An explainable interval-valued wind power prediction method and system

Abstract

The application discloses an explainable interval-valued wind power prediction method and system, which uses feature extraction process, basic model selection strategy and dynamic integration mechanism to improve prediction accuracy and reliability. Firstly, aiming at the complexity of the interval-valued wind power sequence, an explainable multi-objective feature extraction method is designed to extract key trend and fluctuation features and promote in-depth learning of the features and their relationships. Considering the consistency between the features and the model, combined with elastic net regression and Shapley additive explanation (SHAP) method, the contribution of each candidate model to the prediction target is quantified, so that the basic model under different feature sequences is selected in a highly explainable manner. Finally, according to the Shapley value, the weight of the basic model is dynamically adjusted to adapt to the time-varying characteristics of the single model accuracy, and online updating prediction is realized. The model proposed by the application is superior to the benchmark model in multiple evaluation indexes, and shows excellent prediction performance and explainability.

Description

Technical Field

[0001] This invention relates to an interpretable interval value wind power prediction method and system, belonging to the technical field of interval value wind power prediction. Background Technology

[0002] Wind energy, as a green, clean, and renewable energy source, is gradually becoming a core force in the global energy structure. However, grid connection and efficient utilization of wind power face numerous challenges, primarily the randomness, intermittency, and uncertainty of wind power output. These factors make accurate wind power forecasting crucial for maintaining grid stability, optimizing power dispatch, and improving system reliability. Despite the many advantages of wind energy, the unpredictability of wind power places higher demands on the real-time response capabilities and reserve capacity planning of the power system. Accurate wind power forecasting is extremely important to effectively alleviate peak grid pressure, reduce unnecessary energy waste, and improve energy efficiency.

[0003] Physical models predict wind speed and corresponding power output of wind farms by constructing sophisticated numerical models of atmospheric physical processes. A representative example of this type of model is numerical weather prediction (NWP), which relies on complex computational equations to simulate dynamic changes in atmospheric conditions, thereby capturing potential patterns in wind power variations. Although physical models theoretically cover a wide range of forecasting needs from short-term to long-term, they face numerous challenges in practical applications. On the one hand, the construction and computation of physical models heavily depend on high-quality meteorological data and accurate geographical parameters; acquiring and processing this data is often time-consuming, labor-intensive, and costly. On the other hand, due to the extremely high computational complexity of the models and their reliance on expert knowledge, the timeliness of physical models in ultra-short-term and short-term forecasts often fails to meet practical needs, especially when facing rapidly changing atmospheric conditions.

[0004] Unlike physical models, statistical models rely on analyzing the correlation between past and future data to construct a predictive framework. Classical statistical models, such as Kalman filtering (KF) and autoregressive moving average (ARIMA), predict future wind power output by quantifying the inherent structure of time-series data. For example, Tena García et al. divided daily wind turbine power data into high-power and low-power samples. They found that high-power samples exhibit seasonality and demonstrated that ARIMA can effectively predict high-power samples. However, with the rapid development of wind power technology and the continuous expansion of wind farm scale, the volatility and uncertainty of wind power output have increased significantly, making it difficult for traditional statistical models to handle these complex nonlinear relationships.

[0005] Currently, with the widespread adoption of big data, the research trend in wind power forecasting is to explore more advanced and flexible artificial intelligence models. These emerging models not only inherit the advantages of traditional statistical models in time series analysis, but also, through their powerful nonlinear mapping and feature learning capabilities, better capture the intricate nonlinear relationships between wind power generation and various meteorological factors, geographical location, and seasonal variations. Furthermore, under extreme weather conditions or rapidly changing wind conditions, artificial intelligence models exhibit stronger robustness and adaptability.

[0006] Early wind power forecasting research primarily employed single artificial intelligence models, including machine learning and deep learning algorithms, to extract potential information from data. The nonlinear modeling capabilities of AI models have shown significant advantages in handling nonlinear relationships, thus attracting considerable attention. While these models outperform statistical models in handling nonlinear data, several factors still hinder the accuracy of predictions. One such factor is the diversity of information in the raw data, which makes accurately identifying trends and characteristics more complex. Furthermore, unilaterally choosing a single model does not guarantee optimal results in all situations.

[0007] Given the complexity of wind power forecasting and its high accuracy requirements, research using interval-valued time series for wind power forecasting is crucial, as these series provide deeper and broader information. In recent years, interval-valued time series have been widely applied in various fields such as carbon trading price forecasting and crude oil price forecasting. However, research on interval-valued wind power series remains relatively limited. Furthermore, the instability and complexity of wind power generation make point-valued wind power series forecasting quite difficult, while interval-valued wind power series data contain even more potential unknown information. Although this information is valuable, direct modeling is a challenging task. Therefore, feature extraction has become a core strategy for wind power series forecasting, aiming to extract key information through in-depth analysis of the data's intrinsic structure. Among feature extraction methods, decomposition methods have attracted considerable attention due to their ability to deeply reveal the multi-level fluctuations and trends of the original environmental signals and to construct subsequences at different feature scales. Commonly used decomposition techniques include wavelet transform (WT), empirical mode decomposition (EMD) and its variants, and variational mode decomposition (VMD). Among these, the VMD method exhibits strong robustness and high accuracy in component separation. However, the complexity of interval-valued wind power sequences places higher demands on VMD algorithms, especially in selecting optimal parameters for effective feature extraction. Current research mainly focuses on optimizing the number of modes K, but due to the intricate interaction between the number of modes and the penalty coefficient, the penalty coefficient α is also a key factor. With the development of intelligent algorithms, new approaches have been provided to determine the optimal parameters for VMD. Therefore, some scholars have used optimization algorithms to jointly optimize K and α, such as Differential Evolution (DE), Grey Wolf Optimization (GWO), and Whale Optimization (WOA). Research results show that these combined models perform satisfactorily in high-accuracy wind power prediction. Furthermore, it must be recognized that the success of the optimization results depends on the choice of the fitness function. An inappropriate fitness function will affect the performance of the optimization algorithm. Entropy is a reliable indicator of signal uncertainty and complexity, reflecting the internal information content of the signal. Therefore, in optimization algorithms, entropy is often chosen as the fitness function to evaluate the decomposition effect. For example, some scholars have used Arithmetic Optimization (AOA) to optimize the penalty coefficient and number of modes in VMD, and used envelope entropy as the fitness function, thereby further improving the decomposition performance of VMD. Some scholars have used permutation entropy to evaluate the problem of optimizing VMD parameters using two optimization algorithms. These methods have achieved significant results in determining VMD parameters, but they all evaluate the decomposition results from a single perspective and may not fully capture the characteristics of the data. Different entropies reflect the characteristics of the signal from different perspectives, so the optimization results obtained by using different entropies as fitness functions are not the same, resulting in poor robustness and low generalization ability of the optimization results.Furthermore, while decomposition techniques can optimize feature extraction and enable in-depth analysis, excessive subsequences can waste computational resources and lead to error accumulation, negatively impacting prediction accuracy. More importantly, the relationships between subsequences often lack interpretability, making it difficult for non-experts to grasp these conclusions.

