise remainder.

The wind speed data acquisition and denoising module 31 specifically includes:

Get raw wind speed data;

Set the number of decompositions; iteratively calculate the set containing all modes and their center frequencies under the determined number of decompositions; obtain the intrinsic mode function components and noise remainder of the wind speed signal;

Reconstruct the intrinsic mode function components of the wind speed signal to obtain the denoised wind speed sequence.

The wind speed data preliminary prediction module 32 is used to establish a long- and short-term neural network prediction model, perform preliminary predictions on the denoised sequence, and obtain preliminary prediction results.

The wind speed data preliminary prediction module 32 specifically includes:

Divide the denoised wind speed sequence into a training set and a test set according to the ratio of 9:1;

Set the network structure of the long-term and short-term neural network, including the number of input layer neurons, the number of hidden layer neurons and the number of output layer neurons;

Training is performed by inputting the training set into the long-short-term neural network;

Input the test set into the trained long-short-term neural network to obtain preliminary prediction results of wind speed.

The preliminary prediction result correction module 33 based on wind speed ramping is used to correct the preliminary prediction results by defining wind speed ramping events and wind speed ramping rates, and obtain revised wind speed prediction results;

The preliminary prediction result correction module 33 based on wind speed climbing specifically includes:

Calculate the wind speed gradient; introduce the definition of wind speed ramp according to the wind speed gradient;

When the absolute value of the gradient is greater than the threshold A, and the positive gradient at the current moment increases beyond the positive gradient threshold B or the negative gradient decreases beyond the negative gradient threshold C, the error at the current moment is used to correct the wind speed prediction value at the next moment;

For the thresholds A, B and C, it is converted into a multi-objective optimization problem, with the goal of minimizing the root mean square error, and PSO is used to solve the optimization problem.

Obtain the LDS module 34, which is used to describe the influence of the atmospheric dynamic system on the wind speed through the Lorenz equation and obtain the LDS.

Obtain LDS module 34, including:

Given the initial conditions (0, 1, 1), solve the Lorenz equation and obtain the three-dimensional LDS;

According to the Chebyshev distance, the three-dimensional LDS is converted into a one-dimensional perturbation sequence.

The module 35 for obtaining the upper and lower limits of the disturbance is used to fit the LDS through the B-spline interpolation method, fix the confidence interval for the fitting result, and obtain the upper and lower limits of the wind speed disturbance interval.

The module 35 for obtaining the upper and lower limits of disturbance specifically includes:

Fit the distribution of LDS through B-spline difference and obtain the B-spline difference fitting function;

Fix the confidence intervals at 90% and 98% respectively, and calculate the upper and lower quantile points of the fitting function;

The upper quantile is set as the upper limit of the interval prediction, and the lower quantile is set as the lower limit of the interval prediction.

The prediction module 36 is used to sum the revised wind speed prediction result and the wind speed disturbance interval to obtain the interval prediction result of wind speed.

The prediction module 36 specifically includes:

By adding and subtracting the upper and lower quantiles under the 90% confidence interval to the corrected wind speed prediction result, the wind speed interval prediction result under the 90% confidence interval is obtained;

By adding and subtracting the upper and lower quantiles under the 98% confidence interval to the corrected wind speed prediction result, the wind speed interval prediction result under the 98% confidence interval is obtained.

Each embodiment in this specification is described in a progressive manner. Each embodiment focuses on its differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For relevant details, please refer to the description in the method section.

This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method and the core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the present invention There will be changes in the specific implementation methods and application scope of the ideas. In summary, the contents of this description should not be construed as limitations of the present invention.

