Wind power prediction method based on dirmo and differentiated objective function

Abstract

The application relates to the field of wind power prediction, and discloses a wind power prediction method and system based on DIRMO and a differentiated target function. In order to solve the defects that a single model is used for prediction in the existing multi-step power prediction, time series modeling and feature processing cannot be simultaneously considered, the prediction accuracy is reduced, the model robustness is insufficient, and the search efficiency is low, the power and wind speed data after preprocessing are denoised, normalized and feature-extracted to obtain multi-scale features; the normalized power and wind speed data are subjected to time series modeling to generate a preliminary prediction result in the future; the multi-step prediction task is divided into several groups, and a multi-output problem is converted into a single-output problem for training and prediction; the hyperparameters of a LightGBM model are automatically optimized, the optimized hyperparameters are used for training the LightGBM models of the groups, and the final power prediction is carried out based on the preliminary prediction result of the GRU and the multi-scale features. The application is mainly used for predicting wind power.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction, and in particular to a wind power prediction method and system based on a cascaded progressive DIRMO strategy and a differentiated objective function. Background Technology

[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 grid dispatching department to rationally arrange power generation plans, reduce reserve capacity requirements, reduce wind and solar curtailment, and support new energy power plants to participate in the electricity market.

[0003] For multi-step power prediction, existing strategies mainly include recursive strategies, direct strategies, Multiple-Input Multiple-Output (MIMO) strategies, and Direct Multi-Output (DIRMO) strategies. Recursive strategies use a single model for iterative prediction, predicting one step at a time and using the prediction result as input for the next step. Direct strategies build an independent model for each prediction step. MIMO strategies use a single model to predict all steps simultaneously. DIRMO strategies decompose multi-step prediction into several groups, training a model for each group to directly predict multiple steps.

[0004] In machine learning models, performance depends heavily on hyperparameter configuration. Commonly used hyperparameter optimization methods include grid search, random search, Bayesian optimization, and heuristic algorithms. Based on different parameter optimization criteria and computational cost requirements, different methods exhibit varying advantages and disadvantages in optimization speed and accuracy. These issues, to some extent, limit the improvement of accuracy in multi-step prediction tasks, necessitating further optimization through technological upgrades.

[0005] Regarding multi-step prediction strategies, recursive prediction strategies suffer from severe error accumulation, especially as prediction durations increase. Direct prediction strategies build independent models for each prediction step, incurring huge computational costs, and the models cannot share information, thus failing to capture the fluctuation characteristics of strongly autocorrelated time series such as wind power. The optimization objective of MIMO prediction strategies is the average error across all steps, but in actual evaluation, the error weights for each step are different, and as prediction durations increase, too many output ports make model parameter training difficult. The DIRMO strategy combines direct prediction and MIMO strategies, but its accuracy heavily depends on the number of groups and the step length of each group.

[0006] First, existing methods typically employ a single model for prediction, making it difficult to simultaneously address temporal modeling and feature processing. Deep learning models excel at capturing temporal dependencies but have limited ability to handle complex feature interactions; traditional machine learning models are adept at handling high-dimensional features and nonlinear relationships but lack the ability to model long-term temporal dependencies. A single model cannot fully leverage the advantages of different modeling methods, thus limiting further improvements in prediction accuracy.

[0007] Secondly, new energy power data is affected by various factors such as changes in meteorological conditions and fluctuations in equipment operating status, resulting in a large amount of high-frequency noise and abnormal fluctuations in the raw data. These noise components are mixed with the actual power change trend, directly affecting the model's learning of power change patterns. Existing methods typically use the raw data directly for modeling or only employ simple smoothing processes such as moving averages and median filtering. While these methods can reduce the impact of noise to some extent, they also smooth out useful signal details, leading to information loss. The presence of noisy data interferes with the model training process, causing the model to learn incorrect patterns, reducing generalization ability, and ultimately resulting in decreased prediction accuracy and insufficient model robustness.

