Enhanced wind power prediction method and device based on deep learning, equipment and storage medium

Abstract

The invention provides an enhanced wind power prediction method and device based on deep learning, equipment and a storage medium, and the method comprises the steps: collecting original wind power data, and carrying out the CEEMDAN mode decomposition, and obtaining a plurality of intrinsic mode components; evaluating the intrinsic mode component by using an IMF collaborative entropy balance index to obtain an evaluation result; based on the evaluation result, obtaining an optimized intrinsic mode component; a wind power prediction model is constructed, a result is output, the Kolmogorov-Arnold network extracts nonlinear features, and the long and short term memory network performs time sequence modeling. According to the method, parameters are calibrated by using a Newton-Rafson optimizer, decomposition dynamic optimization is realized, and the generalization ability is enhanced; and a Kolmogorov-Arnold network and a long and short term memory network are combined, so that the limitation of a traditional model is broken through. And the capability of responding to wind speed change and the adaptability to extreme weather and sudden change wind speed are improved.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and in particular to an enhanced wind power prediction method, apparatus, device, and storage medium based on deep learning. Background Technology

[0002] In recent years, model structures and technical approaches have converged and developed. On the one hand, traditional time series forecasting methods such as ARIMA are still used as basic modeling tools in some systems; on the other hand, machine learning methods such as artificial neural networks, support vector regression, and XGBoost have become mainstream, capable of handling complex data relationships. Deep learning has improved forecasting performance, and recurrent neural networks and their variants have advantages in capturing long-term dependencies. Simultaneously, researchers are attempting to combine different models to form hybrid architectures to balance stability and accuracy. Although various forecasting systems have been deployed, most schemes rely on historical data and lack real-time meteorological feedback mechanisms. Furthermore, existing models are mostly based on fixed structural designs, and their adaptability and generalization capabilities under different regional and climatic conditions need to be strengthened, pointing the way for future technological improvements. Existing technologies have made significant progress in wind power forecasting, but problems remain. Traditional time series forecasting methods are based on linear assumptions, making it difficult to capture the nonlinear characteristics of wind power generation, resulting in limited forecast accuracy. While machine learning models can identify complex patterns and handle nonlinear relationships, they are highly dependent on feature selection, and extracting dynamic features and time-related information is challenging, requiring extensive sample learning. Deep learning methods have advantages in processing time series data, but also face challenges such as large data requirements and difficulty in feature extraction.

[0003] In the field of ultra-short-term forecasting, existing models fall short. For example, the mean absolute error of 15-minute forecasts is relatively high, the ability to respond quickly to changes in wind speed needs improvement, and adaptability to extreme weather or sudden wind speed changes is also a challenge. Although some studies have attempted to integrate new technologies to optimize the ability to respond to sudden wind speed changes, further exploration is needed. Summary of the Invention

[0004] This invention aims to at least address the technical problem of low prediction accuracy in existing technologies, and innovatively proposes an enhanced wind power prediction method, device, equipment, and storage medium based on deep learning.

[0005] To achieve the above-mentioned objectives of this invention, this invention provides an enhanced wind power prediction method based on deep learning, the method comprising: Raw wind power data was collected and CEEMDAN mode decomposition was performed to obtain multiple intrinsic mode components. The intrinsic mode components were evaluated using the IMF co-entropy balance index to obtain the evaluation results; Based on the evaluation results, the parameters of CEEMDAN were adaptively calibrated using a Newton-Raphson optimizer to obtain optimized intrinsic mode components. A wind power prediction model is constructed based on the Kolmogorov-Arnold network and the long short-term memory network. The optimized intrinsic mode components are input into the wind power prediction model, and the wind power prediction model is used to predict wind power and output the prediction results. The Kolmogorov-Arnold network is used to extract nonlinear features from the optimized intrinsic mode components, and the long short-term memory network is used to perform time-series modeling on the extracted nonlinear features.

[0006] On the other hand, the present invention also proposes a deep learning-based enhanced wind power prediction device, the device being used to execute the aforementioned deep learning-based enhanced wind power prediction method; the device includes: The data acquisition module is used to collect raw wind power data; A mode decomposition module, connected to the acquisition module, is used to perform CEEMDAN mode decomposition to obtain multiple intrinsic mode components. An evaluation module, connected to the mode decomposition module, is used to evaluate the intrinsic mode components using the IMF cooperative entropy balance index to obtain evaluation results. An optimization module, connected to the evaluation module, is used to adaptively calibrate the parameters of CEEMDAN using a Newton-Raphson optimizer based on the evaluation results, and obtain multiple optimized intrinsic mode components. The feature extraction module is used to perform nonlinear feature extraction on the optimized intrinsic mode components based on the Kolmogorov-Arnold network to obtain the corresponding feature vectors. The time series modeling module, connected to the feature extraction module, is used to perform time series modeling on the extracted nonlinear features based on the long short-term memory network to generate a predicted sequence of wind power. The aggregation output module, connected to the time series modeling module, is used to aggregate the prediction sequence and output the final wind power prediction result.

[0007] In another aspect, the present invention provides a computer device comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the deep learning-based enhanced wind power prediction method when executing the executable instructions.

[0008] In another aspect, the present invention provides a computer-readable storage medium comprising: A memory on which computer programs are stored; A processor is configured to execute the program in the memory to implement the deep learning-based enhanced wind power prediction method.

[0009] The beneficial effects of this invention are as follows: This invention can accurately capture the nonlinear characteristics of wind power generation and screen key modal components by using CEEMDAN mode decomposition combined with IMF collaborative entropy balance index evaluation. Through adaptive calibration of CEEMDAN parameters using a Newton-Raphson optimizer, the decomposition process is dynamically optimized, enhancing the generalization ability under different regional and climatic conditions. Furthermore, by combining the Kolmogorov-Arnold network (KAN) and the Long Short-Term Memory (LSTM) network, KAN is responsible for extracting the deep nonlinear features of each modal component, while LSTM performs temporal modeling and dynamic prediction. The synergy of these two methods overcomes the limitations of traditional models in feature extraction and temporal correlation capture. Finally, through multi-index evaluation using mean absolute error, root mean square error, and mean absolute percentage error, this method is verified to significantly reduce prediction errors in ultra-short-term prediction scenarios, improve the ability to quickly respond to wind speed changes, and enhance adaptability to extreme weather and sudden wind speed changes.

