Short-term wind-photovoltaic power prediction method based on high-dimensional multivariate meteorological data fusion

Abstract

The present invention belongs to the field of new-energy power prediction of electric power systems. Disclosed is a short-term wind-photovoltaic power prediction method based on high-dimensional multivariate meteorological data fusion. In the present invention, an acquired meteorological data set ERA5 is fully utilized, a massive number of meteorological factors are greatly expanded in terms of physical causality, spatial scales, and temporal scales, and characteristic relationships between the meteorological factors and new-energy power are fully explored, thereby improving the prediction accuracy. Power-related meteorological factors are greatly expanded by using a physical causality method and a spatial longitude and latitude envelope method, so as to form a high-dimensional multivariate meteorological factor set. Dimensionality reduction is performed on ultra-high-dimensional meteorological data for comprehensive correlation analysis and fusion to prevent correlations and information overlaps between multi-dimensional meteorological data from affecting a power prediction model; and then, a sliding time window method is used to form a power prediction sample set. The prediction model is fitted according to meteorological data and actual power station operation data, and is applied to actual engineering projects in Dali, Yunnan for verification. The prediction accuracy of the present invention far exceeds that of conventional prediction methods, thus showing relatively strong robustness and universality.

Description

A short-term wind and solar power prediction method based on high-dimensional multivariate meteorological data fusion Technical Field

[0001] The present invention relates to the field of new energy power prediction for electric power systems, and in particular to a short-term wind and solar power prediction method based on high-dimensional multivariate meteorological data fusion. Background Art

[0002] Wind and solar power generation are characterized by intermittent, highly variable, and highly random characteristics, negatively impacting grid connection. Predicting wind and solar power is one way to mitigate this impact. Therefore, high-precision, low-latency wind power prediction technology has become a major research focus. As research on wind and photovoltaic power prediction continues to deepen, the following practical issues have emerged:

[0003] 1. Unstable data sample quality: Wind and solar power prediction models require extensive historical data samples, such as meteorological and equipment data, for training and validation. These data are affected by numerous factors, including meteorological, human, equipment, and scheduling. Due to the incompleteness and accuracy of my country's numerical weather prediction system, the limited number of data types monitored, the significant differences in the timescales of different data types, and the difficulty in obtaining power-related data, data quality is unstable and exacerbates the uncertainty of prediction results. Furthermore, due to the presence of numerous uncontrollable factors during the data collection process, the quality of uploaded output data from wind and solar power stations varies greatly, making it difficult to guarantee reliability and thus unusable. The dual instability of meteorological and output data sample quality severely impacts prediction model parameter estimation and future power forecasting. 2. Differences in model applicability: Regional weather conditions, geographical factors, applications with varying timescales, and varying power plant equipment and wind and photovoltaic farm layouts can also affect the applicability of prediction models. Customized prediction models with greater applicability are needed for different regions and sites. 3. Prediction accuracy does not meet practical requirements: The accuracy of wind and photovoltaic power generation predictions significantly impacts power generation efficiency, operating costs, and grid stability. The accuracy of wind and photovoltaic power forecasts has a significant impact on power generation efficiency, operating costs, and grid stability. However, current forecast accuracy falls short of actual requirements. Wind power forecast errors are particularly pronounced during peak and low load periods, easily leading to oversupply on the grid and making it difficult to formulate reasonable dispatch plans.

[0004] To address these issues, this paper, supported by the National Natural Science Foundation of China (52079014) and the Fundamental Research Funds for Central Universities (DUT22QN224, DUT22JC21), proposes a short-term wind and solar power forecasting method based on high-dimensional multivariate meteorological data fusion. This method is then tested on a wind and solar power station in Dali, Yunnan Province. Results demonstrate that this method effectively improves wind and solar power forecasting accuracy and exhibits strong robustness and universal applicability.

[0005] Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a short-term wind and solar power prediction method based on high-dimensional multivariate meteorological data fusion, which can effectively improve the wind and solar power prediction accuracy and has strong robustness, universality and practical value.