[0008] Third, machine learning models such as Lightweight Gradient Boosting Machines (LightGBM) typically contain a large number of hyperparameters, the configuration of which has a decisive impact on model performance. Existing methods mainly rely on manual hyperparameter tuning or simple grid search for hyperparameter optimization. Manual tuning requires researchers to repeatedly try different parameter combinations based on experience. This process is not only time-consuming and labor-intensive, but also highly dependent on the professional knowledge and experience of the tuners, making it difficult to guarantee finding the globally optimal configuration. While grid search can systematically traverse the parameter space, its computational cost increases exponentially with the parameter dimension. Although random search alleviates the computational cost problem to some extent, its search efficiency is still low, often requiring a large amount of sampling 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 differentiated objective functions that has high search efficiency, can automatically find the optimal hyperparameter configuration, and avoids the subjectivity and inefficiency of manual parameter tuning. Summary of the Invention

[0010] To address the shortcomings of existing multi-step power prediction methods that rely on a single model, making it difficult to simultaneously consider time-series modeling and feature processing, resulting in decreased prediction accuracy, insufficient model robustness, and low search efficiency, this invention provides a wind power prediction method and system based on DIRMO and differentiated objective functions. This method offers high search efficiency, automatically finds the optimal hyperparameter configuration, and avoids the subjectivity and inefficiency of manual parameter tuning.

[0011] The wind power prediction method based on DIRMO and differentiated objective function described in this invention includes the following steps:

[0012] S1. Collect historical power data, wind speed data, and weather forecast data, and perform preprocessing.

[0013] S2. The VMD-CEEMDAN adaptive hybrid decomposition method is used to denoise and extract features from the preprocessed power and wind speed data to obtain multi-scale features.

[0014] S3, Preliminary Forecasting Stage:

[0015] 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.

[0016] S4, Refined Forecasting Stage:

[0017] 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.

[0018] S5, Regional Differentiation Objective Function;

[0019] 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.

[0020] S6, Automatic Hyperparameter Optimization:

[0021] 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.

[0022] S7. Model Training and Prediction:

[0023] 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.

[0024] This invention innovatively combines the advantages of deep learning and ensemble learning, dividing the prediction method into a preliminary prediction stage and a refined prediction stage, wherein:

[0025] In the preliminary prediction stage, a GRU neural network is used for time series prediction to quickly capture the long-term time series dependencies and trend characteristics of the power series and output preliminary prediction results.

[0026] In the refined prediction stage, the LightGBM model is used based on the DIRMO grouping strategy to perform refined prediction. The long-term prediction task is decomposed into multiple independent group prediction tasks. Each group model has the same input feature space and is independent of each other, and can be trained and predicted in parallel.

[0027] Two-stage linkage mechanism: The output of GRU in the preliminary prediction stage serves as one of the important input features of the LightGBM model in the refined prediction stage, realizing information transfer and complementary advantages between the two stages.

[0028] Furthermore: In S1, the preprocessing specifically includes data merging, data temporal resolution unification, and data structure definition;

[0029] 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.

[0030] 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.

[0031] 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.

[0032] Furthermore: In S2, the VMD-CEEMDAN adaptive hybrid decomposition method achieves deep denoising and multidimensional feature extraction of power data through various adaptive strategies. The denoising and feature extraction of the VMD-CEEMDAN adaptive hybrid decomposition method includes:

[0033] S21. Based on the energy entropy minimization criterion, adaptively select the optimal mode number K of VMD from the candidate mode number set;

[0034] S22. By using CEEMDAN pre-decomposition, the power sequence is decomposed into multiple intrinsic mode functions (IMFs) and a residual term, thus achieving the separation of trend and fluctuation.

[0035] 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.

[0036] S24. An energy entropy-guided mode number determination method is adopted, adaptively determining the mode number based on the rate of change of signal energy entropy and energy concentration to avoid over-decomposition or under-decomposition. A stopping criterion is set based on the energy entropy increment, energy concentration, and maximum mode number. Low-value noise components are removed and effective signal components are retained through component importance assessment and selective reconstruction (energy proportion, correlation, and kurtosis). Different threshold functions are used for different mode components and different frequency components for adaptive threshold denoising, removing noise while retaining useful signals.

[0037] Furthermore, in S3, the specific process of time series modeling includes:

[0038] S31. Construct a multi-layer GRU deep neural network;

[0039] 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.

[0040] S33. After training, the GRU model outputs preliminary predictions for the future specified step size.