[0010] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0011] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of an enhanced wind power prediction method based on deep learning according to the present invention; Figure 2 This is a structural framework diagram of an enhanced wind power prediction method based on deep learning according to the present invention. Figure 3 This is a flowchart illustrating the process of calculating the IMF collaborative entropy balance index according to the present invention. Figure 4 This is a schematic diagram of the Newton-Raphson optimizer of this invention; Figure 5 This is a structural diagram of the wind power prediction model of the present invention; Figure 6 This is a schematic diagram of some of the French wind power generation data in this invention; Figure 7 This is a schematic diagram of some of the wind power generation data in Türkiye in this invention; Figure 8 This invention relates to Entropy-based Fitness Functions; Figure 9 This invention relates to Energy-based Fitness Functions; Figure 10 This invention presents Statistical-feature-based Fitness Functions; Figure 11 This is a statistical chart showing the performance of different fitness functions of this invention on the MAPE index on the French dataset; Figure 12 This is a statistical chart showing the performance of different fitness functions of this invention on the RMSE index on the French dataset; Figure 13 This is a statistical chart showing the performance of different fitness functions of this invention on the MAE index on the French dataset; Figure 14 This is a statistical chart comparing the prediction curves of different fitness functions of the present invention; Figure 15 This is the test function table of the present invention; Figure 16 This is the parameter table of the optimization algorithm of this invention; Figure 17 This is the statistical results table of the present invention; Figure 18 This is a comparison chart of prediction curves from different optimization algorithms of this invention; Figure 19 This is a comparison table of model indicators on the French wind power dataset of this invention; Figure 20 This is a comparison table of model indicators on the Turkish wind power dataset of this invention; Figure 21 This is a table showing the performance improvement of the model on the French wind power dataset of this invention; Figure 22 This is a table showing the performance improvement of the model on the Turkish wind power dataset of this invention; Figure 23 These are comparison images of the model's effects in different seasons. Figure 24 This is a diagram showing the ablation experiment results of the present invention; Figure 25 This is a schematic diagram of the time cross-validation scheme of the present invention; Figure 26 This is a comparison chart of the time-cross-validation model results on the French wind power dataset of this invention; Figure 27 This is a comparison result of the time cross-validation model on the Turkish wind power dataset of this invention; Figure 28 This is a comparison result diagram of the models of this invention. Detailed Implementation

[0012] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0013] Example 1 A deep learning-based enhanced wind power prediction method, the method comprising: S1. Collect raw wind power data and use it to perform CEEMDAN mode decomposition to obtain multiple intrinsic mode components; In step S1, it is necessary to specify in detail that the raw wind power data should include at least historical wind power sequences, wind speed, wind direction, temperature, air pressure, and future meteorological elements obtained from numerical weather prediction (NWP) models. The data acquisition frequency is set according to the forecasting requirements, typically at the minute level (e.g., 5-15 minutes) or hourly level. After acquisition, data preprocessing is required, including missing value imputation (using linear interpolation or the mean of nearby times), outlier detection and correction (using statistical methods such as the 3σ principle or interquartile range), and data standardization (e.g., Z-score standardization or Min-Max normalization) to eliminate dimensional differences and improve the stability of subsequent processing.

[0014] After preprocessing, the processed wind power sequence is modally decomposed using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). CEEMDAN effectively suppresses mode aliasing and residual noise, common in traditional EMD, by repeatedly adding adaptive white noise of specific amplitudes and performing EMD, followed by ensemble averaging of the decomposition results. The specific decomposition process is as follows: First, the parameters of CEEMDAN are set, including the standard deviation ε of the added white noise (the initial value is usually set to 0.1 to 0.3 times the standard deviation of the original signal) and the number of ensemble averaging iterations N (usually set to 50-200). In each decomposition, adaptive white noise of different implementations (amplitude ε multiplied by the signal standard deviation) is added to the original signal, and then EMD decomposition is performed on the noisy signal to obtain a set of Intrinsic Mode Functions (IMFs) and a residual component. This process is repeated N times. Finally, the IMF components of the corresponding order obtained from each decomposition are ensemble-averaged to obtain the final IMF component set (IMF1, IMF2, ..., IMFk) and a residual component. Through CEEMDAN decomposition, the original complex non-stationary wind power sequence is adaptively decomposed into multiple relatively stationary IMF components with different time scales, facilitating subsequent feature extraction and modeling analysis. The number of IMF components k obtained after decomposition depends on the complexity of the original signal, typically between 6 and 10.

[0015] S2. The intrinsic mode components are evaluated using the IMF co-entropy balance index to obtain the evaluation results; the IMF co-entropy balance index includes minimum Pearson correlation, maximum Pearson correlation and minimum information entropy.

[0016] The expression for calculating the IMF co-entropy balance index is: in, This represents a multi-objective optimization fitness function, mainly used to dynamically balance modal redundancy, sparsity, and signal fidelity in the CEEMDAN decomposition process, thereby guiding the adaptive calibration of CEEMDAN parameters. The total number of intrinsic mode functions (IMFs) obtained from the CEEMDAN decomposition, i.e., the total number of IntrinsicModeFunctions generated after decomposition, is the core dimension reflecting the decomposition results. The Pearson correlation coefficient refers to the correlation coefficient between adjacent IMF components, which quantifies the correlation between the first and second IMF components. The and the first The linear correlation of each IMF; the smaller the value, the lower the redundancy of adjacent modes, which can effectively suppress the mode aliasing problem. Indicates the first Each Intrinsic Mode Function (IMF) is a single mode component obtained after CEEMDAN decomposition, carrying the specific frequency characteristics of the original signal; For the first The information entropy of an IMF measures the complexity and uncertainty of that IMF signal. The lower the value, the more concentrated and sparse the amplitude distribution of the IMF, the less noise it contains, and the more prominent its core features. This represents the original wind power signal (OriginalWindPowerSignal), which is the raw input data without decomposition, noise reduction, or other preprocessing. It is also the reference signal for subsequent decomposition and prediction work.

[0017] In step S2, it is necessary to explain in detail that: first, a Pearson correlation coefficient matrix of all IMF components is constructed, and the minimum and maximum values ​​of the correlation coefficients between each IMF component and other components are calculated; second, the information entropy of each IMF component is calculated based on Shannon entropy theory; finally, the ICEBI value is generated by weighted combination of the three indicators. The lower the ICEBI value, the better the stability and independence of the modal component. During the evaluation, IMF components with ICEBI below the threshold are preferentially selected for subsequent processes to effectively filter out noise-dominated or redundant modes.

