Wind power prediction method based on DIRMO and differentiated objective function

Abstract

The invention discloses a wind power prediction method and system based on DIRMO and a differentiated objective function, and relates to the field of wind power prediction. In order to solve the defects that time sequence modeling and feature processing are difficult to consider at the same time, the prediction precision is reduced, the model robustness is insufficient and the search efficiency is low due to the fact that a single model is adopted for prediction in existing multi-step power prediction, denoising, normalization and feature extraction are carried out on preprocessed power and wind speed data, and multi-scale features are obtained; performing time sequence modeling on the normalized power and wind speed data to generate a future preliminary prediction result; dividing a multi-step prediction task into a plurality of groups, and converting a multi-output problem into a single-output problem for training and prediction; and carrying out automatic optimization on hyper-parameters of the LightGBM models, training each group of LightGBM models by using the optimized hyper-parameters, and carrying out final power prediction based on a GRU preliminary prediction result and multi-scale features. The method is mainly used for predicting the wind power.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power prediction, in particular to a wind power prediction method and system based on a cascaded progressive DIRMO strategy and a differentiated objective function. BACKGROUND

[0002] With the development of new energy power generation capacity, new energy power prediction has become a key technology for the safe and stable operation of the power system. Accurate power prediction can help the power grid dispatching department to reasonably arrange the power generation plan, reduce the reserve capacity demand, reduce the phenomenon of curtailment of wind and light, and support new energy stations to participate in the power market transaction.

[0003] For future multi-step power prediction, existing strategies mainly include recursive strategy, direct strategy, multiple-input multiple-output (MIMO) strategy, and direct multi-output (DIRMO) strategy. The recursive strategy uses a single model to iteratively predict, predicting one step at a time and using the prediction result as the input for the next step. The direct strategy builds an independent model for each prediction step. The MIMO strategy uses a single model to predict all steps simultaneously. The DIRMO strategy decomposes multi-step prediction into several groups, each group training a model to directly predict multiple steps.

[0004] In the use of machine learning models, their performance depends on the hyperparameter configuration. Commonly used hyperparameter optimization methods include grid search, random search, Bayesian optimization, and heuristic algorithms. Based on different parameter optimization standards and computing power cost requirements, different methods have their own advantages and disadvantages in terms of optimization speed and accuracy. These problems to some extent limit the improvement of the accuracy of multi-step prediction tasks, and further optimization is urgently needed through technical upgrading.

[0005] In terms of prediction strategy for multi-step prediction, the recursive prediction strategy has a serious error accumulation phenomenon, which becomes more pronounced as the prediction length increases. The direct prediction strategy establishes an independent model for each prediction step, resulting in a huge computational time overhead, and the models cannot share information, so they cannot capture the volatility characteristics of time series such as wind power. The optimization goal of the MIMO prediction strategy is the average error of all steps, but the actual assessment has different step error weights, and when the prediction length increases, too many output ports will make it difficult to train model parameters. The DIRMO strategy combines the direct prediction strategy and the MIMO strategy, but its accuracy is heavily dependent on the number of groups and the step length of each group.

[0006] Firstly, existing methods usually use a single model for prediction, which is difficult to balance time series modeling and feature processing at the same time. Deep learning models are good at capturing time series dependencies, but their ability to handle complex feature interactions is limited. Traditional machine learning models are good at handling high-dimensional features and non-linear relationships, but they lack the ability to model long-term time series dependencies. A single model cannot fully leverage the strengths of different modeling methods, limiting further improvement in prediction accuracy.

[0007] Secondly, new energy power data is affected by many factors such as changes in weather conditions and fluctuations in equipment operating state, and there is a lot of high-frequency noise and abnormal fluctuations in the original data. These noise components are mixed with the real power change trend, directly affecting the model's learning of the power change pattern. Existing methods usually directly use the original data for modeling, or only use simple moving average, median filtering and other smoothing processing. These methods can reduce the impact of noise to some extent, but they also smooth out useful signal details, resulting in information loss. The existence of noisy data can interfere with the model training process, causing the model to learn incorrect patterns, reducing the generalization ability, and ultimately leading to a decrease in prediction accuracy and insufficient model robustness.

[0008] Thirdly, machine learning models such as Light Gradient Boosting Machine (LightGBM) usually contain a large number of hyperparameters, and the configuration of these hyperparameters has a decisive impact on model performance. Existing methods mainly rely on manual parameter tuning or simple grid search for hyperparameter optimization. Manual parameter tuning requires researchers to repeatedly try different parameter combinations based on experience, which not only consumes time and effort, but also highly depends on the expertise and experience accumulation of the parameter tuning personnel, making it difficult to guarantee finding the global optimal configuration. Although grid search can systematically traverse the parameter space, its computational overhead grows exponentially with the parameter dimension. Random search, while alleviating the computational overhead problem to some extent, still has low search efficiency and often requires a large number of samples to find a good parameter configuration. Intelligent optimization methods such as Bayesian optimization perform well in low-dimensional spaces, but their performance drops significantly in high-dimensional parameter spaces.

[0009] Therefore, there is a need for a wind power prediction method and system based on DIRMO and a differentiated objective function that has high search efficiency and can automatically find the optimal hyperparameter configuration, avoiding the subjectivity and inefficiency of manual parameter tuning. SUMMARY

[0010] The present application provides a wind power prediction method and system based on DIRMO and a differentiated objective function, which has high search efficiency and can automatically find the optimal hyperparameter configuration, avoiding the subjectivity and inefficiency of manual parameter adjustment.

[0011] The wind power prediction method based on DIRMO and a differentiated objective function provided by the present application comprises the following steps: S1, collecting historical power data, wind speed data and weather forecast data, and preprocessing; S2, using a VMD-CEEMDAN adaptive hybrid decomposition method to denoise and extract features from the preprocessed power and wind speed data, obtaining multi-scale features; S3, preliminary prediction stage: Using a gated recurrent unit (GRU) to model the time series of normalized power and wind speed data, capturing the time series trend and long-term dependence of the power sequence, and generating a preliminary prediction result for the future; S4, fine prediction stage: Based on the DIRMO grouping strategy, the multi-step prediction task is divided into several groups, each group constructs a LightGBM model, and the data reconstruction method is used to convert the multi-output problem into a single-output problem for training and prediction; S5, regional differentiated objective function; According to different evaluation standards in different regions, a regionally differentiated custom objective loss function is designed as the training target of the LightGBM model, and its first-order gradient and second-order Hessian matrix are derived; the training target of the LightGBM model is completely consistent with the actual evaluation index; S6, automatic optimization of hyperparameters: Using an adaptive differential evolution algorithm to automatically optimize the hyperparameters of the LightGBM model, and the optimization target is the regional differentiated evaluation index on the validation set; S7, model training and prediction: Using the optimized hyperparameters to train each group of LightGBM models, and based on the GRU preliminary prediction result and multi-scale features, the final power prediction is performed.