[0041] Further: In S4, the specific steps of the DIRMO packet strategy include:

[0042] S41. Divide the total 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 the step order between different groups and within groups by adding group features; decompose the long-term prediction task into multiple sub-tasks, each group containing a fixed number of steps, and train and predict independently for each group model.

[0043] S42. Construct multi-class features; improve prediction accuracy by extracting features at different time scales;

[0044] S43. Using a data reconstruction method, the multi-step output of each group is flattened into a one-dimensional vector, and features are copied for each prediction step and step identifier features are added, enabling the LightGBM model to handle multi-step prediction tasks in a single-output form. Since the input feature spaces of each group of models are the same and independent of each other, fully parallel training and prediction can be achieved.

[0045] Furthermore: In S6, the optimization process of the adaptive differential evolution algorithm is as follows:

[0046] S61. Initialize the population, with each individual representing a set of hyperparameter configurations;

[0047] S62. Calculate the population diversity and normalize it to obtain the diversity index D;

[0048] S63. Calculate the current iteration progress and adjust the crossover probability CR;

[0049] 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;

[0050] 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;

[0051] S66. Introduce a continuous, unimproved counter to prevent overfitting and improve efficiency.

[0052] Furthermore: In S7, the training and prediction process of the LightGBM model includes:

[0053] 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.

[0054] S72. Using the optimal hyperparameter configuration obtained by the ADE algorithm, train a LightGBM model for each group of the DIRMO strategy.

[0055] 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.

[0056] 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;

[0057] 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.

[0058] The wind power prediction system of the present invention for implementing the wind power prediction method based on DIRMO and differentiated objective function 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.

[0059] The data acquisition module is used to collect historical power data, wind speed data, and weather forecast data.

[0060] 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.

[0061] 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 dependency of the power sequence, and generate preliminary prediction results for the future.

[0062] 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.

[0063] 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.

[0064] 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.

[0065] 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.

[0066] Furthermore, the preliminary prediction module includes a GRU time series prediction module, which is used to construct a multi-layer GRU deep neural network model and optimize the multi-layer GRU deep neural network model by using mean squared error as the loss function during training.

[0067] Furthermore: the refined prediction module includes a DIRMO grouping module, a multi-scale feature construction module, and a data reconstruction module;

[0068] 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.

[0069] 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.

[0070] The data reconstruction module is used to transform multi-step prediction targets into single-step predictions.

[0071] The beneficial effects of this invention are:

[0072] This invention addresses the dual-specification assessment requirements for renewable energy power prediction. It employs a cascaded progressive DIRMO (Direct Multi-Output) prediction strategy combined with VMD-CEEMDAN (Variational Mode Decomposition-Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) adaptive hybrid decomposition technology to construct a complete power prediction methodology. This system overcomes limitations in existing prediction techniques, such as severe error accumulation, inconsistent assessment indicators, high noise interference, and difficulty in hyperparameter tuning. The benefits of this invention include:

[0073] 1. Innovative denoising and feature extraction using VMD-CEEMDAN adaptive hybrid decomposition;

[0074] This invention innovatively proposes a VMD-CEEMDAN adaptive hybrid decomposition method, overcoming the limitations of traditional decomposition methods. It achieves adaptive separation of trend and fluctuation through CEEMDAN pre-decomposition, adaptively selecting the optimal number of modes K from multiple candidate modes based on the signal energy entropy minimization criterion, and dynamically adjusting the decomposed mode number using an adaptive mode number determination strategy guided by signal energy entropy. A hierarchical adaptive VMD decomposition strategy is employed for different frequency components, combined with an improved adaptive threshold denoising method. Low-value noise components are removed through component importance assessment and selective reconstruction, providing high-quality input features for subsequent prediction models and significantly improving the model's ability to capture complex power fluctuation patterns and its robustness to outlier data.