[0018] S3. Based on the evaluation results, the parameters of CEEMDAN are adaptively calibrated using the Newton-Raphson optimizer to obtain multiple optimized intrinsic mode components. The optimized intrinsic mode components include: S301. Define the objective function based on the evaluation results and initialize the CEEMDAN parameters; In step S301, it is necessary to explain in detail that the objective function is defined based on the value of the IMF co-entropy balance index, specifically minimizing the ICEBI value to optimize the independence and stability of the modal components; at the same time, the key parameters of CEEMDAN are initialized, including setting the white noise standard deviation to 0.2 times the standard deviation of the original wind power sequence (adjustable to 0.1-0.3 times), setting the ensemble average number N to 100 (range 50-200), and setting the initial iteration point of the Newton-Raphson optimizer, including the step size factor and convergence tolerance, to ensure efficient convergence of the optimization process.

[0019] S302. Apply the Newton-Raphson optimizer and the position update formula for iterative optimization. Stop the optimization when the objective function converges to the preset threshold and output the optimized parameter values. The expression for the position update formula is: in, Indicates the first During the nth iteration, the 1st The updated position of each individual (or solution). This represents a random number within the interval (0,1), used to enhance the randomness of the algorithm's search. , , Let represent the three distinct reference solutions corresponding to the nth individual in the population at the IT-th iteration.

[0020] In step S302, it is necessary to explain in detail that the Newton-Raphson optimizer iteratively updates the CEEMDAN parameters to gradually approach the minimum value of the objective function. In each iteration, the gradient vector (i.e., the first derivative) and the Hessian matrix (i.e., the second derivative matrix) of the objective function at the current parameter value are first calculated, and then the parameter values ​​are adjusted using the position update formula. Specifically, the parameter update amount Δθ is determined by the product of the inverse of the Hessian matrix and the gradient vector, i.e., Δθ = -H⁻¹·∇f(θ), where H is the Hessian matrix and ∇f(θ) is the gradient vector. To improve computational efficiency, a quasi-Newton method (such as the BFGS algorithm) can be used to approximate the Hessian matrix, avoiding the direct solution of the inverse matrix. During the iteration, the step size factor is dynamically adjusted through a line search strategy to ensure that each update reduces the objective function value. The convergence tolerance is set to 10⁻⁶ to 10⁻⁴. When the change in the objective function value or the parameter update amount is less than this threshold, the optimization process terminates. The final optimized parameters include the white noise standard deviation ε and the ensemble average number N. For example, ε is optimized to 0.18 times the original signal standard deviation, and N is adjusted to 120 times. At this point, the ICEBI value is minimized, and the independence and stability of the modal components are significantly improved. The optimized IMF component set (IMF1', IMF2', ..., IMFk') will be used as input for subsequent feature extraction, and its number k may be reduced to 4 to 7 due to the filtering of redundant modes.

[0021] S303. Re-execute CEEMDAN decomposition using the optimized CEEMDAN parameters to generate multiple optimized intrinsic mode components.

[0022] The optimized CEEMDAN parameters (such as the white noise standard deviation ε and the ensemble averaging order N) obtained in step S302 are substituted into the CEEMDAN decomposition process to re-decompose the original wind power sequence into modes. The decomposition process is consistent with step S1, that is, adaptive white noise with an amplitude of ε multiplied by the signal standard deviation is added to the original signal, and EMD decomposition is performed to obtain a set of IMF components and a residual component. This process is repeated N times, and the ensemble averaging of the IMF components of the corresponding order is performed. Since the parameters have been adaptively calibrated by the Newton-Raphson optimizer, the IMF components (IMF1', IMF2', ..., IMFk') generated by this decomposition are significantly better than the initial decomposition results in terms of independence and stability, and the number k may be reduced due to the filtering of redundant modes (usually 4 to 7). These optimized IMF components will be used as inputs for subsequent feature extraction to provide a more accurate modal representation for the KAN-LSTM hybrid model.

[0023] S4. A wind power prediction model is constructed based on the Kolmogorov-Arnold network and a long short-term memory network. Multiple optimized intrinsic mode components are input into the wind power prediction model, and the model is used to predict wind power, outputting the prediction results. Specifically, the Kolmogorov-Arnold network is used to extract nonlinear features from the optimized intrinsic mode components, and the long short-term memory network is used to perform time-series modeling on the extracted nonlinear features. S401. Input the optimized multiple intrinsic mode components into the Kolmogorov-Arnold network respectively to perform nonlinear feature extraction and obtain the corresponding feature vectors. In step S401, it is necessary to explain in detail that the optimized IMF components (IMF1', IMF2', ..., IMFk') are input into the respective branch channels of the Kolmogorov-Arnold network (KAN). The KAN network performs deep feature extraction on each IMF component through multi-layer nonlinear transformations. Its core structure includes an input layer, hidden layers, and an output layer. The input layer receives the time-series data of a single IMF component (dimension T×1, where T is the time step); the hidden layers use multiple sets of Kolmogorov-Arnold polynomial basis functions to perform nonlinear mapping on the input data. Each set of basis functions generates a high-dimensional feature representation through weighted summation (e.g., expanding the 1-dimensional input to a D-dimensional feature space, where D is usually set to 8-16); the output layer concatenates and fuses the features from each hidden layer, finally outputting the feature vector (dimension T×D) corresponding to the IMF component. For example, after processing by the KAN network, IMF1' generates feature vector V1 (dimension T×D), IMF2' generates V2, and so on. The feature vectors (V1, V2, ..., Vk') of all IMF components will serve as input for subsequent time series modeling. The advantage of the KAN network lies in its powerful nonlinear fitting capability, which can automatically capture the complex fluctuation patterns hidden in the IMF components (such as periodicity, trends, and abrupt changes), providing more discriminative feature representations for subsequent time series modeling. It is worth noting that the parameters of the KAN network (such as the number of basis functions and the dimension of hidden layers) need to be determined through cross-validation to ensure the generalization of feature extraction.