[0012] The present application innovatively combines the advantages of deep learning and ensemble learning, and divides the prediction method into a preliminary prediction stage and a fine prediction stage, wherein: The preliminary prediction stage uses a GRU neural network for time series prediction, quickly capturing the long-term time series dependence and trend characteristics of the power sequence, and outputting a preliminary prediction result.

[0013] The fine-grained prediction stage uses a LightGBM model based on the DIRMO grouping strategy for fine-grained prediction, decomposes the long-term prediction task into multiple independent grouping prediction tasks, and each group model has the same input feature space and is independent of each other, and can be completely parallel trained and predicted.

[0014] Two-stage linkage mechanism: the GRU output result of the preliminary prediction stage is used as one of the important input features of the LightGBM model of the fine-grained prediction stage, realizing information transmission and complementary advantages between the two stages.

[0015] Further, in S1, the preprocessing specifically includes data merging, data time resolution unification, and data structure definition; The data merging is an inner join merging of the three data sources according to the timestamp, ensuring that there is complete power, wind speed and weather data at each time point; The data time resolution unification is to unify all data according to a preset time resolution, ensuring that all data are consistent with the time granularity of power prediction; The data structure definition forms a preprocessed data set containing timestamps, power values, wind speed values and multiple weather features from the data with unified time resolution.

[0016] Further, in S2, the VMD-CEEMDAN adaptive hybrid decomposition method realizes deep denoising and multi-dimensional feature extraction of power data through multiple adaptive strategies, and the denoising and feature extraction of the VMD-CEEMDAN adaptive hybrid decomposition method includes: S21, adaptively selecting the optimal mode number K of VMD from the mode number candidate set based on the energy entropy minimization criterion; S22, decomposing the power sequence into multiple intrinsic mode functions (IMF) and a residual term through CEEMDAN pre-decomposition, realizing the separation of trend and fluctuation; S23, adopting a hierarchical VMD decomposition strategy for different frequency components, including long-term trend components, medium-term periodic components and short-term fluctuation components, and setting different mode numbers, penalty factors and denoising strategies; S24, using an energy entropy guided mode number determination method to adaptively determine the mode number according to the signal energy entropy change rate and energy concentration degree, avoiding over-decomposition or insufficient decomposition; setting a stop criterion according to the energy entropy increment, energy concentration degree and maximum mode number; removing low-value noise components and retaining effective signal components through component importance evaluation and selective reconstruction (three dimensions of energy proportion, correlation and kurtosis); and adaptively threshold denoising using different threshold functions for different modal components and different frequency components, while retaining useful signals and removing noise.

[0017] Further: in S3, the specific process of the time series modeling includes: S31, a multi-layer GRU deep neural network is constructed; S32, mean square error is used as a loss function to optimize the prediction accuracy, and an early stopping strategy is used on the validation set to prevent overfitting; S33, after training, the GRU model outputs the preliminary prediction results of the future specified steps.

[0018] Further: in S4, the specific steps of the DIRMO grouping strategy include: S41, the total prediction steps F are divided into G groups, each group contains S steps, and F=G×S; an independent LightGBM model is constructed for each group, the input features of each group model are the same, and different groups and step sequences are distinguished by adding group features; the long-term prediction task is divided into multiple parts, each group contains a fixed number of steps, and each group model is independently trained and predicted.

[0019] S42, multi-class features are constructed; by extracting features at different time scales, the prediction accuracy is improved; S43, using data reconstruction method, the multi-step output of each group is flattened into one-dimensional vector, and the features are copied for each prediction step and step identification features are added, so that the LightGBM model can handle the multi-step prediction task in single output form. Since the input feature spaces of each group model are the same and independent of each other, completely parallel training and prediction can be realized.

[0020] Further: in S6, the optimization process of the self-adaptive differential evolution algorithm is: S61, initialize the population, each individual represents a group of hyperparameter configurations; S62, calculate the diversity of the population and perform normalization to obtain the diversity index D; S63, calculate the current iteration progress and adjust the crossover probability CR; S64, calculate the roulette probability according to the historical success rate, and adaptively select the mutation strategy; generate a mutation vector according to the selected strategy, and perform binomial crossover according to the crossover probability CR to generate a trial vector; S65, train the LightGBM model using the hyperparameters corresponding to the trial vector, and evaluate it on the validation set using the regional difference objective function designed in S5; compare the fitness of the trial vector and the target vector, retain the optimal individual, and update the historical success rate of the selected strategy; S66, introduce a continuous no-improvement counter to prevent overfitting and improve efficiency.

[0021] Further: in S7, the training and prediction process of the LightGBM model includes: S71, group the required predicted steps, set the optimization range of all hyperparameters for the LightGBM model of each group, use the ADE algorithm, and use the regional differentiation index on the validation set as the optimization target to iteratively find the optimal hyperparameter configuration; S72, using the ADE algorithm to optimize the obtained optimal hyperparameter configuration, training a LightGBM model for each group of the DIRMO strategy; S73, using a GRU model to perform coarse prediction on the normalized power and wind speed data, and using the preliminary prediction result of the future period output by the GRU model as the GRU enhanced feature of the LightGBM model; S74, constructing multi-class features for each prediction step, the multi-class features including historical statistical features, multi-scale features, weather features, time features, interaction features, DIRMO group features and GRU enhanced features; S75, for each group, extracting the features of each prediction step in the group, using the data reconstruction method to set the corresponding intra-group step identifier for each prediction step, generating a plurality of prediction samples, then loading the LightGBM model of the group, predicting these prediction samples respectively to obtain a plurality of prediction values, and finally sorting and merging to form the final prediction result.

[0022] The wind power prediction system for implementing the wind power prediction method based on DIRMO and a differentiated target function comprises a data acquisition module, a data preprocessing module, a preliminary prediction module, a refined prediction module, a regional differentiation module, an automatic hyperparameter optimization module and a model prediction module. The data acquisition module is used to acquire historical power data, wind speed data and weather forecast data. The data preprocessing module is used to merge and unify the time resolution of the acquired data, and form a preprocessing data set from the preprocessed data. The preliminary prediction module is used to perform time series modeling on the normalized power and wind speed data, capture the time series trend and long-term dependence of the power sequence, and generate a preliminary prediction result for the future. The refined prediction module is used to divide the multi-step prediction task into several groups, construct a LightGBM model for each group, and use the data reconstruction method to convert the multi-output problem into a single-output problem for training and prediction. The regional differentiation module is used to design a regional differentiation custom target loss function as the training target of the LightGBM model according to different evaluation standards of different regions, and derive the first-order gradient and second-order Hessian matrix thereof, so that the training target of the LightGBM model is completely consistent with the actual evaluation index. The hyperparameter automatic optimization module is configured to automatically optimize the hyperparameters of the LightGBM model, and the optimization target is the regional differentiated evaluation index on the validation set. The model prediction module is configured to train each group of LightGBM models using the optimized hyperparameters, and perform final power prediction based on the GRU preliminary prediction result and the multi-scale features.