[0075] 2. The complementary advantages and error control of cascaded progressive DIRMO strategies;

[0076] This invention achieves complementary advantages between the GRU neural network and the gradient boosting tree of the LightGBM model through a cascaded progressive DIRMO prediction strategy, forming a collaborative prediction framework of preliminary prediction and refined prediction. The output of the preliminary prediction stage serves as the input feature for the refined prediction stage, realizing close linkage between the two stages. Each group of models has the same input feature space and is independent of the others, allowing for fully parallel training and prediction, effectively avoiding the error accumulation problem of recursive strategies and significantly improving the stability of long-term predictions. Compared to direct strategies, it significantly reduces the number of models and training time, and compared to multi-input multi-output strategies, its optimization objective is more consistent with the prediction objective. This prediction strategy fully leverages the respective advantages of deep learning and ensemble learning, achieving comprehensive optimization of prediction accuracy, computational efficiency, and model stability.

[0077] 3. Intelligent automatic optimization eliminates the need for manual parameter tuning;

[0078] This invention employs an adaptive differential evolution algorithm to achieve intelligent automatic optimization of hyperparameters in the LightGBM model, avoiding the problem of the model failing to reach its optimal state due to manually selected empirical parameters. The ADE algorithm effectively balances global and local search capabilities through three adaptive strategies. Premature convergence is effectively avoided through a population diversity maintenance mechanism, and regionally differentiated ACC is used as the fitness function to ensure that the optimization objective is highly consistent with practical application requirements. This method significantly improves model performance while reducing reliance on professional parameter tuning experience, thus enhancing the method's generalizability and practicality.

[0079] 4. Broad business adaptability and application value;

[0080] This invention addresses the practical needs of renewable energy power forecasting, providing multi-scenario adaptable business analysis functions with high business scalability and flexibility. The method supports various forecasting scenarios, including ultra-short-term, short-term, and medium-term forecasts, meeting diverse business requirements such as power market trading, grid dispatching, and operation and maintenance optimization. Through regionally differentiated objective function design, it can adapt to different regional assessment standards. Adaptive multi-scale decomposition and intelligent hyperparameter optimization can adapt to power fluctuation characteristics under different wind farms, seasons, and meteorological conditions. The method provides high-precision power forecasting results for renewable energy power plants, supporting their participation in power market trading and reducing assessment costs. Simultaneously, it provides reliable forecasting information for grid dispatching, supporting the safe and stable operation of the grid and the absorption of renewable energy, demonstrating broad application prospects and significant socio-economic value. Attached Figure Description

[0081] Figure 1 This is a flowchart of the prediction method;

[0082] Figure 2 This is a schematic diagram of the cascaded progressive prediction model structure;

[0083] Figure 3 This is a flowchart of the adaptive differential evolution algorithm;

[0084] Figure 4 This is a comparison curve of the 24-hour (short-term) forecast results for region B; where the horizontal axis is the sample number I, the vertical axis is the wind power Y (unit: MW), the blue line represents the actual value, and the red line represents the forecast value;

[0085] Figure 5 This is a scatter plot comparing the 24-hour (short-term) prediction sample accuracy of regions A and B. The left plot shows the accuracy results under the evaluation indicators for region A, and the right plot shows the accuracy results under the evaluation indicators for region B. The horizontal axis represents the sample number I, and the vertical axis represents the sample prediction accuracy P. Detailed Implementation

[0086] The following are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. The embodiments described below are only for explaining the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention should be determined by the scope of the claims. The embodiments of the present invention are described in detail below. In order to facilitate the description of the present invention and simplify the description, the technical terms used in the specification of the present invention should be interpreted broadly, including but not limited to conventional alternatives not mentioned in this application, and including both direct and indirect implementation methods.

[0087] Example 1

[0088] Combination Figures 1-5 This 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.

[0089] 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:

[0090] S1. Data acquisition and preprocessing methods;

[0091] 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:

[0092] S11, Input data acquisition;

[0093] Three types of data are obtained from the data acquisition system of new energy power plants:

[0094] ①Historical power data: Includes timestamps and power values ​​(unit: MW), recording the actual power generation of the power plant.

[0095] ② Wind speed data: Includes timestamps and wind speed values ​​(unit: m / s), recording the measured wind speed at the site.

[0096] 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.

[0097] S12, Data preprocessing;

[0098] ① Data Merging: Perform an inner join on the three data sources based on timestamps to ensure complete power, wind speed, and meteorological data for each time point. The inner join merging includes data merging and handling of missing values.

[0099] ② Unified data time resolution: All data are unified to a 15-minute time resolution, consistent with the time granularity of power prediction.