[0024] S402. Input the feature vector into the long short-term memory network, perform time series modeling, and generate a predicted sequence of wind power. In step S402, it is necessary to explain in detail that the feature vectors (V1, V2, ..., Vk') of each IMF component extracted in step S401 are aligned according to the time step and concatenated into a three-dimensional input tensor (dimension T×k'×D), where T is the time step, k' is the number of optimized IMF components, and D is the feature dimension. This tensor is input to the input layer of a Long Short-Term Memory (LSTM) network. The LSTM recursively processes the temporal data through a gating mechanism (input gate, forget gate, output gate). Specifically, at each time step t, the LSTM unit receives the current input feature vector Vt (dimension 1×D) and the hidden state ht-1 and cell state Ct-1 from the previous time step. It updates the current cell state Ct (storing long-term dependency information) and the hidden state ht (passing short-term context) through nonlinear transformations. For example, the input gate uses the sigmoid function to determine how much new information to retain, the forget gate uses the sigmoid function to filter out old information to be discarded, and the output gate uses the sigmoid function to control the output ratio, ultimately generating the predicted value pt for the current time step. The predicted values ​​(p1, p2, ..., pT) at all time steps are concatenated sequentially to form a predicted wind power sequence (dimension T×1). To capture multi-scale temporal dependencies, the LSTM layer is typically set to 2-3 layers, with 64-128 neurons per layer. Overfitting is prevented using Dropout (rate set to 0.2-0.5) and gradient clipping (threshold set to 1.0). The final output predicted sequence reflects the fluctuation trend of wind power in future periods, and its accuracy is verified by subsequent evaluation metrics (such as root mean square error RMSE and mean absolute error MAE).

[0025] S403. Aggregate the predicted sequences and output the final wind power prediction result.

[0026] In step S403, it is necessary to explain in detail that the wind power prediction sequences (p1, p2, ..., pT) generated in step S402 are summed along the time dimension to obtain the total wind power prediction value Ptotal (dimension 1×1) for the next T times. Specifically, since steps S1 to S3 have decomposed the original wind power sequence into multiple independent modal components, and extracted features through a KAN network and modeled the temporal dependencies through an LSTM network, the prediction sequences (V1→p1, V2→p2, ..., Vk'→pk') corresponding to each IMF component represent the power changes at different time scales (such as second-level fluctuations, minute-level trends, and hour-level cycles). Through the summation operation, these multi-scale prediction results are integrated into a global prediction value, i.e., Ptotal=Σ(pt) (t=1 to T), where pt is the prediction value at the t-th time step. For example, if the decomposition yields four IMF components (k'=4), and the LSTM output predicts a sequence length of T=24 for each component (corresponding to the next 24 hours), then the final prediction result Ptotal is the cumulative value of wind power over 24 hours. To improve the intuitiveness of the prediction results, Ptotal can be split into a discrete prediction point sequence (P1, P2, ..., PT) according to the original time resolution (e.g., one point every 15 minutes), with each point corresponding to the power prediction value at a specific time. Furthermore, to quantify prediction uncertainty, the upper and lower bounds of the prediction interval (e.g., Ptotal±σ, where σ is the standard deviation) can be output simultaneously. This interval is generated through the variance of the LSTM hidden states or Monte Carlo sampling. The final output prediction result (discrete point sequence or cumulative value) will serve as the basis for wind farm dispatch decisions, and its accuracy and stability will be verified through subsequent evaluation indicators (e.g., RMSE, MAE, R²) and visual comparisons (e.g., the overlap between the predicted curve and the actual curve).

[0027] As an optional embodiment of the present invention, the wind power prediction model may be evaluated using mean absolute error, root mean square error, and mean absolute percentage error.

[0028] It's important to note that Mean Absolute Error (MAE) measures prediction accuracy by averaging the absolute errors between the predicted and actual values; a smaller MAE value indicates a closer prediction to the true value. Root Mean Square Error (RMSE), on the other hand, is calculated by averaging the squares of the errors between the predicted and actual values ​​and then taking the square root. It amplifies the impact of larger errors and better reflects the stability of the prediction. Mean Absolute Percentage Error (MAPE) expresses the prediction error as a percentage, directly reflecting the degree to which the predicted value deviates from the true value. By combining these three evaluation metrics, the performance of wind power prediction models can be comprehensively evaluated, including prediction accuracy, stability, and the degree of bias.

[0029] Experimental process In this invention, this embodiment uses datasets from two wind farms located in different countries for performance testing. The data was collected every 10 minutes over a year. The first dataset comes from a wind farm in France with an installed capacity of 2050 kW (hereinafter referred to as Wind Farm 1), while the second dataset comes from a wind farm in Turkey with an installed capacity of 3600 kW (hereinafter referred to as Wind Farm 2). This study utilizes data recorded every 10 minutes from these two datasets over a year. Figure 6 and Figure 7 It displays wind power generation data records from datasets from France and Türkiye.

[0030] First, 80% of the data was used for model training, while the remaining 20% ​​served as a test set to evaluate the model's predictive performance. This method establishes and validates the model's effectiveness. The data curves illustrate that wind energy data exhibits non-stationary and non-linear characteristics, posing challenges to accurate wind energy forecasting. Therefore, employing appropriate models and techniques to improve prediction accuracy is crucial. The method proposed in this study demonstrates superiority in handling these complex characteristics.

[0031] Experimental setup This embodiment uses the PyTorch framework, which is developed in Python 3.10. The experiment was conducted on a Linux-based server equipped with multiple NVIDIA RTX 4090 GPUs to support computationally intensive tasks.

[0032] For the implementation of CEEMDAN, this embodiment uses the pyEMD package, while the implementation of EWT uses the ewtpy package. When implementing LSTM, this embodiment chooses Keras as the preferred tool. When configuring LSTM, this embodiment selects the Adam optimizer, which is renowned for its superior performance compared to other stochastic optimization methods. Following established guidelines, this embodiment configures the learning rate to 0.001 and performs training for 100 epochs. Specific parameter configurations are shown in Table 1.

[0033] Table 1 Experimental Configuration Table Evaluation indicators To evaluate the predictive performance of the model, this study employed multiple indices for quantitative analysis of point predictions, including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). The MAPE index assesses the goodness of fit of the model; a smaller value indicates a better fit. The RMSE index reflects the degree of deviation between predicted and actual values; a smaller value indicates higher prediction accuracy. The calculation formulas are as follows: , , ; in, Mean Absolute Error (MAE) is the average of the absolute values ​​of prediction errors, used to measure the average deviation between predicted and actual values. This represents the sample size, i.e., the total number of observation data points involved in the error calculation. Indicates the first The true value (actual observed value) of each sample. Indicates the first The predicted value of each sample (the estimated value output by the model). The root mean square error (RMSE) is the square root of the average of the sum of squared prediction errors. It is more sensitive to larger errors and reflects the degree of dispersion of the errors. Mean Absolute Percentage Error (MEPE) is the average of the absolute percentages of the prediction error. It reflects the relative error of the predicted value to the actual value in percentage form, and more intuitively reflects the prediction accuracy. In this study, the percentage improvement was also used to evaluate the enhancement effect of the proposed prediction method compared to other prediction methods. The formula for calculating the percentage improvement is as follows: , ; ;in, This indicates the percentage improvement in the RMSE metric; similarly, it quantifies the degree of improvement in RMSE. This represents the root mean square error before optimization. This represents the optimized root mean square error. This indicates the percentage increase in the MAPE indicator, quantifying the degree of improvement in MAPE. This represents the mean absolute percentage error before optimization. This represents the optimized mean absolute percentage error. This represents the percentage improvement in the MAE metric, used to quantify the degree of improvement in MAE after model optimization. This represents the mean absolute error before optimization (baseline model). It represents the mean absolute error after optimization (improved model).