[0008] While hybrid models perform well in wind power forecasting, there is still room for improvement in model selection. Many researchers, in pursuit of novel and complex forecasting models, often overemphasize the innovation of the model itself while neglecting the necessary correspondence between features and the model. This disconnect can lead to poor model performance and wasted resources. Furthermore, different models have varying degrees of importance for features; choosing only one type of model may result in poor robustness and reliability. Therefore, selecting the correct base model is crucial for hybrid forecasting. To address this gap, some studies have attempted to introduce error metrics as a criterion for selecting base models. However, this approach lacks a unified and clear measurement standard. Different researchers may choose different error metrics based on their own understanding, making direct comparison of results difficult and failing to ensure the universal applicability of the selected model across different scenarios. Moreover, the black-box nature of artificial intelligence models leads to poor interpretability; relying solely on error metrics for model selection focuses only on numerical relationships, ignoring the correlation between the model's internal features and actual physical processes. Therefore, while the predictions may be accurate, they often lack sufficient transparency and credibility.

[0009] Despite the significant achievements of previous research on wind power prediction, there are still some aspects that need further optimization and improvement.

[0010] (1) Current wind power prediction models mainly focus on point-value wind power sequences. However, this approach uses only one value to describe the wind power sequence at a specific moment, ignoring the inherent unpredictability and volatility of wind power, thus limiting the effectiveness of decision-making. Although existing studies have quantified potential variability by providing confidence intervals for point-value predictions, these intervals still rely on point-value predictions and cannot fully represent the dynamic range of wind power. In contrast, interval-valued time series can fully reflect the actual range of wind power changes within a given time period by specifying upper and lower limits, effectively reducing information loss caused by uncertainty and randomness. Therefore, constructing interval-valued wind power prediction models has become a promising direction for improving prediction accuracy and practicality. Despite some initial attempts and achievements, existing interval-valued wind power prediction models still have considerable room for improvement in terms of feature extraction accuracy, model selection optimization, and the interpretability of prediction results.

[0011] (2) The high complexity of interval-valued wind power sequences poses a significant challenge to feature extraction. Although VMD can delve into the intrinsic structure of wind power sequences, its performance is limited by parameter selection. To address this issue, researchers have attempted to optimize VMD parameters using intelligent optimization algorithms, but often overlook the crucial role of the fitness function in the decomposition effect. Most studies use only a single entropy function as the fitness function; this one-sided optimization may not fully capture data features, leading to optimization results that deviate from actual needs. In contrast, combining multiple entropy functions can reflect signal features from multiple dimensions, enhancing the robustness of the algorithm. Although decomposition techniques help optimize feature extraction and facilitate in-depth analysis, the relationships between the resulting subsequences often lack interpretability. Therefore, there is an urgent need to develop a more comprehensive approach to improve the accuracy and interpretability of feature extraction to better serve the analysis and prediction of interval-valued wind power sequences.

[0012] (3) In recent years, the booming development of wind power prediction has led to the diversification of algorithm applications, highlighting the complexity and depth of the model selection problem. Given the multifaceted nature of wind power fluctuations, selecting the right model is crucial. However, a major challenge lies in the mismatch between features and models; the sensitivity of models to features is often overlooked, resulting in the use of the same model for prediction on different feature sets. Traditional research focuses on optimizing the predictive performance of models while neglecting interpretability analysis, thus hindering a comprehensive understanding of model logic, generalization potential, and adaptability to different scenarios. This limitation restricts the comprehensiveness and long-term feasibility of model selection. Furthermore, current research on ensemble modeling ignores the time-varying accuracy of individual models, failing to capture their optimal predictive capabilities over time, thus hindering further improvements in prediction accuracy. Summary of the Invention

[0013] Purpose of the invention: In view of the problems and shortcomings of the existing technology, the present invention proposes an interpretable interval value wind power prediction method and system.

[0014] Technical Solution: An interpretable interval-value wind power prediction method that integrates multi-objective feature extraction, base model selection, and dynamic ensemble; including the following steps:

[0015] Step 1, construct the interval value wind power sequence:

[0016] Extract the maximum and minimum wind power values ​​within each set period to create an interval-valued wind power sequence.

[0017] Step 2, Multi-target feature extraction:

[0018] An interpretable multi-objective feature extraction method is used to identify the most predictive features from interval-valued wind power sequences. The feature extraction process consists of two stages. In the first stage, the Multi-Objective Whale Optimization Algorithm (MOWOA) is combined with Variational Mode Decomposition (VMD) technology to optimize the parameters of the VMD. In the second stage, the K-Shape algorithm is used to intelligently cluster the intrinsic mode components (IMFs) obtained from the first stage decomposition to extract the trend and fluctuation features of the wind power sequence.

[0019] Step 3, Selection of an interpretable base model:

[0020] By combining elastic net regression with the SHAP method, the contribution of each candidate model to the prediction target was quantified, and candidate models that exceeded the average contribution threshold were selected as the base models.

[0021] Step 4, Dynamic Integration:

[0022] By using the selected base model, a dynamic weighting mechanism is implemented on the Shapley value of each data point in the interval wind power series to obtain the wind power prediction results.

[0023] In step 1, a period of 1 hour is set, and the maximum and minimum wind power values ​​for each hour are extracted to construct an interval-valued wind power sequence, reflecting the dynamic range of short-term wind power changes. The interval-valued wind power sequence is defined as follows:

[0024]

[0025] Where, x H (t) and x L (t) represents the upper and lower bounds of the period, i.e. the maximum and minimum wind power values ​​in hour t.

[0026] The interval-valued wind power series is represented as the interval center sequence c(t) and the interval radius sequence r(t) to reveal the short-term trend and fluctuation of wind power. The center position of the interval, i.e., the interval center sequence, is represented by c(t), and the extent of the interval expansion, i.e., the interval radius sequence, is represented by r(t).

[0027] c(t)=(x H (t)+x L (t)) / 2;r(t)=(x H (t)-x L (t)) / 2 (2)

[0028] Finally, the interval-valued wind power sequence is expressed as:

[0029]

[0030] Where n is the length of the interval value wind power sequence.

[0031] In step 2, an interpretable multi-objective feature extraction method is used to find the most predictable features from the interval value wind power sequence. The feature extraction process includes two stages. In the first stage, the multi-objective whale optimization algorithm is combined with variational mode decomposition technology. In the second stage, the K-Shape algorithm is used to intelligently cluster the intrinsic mode components (IMF) obtained from the first stage decomposition.

[0032] In the first stage, variational mode decomposition (VMD) is used to extract features from the vertical dimension of the wind power sequence. For the two parameters of the VMD method: the number of modes K and the penalty coefficient α, a multi-objective function is constructed for optimization to quantitatively evaluate the decomposition effect. When constructing the multi-objective function, four entropy functions are integrated: sample entropy (SE), approximate entropy (AE), envelope entropy (EE), and permutation entropy (PE), and MOWOA is used to adjust the parameters to comprehensively evaluate and minimize the average information entropy of the modal components. In the second stage of feature extraction, the K-Shape clustering algorithm is introduced to intelligently cluster the multiple intrinsic modal components (IMFs) obtained in the first stage. The number of K-Shape clusters is set to 2 to extract the trend and fluctuation features of the wind power sequence in an interpretable manner. The specific execution steps are as follows:

[0033]

[0034] Where, p i (t) and q i (t) represents the IMF in two different clusters. The subsequences obtained by summing the IMFs in the clusters that reflect the trend characteristics in c(t) and r(t) are named t. c (t) and t r (t). For t c (t) and t r Reconstructing (t) yields the trend sequence and residual sequence of the interval-valued wind power sequence.