[0023] Further, the preliminary prediction module includes a GRU time series prediction module configured to construct a multi-layer GRU deep neural network model, and use mean square error as a loss function to optimize the multi-layer GRU deep neural network model during the training process.

[0024] Further, the fine prediction module includes a DIRMO grouping module, a multi-scale feature construction module and a data reconstruction module. The DIRMO grouping module is configured to group the total prediction steps in steps, and construct an independent LightGBM model for each group, and each LightGBM model is responsible for predicting the output of each step in the group. The multi-scale feature construction module is configured to construct multi-scale features, and the multi-scale features include time domain statistical features, frequency domain features and correlation features between components. The data reconstruction module is configured to convert the multi-step prediction target into a single-step prediction.

[0025] The beneficial effects of the present application are: The present application aims at the double regulation examination requirements of new energy power prediction, and through the cascading progressive DIRMO (Direct Multi-Output, direct multi-output strategy) prediction strategy combined with the VMD-CEEMDAN (Variational Mode Decomposition-Complete Ensemble Empirical Mode Decomposition with Adaptive Noise, variational mode decomposition-complete ensemble empirical mode decomposition with adaptive noise) adaptive hybrid decomposition technology, a complete power prediction method system is constructed, and the limitations in the existing prediction technology are solved, such as serious error accumulation, inconsistent examination indexes, large noise interference, and difficult hyperparameter optimization. The effects that can be achieved by the present application include: 1. The innovative denoising and feature extraction of VMD-CEEMDAN adaptive hybrid decomposition; The application innovatively proposes a VMD-CEEMDAN adaptive hybrid decomposition method, which breaks through the limitations of traditional decomposition methods. Through CEEMDAN pre-decomposition, the trend-wave adaptive separation is realized, the optimal modal number K is adaptively selected from multiple candidate modal numbers based on the signal energy entropy minimization criterion, and the adaptive modal number determination strategy guided by signal energy entropy is used to dynamically adjust the decomposition modal number. A hierarchical adaptive VMD decomposition strategy is used for different frequency components, combined with an improved adaptive threshold denoising method, and through component importance evaluation and selective reconstruction, low-value noise components are removed, providing high-quality input features for subsequent prediction models, and greatly improving the model's ability to capture complex power fluctuation patterns and robustness to abnormal data.

[0026] 2. Advantageous complementarity and error control of the cascaded progressive DIRMO strategy The application realizes the advantageous complementarity of GRU neural network and LightGBM model gradient boosting tree through the cascaded progressive DIRMO prediction strategy, forming a preliminary prediction-fine prediction collaborative prediction framework. The output result of the preliminary prediction stage is used as the input feature of the fine prediction stage, realizing the close linkage of the two stages. The input feature spaces of each group of models are the same and independent of each other, and can be completely parallel trained and predicted, effectively avoiding the error accumulation problem of the recursive strategy, and significantly improving the stability of long-term prediction. Compared with the direct strategy, the number of models and the training time are greatly reduced, and compared with the multi-input multi-output strategy, the optimization target and the prediction target are more consistent. The prediction strategy fully utilizes the respective advantages of deep learning and ensemble learning, and realizes the comprehensive optimization of prediction accuracy, computational efficiency and model stability.

[0027] 3. Intelligent automatic optimization and artificial parameter tuning liberation The application uses an adaptive differential evolution algorithm to realize intelligent automatic optimization of the LightGBM model hyperparameters, avoiding the problem that the model cannot achieve the optimal state due to the selection of empirical parameters by humans. The ADE algorithm effectively balances the global search and local search capabilities through three adaptive strategies. The population diversity maintenance mechanism effectively avoids the premature convergence problem, and the regional differentiation ACC is used as the fitness function to ensure that the optimization target is highly consistent with the actual application requirements. This method significantly improves the model performance while reducing the dependence on professional parameter tuning experience, improving the generalizability and practicality of the method.

[0028] 4. Wide business adaptation capability and application value The method provided by the application meets the actual demand of new energy power prediction, and provides a business analysis function suitable for multiple scenes, and has high business expansibility and flexibility. The method supports multiple prediction scenes such as ultra-short-term, short-term and medium-term prediction, and meets multiple business demands such as power market transaction, power grid dispatching and operation optimization. The method can adapt to different regional evaluation standards through the design of a regional differentiated objective function, and can adapt to the power fluctuation characteristics under different wind farms, different seasons and different weather conditions through adaptive multi-scale decomposition and intelligent super parameter optimization. The method can provide high-precision power prediction results for new energy stations, support the stations to participate in the power market transaction, reduce the evaluation cost, provide reliable prediction information for power grid dispatching, support the safe and stable operation of the power grid and the consumption of new energy, and has a wide application prospect and significant social and economic value. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is a prediction method flowchart; Figure 2 is a cascaded progressive prediction model structure schematic diagram; Figure 3 is a self-adaptive differential evolution algorithm flowchart; Figure 4 is a comparison curve diagram of 24-hour (short-term) prediction results of region B, wherein the abscissa is the sample serial number I, the ordinate is the wind power Y (unit: MW), the blue line represents the true value, and the red line represents the prediction value; Figure 5 is a comparison diagram of 24-hour (short-term) prediction sample accuracy scatter distribution of region A and region B, wherein the left graph is the accuracy result under the evaluation index of region A, the right graph is the accuracy result under the evaluation index of region B, the abscissa is the sample serial number I, and the ordinate is the sample prediction accuracy P. DETAILED DESCRIPTION

[0030] The following is only a preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered within the protection scope of the application. The following examples are only used to explain the application, and cannot be explained as a limitation of the application, and the protection scope of the application should be subject to the protection scope of the claims. The embodiments of the application are described in detail below. In order to facilitate the description of the application and simplify the description, the technical terms used in the specification of the application should be interpreted broadly, including but not limited to conventional replacement schemes not mentioned in the application, and including direct implementation and indirect implementation.