[0034] Comparison of different fitness functions This paper systematically compares various fitness functions, providing scientific guidance for selecting the optimal fitness function in specific application scenarios. This embodiment explores a series of fitness functions, including entropy-based functions (such as envelope entropy, information entropy, permutation entropy, sample entropy, etc.), energy-based functions (such as energy entropy, minimum energy difference, residual error exponent, etc.), and other statistical features (such as kurtosis, envelope kurtosis factor, envelope spectral peak factor, etc.). Each fitness function has its unique mathematical expression and application scenarios, exhibiting significant differences under different conditions.

[0035] To further enhance the understanding and application of these fitness functions, this embodiment summarizes their mathematical formulas and basic principles in detail. This embodiment covers the specific mathematical expressions of various fitness functions and briefly introduces the working mechanism of each fitness function. Specific details are as follows... Figure 8-10 As shown.

[0036] This embodiment clearly demonstrates the superiority of the ICEBI method through performance analysis of three key metrics, MAPE, RMSE, and MAE, on the French dataset. Figure 11 The performance of different fitness functions on the MAPE metric is described. The results show that ICEBI significantly outperforms other methods in reducing prediction error. Figure 12 This conclusion was further validated by examining the RMSE metric. Here, ICEBI also performed exceptionally well, demonstrating high accuracy and stability when handling data fluctuations. Finally, Figure 13 The MAE metric was used to demonstrate the performance of various fitness functions, further confirming the advantages of ICEBI, especially in minimizing absolute error.

[0037] like Figure 14 As shown in the combined results of these three figures, it is clear that the ICEBI method demonstrates significant superiority across multiple performance metrics, making it highly suitable for improving prediction accuracy and model stability.

[0038] Comparison with other optimization methods In numerous existing studies, benchmark functions are commonly used for numerical validation of various algorithms. These functions include unimodal and multimodal types, designed to evaluate the algorithm's performance in global search and overcoming local optima. In this study, several unimodal and multimodal benchmark functions were selected to comprehensively evaluate the performance of the NRBO algorithm in solving different modal problems. Figure 15 Specific details of the selected benchmark function are provided.

[0039] In addition, this study selected five other representative algorithms, such as the Grey Wolf Optimizer (GWO) algorithm, the Sparrow Optimizer (SSA) algorithm, the Balanced Optimizer (EO) search algorithm, the Whale Optimizer (WOA) algorithm, and the RKO Optimizer algorithm, as comparison algorithms. The parameter settings for each algorithm are as follows: Figure 16 As shown.

[0040] To ensure the objectivity of the NRBO algorithm solution performance, each test function was solved 30 times, and statistical analysis was performed on the maximum and minimum values, mean and standard deviation of the results. The statistical results are as follows: Figure 17 As shown.

[0041] Experimental results demonstrate that the NRBO algorithm exhibits significant advantages in multiple benchmark function tests. Particularly in the evaluation of F1, F2, F3, F5, and F11 functions, NRBO excels in key metrics such as maximum (Max), minimum (Min), mean, and standard deviation (STD). For example, in the F1 function test, NRBO achieved ideal results (0.00E+00) on all four metrics, showcasing its high stability and accuracy. Similarly, in the F2 function test, although NRBO is not absolutely optimal on all metrics, its performance is still commendable, especially in terms of mean and standard deviation, which are close to ideal values, highlighting its efficiency and stability. For the more complex F3 function, NRBO again demonstrates its superior performance, achieving the best results in maximum, minimum, mean, and standard deviation, indicating its excellent performance even when handling complex problems.

[0042] like Figure 18 As shown, in the F5 function test, NRBO achieved an ideal score of 0.00E+00 on all evaluation metrics, further validating the accuracy and reliability of its solution. Finally, in the F11 function test, NRBO maintained its consistently high performance, with all metrics reaching optimal values, demonstrating its ability to solve various types of optimization problems. Overall, NRBO not only demonstrated superior performance in different benchmark function tests, but also exhibited higher stability and accuracy compared to other algorithms such as SSA, GWO, EO, WOA, and RKO, making it an efficient and reliable optimization algorithm.

[0043] Overall performance comparison This paper comprehensively evaluates the effectiveness of the proposed method through comparative analysis with various prediction models. These models include Support Vector Regression (SVR), Artificial Neural Networks (ANN), Random Forest (RF), Long Short-Term Memory (LSTM) networks, and several LSTM networks incorporating decomposition techniques (such as EMD-LSTM, EEMD-LSTM, CEEMDAN-LSTM, and CEEMDAN-EWT-LSTM). To ensure the reliability and adaptability of the model in practical applications, this paper employs two different data segmentation methods: monthly segmentation and seasonal segmentation.

[0044] Case 1: Performance Comparison of Models in Different Months The fundamental advantage of LSTM: Compared to other single prediction methods (SVR, ANN, and RF), Long Short-Term Memory networks exhibit stronger learning capabilities and better capture complex patterns and long-term dependencies in time-series data. For example... Figure 19 and 20 As shown, LSTM has the lowest mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE) values ​​in both datasets, which confirms its superiority as a base prediction model.

[0045] Effectiveness of the hybrid decomposition method: By decomposing the original wind field data into multiple relatively stable subsequences, the hybrid decomposition method effectively improves the performance of wind power generation prediction, thereby enhancing the quality of the input data for the prediction model. For example, in the French dataset, the average MAPE value of the hybrid decomposition method is significantly lower than that of the single prediction method, demonstrating the effectiveness of the decomposition technique in improving prediction accuracy.

