Explainable wind power prediction method based on feature matrix iterative reconstruction

Abstract

The present application relates to the field of new energy power prediction, and proposes an interpretable wind power prediction method based on feature matrix iterative reconstruction, considers the hidden information contained in the model prediction error, mines the maximum path of error generation through the time fusion transformer TFT model with interpretability, dynamically reconstructs the feature matrix, solves the optimal feature matrix, and thus improves the prediction accuracy of the model; the maximum path of error generation is explained from the local and global perspectives respectively, and a traceability mechanism of prediction error is established, so that the present application can effectively improve the safety and reliability of wind power prediction.

Description

Technical Field

[0001] This application relates to the field of new energy power prediction technology, and in particular to an interpretable wind power prediction method based on feature matrix iterative reconstruction. Background Technology

[0002] In recent years, with the development of renewable energy, the wind power industry has experienced unprecedented growth. However, due to the influence of wind speed, wind power generation exhibits significant randomness, intermittency, and uncontrollability. Large-scale wind power grid connection poses a significant challenge to the stable operation of the power system. Therefore, improving the accuracy of wind power forecasting is of great importance for the safe, stable, and economical operation of the power system.

[0003] With the booming development of the wind power industry, wind power prediction technology has been widely researched and applied. Based on different model theories, traditional wind power prediction can generally be divided into statistical models and physical models. Compared to physical models, statistical models, which directly build models based on wind power data, exhibit good prediction accuracy and wide applicability in short-term and ultra-short-term predictions. In recent years, deep learning models derived from next-generation artificial intelligence technology have gradually become popular in wind power / wind speed prediction, further expanding statistical models. However, existing ultra-short-term wind power prediction models generally have common limitations. These methods mainly rely on historical data from individual wind farms without considering the spatial correlation between different wind farms within a region. Since adjacent wind farms experience highly similar environmental and meteorological conditions, their wind speed changes are also strongly correlated. Reasonably utilizing this spatiotemporal correlation can effectively improve the accuracy of wind power prediction.

[0004] Furthermore, because wind power forecasting machine learning models contain activation functions such as Sigmoid and tanh, and their nonlinear components exhibit strong nonconvexity, they are generally "black box" models, lacking interpretability. This results in low reliability of the model predictions in practical applications, failing to meet the reliability requirements of modern power systems for energy regulation. Therefore, improving the interpretability of the prediction results while enhancing the accuracy of power output forecasting models is of great significance. Existing interpretable wind power forecasting technologies focus on combining deep learning models with attention mechanisms, achieving some interpretability by exploring the correlation between model input and output. However, they neglect the multi-morphic interactive features contained in the model prediction errors. Model errors reflect the differences between predicted and actual observed values. These differences contain rich patterns and information about changes in the spatiotemporal dimensions. By rationally utilizing the inherent generation mechanism of errors and achieving interpretable representation of error sources, more reliable and accurate prediction information can be provided in practical applications. Summary of the Invention

[0005] To overcome the shortcomings of the existing technologies, the technical problems to be solved by this invention are twofold: firstly, existing wind power prediction methods fail to fully consider the spatial correlation between different wind farms in a wind power cluster, resulting in insufficient prediction accuracy; secondly, existing interpretable wind power prediction modeling methods often neglect the interpretability of prediction errors. This lack of systematic understanding of error sources leads to low reliability of model prediction results in practical applications, failing to 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 the 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 mutual information theory, the maximum correlation-minimum redundancy (mRMR) principle is used to select the feature factors with the greatest correlation to wind power and the least redundancy among them from the historical wind power and meteorological data, and construct the initial feature matrix.

[0008] Step B. Establish an initial wind power prediction model based on a 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, predictive modeling of the prediction error is performed based on the initial feature matrix. According to the allocation of attention weights in each time step, the periodic error maximization path is obtained, and feature vectors representing the error maximization path are generated.

[0010] Step D. Use the distance between attention weight vectors in each time step to characterize the temporal trend change of each feature vector in the initial feature matrix; extract the vector segments 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 the reconstructed feature matrix; quantify the contribution of each feature in the reconstructed feature matrix to the wind power prediction task based on the Shapley SHAP value method, and remove the feature vector with the smallest contribution.