Claims (10)

1. A short-term wind speed prediction method based on wind speed characteristics, characterized in that the prediction method includes: obtaining an original wind speed sequence; performing variational mode decomposition (VMD) on the wind speed sequence to obtain a denoised sequence and a noise remainder; Establish a long-short-term neural network prediction model LSTM to make preliminary predictions on the denoised sequence and obtain preliminary prediction results; by defining the wind speed climbing event WSR and wind speed climbing rate, the preliminary prediction results are corrected and the revised wind speed prediction results are obtained ; Describe the influence of the atmospheric dynamic system on wind speed through the Lorenz equation, and obtain the Lorenz disturbance sequence LDS; fit the LDS through the B-spline interpolation method, and fix the confidence interval of the fitting result to obtain the upper and lower limits of the wind speed disturbance interval; Sum the corrected wind speed prediction results and wind speed disturbance intervals to obtain the wind speed interval prediction results; Wherein, obtaining the denoising sequence and noise remainder specifically includes: obtaining original wind speed data; setting the number of decompositions; iteratively calculating a set containing all modes and their center frequencies under the determined number of decompositions; obtaining The intrinsic mode function components and noise remainder of the wind speed signal; reconstruct the intrinsic mode function components of the wind speed signal to obtain the denoised wind speed sequence; The above-mentioned definition of wind speed climbing event and wind speed climbing rate is used to correct the preliminary prediction results and obtain the revised wind speed prediction results, which specifically includes: calculating the wind speed gradient; introducing the definition of wind speed climbing according to the wind speed gradient; when the absolute value of the gradient is greater than Threshold 1, and when the positive gradient at the current moment increases beyond the positive gradient threshold 2 or the negative gradient decreases beyond the negative gradient threshold 3, the error at the current moment is used to correct the wind speed prediction value at the next moment; for threshold 1, threshold 2 and threshold 3. Convert it into a multi-objective optimization problem, with the goal of minimizing the root mean square error, and use the particle swarm algorithm (PSO) to solve the optimization problem. 2. The short-term wind speed prediction method according to claim 1, characterized in that: establishing a long-term and short-term neural network prediction model, making preliminary predictions on the denoising sequence, and obtaining preliminary prediction results. The specific process includes: denoising The wind speed sequence is divided into a training set and a test set according to the ratio of 9:1; set the network structure of the long-term and short-term neural network, including the number of input layer neurons, the number of hidden layer neurons and the number of output layer neurons; The training set is input into the long and short-term neural network for training; the test set is input into the trained long and short-term neural network to obtain the preliminary prediction results of wind speed. 3. The short-term wind speed prediction method according to claim 1, characterized in that the influence of the atmospheric dynamic system on the wind speed is described through the Lorenz equation, and the LDS is obtained, which specifically includes: given the initial conditions (0, 1, 1), solving Lorenz equation, obtain three-dimensional LDS; according to the Chebyshev distance, the three-dimensional LDS is converted into a one-dimensional perturbation sequence. 4. The short-term wind speed prediction method according to claim 1, characterized in that the LDS is fitted by B-spline interpolation method, and the confidence interval of the fitting result is fixed to obtain the upper limit and lower limit of the wind speed disturbance interval, Specifically, it includes: fitting the distribution of LDS through B-spline interpolation to obtain the B-spline interpolation fitting function; fixing the confidence intervals at 90% and 98% respectively, and calculating the upper and lower quantile points of the fitting function; The upper quantile is set as the upper limit of the interval prediction, and the lower quantile is set as the lower limit of the interval prediction. 5. The short-term wind speed prediction method according to claim 1, characterized in that the summing of the corrected wind speed prediction results and the wind speed disturbance interval to obtain the interval prediction result of the wind speed specifically includes: by adding a 90% confidence interval The upper and lower quantile points below are added and subtracted to the corrected wind speed prediction result to obtain the wind speed interval prediction result under the 90% confidence interval; by adding and subtracting the upper and lower quantile points under the 98% confidence interval to the corrected wind speed prediction result, The wind speed interval prediction results with 98% confidence interval are obtained. 