[0046] like Figure 21 and 22 As shown, CEEMDAN decomposition has advantages: as an improved version of EMD and EEMD, CEEMDAN exhibits superior decomposition capabilities. For example, in the French dataset, the average MAPE value of CEEMDANLSTM is lower than that of EMDLSTM and EEMDLSTM, indicating that CEEMDAN can more accurately identify and handle modality mixing problems in the data.

[0047] Contributions of EWT Denoising and KAN: The proposed method significantly improves the prediction accuracy of CEEMDAN-LSTM by combining EWT denoising techniques and KAN optimization. Experimental results show that the proposed method improves the prediction accuracy of CEEMDAN-LSTM by more than 17%, demonstrating the effectiveness of EWT denoising in reducing the negative impact of randomness and irregularity in high-frequency subsequences, and the significant contribution of KAN in improving the model's nonlinear approximation ability and generalization performance.

[0048] Overall superiority of the proposed method: The proposed method achieves the best prediction results among all prediction methods. Its average MAPE value is less than 1.5% in both datasets, demonstrating its excellent prediction ability. By systematically combining CEEMDAN decomposition, EWT denoising, KAN, and LSTM prediction techniques, this method effectively solves the complex wind power generation prediction problem and significantly outperforms other benchmark methods.

[0049] In summary, the prediction method proposed in this study demonstrates significant superiority in the field of ultra-short-term wind power prediction. This method significantly improves prediction accuracy by fully utilizing the advantages of the hybrid decomposition model and combining Empirical Mode Decomposition (EWT) denoising technology with the KAN optimization algorithm. Experimental results fully verify the efficiency and advancement of this method in solving complex wind power prediction problems, providing a new technology with broad application prospects and promotion potential for the field of ultra-short-term wind power prediction.

[0050] Case 2: For example Figure 23 As shown, the model performance is compared across different seasons. Spring: In the spring, the NRBOCEEMDANEWTKANLSTM model performed exceptionally well in wind power forecasting, which is crucial for the effective management and optimization of the power system. The model excelled across three key performance metrics: MAPE of 0.87%, RMSE of 31.09, and MAE of 17.83. Compared to the suboptimal NRBOCEEMDANEWTLSTM model, it improved MAPE by 2.25%, RMSE by 3.06%, and MAE by 2.62%. Compared to the baseline model EMDLSTM, the performance improvement was even more significant, with MAPE increasing by 46.63%, RMSE by 43.33%, and MAE by 46.63%.

[0051] Summer: With the arrival of summer, the NRBOCEEMDANEWTKANLSTM model continues to maintain its leading position in wind power forecasting. Its key performance indicators are: MAPE 1.361%, RMSE 50.44, and MAE 28.46. Compared to the suboptimal model NRBOCEEMDANEWTLSTM, this model improves MAPE by 5.35%, RMSE by 2.63%, and MAE by 3.40%. Compared to the baseline model EMDLSTM, the performance improvement is even more significant, with MAPE increasing by 35.19%, RMSE by 32.69%, and MAE by 33.70%.

[0052] Autumn: Entering autumn, the NRBOCEEMDANEWTKANLSTM model continues to perform well in wind power forecasting. Its MAPE is only 0.86%, RMSE is 28.08, and MAE is 17.65. Compared to the suboptimal model NRBOCEEMDANEWTLSTM, its MAPE improves by 1.15%, RMSE by 3.49%, and MAE by 1.12%. Compared to the baseline model EMDLSTM, the performance improvement is even more significant, with MAPE increasing by 48.81%, RMSE by 49.55%, and MAE by 48.84%.

[0053] Winter: During winter, the NRBOCEEMDANEWTKANLSTM model also demonstrated superior wind power forecasting capabilities, which is crucial for ensuring the efficient operation of the power system. Its key performance indicators are: MAPE of 1.32%, RMSE of 40.28, and MAE of 27.31. Compared to the suboptimal model NRBOCEEMDANEWTLSTM, this model improves MAPE by 1.15%, RMSE by 3.49%, and MAE by 1.12%. Compared to the baseline model EMDLSTM, the performance improvement is even more significant, with MAPE increasing by 48.81%, RMSE by 49.55%, and MAE by 48.84%.

[0054] In summary, the NRBOCEEMDANEWTKANLSTM model demonstrates superior performance in seasonal wind energy forecasting, becoming an important tool for managing and optimizing power systems under diverse meteorological conditions. This model significantly outperforms its competitors in key performance indicators, including mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE). Compared to the suboptimal model NRBOCEEMDANEWTLSTM and the baseline model EMDLSTM, it shows a significant performance improvement. Specifically, in spring, the model achieves a MAPE of 0.87%; in summer, its performance remains robust with a MAPE of 1.36%; in autumn, it demonstrates robustness with a MAPE of 0.86%; and even in winter, facing extremely challenging weather conditions, the model provides a reliable MAPE of 1.32%. These results highlight the model's adaptability and accuracy in dealing with diverse and unpredictable weather patterns, thereby improving the stability and efficiency of power systems. This study, by focusing on the French dataset, emphasizes the importance of rigorous testing under complex and variable climatic conditions to ensure the reliability and effectiveness of advanced forecasting models.

[0055] ablation experiment The ablation experiments showed that introducing each model component significantly improved predictive performance, with different components exhibiting complementary effects across various indices and months. The baseline LSTM model showed average MAPE, RMSE, and MAE values ​​of 3.676%, 117.573, and 75.376, respectively. By progressively integrating the CEEMDAN, EWT, NRBO, and KAN components, the final NRBO-CEEMDAN-EWT-KAN-LSTM model achieved optimized average indices of 1.102%, 37.475, and 22.812, representing overall error reductions of 68.61%, 66.72%, and 68.74%, respectively. These findings confirm that each technical module enhances model performance through different mechanisms.

[0056] like Figure 24 As shown, specifically, the CEEMDAN decomposition technique brings the most significant improvement compared to the baseline LSTM, reducing the average MAPE by 34.88%, RMSE by 39.19%, and MAE by 35.15%. This enhancement stems from its multi-scale feature decoupling capability, which accurately captures frequency domain components. The subsequent addition of EWT further refines feature extraction, achieving additional reductions in MAPE and RMSE on top of CEEMDAN-LSTM, by 21.01% and 16.99%, respectively, showing particularly strong local feature extraction capabilities in months such as January and May. The NRBO optimization algorithm contributes 4.49% improvement in MAPE and 4.22% improvement in RMSE through parameter tuning, demonstrating significant effectiveness during seasonal fluctuations (e.g., February and July), highlighting its adaptability to complex temporal patterns. The final KAN network ensemble provides complementary nonlinear modeling capabilities, further reducing the annual average MAPE by 0.052 percentage points, with marginal gains observed during stable data periods such as September and December.