[0035] T(t) = [T L (t),T H (t)]=[t c (t)-t r (t),t c (t)+t r (t)] (5)

[0036] R(t)=[R L (t),R H (t)]=[x L (t)-T L (t),x H (t)-TH (t)] (6)

[0037] T L (t),T H (t) represent the minimum and maximum values ​​of the wind power trend sequence within each hour, respectively; R L (t),R H (t) represent the minimum and maximum values ​​of the wind power residual sequence within each hour, respectively.

[0038] In step 3, the selection of an interpretable basic model:

[0039] In the process of selecting the base model, elastic net regression and SHAP method were combined to quantify the contribution of each candidate model to the prediction target, and candidate models that exceeded the average contribution threshold were selected as the base models.

[0040] First, nonlinearity tests were performed on all trend sequences and residual sequences to confirm their nonlinear characteristics. Five candidate models were selected to create a model pool, including deep learning models LSTM and GRU, and machine learning models BPNN, SVR, and ELM. Subsequently, the Elastic Network Regression Model was used to effectively balance model complexity and prediction accuracy through its regularization properties to capture the complex correlation patterns between the prediction results of the five candidate models and the final prediction target. To analyze the intrinsic contribution of these candidate models, the SHAP method was used to quantitatively evaluate the marginal contribution of each candidate model to the overall output. The Shapley value of each candidate model was calculated using formula (7). Then, the average Shapley value of all candidate models was calculated as a threshold, i.e. This is used to identify candidate models that exceed this threshold, thus forming the final base model set. This refers to the average Shapley score of the candidate models.

[0041] These are the Shapley values ​​for the five candidate models.

[0042] Shapley value of candidate model i The calculation method is as follows:

[0043]

[0044] In the above formula, F represents the feature set used by the elastic net regression model, i.e., five candidate models. S is a subset of F that does not contain candidate model i. x (S) represents the elastic net regression model's prediction for sample x when only the feature set of S (excluding i) is used, where sample x refers to the trend sequence and residual sequence. x(S∪{i}) represents the prediction of sample x using S and candidate model i. |S| represents the element count in S, and |F| represents the total number of elements in f. It is a weighting factor.

[0045] Step 4, Dynamic Integration:

[0046] By using the selected base model, a dynamic weighting mechanism is implemented on the Shapley value of each data point in the interval wind power series to obtain the wind power prediction results.

[0047] The SHAP method is introduced in the weighting process of the base models in the ensemble. A dynamic weighted ensemble strategy based on Shapley values ​​is proposed, allowing for flexible adjustment of the weights of different base models based on their actual performance on different data points during the base model ensemble process, enabling real-time prediction updates. In this way, base models that perform better and contribute more to specific data points will be given higher weights, thus occupying a more important position in the final prediction results. The specific method is as follows:

[0048]

[0049] in This represents the weight of the base model m in the prediction result at time t. This represents the absolute value of Shapley in the prediction result of the basic model m at time t. It is the sum of the absolute values ​​of Shapley for all selected basic models at time t. This represents the prediction result of the basic model m at time t.

[0050] After performing the above calculations to obtain the prediction results for each trend sequence and residual sequence, the predicted trend sequence and residual sequence are integrated to obtain the final prediction result. The calculation formula is as follows:

[0051]

[0052] in, This refers to the predicted result of the interval value wind power sequence within time t. These are the prediction results for the minimum and maximum values ​​within time t, respectively. These are the minimum values ​​of the trend and residual sequences within time t, respectively. These are the maximum values ​​of the trend and residual sequences at time t, respectively.

[0053] An interpretable interval-value wind power prediction system includes the following modules:

[0054] (1) Constructing an interval-valued wind power sequence and a multi-objective feature extraction module: The sampled wind power sequence is divided according to a set period, and the maximum and minimum wind power values ​​of the set period are extracted to generate an interval-valued wind power sequence. Then, a multi-objective whale algorithm based on a multivariate entropy fitness function evaluation system is constructed to optimize VMD technology, intelligently adjusting the number of VMD modes and the penalty coefficient. The K-Shape algorithm is used to identify the trend and fluctuation characteristics in the interval-valued wind power sequence in an interpretable manner, and these features are integrated to enhance the interpretability of the model and reduce data complexity.

[0055] (2) Explainable Base Model Selection Module: First, nonlinearity tests are performed on all trend and residual sequences to confirm their nonlinear characteristics. Then, a candidate model pool containing LSTM, GRU, BPNN, SVR, and ELM is constructed. Next, elastic network regression and the SHAP method are used to quantitatively evaluate the marginal contribution of each candidate model to the overall prediction, thereby obtaining a ranking of the importance of each candidate model to the final prediction result. Based on this, the average Shapley value of each candidate model is calculated as a threshold, and models exceeding this threshold are selected to form the final base model set.

[0056] (3) Dynamic integration module: Based on the Shapley value, the weight of each basic model is flexibly adjusted according to its actual performance on different data points during the integration process, so as to realize real-time update prediction.

[0057] (4) Interval Reconstruction and Model Evaluation Module: The predicted trend and residual upper and lower bounds are integrated into an interval value wind power sequence to obtain the interval value wind power sequence prediction results for a set period.

[0058] The implementation process and methods of the system are the same and will not be described again.

[0059] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the interpretable interval value wind power prediction method as described above.

[0060] A computer-readable storage medium storing a computer program that performs the interpretable interval value wind power prediction method as described above.

[0061] Beneficial effects: Compared with the prior art, the interpretable interval value wind power prediction method and system provided by the present invention have the following advantages:

[0062] (1) A novel interval-value wind power prediction system is proposed, which integrates a robust multi-objective feature extraction mechanism, an interpretable basic model selection strategy, and a flexible dynamic weight allocation method. This comprehensive framework effectively utilizes the inherent variability and uncertainty of interval-value wind power sequences, greatly improving the reliability and credibility of the model prediction output.

[0063] (2) An interpretable multi-objective feature extraction method is proposed to simplify interval-valued wind power sequences and reveal subtle features. To extract smoother and more significant patterns from interval-valued wind power sequences, this study introduces a multi-objective whale algorithm based on a multivariate entropy fitness function evaluation system to optimize VMD technology. By intelligently adjusting the number of modes and penalty coefficients in VMD, the robustness and accuracy of mode extraction are significantly improved. Considering the complex and unknown relationships between intrinsic modal components, the K-Shape algorithm is used to identify trend and fluctuation features in the wind power sequence in an interpretable manner, and these features are integrated to enhance the interpretability of the model and reduce data complexity. This method enhances the comprehensiveness and typicality of feature extraction, thus laying a solid foundation for subsequent wind power sequence analysis and prediction.