[0031] Embodiment 1 In combination Figures 1-5This embodiment discloses a new energy power prediction method based on a cascaded progressive DIRMO strategy and a regionally differentiated objective function. This method employs a cascaded progressive architecture combining coarse prediction using a gated recurrent unit (GRU) with fine prediction using a LightGBM model. It integrates DIRMO grouping strategies, regionally differentiated custom objective functions, Variational Mode Decomposition-Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (VMD-CEEMDAN) adaptive hybrid decomposition, adaptive differential evolution algorithm optimization, and multi-type feature engineering techniques to achieve high-precision and high-efficiency prediction of new energy power.

[0032] This method comprises three main modules: data preprocessing, preliminary prediction, and refined prediction. The data preprocessing module handles data acquisition, VMD-CEEMDAN adaptive hybrid decomposition denoising, and normalization. The preliminary prediction stage uses a GRU model to quickly capture temporal dependencies and output preliminary prediction results. The refined prediction stage employs a DIRMO grouping strategy to decompose the multi-step prediction task into multiple groups, with each group training a LightGBM model. The two stages are closely linked through feature transfer: the GRU output from the preliminary prediction stage serves as a crucial input feature for the LightGBM model in the refined prediction stage, achieving information fusion and complementary advantages between the two stages. Each model group is trained using a region-specific custom objective function, and hyperparameters are automatically optimized using an adaptive differential evolution algorithm. The core technical steps of this method are detailed below: S1. Data acquisition and preprocessing methods; This method collects historical power data, wind speed data, and weather forecast data. After data merging, missing value processing, and format standardization, it provides a complete data foundation for subsequent modeling. The specific implementation steps are as follows: S11, Input data acquisition; Three types of data are obtained from the data acquisition system of new energy power plants: ①Historical power data: Includes timestamps and power values ​​(unit: MW), recording the actual power generation of the power plant.

[0033] ② Wind speed data: Includes timestamps and wind speed values ​​(unit: m / s), recording the measured wind speed at the site.

[0034] ③ Meteorological forecast data: includes timestamps and 8 meteorological dimensions, namely air pressure, temperature, precipitation, dew point, meridional wind, zonal wind, net radiation, and total radiation.

[0035] S12, data preprocessing; ①Data merging: inner join the three data sources according to the timestamp to ensure complete power, wind speed and weather data at each time point. The inner join includes data merging and missing value processing.

[0036] ②Data time resolution unification: unify all data to 15-minute time resolution to keep consistent with the time granularity of power prediction.

[0037] ③Data structure definition: each row of preprocessed data represents a complete observation record at a time point, including 1 timestamp, 1 power value, 1 wind speed value and 8 weather features, totaling 11 fields.

[0038] The data with unified time resolution is formed into a preprocessed dataset containing timestamp, power value, wind speed value and multiple weather features.

[0039] That is, through S1, the original data is merged and the time resolution is unified to obtain the preprocessed dataset.

[0040] S2, VMD-CEEMDAN adaptive hybrid decomposition and feature extraction method; This method innovatively combines the adaptability of variational mode decomposition (VMD) and the multi-scale of complete ensemble empirical mode decomposition (CEEMDAN) to propose a VMD-CEEMDAN adaptive hybrid decomposition method.

[0041] First, the power and wind speed data are decomposed into intrinsic mode functions (IMF) of different frequencies by CEEMDAN to realize trend-fluctuation separation; then the number of modes K is adaptively determined according to the signal energy entropy; then a hierarchical adaptive VMD decomposition strategy is used for different frequency components, and the decomposition mode number is adaptively determined according to the signal energy entropy; an improved adaptive threshold denoising method is used to denoise each modal component; finally, through component importance evaluation and selective reconstruction, low-value noise components are removed, and multi-scale time-frequency correlation features are extracted for subsequent modeling. Compared with traditional decomposition methods, this method can better adapt to signal characteristics, achieve more precise noise removal and more abundant feature extraction. The low-value noise components are evaluated comprehensively by energy ratio, correlation coefficient with the original signal and kurtosis, and their importance scores are determined, expressed as: ; wherein, is the comprehensive importance score of the i-th component, is the energy of the i-th component, For total energy, Let be the correlation coefficient between the i-th component and the original signal. Let be the kurtosis of the i-th component. This represents the maximum kurtosis deviation of all components. All are weighting coefficients, and satisfy the following conditions: When a component has a small energy proportion, weak correlation with the original signal, and kurtosis close to a standard normal distribution, it indicates that the component is mainly random noise rather than an effective signal. When the overall importance score is... Below the preset threshold When a component is identified as a low-value noise component, it is removed; components with scores higher than or equal to a preset threshold are considered valid signal components, which are retained and used in subsequent reconstruction.

[0042] The specific implementation steps are as follows: S21. Adaptive mode number K selection; Unlike traditional methods that fix the number of modes, this method automatically selects the optimal number of modes K based on the time-frequency characteristics of the power data.

[0043] ① Construct a candidate set of modal numbers: including 10 candidate modal numbers from K=3 to K=12.

[0044] ② Calculate the signal energy entropy index: For each candidate mode number K, perform VMD decomposition on the power sequence and calculate the decomposed signal energy entropy E: ; in, , , ; Where E is the signal energy entropy, K is the mode number, and i is the index of the mode component. It represents the energy percentage of the i-th modal component, and T is the total number of time points in the power sequence. It is the amplitude of the i-th mode component obtained from the decomposition at time t. It is the total energy of all modal components.

[0045] ③ Selecting the optimal number of modes: Choose the mode number K that minimizes the signal energy entropy as the optimal number of modes. A smaller signal energy entropy indicates a more concentrated signal energy and a better decomposition effect. Let the optimal number of modes be denoted as K. .

[0046] S22, CEEMDAN pre-decomposition and trend-volatility separation; Pre-decomposition is performed using Complete Set Empirical Mode Decomposition (CEEMDAN) to separate long-term trends and short-term fluctuations.

[0047] ① CEEMDAN decomposition: the preprocessed power sequence is decomposed into a plurality of and 1 residual component: ; wherein, is the power value of the preprocessed power sequence at time t, n is the number of intrinsic mode functions (IMF) obtained by decomposition, is the amplitude of the i th intrinsic mode function component at time t, is the amplitude of the n th residual component at time t; ② Trend-fluctuation separation: according to the frequency characteristics of , the components are divided into three categories: a) long-term trend component: ; b) medium-term periodic component: ; c) short-term fluctuation component: ; wherein, is the long-term trend component, is the medium-term periodic component, is the short-term fluctuation component; ③ Component characteristic analysis: the Hurst index H is calculated for the long-term trend component T(t), the medium-term periodic component C(t) and the short-term fluctuation component S(t) obtained by decomposing the power sequence, respectively, to judge the persistence and predictability of each component: ; wherein, R is the range of the currently calculated component, S is the standard deviation of the currently calculated component, and L is the number of data points of the currently calculated component, indicates that the sequence has persistence, indicates that the sequence has anti-persistence (mean reversion). The Hurst index, also known as the Hurst index or Hurst-Vit index, is a statistical index for analyzing time series data.