[0057] It is worth noting that the effectiveness of the components varied across different months. For example, EWT reduced MAPE by 5.4% in January but only by 1.02% in June, which may be related to the noise characteristics of the data. NRBO achieved a 32.29% reduction in RMSE in February but only improved by 0.62% in August, reflecting the effectiveness of parameter optimization dependent on data distribution. KAN produced continuous improvements in most months, but its effectiveness diminished in June and August, suggesting a potential threshold in balancing model complexity and data information content. Overall, these technical modules form an incremental optimization path: eigenvalue decomposition techniques drive the main performance gains, while optimization algorithms and architectural enhancements further tap into the model's potential. This combined strategy effectively balances global pattern capture and local fluctuation adaptation in time series forecasting.

[0058] Time series cross-validation To obtain robust measurements of model performance, this embodiment implements a time-series cross-validation scheme on datasets from France and Turkey. In this study, a 3-fold time-series cross-validation scheme was applied to the dataset covering the first four months (January to April). The 3-fold time-series cross-validation scheme is as follows: Figure 25 As shown.

[0059] Figure 26 and 27 The results of time series cross-validation on French and Turkish datasets are presented. The proposed method achieves the lowest error on both datasets. Figure 16 The proposed method achieves an average improvement of 67.01% on the French dataset and 78.61% on the Turkish dataset, as reported in the study. Overall, based on time-series cross-validation analysis, the proposed method demonstrates superior predictive accuracy.

[0060] Model comparison experiment In this section, the proposed model is compared and analyzed with other cutting-edge techniques in other fields, including Empirical Mode Decomposition-Elman Neural Network (EMD-ENN) and Integrated Empirical Mode Decomposition Bidirectional Long Short-Term Memory Network (EEMD-BO-LSTM). Comparative experiments on the Turkish dataset reveal the significant performance advantages of the proposed model and method. Figure 28 The results of the comparative experiments were recorded in detail. Figure 28 The comparative experimental results clearly demonstrate that the proposed model exhibits superior performance compared to EMD-ENN and EEMD-BO-LSTM, achieving the lowest mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE). Specifically, the proposed model achieves an average improvement of 38.30% in MAPE compared to EMD-ENN and an average improvement of 28.05% compared to EEMD-BO-LSTM. These experimental results fully demonstrate the effectiveness and superiority of the proposed method, especially when applied to the Turkish dataset, compared to the state-of-the-art methods currently available in that country.

[0061] Given the disordered and irregular nature of IMF1 sequences, the accuracy and stability of prediction models may be negatively impacted. To mitigate the difficulties in the prediction process, this study employs Empirical Wavelet Transform (EWT) technology for denoising IMF1 sequences. The application of EWT significantly reduces the randomness and volatility of IMF1 sequences, thereby simplifying the sequence modeling process and further enhancing the learning ability for prediction accuracy. After data preprocessing, this study uses a Long Short-Term Memory (LSTM) network to predict all sequences and summarizes the prediction results to obtain the final prediction output.

[0062] Example 2 An enhanced wind power prediction device based on deep learning, the device being used to execute the aforementioned enhanced wind power prediction method based on deep learning; the device comprising: The data acquisition module is used to collect raw wind power data; A mode decomposition module, connected to the acquisition module, is used to perform CEEMDAN mode decomposition to obtain multiple intrinsic mode components. An evaluation module, connected to the mode decomposition module, is used to evaluate the intrinsic mode components using the IMF cooperative entropy balance index to obtain evaluation results. An optimization module, connected to the evaluation module, is used to adaptively calibrate the parameters of CEEMDAN using a Newton-Raphson optimizer based on the evaluation results, and obtain multiple optimized intrinsic mode components. The feature extraction module is used to perform nonlinear feature extraction on the optimized intrinsic mode components based on the Kolmogorov-Arnold network to obtain the corresponding feature vectors. The time series modeling module, connected to the feature extraction module, is used to perform time series modeling on the extracted nonlinear features based on the long short-term memory network to generate a predicted sequence of wind power. The aggregation output module, connected to the time series modeling module, is used to aggregate the prediction sequence and output the final wind power prediction result.

[0063] In this embodiment, the acquisition module employs a sensor array, with these sensors positioned at key locations within the wind farm to collect raw wind power data in real time and accurately, including multi-dimensional information such as wind speed, wind direction, temperature, and air pressure. The mode decomposition module utilizes the advanced CEEMDAN algorithm to perform detailed mode decomposition on the acquired complex wind power data, breaking down the raw signal into multiple intrinsic mode components with different frequency characteristics to facilitate a deeper analysis of the inherent patterns in the wind power data. The evaluation module then comprehensively evaluates the decomposed intrinsic mode components, using the innovative IMF co-entropy balance index to quantify the interrelationships and energy distribution among the components. Based on the evaluation results, the optimization module uses a Newton-Raphson optimizer to adaptively calibrate the CEEMDAN parameters. Through continuous iterative optimization, it ensures that the decomposed intrinsic mode components are more accurate and stable, thereby improving the overall accuracy of the prediction model. The feature extraction module utilizes the powerful nonlinear mapping tool, the Kolmogorov-Arnold network, to perform deep feature extraction on the optimized intrinsic mode components, capturing hidden patterns and complex relationships in the data to generate highly representative feature vectors. The temporal modeling module uses these feature vectors as input and employs a Long Short-Term Memory (LSTM) network for temporal modeling, fully leveraging its advantage in processing long-sequence data to capture the trends and periodic characteristics of wind power changes over time, generating accurate wind power prediction sequences. Finally, the aggregation output module organically aggregates the prediction sequences generated by each temporal modeling module, using weighted averaging or other fusion strategies to output the final wind power prediction result.

[0064] Example 3 A computer device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement a deep learning-based enhanced wind power prediction method in Embodiment 1 when executing the executable instructions.

[0065] It should be noted that the computer device includes a processor, a memory, and may also include one or more of a multimedia component, an input / output (I / O) interface, and a communication component.

[0066] The processor controls the overall operation of the computer device to complete all or part of the steps in the above-mentioned automated testing method for factory equipment based on big data.