[0064] (3) A novel method for selecting basic models was explored. The elastic network regression model was combined with the SHAP interpretation tool and applied to the basic model selection process. This method not only enriches the dimensions to be considered in the model selection stage, but also provides a more comprehensive and in-depth perspective on the performance evaluation and contribution measurement of each basic model through the detailed feature contribution information provided by SHAP.

[0065] (4) A dynamic weighting mechanism combining the core principles of SHAP was developed. As a leading model interpretation tool, SHAP accurately quantifies the contribution of input features to model prediction, clarifies the complex relationship between features and results, and improves the interpretability of the model. Based on this, SHAP is used to analyze the model contribution of specific data points in real time, allowing for dynamic adjustment of weight allocation. This ensures that the weights are closely aligned with the current features of the dataset, and optimizes model prediction through an online update strategy. Attached Figure Description

[0066] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0067] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0068] An interpretable interval-value wind power prediction method integrates multi-objective feature extraction and base model selection with dynamic ensemble; it includes the following steps:

[0069] Step 1, construct the interval value wind power sequence:

[0070] Extract the maximum and minimum wind power values ​​per hour to create an interval-valued wind power sequence, thereby gaining a more comprehensive understanding of how wind power changes over time.

[0071] The interval-valued wind power sequence is generated by dividing the original wind power time series with a sampling frequency of 15 minutes into multiple time windows of 1 hour, extracting features from the power values ​​within each window, and generating a sequence consisting of the interval center and the interval radius. Each interval represents the fluctuation range and average level of wind power within that time period. The interval-valued wind power sequence constructed by formulas (1) to (3) compresses the original high-frequency data into a description of fluctuation range. The interval center sequence c(t) represents the overall trend of the data, and the interval radius sequence r(t) represents the uncertainty of the data. This method separates uncertainty (interval radius) from trend (interval center), and can more effectively approximate the complex features of the interval-valued wind power sequence. The significance of the interval is that it is a symmetrical interval defined by the center point and the radius, providing an efficient and interpretable data foundation for subsequent analysis.

[0072] To improve wind power generation efficiency and ensure stable grid connection, a new data format, namely the interval-valued wind power series, has been introduced. Compared with the traditional point-valued wind power series, this format can effectively capture the fluctuation range of data within a specific time period, providing a richer and more comprehensive perspective. By accurately extracting the maximum and minimum wind power values ​​for each hour, an interval-valued wind power series is constructed, reflecting the dynamic range of short-term wind power changes. The interval-valued wind power series can be defined as follows:

[0073]

[0074] Where, x H (t) and x L (t) represents the maximum and minimum wind power in hour t, respectively.

[0075] Furthermore, to more clearly reveal the trend and fluctuation of short-term wind power, the interval-valued wind power sequence can also be represented as an interval center sequence and an interval radius sequence. The center position of the interval, i.e., the interval center sequence, is represented by c(t), and the extent of the interval expansion, i.e., the interval radius sequence, is represented by r(t).

[0076] c(t)=(x H (t)+x L (t)) / 2;r(t)=(xH (t)-x l (t)) / 2 (2)

[0077] Finally, the interval-valued wind power sequence can be expressed as:

[0078]

[0079] Where n is the length of the interval value wind power sequence.

[0080] Step 2, Multi-target feature extraction:

[0081] This paper proposes an interpretable multi-objective feature extraction method to identify the most predictive features from interval-valued wind power sequences. The feature extraction process consists of two stages. In the first stage, the Multi-Objective Whale Optimization Algorithm (MOWOA) is combined with Variational Mode Decomposition (VMD) to minimize data redundancy and complexity while retaining key information. In the second stage, the K-Shape algorithm is used to intelligently cluster the Intrinsic Mode Components (IMFs) obtained from the first stage decomposition, ensuring that the most predictive features are extracted from the interval-valued wind power sequences. In wind power sequence modeling, directly processing the raw data often limits the ability to fully utilize its inherent information, leading to reduced prediction performance. To address this issue, an interpretable multi-objective feature extraction strategy is proposed to capture the complex features in wind power sequences. Initially, the VMD method is used to obtain features from the vertical dimension of the wind power sequence. The effectiveness of VMD is highly dependent on two parameters: the number of modes K and the penalty coefficient α. An excessively large K value may lead to spurious modes, while an excessively small K value may not adequately represent the data features. Similarly, α directly affects the bandwidth of modal components; an excessively small α value may lead to modal aliasing, while an excessively large α value may ignore local details. Existing studies typically determine optimal parameters through repeated trials, lacking objectivity. Therefore, this invention focuses on optimizing VMD parameters by constructing a multi-objective function to quantitatively evaluate the decomposition effect. Given the limitations of a single entropy function, this invention integrates four entropy functions: sample entropy (SE), approximate entropy (AE), envelope entropy (EE), and permutation entropy (PE), and uses MOWOA to adjust the parameters, comprehensively evaluating and minimizing the average information entropy of modal components to maximize information extraction efficiency.

[0082] Four entropies are defined as optimization objectives: In each iteration, the four entropies of each intrinsic modal component (IMF) are calculated in the fitness function, and the average entropy value of all IMFs is used as the four independent fitness function values. Multi-objective optimization is performed using MOWOA: 50 whale individuals (num_whales) are generated, with each whale's position corresponding to two VMD parameters K and α, i.e., the position vector of each whale individual is [K, α]. The VMD parameters K and α are updated according to the shrinking encirclement, bubble net attack, and random search strategies of the whale optimization algorithm, generating multiple candidate solutions. Non-dominated solutions are maintained and stored using the Pareto front in multi-objective optimization. After each iteration, the new solution is compared with the archived solutions, and non-dominated solutions are retained to ensure that the final solution set is Pareto optimal. Optimal parameters are evaluated and selected: By analyzing the Pareto front, parameters that minimize the four entropies are selected to obtain the best VMD decomposition results, ensuring that the intrinsic modal components have low complexity and high information content.

[0083] In the second stage of feature extraction, the K-Shape clustering algorithm is introduced to reduce the time consumption and error accumulation associated with directly predicting intrinsic mode components (IMFs). Wind power series data typically exhibits a bimodal characteristic, including trends and fluctuations. Trends reflect the overall changes in wind power output over time, such as diurnal and seasonal variations, while fluctuations show rapid changes in the short term, potentially stemming from factors such as weather changes and wind speed fluctuations. Therefore, in this embodiment, the number of K-Shape clusters is set to 2 to extract the trend and fluctuation features of interval-valued wind power series in an interpretable manner. The specific execution steps are as follows:

[0084]

[0085] Where, p i (t) and q i (t) represent the intrinsic mode components (IMFs) in two different clusters. The subsequences obtained by summing the IMFs of the clusters that reflect the trend characteristics in c(t) and r(t) are named t, respectively. c (t) and t r (t). For t c (t) and t r Reconstructing (t) yields the trend sequence and residual sequence of the interval-valued wind power sequence.

[0086] T(t) = [T L (t),T H (t)]=[t c (t)-t r (t),t c (t)+t r (t)] (5)

[0087] R(t)=[R L (t),R H (t)]=[x L (t)-T L (t),x H (t)-T H (t)] (6)

[0088] T L (t),T H (t) represent the minimum and maximum values ​​of the wind power trend sequence within each hour, respectively; R L (t),R H (t) represent the minimum and maximum values ​​of the wind power residual sequence within each hour, respectively.