[0048] S23, hierarchical adaptive VMD decomposition; Different VMD decomposition strategies are used for different frequency components obtained by CEEMDAN decomposition.

[0049] According to the characteristics of different frequency components, the following parameters are defined: a) is the energy entropy increment stop threshold of the long-term trend component, and the value range is 0.001~0.01, when the energy entropy increment is less than the value, it indicates that the decomposition has no significant improvement on the trend extraction; b) is the energy entropy increment stop threshold of the short-term fluctuation component, and the value range is 0.005~0.02; c) The energy concentration threshold for the intermediate periodic component is set at 0.3 to 0.5. When the energy concentration exceeds this value, it indicates that the main periodic component has been effectively extracted. d) , , These are the VMD penalty factors corresponding to the long-term trend component, the medium-term cycle component, and the short-term volatility component, respectively. e) K_max is the maximum allowed number of modes, with a value range of 8 to 15; ① Long-term trend component T(t): a) Determining the number of modes: Starting with an initial number of modes K=2, gradually increase the number of modes and calculate the signal energy entropy E(K), and calculate the energy entropy increment. ,when Stop at this time; b) Noise reduction strategy: Conservative noise reduction to preserve complete trend information; c) Penalty factor selection: Larger penalty factor ( ), to ensure smooth decomposition; ②Medium-term periodic component C(t): a) Determining the number of modes: Starting with an initial number of modes K=2, gradually increase the number of modes and calculate the energy concentration. ,in Let i be the energy percentage of the i-th modal component, when Stop

[0050] b) Noise reduction strategy: Moderate noise reduction while preserving periodic characteristics; c) Selection of penalty factor: Medium penalty factor To balance the accuracy and smoothness of the decomposition process; ③ Short-term fluctuation component S(t): a) Determining the number of modes: Starting with an initial number of modes K=2, gradually increase the number of modes until the maximum number of modes is reached. Or signal energy entropy increment Stop at this time; b) Noise reduction strategy: Aggressive noise reduction, employing an adaptive thresholding method to remove high-frequency noise, with a noise reduction threshold... Where N is the signal length, The noise standard deviation is estimated using the median absolute deviation (MAD) method: ; in, The first level of detail coefficients, median indicates the median, and 0.6745 is the conversion coefficient between MAD and standard deviation under Gaussian distribution; c) penalty factor selection: smaller penalty factor , improve the decomposition accuracy; S24, signal energy entropy guided modal number adaptive determination; This step describes the general method framework of component modal number determination in S23, which does not fix the modal number, but automatically determines the optimal modal number according to the signal energy distribution. The following general parameters are defined: a) is the general energy entropy increment stop threshold, and the corresponding value is taken according to the component type; b) is the general energy concentration stop threshold, and the corresponding value is taken according to the component type; c) is the energy concentration when the modal number is K; d) E(K) is the energy entropy increment when the modal number is K; ① Step-by-step decomposition and calculation of signal energy entropy: For modal number K, VMD decomposition is performed, and the signal energy entropy is calculated, and the energy concentration is calculated, where is the energy proportion of the i-th modal component, and the energy entropy increment is calculated, where is the signal energy entropy when the modal number is K-1; ② Stopping criterion: stop the decomposition when any of the following conditions is met: a) The signal energy entropy increment is less than the threshold: ; b) The energy concentration exceeds the threshold: ; c) The maximum modal number is reached: ; ③ Optimal modal number selection: select the modal number that minimizes the signal energy entropy and maximizes the energy concentration as the optimal modal number.

[0051] That is, through S2, the VMD-CEEMDAN decomposition is performed on the preprocessed data set to obtain multi-scale features.

[0052] S3, preliminary prediction stage: GRU time series prediction method; In the preliminary prediction stage, a Gated Recurrent Unit (GRU) is used to perform time-series modeling on the normalized power and wind speed data. This quickly captures the temporal trend and long-term dependencies of the power sequence, generating preliminary predictions for the future. The GRU network, through its reset and update gate mechanisms, can effectively learn long-term temporal dependencies. Compared to LSTM, it has a simpler structure and faster training speed, making it suitable for processing the temporal characteristics of power data. The core role of this stage is to provide preliminary prediction results as input features for the refined prediction stage. The specific implementation steps are as follows: S31. Construct a multi-layer GRU deep neural network, including: ① Input layer: Receives historical time window data, with input features being normalized power and wind speed; ②GRU layer: A multi-layer GRU structure is adopted to extract temporal features layer by layer, and long-term dependencies are learned through reset gate and update gate mechanisms; ③ Dropout layer: Add a Dropout layer between GRU layers to prevent overfitting and improve the model's generalization ability; ④ Output layer: Fully connected layer, outputs preliminary prediction results for the next several steps; S32. The mean squared error (MSE) is used as the loss function to optimize the prediction accuracy, and an early stopping strategy is used on the validation set to prevent overfitting. S33. After training, the GRU model outputs preliminary predictions for a specified future step. These preliminary predictions contain the temporal trends and long-term dependencies learned by the GRU, serving as a crucial link between the two stages. These preliminary predictions will be used as one of the important input features of the LightGBM model in the refined prediction stage, and will be input into the LightGBM model along with other features to achieve information transfer and feature enhancement between the two stages.

[0053] S4. Refined Forecasting Stage: DIRMO Group Forecasting Method; The refined prediction stage uses the GRU output from the initial prediction stage as one of the important input features. Combined with rich feature engineering, a DIRMO grouping strategy is employed to decompose the multi-step prediction task into several sub-tasks. Each sub-task trains only one LightGBM model to execute the MIMO prediction strategy. The models in each group have the same input feature space and can be trained and predicted in parallel, avoiding the error accumulation problem of recursive strategies. At the same time, it significantly reduces the number of models compared to direct strategies. The specific implementation steps are as follows: S41, DIRMO grouping strategy design; The prediction task with a total of F prediction steps is decomposed into G groups, each containing S steps, satisfying the relation An independent LightGBM model is built for each group, and each LightGBM model is responsible for predicting the S-step output of the group. The input features of each group model are the same, but they are distinguished by DIRMO group features for different groups and different steps within the group, enabling the model to learn the characteristics of different prediction steps.

[0054] S42, feature construction method; Multi-class features are constructed to fully exploit data value, and by extracting features at different time scales, the prediction model can learn and understand the multi-level rules of power changes, thereby improving the modeling ability and prediction accuracy of complex fluctuation patterns. Including: ① Historical statistical features: statistical features (mean, standard deviation, maximum, minimum, rate of change, correlation coefficient, etc.) are extracted from historical power and wind speed data, reflecting the overall distribution and trend of historical data; ② Multi-scale features: time-frequency domain features extracted from power and wind speed data by VMD-CEEMDAN adaptive hybrid decomposition method, including time domain statistical features (mean, standard deviation, skewness, kurtosis, etc.), frequency domain features (main frequency, spectral centroid, spectral bandwidth, etc.) and component correlation features.