[0100] ③ Data structure definition: Each row of data after preprocessing represents a complete observation record at a time point, containing 1 timestamp, 1 power value, 1 wind speed value, and 8 meteorological features, for a total of 11 fields.

[0101] The data, after being standardized to a uniform time resolution, are used to form a preprocessed dataset containing timestamps, power values, wind speed values, and multiple meteorological features.

[0102] In other words, S1 is used to perform operations such as merging the original data and unifying the time resolution to obtain a preprocessed dataset.

[0103] S2, VMD-CEEMDAN adaptive hybrid decomposition and feature extraction method;

[0104] This method innovatively combines the adaptability of variational mode decomposition (VMD) with the multi-scale nature of complete set empirical mode decomposition (CEEMDAN), proposing a VMD-CEEMDAN adaptive hybrid decomposition method.

[0105] First, power and wind speed data are decomposed into Intrinsic Mode Functions (IMFs) of different frequencies using CEEMDAN to achieve trend-fluctuation separation. Then, the number of VMD modes K is adaptively determined based on signal energy entropy. Next, a hierarchical adaptive VMD decomposition strategy is applied to different frequency components, with the number of decomposed modes determined adaptively based on signal energy entropy. An improved adaptive threshold denoising method is used to denoise each mode component. Finally, low-value noise components are removed through component importance assessment and selective reconstruction, 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, achieving more refined noise removal and richer feature extraction. The low-value noise components are comprehensively evaluated using three indicators: energy proportion, correlation coefficient with the original signal, and kurtosis, and are determined using a comprehensive importance score, expressed as:

[0106] ;

[0107] in, Let i be the overall importance score of the i-th component. Let i be the energy corresponding to 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.

[0108] The specific implementation steps are as follows:

[0109] S21. Adaptive mode number K selection;

[0110] 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.

[0111] ① Construct a candidate set of modal numbers: including 10 candidate modal numbers from K=3 to K=12.

[0112] 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:

[0113] ;

[0114] in, , , ;

[0115] 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.

[0116] 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. .

[0117] S22, CEEMDAN pre-decomposition and trend-volatility separation;

[0118] Pre-decomposition is performed using Complete Set Empirical Mode Decomposition (CEEMDAN) to separate long-term trends and short-term fluctuations.

[0119] ①CEEMDAN decomposition: The preprocessed power sequence is decomposed into multiple... And one residual component:

[0120] ;

[0121] in, is the power value of the preprocessed power sequence at time t, and n is the number of intrinsic mode functions (IMFs) obtained from the decomposition. It is the amplitude of the i-th intrinsic mode function component at time t. It is the amplitude of the nth-order residual component at time t;

[0122] ② Trend-volatility separation: based on Based on the frequency characteristics, the components are divided into three categories:

[0123] a) Long-term trend component: ;

[0124] b) Medium-term cycle component: ;

[0125] c) Short-term volatility components: ;

[0126] in, It is a long-term trend component. It is a medium-term cycle component. It is a short-term fluctuation component;

[0127] ③ Component characteristic analysis: Calculate the Hurst exponent H for the long-term trend component T(t), medium-term cycle component C(t), and short-term fluctuation component S(t) obtained from the power sequence decomposition above, and determine the persistence and predictability of each component:

[0128] ;

[0129] Where R is the range of the current calculated component, S is the standard deviation of the current calculated component, and L is the number of data points for the current calculated component. This indicates that the sequence has persistence. The Hurst exponent, also known as the Hurst index or Hurst-Williams index, is a statistical indicator used to analyze time series data, indicating that the series exhibits anti-persistence (mean regression).

[0130] S23, Hierarchical adaptive VMD decomposition;

[0131] Different VMD decomposition strategies were used for the different frequency components obtained from CEEMDAN decomposition.

[0132] The following parameters are defined to account for the characteristics of different frequency components:

[0133] a) The threshold value for stopping the energy entropy increment of the long-term trend component is 0.001 to 0.01. When the energy entropy increment is less than this value, it means that continuing the decomposition will not significantly improve the trend extraction.

[0134] b) The threshold value for stopping the energy entropy increment of short-term fluctuation components is 0.005~0.02.

[0135] 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.

[0136] 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.