[0089]

[0090]

[0091] The meanings of the formulas and their parameters in Algorithm 1 are as follows:

[0092] (a).X(k+1)=X * (k)-A·D

[0093] (b). X(k+1)=x rand (k)-A·D

[0094] (c)

[0095] In formulas (a), (b), and (c), D = |C·X * (k)-X(k)|,D′=|X * (k)-X(k)|. k is the number of iterations; X * X(k) is the current optimal humpback whale position vector; X(k) represents the current position of the humpback whale, and X(k+1) represents the position of the humpback whale in the next iteration. rand It is a random humpback whale; A and C are coefficient vectors used to adjust the search direction and step size; b is the spiral shape constant. l is a random number between [-1, 1], and p is a random number between [0, 1].

[0096] (d).SBD(x,y)=1-max(NCC c (x,y))

[0097] Where x and y are two time series data, x = (x1, x2, ..., xy) n ),y=(y1,y2,…,y n ); NCC c(x,y) represents the normalized cross-correlation coefficient of the two sequences.

[0098] (e)

[0099] T(t) = [T L (t),T H (t)]=[t c (t)-t r (t),t c (t)+t r (t)]

[0100] (f).

[0101] R(t)=[R L (t),R H (t)]=[x L (t)-T L (t),x H (t)-T H (t)]

[0102] In formulas (e) and (f), t c (t) and t r (t) reflects the trend characteristics of c(t) and r(t) respectively; T L (t),T H (t) represent the minimum and maximum values ​​of the wind power trend sequence within each hour, respectively; R L (t),R H (t) refers to the minimum and maximum values ​​of the wind power residual sequence within each hour, respectively.

[0103] Step 3, Selection of an interpretable base model:

[0104] In the process of selecting the base model, elastic net regression is combined with the SHAP method to quantify the contribution of each candidate model to the prediction target, and the candidate model that exceeds the average contribution threshold is selected as the base model.

[0105] To develop an efficient and robust predictive modeling system, we propose a module for selecting the base model, which combines Elastic Net Regression with the SHAP method to accurately identify the base model most influential on the prediction target. First, nonlinear tests are performed on all trend and residual sequences to confirm their nonlinear characteristics. Since statistical models cannot capture these characteristics, five common artificial intelligence models are selected to create a model pool, including deep learning models LSTM and GRU, and machine learning models BPNN, SVR, and ELM. Subsequently, this strategy relies on the Elastic Net Regression model, which effectively balances model complexity and prediction accuracy through its regularization properties, to capture the complex correlation patterns between the prediction results of the five candidate models and the final prediction target. To analyze the intrinsic contributions of these models, the SHAP method is used to quantitatively evaluate the marginal contribution of each candidate model to the overall output. By calculating the Shapley value of each candidate model using formula (7), we can gain insight into their impact and relative importance on the final prediction result. Then, the average Shapley value of all candidate models is calculated as a threshold, i.e. This is used to identify models that significantly exceed this value, thus forming the final set of base models. This strategy ensures the superior predictive performance of the selected base models while maintaining the generalization ability and stability of the ensemble learning models by limiting the size of the set, thereby helping to reduce the risk of overfitting due to excessive complexity. This refers to the average Shapley score of the candidate models. These are the Shapley values ​​for the five candidate models.

[0106] Shapley value of candidate model i The calculation is as follows:

[0107]

[0108] In the above formula, F represents the feature set used by the elastic net regression model, i.e., five candidate models. S is a subset of F that does not contain candidate model i. x (S) represents the elastic net regression model's prediction for sample x when only the feature set of S (excluding i) is used. x (S∪{i}) represents the prediction of sample x when using S and candidate model i.

[0109] |S| represents the element count in S, and |F| represents the total number of elements in F. It is a weighting factor that takes into account all possible combinations and orders of features to ensure a fair distribution of contributions.

[0110] Step 4, Dynamic Integration:

[0111] A dynamic weighting mechanism is implemented using the Shapley value at each data point based on the selected base model, thereby obtaining more accurate and robust wind power prediction results. Each data point represents the hourly value of the interval wind power sequence created in step 1.

[0112] Traditional ensemble learning methods typically assign fixed weights to a set of well-performing base models using a static approach after selection. While this method is easy to implement, it overlooks a crucial factor: the contribution of each base model is not constant across different time points or data samples. To more accurately capture and utilize the dynamic changes in contribution rates, this embodiment introduces the SHAP method during the weighting process of the base models in the ensemble. This embodiment proposes a dynamic weighted ensemble strategy based on Shapley values. This dynamic weight allocation mechanism allows for flexible adjustment of the weights of different base models during the ensemble process based on their actual performance on different data points, enabling real-time updates to predictions. Thus, base models that perform better and contribute more significantly to specific data points will be given higher weights, thereby occupying a more important position in the final prediction result. The specific method is as follows:

[0113]

[0114] in This represents the weight of the base model m in the prediction result at time t. This represents the absolute value of Shapley in the prediction result of the basic model m at time t. It is the sum of the absolute values ​​of Shapley for all selected basic models at time t. This represents the prediction result of the basic model m at time t.

[0115] After performing the above calculations to obtain the prediction results for each trend sequence and residual sequence, the predicted trend sequence and residual sequence are integrated to obtain the final prediction result. This is achieved by reconstructing the hourly maximum and minimum values ​​of the wind power sequence using the predicted maximum and minimum values ​​of the trend sequence and residual sequence. The calculation formula is as follows:

[0116]

[0117] in, This refers to the predicted result of the interval value wind power sequence within time t. These are the prediction results for the minimum and maximum values ​​within time t, respectively. These are the minimum values ​​of the trend and residual sequences within time t, respectively. These are the maximum values ​​of the trend and residual sequences at time t, respectively.

[0118] An interpretable interval-value wind power prediction system includes the following modules:

[0119] (1) Constructing an interval-valued wind power sequence and a multi-objective feature extraction module: The wind power sequence with a sampling frequency of 15 minutes is divided into 1-hour intervals, and the maximum and minimum wind power values ​​for each hour are extracted to generate an interval-valued wind power sequence. Then, a multi-objective whale algorithm based on a multivariate entropy fitness function evaluation system is constructed to optimize VMD technology and intelligently adjust the number of VMD modes and the penalty coefficient.

[0120] Based on the Multi-Objective Whale Optimization Algorithm (MOWOA), this method constructs a multi-dimensional evaluation system (i.e., a multivariate entropy fitness function evaluation system) by integrating multiple entropy functions (sample entropy, permutation entropy, envelope entropy, and approximate entropy) to dynamically optimize the two core parameters of VMD—the number of modes (K) and the penalty coefficient (α). Traditional methods typically rely on a single entropy index to measure the quality of modal components, which can easily lead to overfitting or information loss in the decomposition results due to the single evaluation dimension. In contrast, this method uses a multi-objective optimization framework to simultaneously minimize multiple entropy objectives, quantifying the quality of modal components from different dimensions. This comprehensively captures the complex characteristics of the signal, ensuring that the decomposed modal components possess low redundancy, high interpretability, and physical meaning.