[0011] Step E. Generate and remove the feature vectors with the smallest contribution in multiple rounds until the iteration stopping 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 feature matrix provided in this application considers the latent information contained in the model prediction error. It mines the error maximization path through an interpretable time fusion converter (TFT) model and dynamically reconstructs the feature matrix to obtain the optimal feature matrix, thereby improving the prediction accuracy of the model. The method also provides interpretable analysis of the error maximization path from both local and global perspectives and establishes a source tracing mechanism for prediction errors. Therefore, this invention can effectively improve the safety and reliability of wind power prediction. Attached Figure Description

[0013] The above is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] Figure 1 A flowchart illustrating the interpretable wind power prediction method based on feature matrix iterative reconstruction provided by this invention.

[0015] Figure 2 This is a schematic diagram of the error tracing mechanism for the interpretable wind power prediction model provided by the present invention;

[0016] Figure 3 This is a schematic diagram of the process for obtaining the optimal feature matrix based on iterative dynamic reconstruction of the feature matrix, as provided by the present invention. Detailed Implementation

[0017] The invention will be further described below with reference to the accompanying drawings.

[0018] See Figure 1 The flowchart of the interpretable wind power prediction method based on feature matrix iterative reconstruction provided in this application is shown in the figure.

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

[0020] The design uses historical meteorological and power data from various wind farms as input to perform ultra-short-term wind power prediction for target wind farms in the target area. Firstly, based on mutual information theory, the maximum correlation-minimum redundancy (mRMR) principle is used to select meteorological factors that have the highest correlation with wind power but the lowest redundancy among themselves.

[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] In the formula, 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 I(x) represents the mutual information value between meteorological feature i and the wind power sequence of the target electric field. j ,x i ) represents 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 that are highly correlated 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 each feature x in the initial dataset S. i The mean of the mutual information between wind power and S, where R is the magnitude of the mutual information between each feature in S. m x is the number of features selected in dataset S. i and x j These are the time series of meteorological features i and j, respectively.

[0027] A basic dataset containing meteorological features and wind power of each wind farm was obtained through feature filtering. Assuming the dataset consists of N time series of length T, feature vectors were generated after normalizing the time series, resulting in the initial feature matrix X'. (T×N) The data normalization principle is shown in equation (5):

[0028]

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

[0030] Step B. Establish an initial wind power prediction model based on a 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, we establish multiple deep learning prediction models to study the mapping mechanism between input features and target electric field power. Assuming we select k basic deep learning models, we train and model each model using feature matrices, and obtain model parameters through gradient descent. Based on the trained deep learning models, we obtain the predicted value of the target wind farm power. This mapping relationship can be expressed as:

[0032]

[0033] In the formula, j = {1,2,...,k}, g j This is the j-th basic deep learning model trained based on the feature matrix X; Indicates based on model g j The predicted power value of the target wind farm is obtained. The prediction errors of each basic model are calculated.

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

[0035]

[0036] In the formula: y and Let P be the normalized actual value of the target wind farm power and the predicted values ​​of each model; N The total rated installed capacity of the wind farm cluster; I NRMSE and I NMAE These are the root mean square error and the mean absolute error after normalization, respectively.

[0037] Step C. Using the interpretable Time Fusion Transformer (TFT) model, predictive modeling of the prediction error is performed based on the initial feature matrix. According to the allocation of attention weights in each time step, the periodic error maximization path is obtained, and feature vectors representing the error maximization path are generated.

[0038] The characteristic matrix X′ (T×N) As input, the error sequence e of each power prediction model is used based on the interpretable TFT model. i Perform modeling and output the predicted values ​​of the errors of each model. and the corresponding model attention weight matrix Y j, where T is the length of the time series. The contribution of the input features to the error prediction task can be obtained by multiplying the attention weight matrix by the original input feature matrix. The specific calculation method is expressed as follows:

[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 contribution of input features, Y t T Let Y represent the transpose of the t-th column of the attention weight matrix Y. This operation extracts all weights in Y relevant to time step t, and multiplies them by each row of the input feature matrix X′, resulting in a vector of length N representing the contribution of each input feature to the error at time step t. Finally, the input feature contribution matrix C∈R is obtained. T×N Based on the contribution of each feature to the error at each time step, the error-maximizing path can be represented as:

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

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

[0044] Based on the error maximization path, an error tracing mechanism is established as follows: Figure 2As shown. From a local perspective, the feature segments contained in the path leading to the maximization of prediction error in a single model are correlated and matched with various interpretable regular features (such as the spatiotemporal variation patterns of wind power such as seasonality and periodicity). The magnitude characteristics of the prediction error are used to analyze the interpretability of the cause of the error. From a global perspective, based on the set of paths that maximize the prediction contribution of all basic models, and considering the prior knowledge of specific events or extreme weather, the error sources associated with extreme meteorological features and temporal fluctuation features of the prediction model are obtained through rank correlation set theory analysis and correlation mining methods. Errors from the same source are classified, and the source analysis of the error is performed through the trend characteristics of the prediction error.