[0007] Technical solution of the present invention:

[0008] A short-term wind and solar power prediction method based on high-dimensional multivariate meteorological data fusion is characterized by the following steps:

[0009] (1) Construct a set of meteorological factors related to power generation power of power plants S m,k , expand the possible meteorological factors from the physical causes and spatial scales; assume that the longitude and latitude of the power station are (N, E), for the selected meteorological data set, if its latitude spatial resolution is n and its longitude spatial resolution is e; the power station is centered on itself and is enclosed in the spatial grid area (N-in, N+in) and (E-je, E+je), the spatial latitude and longitude resolution expansion scales are i and j, and the expression is as follows:

[0010] Where: m represents the meteorological factor sequence related to the power generation of the power station, which is M in total; k represents the number of meteorological points when the spatial resolution is expanded to scale i and j, which is K in total; according to the Cartesian product, the meteorological factor sequence set S m,k There are M×K meteorological factor sequences in total;

[0011] (2) A method for constructing a power prediction sample set is used. The meteorological factor sequence set is constructed using the power generation power-related meteorological factor sequence and the latitude and longitude location of the power station as feature inputs. Comprehensive correlation analysis, ultra-high-dimensional data dimensionality reduction and fusion, and time scale expansion are used to form a power prediction sample set that is independent of each other, covers the original information, and has an expanded time dimension. The specific steps are as follows:

[0012] Step 1. Calculate the high-dimensional multivariate meteorological factor set S related to the power generation of the power station according to formula (1) m,k , conduct correlation analysis on power generation of power station and meteorological factor sequence one by one, and take weighted average of correlation coefficients;

[0013] Where: ρ (Smk,f) Represents a meteorological factor sequence S in the meteorological factor sequence set mk The comprehensive correlation coefficient with the power station output sequence F; I is the number of selected correlation coefficients; the power station output sequence set F in the time series T = [f1, f2, ..., f T ] T and meteorological factor sequence set S m,k According to formula (2), the comprehensive correlation coefficient set of meteorological factors is obtained:

[0014] Screen out D meteorological factors with correlation coefficients greater than 0.6. If ρ mk ≥0.6, then S m,k ∈S D , and obtain the corresponding strongly correlated meteorological factor sequence set S D ={S1,…,S D}, the two-dimensional meteorological factor set S m,k Reduced to one-dimensional S D ;

[0015] Step 2. If S D The sample set of time series length T is X=[x1,x2,…,x T ] T , x i ∈R D , where x i is the i-th sample, T is the total number of samples, D is the characteristic dimension of the high-dimensional sample, that is, the number of meteorological factors that are strongly correlated with the output sequence, R D is a set of real numbers of D-dimensional strongly correlated meteorological factors; the d-dimensional sample set after dimensionality reduction is Y=[y1,y2,…,y T ] T ,y i ∈R d ; The goal is to minimize the residual variance μ of the sample set before and after dimensionality reduction, see formula (4); D X and D Y are the Euclidean distance matrices of X and Y respectively, and determine the optimal target dimension d;

[0016] Get the optimal d-dimensional meteorological factor data series set Y within the time series length T after dimensionality reduction d :

[0017] Step 3. Take the power station output itself as part of the power prediction sample set and combine it with the dimensionality-reduced meteorological factor data sequence set Y d Together they constitute the power prediction sample set Y d+1 =[y1,y2,...,y d,f], according to the time window w for Y d+1 Expand to form the power prediction input set Y d+1,in , corresponding to the prediction model output set Y out : Y out =[f w+1 ,f w+2 …,f T ] T (7)

[0018] Where: y w+i is a two-dimensional matrix of w×(d+1), corresponding to the output data f w+i ; where i = 1, 2, ..., Tw; Y out The power station output sequence set F is intercepted and the time sequence is aligned with the power forecast input set Y d+1,in correspond.

[0019] (3) Use the sample set to fit the input-output set, that is, the relationship between the output data at each moment t and the w×(d+1) power prediction input set in the previous w time broadband windows. Then, input any moment i y w+i The input set can output the output data f at that moment w+i。

[0020] Where: W in is a weight matrix, B in is a bias matrix.

[0021] The results of the present invention have the following beneficial effects: (1) high accuracy, which is much higher than the common machine learning algorithms such as BP artificial neural network, XGBoost and random forest, and can meet the actual use needs; (2) physically explainable, deeply explore the relationship between power generation and various related meteorological factors, find that there is a lag and instantaneous relationship between strong meteorological factors and power, and explain it with the geographical environment in which they are located; (3) good robustness, with little impact from abnormal data points, and is not afraid of the impact of factors such as unit power restrictions, data collection errors and human force majeure; (4) strong adaptability, can simultaneously meet the power prediction of wind and solar power stations, is a more universal prediction method with strong practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] FIG1 is a diagram showing the overall solution framework of the method of the present invention;

