Inplausible wind power prediction method based on feature matrix iterative reconstruction

Abstract

The invention relates to the field of new energy power prediction, and provides an interpretable wind power prediction method based on feature matrix iterative reconstruction, which considers hidden information contained in a model prediction error, generates a maximized path through an interpretable time fusion converter TFT model mining error, and dynamically reconstructs a feature matrix. The optimal feature matrix is solved, so that the prediction precision of the model is improved; according to the method, explainable analysis is carried out on the error generation maximization path from local and global angles, and a traceability mechanism of the prediction error is established, so that the safety degree and the credibility of wind power prediction can be effectively improved.

Description

Technical Field

[0001] The present application relates to the technical field of renewable energy power prediction, and in particular to an interpretable wind power prediction method based on iterative reconstruction of a feature matrix. Background Art

[0002] In recent years, with the development of renewable energy, the wind power industry has shown an unprecedented development speed. However, affected by wind speed, wind power generation has obvious randomness, intermittency and uncontrollability. Large-scale wind power grid connection has brought great challenges to the stable operation of the power system. Therefore, improving the accuracy of wind power prediction is of great significance to the safe, stable and economical operation of the power system.

[0003] With the vigorous development of the wind power industry, the prediction technology for wind power has also been widely studied and applied. According to different model theories, traditional wind power prediction can generally be divided into statistical models and physical models. Compared with physical models, statistical models directly build models for wind power data, showing good prediction accuracy and wide applicability in short-term and ultra-short-term predictions. In recent years, deep learning models derived from the new generation of artificial intelligence technology have gradually become popular in the prediction of wind power / wind speed, and statistical models have been further expanded. However, existing ultra-short-term wind power prediction models generally have common limitations. These methods mainly rely on the historical data of a single wind farm, without considering the spatial correlation between different wind farms in a region. Since the environmental and meteorological conditions of adjacent wind farms are highly similar, there is also a strong correlation between their wind speed changes. Reasonable use of this spatiotemporal correlation can effectively improve the accuracy of wind power prediction.

[0004] In addition, since the machine learning model for wind power prediction contains activation functions such as Sigmoid and tanh, the nonlinear link has strong non-convexity and is generally a "black box" model. It lacks interpretability, resulting in low credibility of the model prediction results in practical applications, which cannot meet the reliability requirements of modern power systems for energy regulation. Therefore, while improving the accuracy of the output prediction model, it is of great significance to obtain the interpretability of its prediction results. The existing interpretable wind power prediction technology focuses on combining deep learning models with attention mechanisms. By exploring the correlation between model input and output, the interpretability of the model is achieved to a certain extent, but the multi-modal interactive features contained in the model prediction error are ignored. The model error reflects the difference between the predicted value and the actual observed value. These differences contain rich laws and information of changes in the spatiotemporal dimension. Reasonable use of the inherent generation mechanism of the error to achieve interpretable representation of the error source can provide more reliable and accurate prediction information in practical applications. Summary of the invention

[0005] In order to overcome the deficiencies of the above-mentioned prior art, the technical problem to be solved by the present invention is that, on the one hand, the existing wind power prediction method fails to fully consider the spatial correlation between different wind farms in the wind power cluster, and the prediction accuracy needs to be further improved; on the other hand, the existing interpretable wind power prediction modeling method often ignores the interpretability of the prediction error. The lack of a systematic understanding of the error traceability leads to low credibility of the model prediction results in practical applications, which cannot meet the reliability requirements of modern power systems for energy regulation.