[0055] ③ Weather features: extract weather forecast data (barometric pressure, temperature, precipitation, dew point, wind direction, radiation, etc.) at the prediction time, reflecting future weather conditions; ④ Time features: extract time attributes (hour, week, month, season, whether it is a weekday, etc.) at the prediction time, capturing the periodicity of power data; ⑤ Interaction features: build interaction features between key variables (such as power x barometric pressure, power x temperature, etc.), capturing the non-linear relationship between variables; ⑥ DIRMO group features: build grouping information features (group number, intra-group position, time offset, etc.), enabling the model to distinguish between different groups and different steps within the group; ⑦ GRU enhanced features (preliminary prediction result features): use the GRU output results of the preliminary prediction stage as input features for the refined prediction stage, including GRU prediction values, prediction value statistical features, trend coefficients, difference features, etc., to realize information transmission and fusion between the two stages; S43, realize multi-output by data reconstruction method; LightGBM model natively only supports single output, while DIRMO strategy requires each model to directly predict the multi-step output of the group. Therefore, the data reconstruction method is used to convert the multi-output problem into a single output problem, and each training sample is reconstructed into multiple sub-samples, each corresponding to a prediction step within the group, solving the contradiction between the standard LightGBM model's single output and the requirement of multi-step output: ①Reconstruction principle: flatten the multi-step prediction target of each group into a one-dimensional vector, duplicate the features for each prediction step, and add step identification features to enable the model to distinguish between different steps within the group; ②Training process: each original sample is expanded into S training samples, each corresponding to a prediction step within the group, and the number of samples is expanded to S times the original number; ③Prediction process: for test samples, set different step identification feature values respectively to generate S prediction samples, and use the model to predict respectively to obtain the S-step prediction results of the group; Thus, by data reconstruction, the multi-output problem is converted into a single-output problem, enabling the LightGBM model to complete the multi-output requirement of the DIRMO strategy.

[0056] S5, regional differentiation target function; The method designs a regional differentiation custom target function according to different evaluation standards of A region and B region, so that the training target of the LightGBM model is completely consistent with the actual evaluation index, thereby directly optimizing the evaluation accuracy ACC. The regional differentiation target function is converted into a LightGBM model custom target function, and the corresponding first-order gradient and second-order gradient (Hessian matrix) are derived, so that the model training process is directly optimized for the evaluation index.

[0057] The A region refers to: an area with an average annual wind speed greater than 6 meters per second and relatively stable wind speed fluctuations; or an area with a wind speed within-day standard deviation less than 2 meters per second and a wind direction change rate less than 15 degrees per hour; or an area with an annual effective generating hour greater than 2000 hours and a wind power curve correlation coefficient with a theoretical power curve higher than 0.85.

[0058] The B region refers to an area with complex wind resource characteristics, specifically: 1. Wind speed fluctuation is severe: the wind speed within-day standard deviation is greater than 3 meters per second, and the gust coefficient (ratio of maximum wind speed to average wind speed) exceeds 1.8; 2. Extreme weather occurs frequently: the number of days affected by extreme weather such as typhoons and monsoons is more than 30 days per year, or the frequency of wind speed sudden change events (hourly wind speed change rate exceeds 5 meters per second) is more than 50 times per year; 3. Low wind or high wind period accounts for a high proportion: the proportion of time periods with wind speed below the cut-in wind speed or above the cut-out wind speed leading to shutdown is more than 25% of the total annual time.

[0059] The specific implementation steps are as follows: ①For A region evaluation indicators, A region uses a weighted RMSE target function, with weights being the proportion of each sample absolute error to total absolute error; expressed by loss function as: ; where, is the loss function of A area, is the true power value of the i-th sample, is the predicted power value of the i-th sample, is the weight of the i-th sample.

[0060] Weight is dynamically calculated according to the current prediction value at each iteration, and the weight distribution is continuously adjusted as the model training progresses, and the calculation formula is: ; First-order derivative: ; Hessian calculation: ; ②For the evaluation index of B area, the B area adopts the normalized MAE objective function, and the denominator of the normalization is the larger value of the true power and the installed capacity in a certain proportion. The loss function is expressed as: ; where, is the loss function of B area, and cap is the installed capacity of the wind farm in the month; First-order derivative: ; Hessian calculation: ; S6, adaptive differential evolution algorithm optimization hyperparameter method; This method uses the adaptive differential evolution algorithm (Adaptive Differential Evolution, ADE) to automatically optimize the hyperparameter configuration of the LightGBM model, and uses the regional differentiation index on the validation set as the optimization target.

[0061] S61, initialization and diversity calculation: Initialize the population (NP individuals), and each individual represents a set of hyperparameter configurations. At the beginning of each iteration, calculate the population diversity and normalize to obtain the diversity index D.

[0062] S62, diversity calculation and adaptive adjustment of scaling factor F: First, calculate the population diversity and perform normalization to obtain the diversity index D. Then adjust the scaling factor F according to the relationship between D and the preset threshold: ① If D is greater than the threshold (indicating that the population diversity is high), set the scaling factor to the minimum value (F=Fmin), focusing on local development; ② If D is less than or equal to a threshold value (indicating that the population diversity is low), the scaling factor is set to a maximum value (F=Fmax) to enhance the mutation disturbance and jump out of the local optimum.

[0063] S63, Cross probability CR adjustment based on iteration progress: The current iteration progress (iter / itermax) is calculated, and the cross probability CR is adjusted using a linear decreasing strategy. It is expressed as: This strategy maintains a high CR at the early stage of iteration to fully explore, and reduces the CR at the later stage to promote convergence.

[0064] S64, Mutation and crossover operation: ① Strategy selection: calculate the roulette probability according to the historical success rate, and adaptively select the mutation strategy (such as rand / 1, best / 1, current-to-best / 1); ② Generate vector: generate mutation vector according to the selected strategy, and perform binomial crossover according to the CR calculated in step S63 to generate trial vector.

[0065] S65, Fitness evaluation and population update: Use the hyperparameters corresponding to the trial vector to train the LightGBM model, and use the regional differentiation objective function designed in S5 to evaluate on the validation set. Compare the fitness of the trial vector and the target vector, keep the optimal individual, and update the historical success rate of the selected strategy.