[0121] The K-Shape algorithm is used to identify trend and fluctuation characteristics in interval-valued wind power series in an interpretable manner, and these characteristics are integrated to enhance the interpretability of the model and reduce data complexity. The intrinsic mode components (IMFs) obtained in the first stage of step 2 are divided into two different clusters using the K-Shape clustering algorithm. The central sequence of each cluster may represent a typical pattern, such as high volatility, stable rise, periodic fluctuation, etc. The components in the two clusters obtained by the K-Shape algorithm are accumulated (i.e., integrated). The two integrated sequences represent the trend and fluctuation characteristics in the wind power time series, respectively. This reduces the number of IMFs obtained in the first stage of step 2, thereby reducing the computational burden and error accumulation. Furthermore, the integrated sequences can be reconstructed and interpreted as the trend and residual sequences in formulas (5) and (6).

[0122] (2) Explainable Base Model Selection Module: First, nonlinearity tests are performed on all trend and residual sequences to confirm their nonlinear characteristics. Then, a candidate model pool containing LSTM, GRU, BPNN, SVR, and ELM is constructed. Next, elastic network regression and the SHAP method are used to quantitatively evaluate the marginal contribution of each candidate model to the overall prediction, thereby obtaining a ranking of the importance of each candidate model to the final prediction result. Based on this, the average Shapley value of each candidate model is calculated as a threshold, and models exceeding this threshold are selected to form the final base model set.

[0123] (3) Dynamic integration module: Based on the Shapley value, the weight of each basic model is flexibly adjusted according to its actual performance on different data points during the integration process, so as to realize real-time update prediction.

[0124] (4) Interval Reconstruction and Model Evaluation Module: The predicted trend and residual upper and lower bounds are integrated into an interval value wind power sequence to obtain the hourly interval value wind power sequence prediction results.

[0125] Example:

[0126] (1) Data description and preprocessing

[0127] This study used wind power data from a wind farm in Guizhou Province, China, spanning from December 1, 2019 to November 30, 2020. The data was divided into four seasons. The sampling frequency was 15 minutes, and missing values ​​were imputed using cubic spline interpolation. The dataset was divided into training, validation, and test sets in a 6:2:2 ratio.

[0128] The irregularity and volatility of wind power data pose challenges to modeling, mainly due to the presence of noise and outliers. To address this issue, a rolling average denoising technique is employed to smooth the raw wind power data, thereby reducing noise interference with model training. Furthermore, the Hampel detector method is used to identify and correct outliers, effectively preserving the basic characteristic information of the data.

[0129] (2) Model performance evaluation

[0130] To comprehensively evaluate the predictive power of the proposed model, four evaluation metrics were used: Mean Absolute Error (IMAE), Root Mean Square Error (IRMSE), Sum of Squared Error (ISSE), and Mean Relative Variance (IARV). Specifically, IMAE, IRMSE, and ISSE are indicators of prediction accuracy, with lower values ​​indicating higher precision. IARV is used to compare the prediction error of the reference model with the time series mean. A closer IARV value to 0 indicates better performance, suggesting that the model's predictions better reflect the dynamics of the actual series.

[0131]

[0132] Where x L (t),x H (t) represent the maximum and minimum wind power values ​​within one hour, respectively. These are the corresponding predicted values, and n represents the length of the interval wind power sequence.

[0133] To more intuitively compare the performance of various models, we added the following metrics, using different error evaluation indicators to represent the degree of improvement of the two methods:

[0134]

[0135]

[0136] Wherein, IMAE1 and IMAE2 represent the mean absolute errors of the baseline model and the model proposed in this invention, respectively.

[0137] (3) Optimal Feature Extraction

[0138] To effectively capture the rich information contained in interval-valued wind power sequences, this invention designs a multi-objective feature extraction method to determine the optimal features. First, the parameters of VMD are optimized using MOWOA to facilitate the decomposition of the original interval-valued wind power sequences and effectively reduce data complexity. The optimized parameter results are shown in Table 1. After sequence decomposition, the K-Shape method is used to analyze features and similarities to further extract features and minimize error accumulation. The number of clusters in the K-Shape method is set to 2, which clearly distinguishes the trend and fluctuation characteristics of the interval-valued wind power sequence data. Therefore, the trend sequence and residual sequence of the interval-valued wind power sequence are constructed using equations (5) and (6). Through these steps, the optimal feature sequence is determined for subsequent modeling.

[0139] Table 2 Optimization parameter combinations for VMD

[0140]

[0141] (4) Interval forecasting

[0142] In predictive analytics, different types of predictive models often perform well on specific datasets. However, traditional statistical models often struggle to effectively capture nonlinear factors in time series data, which are crucial for accurate prediction. Therefore, this embodiment performs nonlinearity testing on the selected optimal feature sequences (i.e., all trend sequences c(t) and residual sequences r(t)) before model selection. The results show that when the embedding dimension is between 2 and 5, all p-values ​​are equal to 0 within the 95% confidence interval. This rejects the null hypothesis and confirms the nonlinear relationship in the data. Given the advantages of artificial intelligence models in handling complex nonlinear data structures, this embodiment narrows the model selection to three representative machine learning models and two deep learning models, forming a model pool.

[0143] Choosing an appropriate base model is crucial to the overall effectiveness and reliability of the integrated model. This embodiment combines elastic net regression with the SHAP method to scientifically identify the base models that have a significant impact on the prediction results. This method quantifies the contribution of each candidate model and clearly illustrates the influence of each feature or model in an intuitive hierarchical diagram using Shapley values. Ultimately, candidate models with Shapley values ​​above the average are selected as base models, demonstrating strong predictive capabilities and complementing each other to improve the accuracy and stability of the integrated model.

[0144] Shapley values ​​effectively explain the contribution of each model to each data point. Therefore, formulas (8) and (9) are used to dynamically weight the prediction results of the basic model, thereby giving full play to the best prediction performance of each model at the corresponding data point. Then, formula (10) is used to reconstruct the interval value wind power prediction results.

[0145] (5) Comparative Analysis

[0146] Four sets of comparative experiments were analyzed in depth to comprehensively evaluate the effectiveness of the proposed model. These experiments included comparisons with prediction models that do not involve data transformation, models without feature extraction, relatively simple weighted prediction models, and models lacking dynamic ensemble. In addition, three recent interval value prediction methods were compared and analyzed. Specific prediction evaluation metrics for each model are detailed in Table 2, and the degree of improvement of each comparative model on various error evaluation metrics is shown in Table 3. The comparative experimental results show that the proposed model has a significant improvement in prediction accuracy, outperforming the baseline model.