[0006] To this end, the present invention proposes an interpretable wind power prediction method based on iterative reconstruction of a feature matrix, the method comprising:

[0007] Step A. Preprocess the historical wind power and meteorological data of each wind farm in the target area: Based on the mutual information theory, the maximum correlation-minimum redundancy (mRMR) principle is adopted to select the characteristic factors with the greatest correlation with wind power and the least redundancy with each other from the historical wind power and meteorological data, and construct the initial characteristic matrix;

[0008] Step B. Establish an initial wind power prediction model based on the deep learning model, substitute the initial feature matrix into the initial wind power prediction model for training, predict the future output of the target wind farm in the target area, and calculate the prediction error of the wind power;

[0009] Step C. Using the interpretable time fusion transformer TFT model, the prediction modeling of the prediction error is performed based on the initial feature matrix, and according to the distribution of attention weights of the model in each time step, the periodic error maximization path is obtained, and the feature vector representing the error maximization path is generated;

[0010] Step D. Use the distance between the attention weight vectors in each time step to characterize the time series trend changes of each feature vector in the initial feature matrix; extract the vector fragments with extreme trend changes in each feature vector, and aggregate them to generate feature vectors containing extreme trend changes; merge and reconstruct the feature vectors representing the error maximization path and the feature vectors containing extreme trend changes with the initial feature matrix to obtain a reconstructed feature matrix; quantify the contribution of each feature in the reconstructed feature matrix in the wind power prediction task based on the Shapley SHAP value method, and eliminate the feature vector with the smallest contribution;

[0011] Step E. Generate and eliminate the eigenvector with the smallest contribution in multiple rounds until the iteration stop condition is met to obtain the optimal feature matrix; construct an ultra-short-term wind power prediction model based on the optimal feature matrix obtained by iteration; linearly superimpose the predicted wind power output in the ultra-short-term wind power prediction model with the predicted value of the prediction error of the TFT model to obtain the final wind power prediction value.

[0012] In summary, the interpretable wind power prediction method based on iterative reconstruction of the characteristic matrix provided by the present application takes into account the potential information contained in the model prediction error, mines the error maximization path through the interpretable time fusion converter TFT model and dynamically reconstructs the characteristic matrix to obtain the optimal characteristic matrix, thereby improving the prediction accuracy of the model; the error maximization path is interpretably analyzed from local and global perspectives, and a tracing mechanism for prediction errors is established. Therefore, the present invention can effectively improve the safety and credibility of wind power prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The above is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0014] Figure 1 A schematic flow chart of an interpretable wind power prediction method based on iterative reconstruction of a characteristic matrix provided by the present invention;

[0015] Figure 2 A schematic diagram of the error tracing mechanism of the explainable wind power prediction model provided by the present invention;

[0016] Figure 3 A schematic diagram of a process for obtaining an optimal feature matrix based on iterative dynamic reconstruction of the feature matrix provided by the present invention. DETAILED DESCRIPTION

[0017] The present invention is further described below in conjunction with the accompanying drawings.

[0018] See also Figure 1 , a flow chart of an interpretable wind power prediction method based on iterative reconstruction of feature matrix provided in this application.

[0019] Step A. Preprocess the historical wind power and meteorological data of each wind farm in the target area. Based on the mutual information theory, the maximum relevance-minimum redundancy (mRMR) principle is used to select the characteristic factors with the greatest correlation with wind power and the least redundancy with each other from the historical wind power and meteorological data, and construct the initial characteristic matrix.

[0020] The design uses the historical meteorological data and historical power data of each wind farm as input to perform ultra-short-term wind power forecasting for the target farm in the target area. First, based on the mutual information theory, the maximum relevance-minimum redundancy (mRMR) principle is used to select meteorological factors with the greatest correlation with wind power and the least redundancy with each other:

[0021] I(p,x i )=H(x i )-H(x i |p) (1)

[0022] I(x j ,x i )=H(x i )-H(x i |x j ) (2)

[0023] Where xi is the time series of meteorological feature i, p is the wind power series of the target electric field, H(.) represents the calculation of entropy, and I(p,x i ) is the mutual information value between meteorological feature i and the target wind power sequence, I(x j ,x i ) is the mutual information value between meteorological feature i and meteorological feature j.