[0137] e) K_max is the maximum allowed number of modes, with a value range of 8 to 15;

[0138] ① Long-term trend component T(t):

[0139] 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;

[0140] b) Noise reduction strategy: Conservative noise reduction to preserve complete trend information;

[0141] c) Penalty factor selection: Larger penalty factor ( ), to ensure smooth decomposition;

[0142] ②Medium-term periodic component C(t):

[0143] 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

[0144] b) Noise reduction strategy: Moderate noise reduction while preserving periodic characteristics;

[0145] c) Selection of penalty factor: Medium penalty factor To balance the accuracy and smoothness of the decomposition process;

[0146] ③ Short-term fluctuation component S(t):

[0147] 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;

[0148] 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:

[0149] ;

[0150] 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;

[0151] c) Penalty factor selection: smaller penalty factor Improve decomposition accuracy;

[0152] S24. Adaptive determination of the number of modes guided by signal energy entropy;

[0153] This step describes a general framework for determining the number of modes for each component in S23. The number of modes is not fixed; the optimal number is automatically determined based on the signal energy distribution. The following general parameters are defined:

[0154] a) The general energy entropy increment stopping threshold is determined based on the component type. value;

[0155] b) The general energy concentration stopping threshold is determined based on the component type. value;

[0156] c) The energy concentration when the number of modes is K;

[0157] d) E(K) is the energy entropy increment when the number of modes is K;

[0158] ① Stepwise decompose and calculate the signal energy entropy:

[0159] For the mode number K, perform VMD decomposition and calculate the signal energy entropy. And calculate energy concentration. ,in Calculate the energy entropy increment for the energy percentage of the i-th modal component. ,in The signal energy entropy is K-1 when the number of modes is K-1.

[0160] ② Stopping Criterion: Decomposition stops when any of the following conditions are met:

[0161] a) The signal energy entropy increment is less than the threshold: ;

[0162] b) Energy concentration exceeds the threshold: ;

[0163] c) Reaching the maximum number of modes: ;

[0164] ③ Optimal mode number selection: Select the mode number that minimizes the signal energy entropy and maximizes the energy concentration as the optimal mode number.

[0165] In other words, by using S2, the preprocessed dataset is decomposed into VMD-CEEMDAN to obtain multi-scale features.

[0166] S3. Preliminary prediction stage: GRU time series prediction method;

[0167] 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:

[0168] S31. Construct a multi-layer GRU deep neural network, including:

[0169] ① Input layer: Receives historical time window data, with input features being normalized power and wind speed;

[0170] 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;

[0171] ③ Dropout layer: Add a Dropout layer between GRU layers to prevent overfitting and improve the model's generalization ability;

[0172] ④ Output layer: Fully connected layer, outputs preliminary prediction results for the next several steps;

[0173] 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.

[0174] 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.

[0175] S4. Refined Forecasting Stage: DIRMO Group Forecasting Method;

[0176] 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:

[0177] S41, DIRMO grouping strategy design;

[0178] 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 that group. The input features of each group model are the same, but the DIRMO group features are used to distinguish between different groups and the asynchronous behavior within groups, so that the model can learn the characteristics of different prediction step sizes.

[0179] S42, Feature construction method;

[0180] By constructing multiple feature classes and fully mining the value of data, and extracting features at different time scales, the prediction model can simultaneously learn and understand the multi-level patterns of power changes, thereby improving its ability to model complex fluctuation patterns and its prediction accuracy. This includes:

[0181] ①Historical statistical characteristics: Extract statistical characteristics (mean, standard deviation, maximum value, minimum value, rate of change, correlation coefficient, etc.) from historical power and wind speed data to reflect the overall distribution and trend of historical data;

[0182] ② Multi-scale features: Time-frequency domain features extracted from power and wind speed data using the VMD-CEEMDAN adaptive hybrid decomposition method, including time-domain statistical features (mean, standard deviation, skewness, kurtosis, etc.), frequency-domain features (dominant frequency, spectral centroid, spectral bandwidth, etc.), and inter-component correlation features.