[0023] Figure 2 is a schematic diagram of the power station latitude and longitude envelope method;

[0024] Figure 3 is a schematic diagram of the division of high-dimensional meteorological data within a time window;

[0025] Figure 4 is the wind power station power prediction input and output set;

[0026] Figure 5 compares the results of wind power station prediction with other methods;

[0027] Figure 6 is a physical explanation of the wind power station power prediction results;

[0028] Figure 7 compares the results of PV power station prediction with other methods;

[0029] Figure 8 shows the physical correlation and robustness of the PV power station prediction results. DETAILED DESCRIPTION

[0030] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0031] Renewable energy power prediction requires the establishment of a relationship with related meteorological factors. In order to deeply explore the characteristic relationship between the two, the ERA5 comprehensive meteorological data set is selected as the input of the new energy power prediction model. Although it can fully explore the meteorological factors related to power, considering the inaccurate deviation of the longitude and latitude of the power station and the influence of factors such as geographical terrain, the power generation prediction of a single meteorological point data is inaccurate. The key meteorological point describing the output of a power station is not necessarily at the location of the power station, or it should be a combination of meteorological data of multiple key meteorological points, as shown in Figure 2. Therefore, a high-dimensional multivariate meteorological factor sequence set S related to power generation is constructed. m,k , significantly expanding the possible meteorological factors from the perspective of physical causes and spatial scale. Assuming the longitude and latitude of the power station are (N, E), and selecting the European comprehensive, stable, and easily accessible meteorological dataset ERA5, if its dimensional spatial resolution is n and its longitude spatial resolution is e, the power station itself can be enclosed in the spatial grid area (N-in, N+in) and (E-je, E+je), with the spatial longitude and latitude resolution expansion scales being i and j, the expression is as follows:

[0032] Where: m represents the meteorological factor sequence related to the power of wind and solar power stations, which is M in total; k represents the number of meteorological points when the spatial resolution is expanded to scale i and j, which is K in total; then according to the Cartesian product, the high-dimensional multivariate meteorological factor set S m,k There are M×K meteorological factor sequences in total.

[0033] For the high-dimensional multivariate meteorological factor set S m,k , it is necessary to screen out relevant and independent meteorological factor sets, and adopt comprehensive correlation analysis, ultra-high-dimensional data dimensionality reduction and time scale expansion to form the input sample set of the power prediction model.

[0034] According to the high-dimensional multivariate meteorological factor sequence set S related to the input power station type power m,kAnd the corresponding power station output sequence set F, the generated power and meteorological factors are analyzed from three aspects according to formula (10), and the correlation coefficients are weighted averaged.

[0035] Where: Respectively represent a meteorological factor sequence S in the meteorological factor sequence set mk The three aspects of correlation and comprehensive correlation coefficient with the output sequence F. mk ,F) represents S mk and the covariance of F; and σ F Represents S mk and the standard deviation of F; E(S mk ) and E(F) represent S mk and the mathematical expectation of F; rg(S mk,i ) and rg(f i ) represent the ascending or descending order of each element, n represents the number of data, d i represents the difference between two orders; sgn is the sign function, S mk,i -S mk,j and f i -f j The same sign is 1, the opposite sign is -1, and 0 is not included in the summation. i represents the difference between two orders; sgn is the sign function, S mk,i -S mk,j and f i -f j The same sign is 1, the opposite sign is -1, and 0 is not included in the sum.

[0036] High-dimensional multivariate meteorological factor set S m,k The comprehensive correlation coefficient set of meteorological factors can be obtained by using the power station output sequence set F according to formula (10):

[0037] Screen out D meteorological factors with correlation coefficients greater than 0.6. If ρ mk ≥0.6, then let S mk ∈S D , and obtain the corresponding strongly correlated meteorological factor sequence set S D ={S1,…,S D}, the two-dimensional meteorological factor set S m,k Reduced to one-dimensional S D .