[0147] Table 2. Model Comparison Results

[0148]

[0149] Table 3. Improvement degree under different error evaluation indicators

[0150]

[0151]

[0152] (6) Comparison with models of untransformed data

[0153] To optimize the utilization of short-term wind power fluctuation data, this embodiment converts the maximum and minimum values ​​of the hourly wind power sequence into interval center and radius sequences, respectively reflecting the central location and extended range of the interval. To verify the necessity of this conversion method, we compared the proposed model with a model that directly uses the maximum and minimum wind power sequences; this experiment is referred to as the "model with unconverted data." Taking spring data as an example, compared with the model with unconverted data, the proposed model improves the IMAE, IRMSE, and ISSE indices by 9.6388%, 18.1707%, and 10.0476%, respectively (see Table 3). Experimental results show that when using the converted dataset, the model proposed in this embodiment exhibits stronger predictive power than the unconverted dataset. This indicates that sequence c(t) is effective in capturing the location characteristics and trends of the interval sequence. Conversely, sequence r(t) visualizes the range extension and small fluctuations of the interval sequence. Therefore, choosing c(t) and r(t) to describe the overall behavior of the sequence not only helps to extract valuable information more effectively, but also improves the predictive power of the model, enabling us to gain a deeper understanding of the complex patterns and trends inherent in the dataset.

[0154] (7) Comparison with models without feature extraction

[0155] Wind power data is nonlinear and non-stationary time series data influenced by various complex factors. Due to its high volatility and uncertainty, directly predicting raw wind power data often fails to effectively capture its underlying patterns and characteristics. During model training, this lack of crucial information negatively impacts the accuracy and reliability of prediction results. Therefore, feature extraction is a critical step in wind power prediction. Through feature extraction, we can identify key variables affecting wind power generation, such as wind direction, wind speed, and temperature, and detect periodic and trend components in the time series. This process provides the model with more refined and higher-quality data input. To determine the effectiveness of the proposed multi-objective feature extraction strategy in wind power prediction, we designed a comparative experiment, directly using interval center and radius sequences to build the model without using MOWOA-VMD and K-Shape methods for feature extraction. This comparative experiment is referred to as the "model without feature extraction."

[0156] As shown in Table 3, compared with the baseline model, the model proposed in this embodiment shows improvements across three different error evaluation metrics. Notably, this improvement varies with seasonal data, indicating that the feature information contained in the wind power data differs seasonally and affects the overall prediction results. This further emphasizes the necessity of feature extraction. Furthermore, the comparative results demonstrate that implementing multi-objective feature extraction can improve the effectiveness of the wind power prediction model. Simultaneously, it significantly reduces the complexity of model training.

[0157] (8) Comparison with interpretable base model selection and dynamic integration modules

[0158] To evaluate the effectiveness of the proposed interpretable base model selection and dynamic integration strategy in wind power forecasting, this embodiment includes two comparative experiments. These experiments aim to determine the capabilities of conventional wind power forecasting models that do not implement interpretable base model selection or dynamic integration, thereby validating the advantages of the new strategy. Both experimental groups used the same dataset as the proposed model and implemented a multi-objective feature extraction method to ensure fairness in the comparison.

[0159] In the first comparative experiment, neither interpretable base model selection nor dynamic ensemble was used. After making predictions using candidate models from the model pool, the final prediction result was obtained by averaging the predictions of the five candidate models. The second group omitted the dynamic ensemble mechanism and instead, after determining the optimal base model, the final prediction result was obtained by averaging the predictions of the selected base model. The results show that the model proposed by the method in this embodiment outperforms the comparative experimental group on all evaluation metrics. Furthermore, as shown in Table 3, the model using interpretable base model selection and dynamic ensemble shows greater improvement than the model using only interpretable base model selection without dynamic ensemble. For example, in autumn, the method combining interpretable base model selection and dynamic ensemble improved the baseline model by more than 38% on three evaluation metrics, while the method using only interpretable base model selection only improved it by about 1%. The results of the comparative experiments show that using SHAP can clearly identify the impact of each model on the prediction results, thereby improving the interpretability of the model and optimizing the prediction accuracy. In addition, by utilizing the Shapley value of each data point in the dynamic ensemble strategy, the best predictive performance of each model at different time points can be effectively utilized.

[0160] (9) Comparison with recent interval value prediction methods

[0161] To further validate the interval value prediction technique introduced in this embodiment, this section compares the proposed model with three recent methods: BEMD-DE-XGBoost, BEMD-MMPE-LSTM-BOA, and MVMD-FML-MMZOA. For a fair comparison, parameters were kept consistent. The analysis is as follows: Compared to our method, the BEMD-DE-XGBoost model shows an increase of over 70% in all error metrics. This significant deterioration is likely due to the method employed by BEMD-DE-XGBoost, which independently predicts each intrinsic modal component generated by the decomposition. This method ignores potential redundancy among intrinsic modal components, hindering the learning process. Furthermore, this method is both time-consuming and prone to error accumulation. Additionally, BEMD-DE-XGBoost relies entirely on XGBoost to predict all subsequences, neglecting the adaptability of the prediction model to specific datasets. In contrast, the BEMD-MMPE-LSTM-BOA framework, based on multivariate multiscale permutation entropy (MMPE) analysis and remodeling of modal components, captures internal factors, thereby reducing the accumulation of estimation errors and significantly lowering computational costs. This framework demonstrates superior performance, showing a significant improvement over BEMD-DE-XGBoost. While component reconstruction effectively reduces prediction complexity, the model still has room for further improvement. For example, a significant drawback of this model is the distorted performance on certain datasets. The empirical analysis above confirms that the proposed hybrid model can efficiently and accurately predict wind power generation data, marking substantial progress in this field. This embodiment provides a new perspective for wind power forecasting, particularly in handling the uncertainty of wind power output. By introducing interval value forecasting methods, it not only enriches the existing theoretical framework for wind power forecasting but also provides an effective tool for practical applications. Future research can consider other factors affecting wind power output fluctuations to provide more accurate and comprehensive forecasts. Simultaneously, exploring the application of interval value time series in forecasting other renewable energy sources will also promote the efficient utilization of renewable energy and the stable operation of the power system.

[0162] Obviously, those skilled in the art should understand that the steps of the interpretable interval value wind power prediction method or the modules of the interpretable interval value wind power prediction system described in the above embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by the computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

Claims

1. An interpretable interval-value wind power prediction method, characterized in that, Includes the following steps: Step 1, construct the interval value wind power sequence: Extract the maximum and minimum wind power values ​​within each set period to create an interval value wind power sequence; Step 2, Multi-target feature extraction: Using an interpretable multi-objective feature extraction method, the most predictive features are identified from interval-valued wind power sequences. The feature extraction process consists of two stages. In the first stage, the multi-objective whale optimization algorithm is combined with variational mode decomposition technology to optimize the parameters of variational mode decomposition. In the second stage, the K-Shape algorithm is used to intelligently cluster the intrinsic modal components obtained from the first stage decomposition to extract the trend and fluctuation characteristics of the wind power sequence. In the first stage, the variational mode decomposition method is used to obtain features from the vertical dimension of the wind power sequence; for the two parameters of the variational mode decomposition method: the number of modes. and penalty coefficient A multi-objective function is constructed to optimize and quantitatively evaluate the decomposition effect. When constructing the multi-objective function, four entropy functions are integrated: sample entropy, approximate entropy, envelope entropy and permutation entropy, and MOWOA is used to adjust the parameters to minimize the average information entropy of the modal components. In the second stage of feature extraction, the K-Shape clustering algorithm is introduced to intelligently cluster multiple intrinsic modal components obtained in the first stage. The number of K-Shape clusters is set to 2, so as to extract the trend and fluctuation features of the wind power sequence in an interpretable way. Step 3, Selection of an interpretable base model: By combining elastic net regression with the SHAP method, the contribution of each candidate model to the prediction target is quantified, and candidate models that exceed the average contribution threshold are selected as the base model. Step 4, Dynamic Integration: By using the selected base model, a dynamic weighting mechanism is implemented on the Shapley value of each data point in the interval wind power series to obtain the wind power prediction results.