[0183] ③ Meteorological characteristics: Extract meteorological forecast data (air pressure, temperature, precipitation, dew point, wind direction, radiation, etc.) at the forecast time to reflect future weather conditions;

[0184] ④ Time characteristics: Extract the time attributes (hour, weekday, month, season, whether it is a weekday, etc.) at the prediction time to capture the periodic patterns of power data;

[0185] ⑤ Interaction Features: Construct interaction features between key variables (such as power × air pressure, power × temperature, etc.) to capture nonlinear relationships between variables;

[0186] ⑥ DIRMO group features: Construct group information features (group number, position within the group, time offset, etc.) to enable the model to distinguish between different groups and asynchronous behavior within the same group;

[0187] ⑦ GRU Enhanced Features (Preliminary Prediction Result Features): The GRU output results in the preliminary prediction stage are used as input features in the refined prediction stage, including GRU predicted values, statistical features of predicted values, trend coefficients, difference features, etc., to realize information transfer and fusion between the two stages.

[0188] S43. Use data reconstruction method to achieve multiple outputs;

[0189] The LightGBM model natively only supports single output, while the DIRMO strategy requires each model to directly predict multiple outputs for that group. Therefore, a data reconstruction method is used to transform the multi-output problem into a single-output problem, reconstructing each training sample into multiple sub-samples, with each sub-sample corresponding to a prediction step within the group. This resolves the contradiction between the standard LightGBM model's requirement of single output and multi-step output.

[0190] ① Reconstruction principle: Flatten the multi-step prediction target of each group into a one-dimensional vector, copy the feature for each prediction step, and add step identification features so that the model can distinguish the asynchronous state within the group;

[0191] ② Training process: Each original sample is expanded into S training samples, and each training sample corresponds to one prediction step within the group. The number of samples is expanded to S times the original number.

[0192] ③ Prediction process: For the test sample, different step identifier feature values ​​are set to generate S prediction samples. The model is used to predict each sample to obtain the S-step prediction result for the group.

[0193] Thus, by using data reconstruction, the multi-output problem is transformed into a single-output problem, enabling the LightGBM model to meet the multi-output requirements of the DIRMO strategy.

[0194] S5, Regional Differentiation Objective Function;

[0195] This method designs a region-specific custom objective function for the different assessment criteria in regions A and B, ensuring that the training objective of the LightGBM model is completely consistent with the actual assessment metrics, thereby directly optimizing the assessment accuracy (ACC). The region-specific objective function is transformed into a custom objective function for the LightGBM model, and the corresponding first and second gradients (Hessian matrix) are derived, allowing the model training process to directly optimize for the assessment metrics.

[0196] Region A refers to areas where: the annual average wind speed is greater than 6 m / s and the wind speed fluctuation is relatively stable; or the daily standard deviation of wind speed is less than 2 m / s and the wind direction change rate is less than 15 degrees / hour; or the annual effective power generation hours are greater than 2000 hours and the correlation coefficient between the wind power curve and the theoretical power curve is higher than 0.85.

[0197] Region B refers to an area with complex wind resource characteristics, specifically manifested as follows:

[0198] 1. Dramatic wind speed fluctuations: The daily standard deviation of wind speed is greater than 3 m / s, and the gust coefficient (the ratio of maximum wind speed to average wind speed) exceeds 1.8;

[0199] 2. Frequent extreme weather events: The average number of days affected by extreme weather such as typhoons and monsoons exceeds 30 days per year, or the frequency of sudden wind speed events (wind speed hourly change rate exceeds 5 m / s) exceeds 50 times per year;

[0200] 3. High proportion of low or high wind periods: The proportion of downtime caused by wind speeds lower than the cut-in wind speed or higher than the cut-out wind speed exceeds 25% of the total annual duration.

[0201] The specific implementation steps are as follows:

[0202] ① For the evaluation indicators of region A, the weighted RMSE objective function is adopted, with the weights being the proportion of each sample's absolute error to the total absolute error; the loss function is expressed as:

[0203] ;

[0204] in, It is the loss function for region A. It is the true power value of the i-th sample. It is the predicted power value of the i-th sample. It is the weight of the i-th sample.

[0205] Weight The weights are dynamically calculated based on the current predicted values ​​during each iteration. As the model training progresses, the weight distribution is continuously adjusted. The calculation formula is as follows:

[0206] ;

[0207] First derivative:

[0208] ;

[0209] Hessian calculation:

[0210] ;

[0211] ② For the evaluation indicators in region B, a normalized MAE objective function is adopted, with the normalized denominator being the larger of the actual power and the installed capacity as a certain ratio. The loss function is expressed as:

[0212] ;

[0213] in, This is the loss function for region B, where cap is the installed capacity of the wind farm in the current month.