[0038] A large amount of multidimensional meteorological data is helpful for describing the characteristics of wind and solar energy. However, the correlation and information overlap between multidimensional meteorological data will increase the difficulty of establishing a mapping relationship between meteorological information and power supply. The high dimensionality of the data will also lead to problems such as reduced versatility of the prediction model, complex calculations, and model overfitting. If S D The sample set of time series length T is: X=[x1,x2,…,x T ] T , x i ∈R D , where x i (i=1,2,…,T) is the i-th sample, T is the total number of samples, D is the characteristic dimension of the high-dimensional sample, that is, the number of factors that are strongly correlated with the output. Let the d-dimensional sample set after dimensionality reduction Y=[y1,y2,…,y T ] T ,y i ∈R d , follow the steps below to fuse high-dimensional meteorological data:

[0039] (1) Calculate the Euclidean distance between each sample point and other points:

[0040] The Euclidean distance D ij Arrange from small to large, preset the neighboring parameters, that is, the k (k<N) nearest neighboring points to the sample point, and get the value of x i The set of neighboring points: Q(i)=[x i1 ,x i2 ,…,x ik ]Euclidean distance for sample x i with x j The similarity measurement result of , the smaller the value, the higher the surface similarity.

[0041] (2) Calculate the linearized reconstruction coefficient matrix W and solve the coefficients in the reconstruction coefficient matrix M with the goal of minimizing the linearized error function. The linearized error function is:

[0042] Where Q(i) is the set of k nearest neighbor samples of i, and the weight coefficient w i j Make normalization restrictions, that is, satisfy the formula:

[0043] Yes, not x i Neighborhood sample x j , let the corresponding weight value w i j =0.

[0044] (3) Perform dimensionality reduction based on the reconstruction coefficient matrix M and the preset target dimension d. Based on the reconstruction coefficient matrix M, the optimization problem is constructed with the goal of minimizing the loss function δ and satisfying the following constraints (see formula (15), y i is the i-th sample x i The result after dimensionality reduction, I d is a d-order unit matrix, and the strongly correlated meteorological factor dataset Y after dimensionality reduction is obtained by solving it.

[0045] (4) The residual variance μ is used to evaluate the data after dimensionality reduction. The smaller the residual variance, the greater the correlation between the data before and after dimensionality reduction. X and D Y are the Euclidean distance matrices of X and Y respectively. The goal is to minimize the residual variance as shown in formula (16), and the optimal target dimension d and neighbor parameter k are determined.

[0046] (5) Using the grid search method, set the range of the target dimension d and the nearest neighbor parameter k, and repeat steps 2 to 3 until the target dimension d and the nearest neighbor parameter k corresponding to the minimum residual variance μ are obtained, which are the optimal dimensionality reduction parameters, as shown in Figure 3. The optimal meteorological factor data series set Y within the time series length T after dimensionality reduction is obtained. d :

[0047] Considering that both the power generation and meteorological factor series are time series data, in order to fully explore the characteristics of meteorological data and renewable energy power generation in the time dimension, for any time series factor y, it is divided into multiple time windows of equal length according to a certain sliding step size. Data analysis and modeling are performed in each window to better understand the changing trend and periodicity of the data, as shown in Figure 3. The power station output is taken as part of the meteorological factor data series set and compared with the meteorological factor data series set Y after dimensionality reduction. d Together they constitute the power prediction sample set Y d+1 =[y1,y2,...,y d ,f], according to the time window w for Y d+1 Expand to form the power prediction input set Y d+1,in , corresponding to the prediction model output set Y out : Y out =[f w+1 ,f w+2 …,f T ] T (19)

[0048] Where: y w+i (i=1,2,…,Tw) is a two-dimensional matrix of w×(d+1), corresponding to the output data fw+i .

[0049] The sample set is used to fit the input-output set, that is, the relationship between the output data at each moment t and the deep learning input set of w×(d+1) in the previous w time broadband windows. Then, for any i moment y w+i The input set can output the output data f at that moment w+i。

[0050] Where: W in is a weight matrix, B in is a bias matrix.

[0051] The method is verified by taking a wind and photovoltaic power station in a certain area of ​​Yunnan as an example. The actual output data and meteorological data from March to May 2020 are used to build the model, and the data of four days in April 2021 are used for testing, with a time scale of 1 hour.

[0052] For the wind power station, according to its power generation principle, 11 relevant meteorological factors are selected, as shown in Table 1. The longitude and latitude of the wind power station are (26.47N, 100.01E). According to formula (9), let i = j = 1, and the number of spatial meteorological points selected is 9, which encloses the power station in the area from 26.25°N to 26.75°N, 99.75°E to 100.25°E. Therefore, the high-dimensional multivariate meteorological factor sequence has 99 dimensions. According to formula (10), a comprehensive correlation analysis is performed. The three-level correlation of each factor is shown in Table 2. From the analysis of single meteorological factors and meteorological points, as shown in Table 3, the correlation of 6 meteorological factors with typical representatives is selected. From this analysis: (1) The meteorological points with good correlation with power station power are not necessarily at the location of the power station. In fact, the correlation between u10, v100, and u100 and power at the power station is the inverse among various meteorological points. This shows the importance of expanding the spatial scale to mine multivariate meteorological factors. (2) The overall correlation between the two wind speed-related meteorological factors i10fg and fg10 is higher than that of other wind speed meteorological factors. It can be seen that the maximum value of the maximum wind speed and the average wind speed is more effective in describing wind power, which illustrates the importance of expanding the physical causes to explore high-dimensional meteorological factors.