[0066] S66, Early stopping mechanism and termination judgment: Introduce a continuous no improvement counter n to prevent overfitting and improve efficiency: ① If the current iteration has improved (i.e. a better individual has been found), reset the counter ; ② If there is no improvement (i.e. the optimal fitness value of the entire population on the validation set does not improve in the current generation), the counter is incremented by 1 ( ); ③ Termination condition: when the maximum number of iterations (iter=itermax) is met or the number of consecutive no improvements reaches the preset threshold (n=N), the algorithm terminates and outputs the optimal hyperparameter configuration; otherwise, jump back to step S61 and continue the loop.

[0067] S7, Model training and prediction method; This method uses the optimal hyperparameter configuration obtained by the ADE algorithm described in S6 to train multiple LightGBM models and make predictions on test samples. Each group of models can be trained and predicted in parallel, fully utilizing the computing resources and improving the efficiency.

[0068] S71, Hyperparameter selection and optimization; In the DIRMO prediction strategy, the required prediction steps are grouped, and for each group of LightGBM models, the optimization range of all hyperparameters (including tree structure parameters, learning rate, feature sampling parameters, regularization parameters, etc.) is set. The ADE algorithm described in S6 is used to verify the regional differentiation index on the validation set as the optimization target, and the optimal hyperparameter configuration is iteratively found.

[0069] S72, model training configuration; Using the optimal hyperparameter configuration obtained by the ADE algorithm, a LightGBM model is trained for each group of the DIRMO strategy. The training configuration of each model includes: ① Basic algorithm: Gradient Boosting Decision Tree (GBDT) algorithm; ② Objective function: Use the regional differentiation custom objective function described in S5; ③ Hyperparameter configuration: Use the optimal hyperparameter value obtained by the ADE algorithm described in S6; ④ Early stopping strategy: Use early stopping mechanism to prevent overfitting, and stop training when the validation set loss does not decrease for multiple rounds in a row; Model prediction process; S73, GRU preliminary prediction; Use the GRU model trained in S3 for rough prediction. Input the historical normalized power and wind speed data, and the GRU model outputs the preliminary prediction results for the future period, which are used as GRU enhanced features for the LightGBM model.

[0070] S74, feature construction; According to the feature construction method described in S4, multiple types of features are constructed for each prediction step. The features include: historical statistical features, multi-scale features, meteorological features, time features, interaction features, DIRMO group features, and GRU enhanced features. Among them, the multi-scale features come from the VMD-CEEMDAN hybrid decomposition described in S2, and the GRU enhanced features are calculated based on the GRU preliminary prediction results obtained in S1.

[0071] S75, grouped prediction; For each group, extract the features of each prediction step in the group, and use the data reconstruction method to set the corresponding intra-group step identifier for each prediction step to generate multiple prediction samples. Then load the LightGBM model for this group, and predict these prediction samples to obtain multiple prediction values. Finally, sort and merge to form the final prediction result.

[0072] As Figure 4As shown in the 24-hour (short-term) prediction result comparison curve of region B, the prediction accuracy of this method in region B is significantly improved, especially in the case of severe wind speed fluctuations and frequent extreme weather. By introducing the regional differentiated objective function, the model can more accurately capture the power variation law under complex wind resource characteristics. In addition, the strategy of optimizing hyperparameters by adaptive differential evolution algorithm further enhances the generalization ability of the model, so that it can maintain stable prediction performance under different meteorological conditions. Experimental results show that compared with the traditional single objective function prediction method, the average absolute error in region B is reduced by more than 10%. This fully verifies the effectiveness and practicality of the combination of regional differentiated objective function and ADE algorithm.

[0073] As shown in Figure 5 The scatter plot of prediction sample accuracy in regions A and B can clearly observe the different performance characteristics of prediction accuracy under two regional evaluation indicators. For region A, the deviation between predicted value and true value is small, which reflects high prediction accuracy and stability. This is due to the effective optimization of the weighted RMSE objective function for region A evaluation indicators, which enables the model to better fit the smooth power variation trend. The accuracy of each sample in region B is relatively low, which is due to the amplification of prediction error in low power period by region B evaluation indicators, but compared with the traditional method using single objective function, the overall accuracy has been significantly improved.

[0074] Further analysis shows that in the experiment using the A region differentiated objective function, the model is particularly accurate in capturing long-term trend components T(t) and medium-term cycle components C(t), verifying the advantages of VMD-CEEMDAN hybrid decomposition strategy in handling multi-scale features. When using the B region differentiated objective function, the prediction effect of short-term fluctuation component S(t) is most significantly improved, indicating that the combination of GRU enhanced features and data reconstruction method effectively alleviates the problem that a single model cannot cope with complex scenarios.

[0075] In addition, through the optimization of hyperparameters by ADE algorithm, the models of the two regions show stronger robustness and generalization ability. Experimental results show that whether in high wind speed or low wind speed period, the model can maintain high prediction consistency, fully proving the potential of adaptive differential evolution algorithm in improving model performance. These achievements provide an innovative and efficient solution for the field of wind power prediction, which has important practical application value.

Claims

1. A wind power prediction method based on DIRMO and differentiated objective function, characterized in that, Includes the following steps: S1. Collect historical power data, wind speed data, and weather forecast data, and perform preprocessing. S2. The VMD-CEEMDAN adaptive hybrid decomposition method is used to denoise, normalize and extract features from the preprocessed power and wind speed data to obtain multi-scale features. S3, Preliminary Forecasting Stage: A gated cyclic unit (GRU) is used to perform time-series modeling on the normalized power and wind speed data, capturing the time-series trend and long-term dependencies of the power sequence, and generating preliminary predictions for the future. S4, Refined Forecasting Stage: Based on the DIRMO grouping strategy, the multi-step prediction task is divided into several groups, and a LightGBM model is built for each group. The data reconstruction method is used to transform the multi-output problem into a single-output problem for training and prediction. S5, Regional Differentiation Objective Function; To address the different assessment criteria in different regions, a custom objective loss function with regional differences is designed as the training objective of the LightGBM model, and its first-order gradient and second-order Hessian matrix are derived; so that the training objective of the LightGBM model is completely consistent with the actual assessment indicators. S6, Automatic Hyperparameter Optimization: The hyperparameters of the LightGBM model are automatically optimized using an adaptive differential evolution algorithm, with the optimization objective being the regional differentiation evaluation index on the validation set. S7. Model Training and Prediction: The optimized hyperparameters were used to train each group of LightGBM models, and the final power prediction was made based on the preliminary prediction results of GRU and multi-scale features.

2. The wind power prediction method based on DIRMO and differentiated objective function according to claim 1, characterized in that, In S1, the preprocessing specifically includes data merging, data time resolution unification, and data structure definition; The data merging process involves performing an inner join on the three data sources based on timestamps to ensure that complete power, wind speed, and meteorological data are available at each point in time. The unified data time resolution means unifying all data according to a preset time resolution to ensure that all data are consistent with the time granularity of power prediction. The data structure definition forms a preprocessed dataset containing timestamps, power values, wind speed values, and multiple meteorological features from the data after unifying the time resolution.