6. A short-term wind speed prediction system, characterized in that the prediction system includes: a wind speed data acquisition and denoising module for obtaining the original wind speed sequence; performing variational mode decomposition on the wind speed sequence to obtain the denoising sequence and noise remainder; a wind speed data preliminary prediction module, used to establish a long and short-term neural network prediction model, perform preliminary predictions on the denoised sequence, and obtain preliminary prediction results; a preliminary prediction result correction module based on wind speed climbing, used to pass Define the wind speed ramp event and wind speed ramp rate, correct the preliminary prediction results, and obtain the revised wind speed prediction results; obtain the LDS module, which is used to describe the impact of the atmospheric dynamic system on the wind speed through the Lorenz equation, and obtain the LDS; obtain the disturbance upper The lower limit module is used to fit the LDS through the B-spline interpolation method, and fix the confidence interval of the fitting results to obtain the upper and lower limits of the wind speed disturbance interval; the prediction module is used to combine the revised wind speed prediction results and the wind speed disturbance interval Perform summation to obtain the interval prediction results of wind speed; The wind speed data acquisition and denoising module specifically includes: obtaining the intrinsic modal component unit for obtaining the original wind speed data; setting the number of decompositions; and iteratively calculating the set sum containing all modes under the determined number of decompositions. Their center frequencies; obtain the eigenmodal function components and noise remainder of the wind speed signal; obtain the denoised wind speed sequence unit, which is used to reconstruct the eigenmodal function components of the wind speed signal and obtain the denoised wind speed sequence; The preliminary prediction result correction module based on wind speed ramping specifically includes: defining a wind speed ramping unit for calculating the wind speed gradient; introducing the definition of wind speed ramping based on the wind speed gradient; defining a correction method unit for calculating the absolute value of the gradient is greater than the threshold 1, and the positive gradient at the current moment increases beyond the positive gradient threshold 2 or the negative gradient decreases beyond the negative gradient threshold 3, use the error at the current moment to correct the wind speed prediction value at the next moment; the threshold unit is used to solve Thresholds 1, 2 and 3, transform it into a multi-objective optimization problem, with the goal of minimizing the root mean square error, and use PSO to solve the optimization problem. 7. The short-term wind speed prediction system according to claim 6, characterized in that the preliminary prediction module of wind speed data specifically includes: a dividing and dividing wind speed sequence unit for dividing the denoised wind speed sequence according to a ratio of 9:1. Divide it into a training set and a test set; set the network structural unit of the long-term and short-term neural network, which is used to include the number of input layer neurons, the number of hidden layer neurons and the number of output layer neurons; the training model unit is used to The training set is input into the long-term and short-term neural network for training; the prediction unit is used to input the test set into the trained long-term and short-term neural network to obtain preliminary prediction results of wind speed. 8. The short-term wind speed prediction system according to claim 6, characterized in that the obtaining LDS module specifically includes: obtaining a three-dimensional LDS unit for given initial conditions (0, 1, 1), solving the Lorenz equation, and obtaining Three-dimensional LDS; obtains one-dimensional LDS unit, which is used to convert three-dimensional LDS into one-dimensional LDS based on the Chebyshev distance. 9. The short-term wind speed prediction system according to claim 6, characterized in that the module for obtaining the upper and lower limits of disturbance specifically includes: a fitting unit for fitting the distribution of LDS through B-spline interpolation and obtaining B samples. Bar interpolation fitting function; quantile point calculation unit, used to fix the confidence interval to 90% and 98% respectively, calculate the upper and lower quantile points of the fitting function; determine the upper and lower limit units of interval prediction, used to divide the upper quantile point Set as the upper limit of the interval prediction, and set the lower quantile as the lower limit of the interval prediction. 10. The short-term wind speed prediction system according to claim 6, wherein the prediction module specifically includes: a lower interval prediction result unit of the 90% confidence interval, which is used to add the upper and lower quantiles of the 90% confidence interval. Subtracted to the corrected wind speed prediction result, the wind speed interval prediction result under the 90% confidence interval is obtained; the interval prediction result unit under the 98% confidence interval is used to add and subtract the upper and lower quantiles under the 98% confidence interval to the corrected For the wind speed prediction results, the wind speed interval prediction results with a 98% confidence interval were obtained.