[0024] Then, the maximum correlation standard maxD(S,p) is used to select features with high correlation with the wind power sequence; based on the minimum redundancy standard min R(S), redundant features are removed:

[0025]

[0026] In the formula, D represents the features x in the initial data set S i and wind power, R is the mutual information size between each feature in S, S m is the number of features selected in the dataset S, x i and x j are the time series of meteorological characteristics i and j respectively.

[0027] Through feature screening, a basic data set containing meteorological characteristic factors and wind power related to each wind farm is obtained. Assuming that the data set consists of N time series with a length of T, the time series is normalized to generate a feature vector, and the initial feature matrix X' is obtained. (T×N) The data normalization principle is shown in formula (5):

[0028]

[0029] Where i = {1, 2, ..., N}, x i is the original data of the i-th feature sequence, x i,min and x i,max are the minimum and maximum values ​​of the i-th feature sequence data, x′ i is the eigenvector after data normalization, that is, 0≤x′ i ≤1.

[0030] Step B. Establish an initial wind power prediction model based on the deep learning model, substitute the initial feature matrix into the initial wind power prediction model for training, predict the future output of the target wind farm, and calculate the prediction error of the wind power.

[0031] First, a variety of deep learning prediction models are established to study the mapping mechanism between input features and target electric field power. Assuming that k basic deep learning models are selected, the feature matrix is ​​used to train and model each deep learning prediction model, and the model parameters are obtained by the gradient descent method. Based on the trained deep learning models, the predicted value of the target wind farm power is obtained. This mapping relationship can be expressed as:

[0032]

[0033] Where j = {1, 2, ..., k}, g j is the jth basic deep learning model trained based on the feature matrix X; Represents the model based on g j The power forecast value of the target wind farm is obtained. The forecast error of each basic model is calculated.

[0034] The normalized root mean square error and the normalized mean absolute error are used as error evaluation indicators, and the model with the highest preliminary prediction accuracy is selected as the basic prediction model. The expression is:

[0035]

[0036] Where: y and are respectively the normalized actual value of the target wind farm power and the predicted value of each model; P N is the total rated installed capacity of the wind farm group; I NRMSE and I NMAE are the root mean square error and mean absolute error after normalization, respectively.

[0037] Step C. Use the interpretable time fusion transformer TFT model to predict the prediction error based on the initial feature matrix, obtain the periodic error maximization path according to the distribution of attention weights of the model in each time step, and generate a feature vector representing the error maximization path.

[0038] The feature matrix X′ (T×N) As input, the interpretable TFT model is used to analyze the error sequence e of each power prediction model. i Modeling and outputting the predicted values ​​of each model error And the corresponding model attention weight matrix Y j, where T is the length of the time series. By multiplying the attention weight matrix with the original input feature matrix, the contribution of the input feature in the error prediction task can be obtained. The specific calculation method is expressed as:

[0039] c t =X′ (T×N) Y t T (T×T) (9)

[0040] C=[X′Y1 T ,X′Y2 T ,…,X′Y T T ] (10)

[0041] In the formula, c t To predict the input feature contribution, Y t T Represents the transpose of the tth column of the attention weight matrix Y. This operation extracts all weights related to time step t in Y and multiplies them with each row of the input feature matrix X′ to obtain a vector of length N, which represents the contribution of each input feature to the error at time step t. Finally, the input feature contribution matrix C∈R T×N According to the contribution of each feature to the error in each time step, the error maximization path can be expressed as:

[0042] q max =(X′Y1 T ) max ,(X′Y2 T ) max ,…,(X′Y T T ) max (11)

[0043] In the formula, q max is the eigenvector containing the error-maximizing path, (X′Y t T )max The feature sequence segment that contributes most to the error generation at time step t.

[0044] According to the error maximization path, an error tracing mechanism is established such as Figure 2As shown in the figure. From a local perspective, the characteristic fragments contained in the path that leads to the maximum prediction error in a single model are correlated and matched with various interpretable regular features (such as seasonal and periodic wind power spatiotemporal variation), and the cause of the error is explained through the amplitude characteristics of the prediction error. From a global perspective, based on the set of paths that maximize the error prediction contribution of all basic models, considering the prior knowledge of specific events or extreme weather, the error sources associated with extreme meteorological characteristics and time series fluctuation characteristics of the prediction model are obtained through the rank-related set theory analysis method and the correlation relationship mining method, the same-source errors are classified, and the error is traced through the trend characteristics of the prediction error.