[0053] A total of 99 meteorological factors were ranked by their comprehensive correlation, and 26 strongly correlated meteorological factors were screened based on a comprehensive correlation coefficient greater than 0.6, as shown in Table 4. Because the data for each meteorological factor within the region and for the same meteorological factor at different meteorological points are highly correlated, and to reduce the complexity of the prediction model and prevent overfitting, the 26 strongly correlated meteorological factors were further reduced in dimension. Two-stage optimization and dimensionality reduction: The first stage: grid search is performed on the target dimension d after dimensionality reduction, and the neighbor parameter k is searched in a smaller range. The purpose is to find the optimal target dimension d as soon as possible; the optimal target dimension d is gradually searched in the range of 1 to 26, and the neighbor parameter k is first searched in a small range of 10 to 20 to determine the target dimension as 10; the second stage: after determining the optimal target dimension d, the neighbor parameter k is gradually determined by bisection, the target dimension d is fixed at 10, the value of the neighbor parameter k is expanded, and after searching in multiple ranges, it is finally determined that when the neighbor parameter k is 86, the residual variance μ takes the minimum value, and finally the 10-dimensional meteorological factor data series set Y is obtained. d Taking the power station output as part of the meteorological factor data series, the window length is set to 96 according to formula (18), and the power prediction input set Y is obtained. d+1,in , and the prediction model output set Y out The corresponding ones are shown in FIG4 .

[0054] The power prediction input and output sets were fitted and tested, and the model prediction results were analyzed from the following aspects: (1) Accuracy: Conventional machine learning models such as XGBoost, BP artificial neural network and random forest were compared. The evaluation indicators of each prediction model are shown in Table 5, and the prediction results are shown in Figure 5. The prediction accuracy is much higher than that of conventional prediction methods. (2) Physical interpretability: Based on the actual output and predicted output curves, the three meteorological factor sequences with the strongest correlation with wind power were added, and the normalized curves are shown in Figure 6. It was found that in the first to second day, that is, the 1-40h period, the actual output and predicted output curves and the three meteorological factor sequences with the strongest correlation had basically the same change trend. In the 40h-121h period, the change trend of the three meteorological factor sequences with the strongest correlation had a lag with the actual output and predicted output curves, indicating the physical interpretability of the method and the necessity of expanding the sample set of the time sliding window algorithm, which can explore the characteristic relationship between the actual output and the strongly correlated meteorological factor sequence over a longer time period. (3) Robustness: As shown in Figure 6, for the abnormal section data, the actual power value is almost zero, while the wind speed is strong during the abnormal period. The possible reasons are power rationing of the units and human force majeure factors. However, the prediction model can correctly handle various abnormal situations in the input data, and no large error occurs in the entire prediction result due to an abnormal situation. It can be seen that the model has strong robustness.

[0055] Similarly, a case study was conducted on a photovoltaic power station to construct a high-dimensional multivariate meteorological factor set. A comprehensive correlation analysis was performed on the 207-dimensional meteorological factors, and 27 strongly correlated meteorological factors were screened out. Table 6 shows the photovoltaic strongly correlated meteorological factors. After a two-stage dimensionality reduction process, the optimal target dimension d = 2 and the nearest neighbor parameter k = 441 were used to reduce the data to a 2-dimensional meteorological factor data sequence set Y d , and construct the power prediction input and output set. Figure 7 shows the comparison results with various prediction models. Table 6 shows the evaluation indicators of various optoelectronic prediction models. Figure 8 shows the physical correlation and robustness of the prediction results.

[0056] By verifying the actual data of wind and solar power stations of different types and locations, the short-term wind and solar power prediction method based on high-dimensional multivariate meteorological data fusion proposed in the present invention is highly accurate, physically explainable, robust, and adaptable. It can simultaneously meet the power prediction of wind and solar power stations. It is a more universal prediction method with strong practical value.