[0045] Step D. Use the distance between attention weight vectors in each time step to characterize the temporal trend change of each feature vector in the initial feature matrix; extract the vector segments 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 the reconstructed feature matrix; quantify the contribution of each feature in the reconstructed feature matrix to the wind power prediction task based on the Shapley SHAP value method, and remove the feature vector with the smallest contribution.

[0046] For each feature, the attention weight matrix Y is calculated. j The distance between attention vectors at each time step and the average pattern are used to quantify significant changes in 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. Further construction... Where τ max The maximum prediction scale is used. To compare the similarity between attention weight vectors, the distance metric proposed by Comaniciu is used:

[0049]

[0050] In the formula, a and b represent two distinct probability vectors, and k(a,b) represents the similarity between vectors a and b. The views at all scales are aggregated as follows:

[0051]

[0052] In the formula, dist(t) represents the set of volatility across all scales at time step t. The extreme eigenvector q is constructed by concatenating the eigenvector segments showing significant changes in time-series volatility characteristics. kUsing the extreme eigenvector q k and the feature vector q containing the error-maximizing path generated in step C. maxi The original feature matrices are merged and reconstructed to obtain the reconstructed feature matrix X′. (T×(N+2)) This paper utilizes cooperative game theory and constructs an interpretable model of feature contributions in wind power prediction based on the SHAP framework. Through multiple iterations, the contribution of each eigenvector within the reconstructed matrix to power prediction is obtained. The two eigenvectors with the lowest contribution values ​​are then removed. The SHAP contribution values ​​follow the following equation:

[0053]

[0054] In the formula, y base f(p) represents the baseline value of the model, i.e., the mean of the target variable for all samples; i ), f(p maxi f(q) k ) represent the SHAP values ​​of the original eigenvector, the eigenvector containing the error-maximizing path, and the extreme eigenvector, respectively, which are the power prediction contribution values. The eigenvector with the lowest contribution value is removed to obtain the dynamically reconstructed feature matrix X. ( " T×N) The visualization process for single-step feature matrix iterative reconstruction is as follows: Figure 3 As shown.

[0055] Step E. Generate and remove the feature vectors with the smallest contribution in multiple rounds until the iteration stopping 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] By generating and removing the feature vectors with the smallest contributions in multiple rounds, the reconstruction of the feature matrix stops when the contribution of each feature in the reconstructed feature matrix reaches a preset expected equilibrium value. At this point, the contribution of each feature in the power prediction task of the target electric field is balanced and effective, thus obtaining 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 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 the feature matrix, characterized in that... This includes the following steps: Step A. Preprocess historical wind power and meteorological data of each wind farm in the target area: Based on mutual information theory, the maximum correlation-minimum redundancy (mRMR) principle is used to select the feature factors with the greatest correlation to wind power and the least redundancy among each other from the historical wind power and meteorological data, and construct an initial feature matrix; Step B. Establish an initial wind power prediction model based on a 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. Step C. Using the interpretable Time Fusion Transformer (TFT) model, predict and model the prediction error based on the initial feature matrix. According to the allocation of attention weights in each time step, obtain the periodic error maximization path and generate a feature vector representing the error maximization path. Step D. Use the distance between attention weight vectors in each time step to characterize the temporal trend change of each feature vector in the initial feature matrix; extract the vector segments 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 the reconstructed feature matrix; quantify the contribution of each feature in the reconstructed feature matrix to the wind power prediction task based on the Shapley SHAP value method, and remove the feature vector with the smallest contribution. Step E. Generate and remove the feature vector with the smallest contribution in multiple rounds until the iteration stopping 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 the feature matrix according to claim 1, characterized in that: In step A, the historical wind power and meteorological data are reduced in dimensionality by selecting meteorological characteristic factors, and the meteorological factors with the greatest correlation to wind power and the least redundancy between them are selected. The initial feature matrix is ​​composed of the feature vectors of wind power and related meteorological factors of each wind farm.

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

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

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

6. The interpretable wind power prediction method based on iterative reconstruction of the feature matrix according to claim 1, characterized in that: In step E, through multiple rounds of iteration, the reconstruction of the feature matrix is ​​stopped when the contribution of each feature in the reconstructed feature matrix reaches the preset expected equilibrium value, and the resulting 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