[0045] Step D. Use the distance between the attention weight vectors in each time step to characterize the time series trend changes of each eigenvector in the initial feature matrix; extract the vector fragments with extreme trend changes in each eigenvector, and aggregate them to generate eigenvectors containing extreme trend changes; merge and reconstruct the eigenvectors representing the error maximization path and the eigenvectors containing extreme trend changes with the initial feature matrix to obtain a reconstructed feature matrix; quantify the contribution of each feature in the reconstructed feature matrix in the wind power prediction task based on the Shapley SHAP value method, and eliminate the eigenvector with the smallest contribution.

[0046] For each feature, by calculating the attention weight matrix Y j The distance between the attention vectors at each time step and the average pattern are used to quantify the significant changes in the temporal dynamics. The average attention pattern at each prediction scale τ is defined as:

[0047]

[0048] In the formula, α(t,i,τ) represents the attention coefficient of feature i at time step t, and T is the length of the feature sequence. where τ max is the maximum prediction scale. In order to compare the similarity between attention weight vectors, the distance metric formula proposed by Comaniciu is used:

[0049]

[0050] In the formula, a and b are represented as two different probability vectors, and k(a,b) represents the similarity between vectors a and b. Aggregate the views of all scales as follows:

[0051]

[0052] Where dist(t) is the set of volatility of all scales at time step t. The extreme feature vector q is formed by concatenating the feature vector segments whose time series volatility characteristics change significantly. k. Using the extreme eigenvector q k and the feature vector q containing the error maximizing path generated according to step C maxi Merge and reconstruct the original feature matrix to obtain the reconstructed feature matrix X′ (T×(N+2)) . Using cooperative game theory, an interpretable model for feature contribution in wind power forecasting tasks is constructed based on the SHAP framework. Through multiple searches, the contribution of each eigenvector in the reconstructed matrix in power forecasting is obtained. The two eigenvectors with the lowest contribution values ​​are screened out. Among them, the SHAP contribution value obeys the following equation:

[0053]

[0054] In the formula, y base is the baseline value of the model, that is, the mean of all sample target variables; f(p i )、f(p maxi )、f(q k ) are the original eigenvector, the eigenvector containing the error maximization path, and the SHAP value of the extreme eigenvector, i.e., the power prediction contribution value. Eliminate the eigenvector with the lowest contribution value and obtain the dynamically reconstructed eigenmatrix X ( ″ T×N) Among them, the visualization process of single feature matrix iterative reconstruction is as follows Figure 3 shown.

[0055] Step E. Generate and eliminate the eigenvector with the smallest contribution in multiple rounds until the iteration stop condition is met to obtain the optimal feature matrix; construct an ultra-short-term wind power prediction model based on the optimal feature matrix obtained by iteration; linearly superimpose the predicted wind power output in the ultra-short-term wind power prediction model with the predicted value of the prediction error of the TFT model to obtain the final wind power prediction value.

[0056] The reconstruction of the feature matrix is ​​stopped by repeatedly generating and eliminating the eigenvector with the smallest contribution until the contribution of each feature in the reconstructed feature matrix reaches the preset expected equilibrium value. At this time, the contribution of each feature in the power prediction task of the target electric field is balanced and effective, so as to obtain the optimal feature matrix and the TFT error correction model trained based on the optimal feature matrix. Finally, the basic prediction model selected in step B is trained using the obtained optimal feature matrix to obtain the ultra-short-term wind power prediction value of the target electric field. The predicted wind power is linearly superimposed with the TFT prediction error to obtain the final wind power prediction value. The normalized root mean square error and the normalized mean absolute error in step B are used as error evaluation indicators to verify the effectiveness of the method.