3. The wind power prediction method based on DIRMO and differentiated objective function according to claim 1, characterized in that, In S2, the denoising and feature extraction of the VMD-CEEMDAN adaptive hybrid decomposition method includes: S21. Based on the energy entropy minimization criterion, adaptively select the optimal mode number K of VMD from the candidate mode number set; S22. By using CEEMDAN, the power sequence is decomposed into multiple intrinsic mode functions (IMFs) and a residual term, thus separating the trend from the fluctuation. S23. A hierarchical VMD decomposition strategy is adopted for different frequency components, including long-term trend components, medium-term cycle components and short-term fluctuation components, and different modal numbers, penalty factors and denoising strategies are set respectively. S24. An energy entropy-guided mode number determination method is adopted, and a stopping criterion is set based on the energy entropy increment, energy concentration, and maximum mode number. Through component importance assessment and selective reconstruction, low-value noise components are removed, and effective signal components are retained.

4. The wind power prediction method based on DIRMO and differentiated objective function according to claim 1, characterized in that, In S3, the specific process of time series modeling includes: S31. Construct a multi-layer GRU deep neural network; S32. The mean squared error is used as the loss function to optimize the prediction accuracy. At the same time, an early stopping strategy is used on the validation set to prevent overfitting. S33. After training, the GRU model outputs preliminary predictions for the future specified step size.

5. The wind power prediction method based on DIRMO and differentiated objective function according to claim 1, characterized in that, In S4, the specific steps of the DIRMO packet strategy include: S41. Divide the total number of prediction steps F into G groups, each group containing S steps, satisfying F=G×S; construct an independent LightGBM model for each group, with the same input features for each group model, and distinguish between different groups and the step order within a group by adding group features. S42. Construct multi-class features; improve prediction accuracy by extracting features at different time scales; S43. Using the data reconstruction method, the multi-step output of each group is flattened into a one-dimensional vector, and the features are copied for each prediction step and step identification features are added, so that the LightGBM model can handle multi-step prediction tasks in a single output form.

6. The wind power prediction method based on DIRMO and differentiated objective function according to claim 1, characterized in that, In S6, the optimization process of the adaptive differential evolution algorithm is as follows: S61. Initialize the population, with each individual representing a set of hyperparameter configurations; S62. Calculate the population diversity and normalize it to obtain the diversity index D; S63. Calculate the current iteration progress and adjust the crossover probability CR; S64. Calculate the roulette wheel betting probability based on the historical success rate, and adaptively select a mutation strategy; generate a mutation vector based on the selected strategy, and perform binomial crossover based on the crossover probability CR to generate an experimental vector; S65. Train the LightGBM model using the hyperparameters corresponding to the experimental vectors, and evaluate it on the validation set using the regional differentiation objective function designed in S5; compare the fitness of the experimental vectors and the target vectors, retain the best individual, and update the historical success rate of the selected strategy; S66. Introduce a continuous, unimproved counter to prevent overfitting and improve efficiency.

7. The wind power prediction method based on DIRMO and differentiated objective function according to claim 1, characterized in that, In S7, the training and prediction process of the LightGBM model includes: S71. Group the required number of prediction steps, set the optimization range of all hyperparameters for each LightGBM model, and use the ADE algorithm with the regional differentiation index on the validation set as the optimization target to iteratively find the optimal hyperparameter configuration. S72. Using the optimal hyperparameter configuration obtained by the ADE algorithm, train a LightGBM model for each group of the DIRMO strategy. S73. Use the GRU model to make coarse predictions on the normalized power and wind speed data, and use the preliminary prediction results of the future time period output by the GRU model as the GRU enhancement features of the LightGBM model. S74. Construct multiple types of features for each prediction step, including historical statistical features, multi-scale features, meteorological features, temporal features, interaction features, DIRMO group features, and GRU enhancement features; S75. For each group, extract the features of each prediction step in the group, use the data reconstruction method to set the corresponding group step number identifier for each prediction step, and generate multiple prediction samples; then load the LightGBM model of the group, predict these prediction samples separately, obtain multiple prediction values, and finally sort and merge them to form the final prediction result.

8. A wind power prediction system for implementing the wind power prediction method based on DIRMO and differentiated objective function as described in any one of claims 1-7, characterized in that, It includes a data acquisition module, a data preprocessing module, a preliminary prediction module, a refined prediction module, a regional differentiation module, a hyperparameter automatic optimization module, and a model prediction module; The data acquisition module is used to collect historical power data, wind speed data, and weather forecast data. The data preprocessing module is used to merge the collected data and unify the time resolution, and to form a preprocessed dataset from the preprocessed data. The preliminary prediction module is used to perform time-series modeling on power and wind speed data, capture the time-series trend and long-term dependency of the power sequence, and generate preliminary prediction results for the future. The refined prediction module is used to divide the multi-step prediction task into several groups, build a LightGBM model for each group, and use the data reconstruction method to transform the multi-output problem into a single-output problem for training and prediction. The regional differentiation module is used to design a regionally differentiated custom target loss function as the training objective of the LightGBM model for different assessment standards in different regions, and derive its first-order gradient and second-order Hessian matrix; so that the training objective of the LightGBM model is completely consistent with the actual assessment indicators. The hyperparameter automatic optimization module is used to automatically optimize the hyperparameters of the LightGBM model, with the optimization target being the regional differentiation evaluation index on the validation set. The model prediction module is used to train each group of LightGBM models using optimized hyperparameters, and to perform final power prediction based on the preliminary prediction results of GRU and multi-scale features.

9. The wind power prediction system based on the DIRMO and differentiated objective function wind power prediction method according to claim 8, characterized in that, The preliminary prediction module includes a GRU time series prediction module, which is used to construct a multi-layer GRU deep neural network model. During the training process, the mean squared error is used as the loss function to optimize the multi-layer GRU deep neural network model.

10. The wind power prediction system based on the DIRMO and differentiated objective function wind power prediction method according to claim 8, characterized in that, The refined prediction module includes a DIRMO grouping module, a multi-scale feature construction module, and a data reconstruction module; The DIRMO grouping module is used to group the total prediction steps into steps and build an independent LightGBM model for each group. Each LightGBM model is responsible for predicting the output of each step of the group. The multi-scale feature construction module is used to construct multi-scale features; the multi-scale features include time-domain statistical features, frequency-domain features, and correlation features between components. The data reconstruction module is used to transform multi-step prediction targets into single-step predictions.

Images