[0057] Table 1 Meteorological factors related to wind power

[0058] Table 2 Three correlations of wind power meteorological factors

[0059] Table 3 Correlation analysis of wind power single meteorological factor and multiple meteorological points

[0060] Table 4 Meteorological factors with strong correlation to wind power

[0061] Table 5 Evaluation indicators of various wind power prediction models

[0062] Table 6 Photoelectric strong correlation meteorological factors

[0063] Table 7 Evaluation indicators of various optoelectronic prediction models

Claims

1. A short-term wind and solar power prediction method based on high-dimensional multivariate meteorological data fusion, characterized in that: The steps include: (1) Construct a set of meteorological factors related to power generation of power stations S m,k , expand the possible meteorological factors from the physical causes and spatial scales; suppose the longitude and latitude of the power station are (N, E), for the selected meteorological data set, if its latitude spatial resolution is n, and its longitude spatial resolution is e; the power station is centered on itself and is enclosed in the (N-in, N+in), (E-je, E+je) spatial grid areas, and the spatial longitude and latitude resolution expansion scales are i and j, and the expression is as follows: Where m represents the meteorological factor sequence related to the power generation of the power station, which is M in total; k represents the number of meteorological points when the spatial resolution is expanded to i and j, which is K in total; according to the Cartesian product, the meteorological factor sequence set S m,k There are M×K meteorological factor sequences in total; (2) A method for constructing a power prediction sample set is used to construct a meteorological factor sequence set based on the meteorological factor sequence related to the power generation of the power station and the latitude and longitude of the power station. A comprehensive correlation analysis, ultra-high-dimensional data dimensionality reduction fusion and time scale expansion are used to form a power prediction sample set that is independent of each other, covers the original information and has an expanded time dimension. The specific steps are as follows: Step 1. Calculate the high-dimensional multivariate meteorological factor set S related to the power generation of the power station according to formula (1) m,k , conduct correlation analysis on power generation of power stations and meteorological factor sequences one by one, and take weighted average of correlation coefficients; Where: Represents a meteorological factor sequence S in the meteorological factor sequence set mk The comprehensive correlation coefficient with the power station output sequence F; I is the number of selected correlation coefficients; The power station output sequence set F=[f1,f2,…,f T ] T and meteorological factor sequence set S m,k According to formula (2), the comprehensive correlation coefficient set of meteorological factors is obtained: Select D meteorological factors with correlation coefficients greater than 0.

6. If ρ mk ≥0.6, then S m,k ∈S D , and obtain the corresponding strongly correlated meteorological factor sequence set S D ={S1,...,S D }, the two-dimensional meteorological factor set S m,k Reduced to one dimension S D ; Step 2. If S D The sample set of time series length T is X = [x1, x2, ..., x T ] T , x i ∈R D , where x i is the i-th sample, T is the total number of samples, D is the characteristic dimension of the high-dimensional sample, that is, the number of meteorological factors that are strongly correlated with the output sequence, R D is a real number set of D-dimensional strongly correlated meteorological factors; the d-dimensional sample set Y after dimension reduction is [y1, y2,…, y T ] T ,y i ∈R d ; The goal is to minimize the residual variance μ of the sample set before and after dimensionality reduction, see formula (4); D X and D Y The Euclidean distance matrices of X and Y respectively, determine the optimal target dimension d； Get the optimal d-dimensional meteorological factor data series set Y within the time series length of T after dimensionality reduction d : Step 3. Take the power station output itself as part of the power prediction sample set and compare it with the dimensionality-reduced meteorological factor data sequence set Y d Together they constitute the power prediction sample set Y d+1 =[y1,y2,...,y d ,f], according to the time window w for Y d+1 Expand to form the power prediction input set Y d+1,in , corresponding to the prediction model output set Y out : Y out =[f w+1 ,f w+2 …,f T ] T (7) Where: y w+i is a two-dimensional matrix of w×(d+1), corresponding to the output data f w+i ; where i = 1, 2, ..., Tw; Y out The power station output sequence set F is intercepted and matched with the power forecast input set Y in time sequence d+1,in correspond; (3) Use the sample set to fit the input-output set, that is, the relationship between the output data at each time t and the power prediction input set of w×(d+1) in the previous w time broadband windows. Then, input any time y w+i The input set can output the output data f at that moment w+i ; Where: W in is a weight matrix, B in is a bias matrix.

Images