Claims

1. An interpretable wind power prediction method based on iterative reconstruction of feature matrix, characterized by , including the following steps: Step A. Preprocessing the historical wind power and meteorological data of each wind farm in the target area: Based on the mutual information theory, the maximum correlation-minimum redundancy (mRMR) principle is adopted to select the characteristic factors with the greatest correlation with wind power and the least redundancy with each other from the historical wind power and meteorological data, and construct an initial characteristic matrix; Step B. establishing an initial wind power prediction model based on a deep learning model, substituting the initial feature matrix into the initial wind power prediction model for training, predicting the future output of the target wind farm in the target area, and calculating the prediction error of the wind power; Step C. using an interpretable time fusion transformer TFT model to perform prediction modeling on the prediction error based on the initial feature matrix, obtaining a periodic error maximization path according to the distribution of attention weights of the model in each time step, and generating a feature vector representing the error maximization path; Step D. Use the distance between the attention weight vectors in each time step to characterize the time series trend changes of each feature vector in the initial feature matrix; extract the vector fragments with extreme trend changes in each feature vector, and aggregate them to generate feature vectors containing extreme trend changes; merge and reconstruct the feature vector representing the error maximization path and the feature vector containing extreme trend changes with the initial feature matrix to obtain a reconstructed feature matrix; quantify the contribution of each feature in the reconstructed feature matrix in the wind power prediction task based on the Shapley SHAP value method, and eliminate the feature vector with the smallest contribution; Step E. Generate and eliminate the eigenvector with the smallest contribution in multiple rounds until the iteration stop condition is met to obtain the optimal feature matrix; construct an ultra-short-term wind power prediction model based on the optimal feature matrix obtained by iteration; linearly superimpose the predicted wind power output in the ultra-short-term wind power prediction model with the predicted value of the prediction error of the TFT model to obtain the final wind power prediction value.

2. The interpretable wind power prediction method based on iterative reconstruction of feature matrix according to claim 1 is characterized in that: In step A, the dimension reduction of the historical wind power and meteorological data is achieved by selecting meteorological characteristic factors, and the meteorological factors with the greatest correlation with wind power and the least mutual redundancy are screened out; The initial characteristic matrix is ​​composed of the characteristic vectors of the wind power of each wind farm and related meteorological factors.

3. The interpretable wind power prediction method based on iterative reconstruction of feature matrix according to claim 1 is characterized in that: In the step B, a plurality of deep learning models are used to predict the ultra-short-term wind power of the target wind farm based on the initial feature matrix, and the model parameters, the predicted values ​​of the output power of each model and the corresponding prediction errors are obtained by the gradient descent method.

4. The interpretable wind power prediction method based on iterative reconstruction of feature matrix according to claim 1 is characterized in that: The TFT model in step C models the prediction error, and the TFT model outputs a predicted value of the prediction error and an attention weight matrix; The influence of different features on the prediction error is represented by the attention allocation weights of different features of the TFT model in each time step.

5. The interpretable wind power prediction method based on iterative reconstruction of feature matrix according to claim 1 is characterized in that: In the step D, a new eigenvector is generated to reconstruct the initial eigenmatrix, and the obtained reconstructed eigenmatrix is ​​the union of the initial eigenmatrix and the new eigenvector; The contribution of each feature in the reconstructed matrix is ​​quantified based on the SHAP value method, and the eigenvector with the lowest contribution in the reconstructed feature matrix is ​​eliminated.

6. The interpretable wind power prediction method based on iterative reconstruction of feature matrix according to claim 1 is characterized in that: In the step E, multiple rounds of iterations are performed, and when the contribution of each feature in the reconstructed feature matrix reaches a preset expected equilibrium value, the reconstruction of the feature matrix is ​​stopped, and the obtained feature matrix is ​​the optimal feature matrix; based on the optimal feature matrix, the ultra-short-term wind power prediction model based on TFT model error correction is constructed.

Images