2. The interpretable interval value wind power prediction method according to claim 1, characterized in that, In step 1, the set period is 1 hour. The maximum and minimum wind power values ​​for each hour are extracted to construct an interval wind power sequence. The interval wind power sequence is defined as follows: in, and These represent the upper and lower bounds of the period, respectively, i.e., the first... t Maximum and minimum wind power values ​​per hour; Representing the interval-valued wind power sequence as the interval center sequence and interval radius sequence This is used to reveal the trend and fluctuation of short-term wind power; the center position of the interval, i.e., the interval center sequence, is used. The extent of the interval expansion, i.e., the interval radius sequence, is represented by... express; Finally, the interval-valued wind power sequence is expressed as: in, It is the length of the interval value wind power sequence.

3. The interpretable interval value wind power prediction method according to claim 2, characterized in that, In step 2, the feature extraction process includes two stages. In the first stage, the multi-objective whale optimization algorithm is combined with variational mode decomposition technology. In the second stage, the K-Shape algorithm is used to intelligently cluster the intrinsic mode components obtained from the first stage decomposition. In the first stage, the variational mode decomposition method is used to obtain features from the vertical dimension of the wind power sequence; for the two parameters of the variational mode decomposition method: the number of modes. and penalty coefficient A multi-objective function is constructed to optimize and quantitatively evaluate the decomposition effect. When constructing the multi-objective function, four entropy functions are integrated: sample entropy, approximate entropy, envelope entropy and permutation entropy, and MOWOA is used to adjust the parameters to minimize the average information entropy of the modal components. In the second stage of feature extraction, the K-Shape clustering algorithm is introduced to intelligently cluster the multiple intrinsic modal components obtained in the first stage. The number of K-Shape clusters is set to 2 to extract the trend and fluctuation features of the wind power sequence in an interpretable manner. The specific implementation formula is as follows: (4) in, and These represent the IMFs in two different clusters; they will reflect... and The subsequences obtained by accumulating the IMFs in the clusters with mesotrality features are named respectively. and ;right and Reconstruction is performed to obtain the trend sequence and residual sequence of the interval value wind power series; (5) (6) These represent the minimum and maximum values ​​of the wind power trend sequence within each hour, respectively. These represent the minimum and maximum values ​​of the wind power residual sequence within each hour, respectively.

4. The interpretable interval value wind power prediction method according to claim 2, characterized in that, In the process of selecting the base model, elastic net regression is combined with the SHAP method to quantify the contribution of each candidate model to the prediction target, and the candidate model that exceeds the average contribution threshold is selected as the base model. First, nonlinearity tests were performed on all trend sequences and residual sequences to confirm their nonlinear characteristics; Five candidate models were selected to create the model pool. The five candidate models include deep learning models LSTM and GRU, as well as machine learning models BPNN, SVR and ELM. Subsequently, an elastic net regression model was used to capture the complex correlation patterns between the prediction results of the five candidate models and the final prediction target. The SHAP method is used to quantitatively evaluate the marginal contribution of each candidate model to the overall output; The Shapley value of each candidate model is calculated using formula (7); then the average Shapley value of all candidate models is calculated as the threshold, i.e. This is used to identify candidate models that exceed the threshold, thus forming the final base model set; among them, This refers to the average Shapley value of the candidate models. These are the Shapley values ​​for the five candidate models; Candidate Model Shapley value The calculation method is as follows: This represents the feature set used in the elastic network regression model. yes A model that does not contain candidate models A subset of; Indicates when only excluding of When considering the feature set, the elastic network regression model applies to the sample. The prediction, the sample This refers to trend series and residual series; Indicates that it is in use and candidate models Time for sample The prediction; express Counting elements in express The total number of elements in; It is a weighting factor.

5. The interpretable interval value wind power prediction method according to claim 2, characterized in that, By using the selected base model, a dynamic weighting mechanism is implemented on the Shapley value of each data point in the interval wind power series to obtain the wind power prediction results; The SHAP method is introduced during the weighting of the base models in the ensemble. A dynamic weighted ensemble strategy based on Shapley values ​​is proposed, which allows for flexible adjustment of the weights of different base models based on their actual performance on different data points during the base model ensemble process, enabling real-time prediction updates. Base models that contribute more to the performance of data points will be assigned higher weights. The formulas for weights and prediction results are as follows: in Representation of the basic model At any moment Weights in the prediction results; Representation of the basic model At any moment The absolute value of Shapley in the prediction results; It is all selected basic models at time 1000. The sum of the absolute values ​​of Shapley; Representation of the basic model At any moment The prediction results; After obtaining the prediction results for each trend sequence and residual sequence through calculation, the predicted trend sequence and residual sequence are integrated to obtain the final prediction result. The calculation formula is as follows: in, Refers to time Interval-valued wind power sequence prediction results within the range; They are time points The prediction results for the minimum and maximum values ​​within the range; They are time points The minimum value of the internal trend and the residual sequence. They are time points The maximum value of the internal trend and residual sequence.

6. An interpretable interval-value wind power prediction system, characterized in that, Includes the following modules: (1) Constructing an interval-valued wind power sequence and a multi-objective feature extraction module: Divide the sampled wind power sequence into a set period, extract the maximum and minimum wind power values ​​of the set period, and generate an interval-valued wind power sequence; then construct a multi-objective whale algorithm based on a multivariate entropy fitness function evaluation system to optimize VMD technology, and intelligently adjust the number of VMD modes and the penalty coefficient; use the K-Shape algorithm to identify the trend and fluctuation features in the interval-valued wind power sequence in an interpretable way, and integrate these features to enhance the interpretability of the model and reduce data complexity; (2) Explainable base model selection module: First, nonlinearity tests are performed on all trend sequences and residual sequences to confirm their nonlinear characteristics; then, a candidate model pool containing LSTM, GRU, BPNN, SVR and ELM is constructed; next, elastic network regression and SHAP methods are used to quantify the marginal contribution of each candidate model in the overall prediction, thereby obtaining the importance ranking of each candidate model to the final prediction result; on this basis, the average Shapley value of each candidate model is calculated as a threshold, and models exceeding this threshold are selected to form the final base model set; (3) Dynamic integration module: Based on the Shapley value, the weights of each basic model are flexibly adjusted according to the actual performance of each model on different data points during the integration process, so as to realize real-time update of prediction; (4) Interval reconstruction and model evaluation module: The predicted trend and residual upper and lower bounds are integrated into an interval value wind power sequence, thereby obtaining the prediction result of the interval value wind power sequence for a set period.

7. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the interpretable interval value wind power prediction method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that performs the interpretable interval value wind power prediction method as described in any one of claims 1-5.

Images