[0067] Memory is used to store various types of data to support the operation of the computer device. This data may include, for example, instructions for any application or method used to operate on the computer device, as well as application-related data. Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0068] The multimedia component may include a screen and an audio component, wherein the screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals; for example, the audio component may include a microphone for receiving external audio signals, the received audio signals may be further stored in memory or transmitted via a communication component; the audio component may also include at least one speaker for outputting audio signals.

[0069] I / O interfaces provide interfaces between the processor and other interface modules, such as keyboards, mice, buttons, etc.; these buttons can be virtual buttons or physical buttons.

[0070] The communication component is used for wired or wireless communication between the computer device and other devices; wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G or 5G, or one or more combinations thereof, and the corresponding communication component may include: Wi-Fi module, Bluetooth module, NFC module, mobile communication module.

[0071] As a preferred embodiment, the computer device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the above-described automated testing method for factory equipment based on big data.

[0072] Example 4 A computer-readable storage medium comprising: A memory on which computer programs are stored; A processor is configured to execute the program in the memory to implement a deep learning-based enhanced wind power prediction method as described in Embodiment 1.

[0073] It should be noted that the electronic device disclosed in this embodiment includes a processor and a memory for storing processor-executable instructions. The processor is configured to implement any of the aforementioned efficient and optimized cloud resource task scheduling methods when executing the executable instructions.

[0074] It should be noted here that the number of processors can be one or more. Furthermore, the electronic device in this embodiment may also include input devices and output devices. The processor, memory, input devices, and output devices can be connected via a bus or other means, without specific limitations herein.

[0075] As a computer-readable storage medium, memory can be used to store software programs, computer-executable programs, and various modules, such as the program or module corresponding to the cloud resource task scheduling method with high efficiency and optimization according to embodiments of this disclosure. The processor executes various functional applications and data processing of the electronic device by running the software programs or modules stored in the memory.

[0076] Input devices can be used to receive input digital numbers or signals. These signals can be key signals related to user settings and function control of the device / terminal / server. Output devices can include display devices such as screens.

[0077] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A deep learning-based enhanced wind power prediction method, characterized in that, The method includes: Raw wind power data was collected and CEEMDAN mode decomposition was performed to obtain multiple intrinsic mode components. The intrinsic mode components were evaluated using the IMF co-entropy balance index to obtain the evaluation results; Based on the evaluation results, the parameters of CEEMDAN were adaptively calibrated using a Newton-Raphson optimizer to obtain optimized intrinsic mode components. A wind power prediction model is constructed based on the Kolmogorov-Arnold network and the long short-term memory network. The optimized intrinsic mode components are input into the wind power prediction model, and the wind power prediction model is used to predict wind power and output the prediction results. The Kolmogorov-Arnold network is used to extract nonlinear features from the optimized intrinsic mode components, and the long short-term memory network is used to perform time-series modeling on the extracted nonlinear features.

2. The deep learning-based enhanced wind power prediction method as described in claim 1, characterized in that, The IMF co-entropy balance index includes minimum Pearson correlation, maximum Pearson correlation, and minimum information entropy.

3. The deep learning-based enhanced wind power prediction method as described in claim 1 or 2, characterized in that, The expression for calculating the IMF co-entropy balance index is: in, Represents the fitness function for multi-objective optimization. This represents the total number of intrinsic mode functions obtained from the CEEMDAN decomposition. This represents the Pearson correlation coefficient between adjacent IMF components. Indicates the first One eigenmode function Indicates the first Information entropy of an IMF This represents the original wind power signal.

4. The deep learning-based enhanced wind power prediction method as described in claim 1, characterized in that, Based on the evaluation results, the parameters of CEEMDAN were adaptively calibrated using a Newton-Raphson optimizer to obtain the optimized intrinsic mode components, including: Based on the evaluation results, define the objective function and initialize the CEEMDAN parameters; The Newton-Raphson optimizer and the position update formula are used for iterative optimization. The optimization stops when the objective function converges to a preset threshold, and the optimized parameter values ​​are output. Re-execute the CEEMDAN decomposition using the optimized CEEMDAN parameters to generate multiple optimized intrinsic mode components.

5. The deep learning-based enhanced wind power prediction method as described in claim 4, characterized in that, The expression for the position update formula is: in, Indicates the first During the nth iteration, the 1st The updated position of each individual This represents a random number within the interval (0,1). , , Let represent the three distinct reference solutions corresponding to the nth individual in the population at the IT-th iteration.

6. The deep learning-based enhanced wind power prediction method as described in claim 1, characterized in that, A wind power prediction model is constructed based on the Kolmogorov-Arnold network and a long short-term memory network. Multiple optimized intrinsic mode components are input into the wind power prediction model, and wind power is predicted using this model. The output prediction results include: The optimized intrinsic mode components are input into the Kolmogorov-Arnold network for nonlinear feature extraction to obtain the corresponding feature vectors. The feature vector is input into the long short-term memory network to perform time series modeling and generate a predicted sequence of wind power. The predicted sequences are aggregated to output the final wind power prediction result.

7. The deep learning-based enhanced wind power prediction method as described in claim 1, characterized in that, The wind power prediction model is evaluated using mean absolute error, root mean square error, and mean absolute percentage error.

8. A deep learning-based enhanced wind power prediction device, characterized in that, The apparatus is used to execute the deep learning-based enhanced wind power prediction method as described in any one of claims 1 to 7; the apparatus includes: The data acquisition module is used to collect raw wind power data; A mode decomposition module, connected to the acquisition module, is used to perform CEEMDAN mode decomposition to obtain multiple intrinsic mode components. An evaluation module, connected to the mode decomposition module, is used to evaluate the intrinsic mode components using the IMF cooperative entropy balance index to obtain evaluation results. An optimization module, connected to the evaluation module, is used to adaptively calibrate the parameters of CEEMDAN using a Newton-Raphson optimizer based on the evaluation results, and obtain multiple optimized intrinsic mode components. The feature extraction module is used to perform nonlinear feature extraction on the optimized intrinsic mode components based on the Kolmogorov-Arnold network to obtain the corresponding feature vectors. The time series modeling module, connected to the feature extraction module, is used to perform time series modeling on the extracted nonlinear features based on the long short-term memory network to generate a predicted sequence of wind power. The aggregation output module, connected to the time series modeling module, is used to aggregate the prediction sequence and output the final wind power prediction result.

9. A computer device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement, when executing the executable instructions, an enhanced wind power prediction method based on deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, include: A memory on which computer programs are stored; A processor for executing the program in the memory to implement the deep learning-based enhanced wind power prediction method according to any one of claims 1 to 7.

Images