[0214] First derivative:

[0215] ;

[0216] Hessian calculation:

[0217] ;

[0218] S6. Adaptive Differential Evolution Algorithm for Hyperparameter Optimization;

[0219] This method uses the Adaptive Differential Evolution (ADE) algorithm to automatically optimize the hyperparameter configuration of the LightGBM model, with the regional differentiation index on the validation set as the optimization objective.

[0220] S61. Initialization and Diversity Calculation:

[0221] Initialize the population (NP individuals), with each individual representing a set of hyperparameter configurations. At the beginning of each iteration, calculate the population diversity and normalize it to obtain the diversity index D.

[0222] S62. Diversity Calculation and Adaptive Adjustment of Scaling Factor F:

[0223] First, population diversity is calculated and normalized to obtain the diversity index D. Then, the scaling factor F is adjusted based on the relationship between D and a preset threshold.

[0224] ① If D is greater than the threshold (indicating high population diversity), then the scaling factor is set to the minimum value (F=Fmin), focusing on local development;

[0225] ② If D is less than or equal to the threshold (indicating low population diversity), the scaling factor is set to the maximum value (F=Fmax) to enhance the mutation perturbation and escape the local optimum.

[0226] S63. Adjustment of crossover probability (CR) based on iteration progress:

[0227] Calculate the current iteration progress (iter / itermax) and adjust the crossover probability CR using a linear decreasing strategy. Expressed as: This strategy maintains a high CR (convergence ratio) in the early stages of iteration to allow for sufficient exploration, and then reduces the CR in the later stages to promote convergence.

[0228] S64. Mutation and Crossover Operations:

[0229] ① Strategy selection: Calculate the probability of roulette based on historical success rates, and adaptively select mutation strategies (such as rand / 1, best / 1, current-to-best / 1).

[0230] ② Generate vectors: Generate mutation vectors according to the selected strategy, and perform binomial crossover based on the CR calculated in step S63 to generate experimental vectors.

[0231] S65. Fitness Assessment and Population Renewal:

[0232] The LightGBM model is trained using the hyperparameters corresponding to the trial vectors, and evaluated on the validation set using a region-differentiated objective function designed with S5. The fitness of the trial vectors and the objective vectors is compared, the best individual is retained, and the historical success rate of the selected strategy is updated.

[0233] S66. Early Stop Mechanism and Termination Judgment:

[0234] Introducing a continuous, unimproved counter n to prevent overfitting and improve efficiency:

[0235] ① If the current iteration shows improvement (i.e., a better individual has been found), reset the counter. ;

[0236] 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), then the counter is incremented by 1. );

[0237] ③ Termination condition: When the maximum number of iterations (iter=itermax) or the number of consecutive no improvements reaches the preset threshold (n=N), the algorithm terminates and outputs the optimal hyperparameter configuration; otherwise, it jumps back to step S61 to continue the loop.

[0238] S7. Model training and prediction methods;

[0239] 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. The models can be trained and predicted in parallel, making full use of computational resources and improving efficiency.

[0240] S71. Hyperparameter selection and optimization;

[0241] In the DIRMO prediction strategy, the required number of prediction steps are grouped. For the LightGBM model in each group, 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 iteratively find the optimal hyperparameter configuration with the regional differentiation index on the validation set as the optimization target.

[0242] S72, Model Training Configuration;

[0243] The optimal hyperparameter configuration obtained through ADE algorithm is used to train a LightGBM model for each group of the DIRMO strategy. The training configuration for each model includes:

[0244] ① Basic Algorithm: Gradient Boosting Decision Tree (GBDT) algorithm is adopted;

[0245] ② Objective function: Use the regional differentiation custom objective function described in S5;

[0246] ③ Hyperparameter configuration: Optimal hyperparameter values ​​obtained using the ADE algorithm described in S6;

[0247] ④ Early stopping strategy: An early stopping mechanism is used to prevent overfitting. Training is stopped early when the validation set loss has not decreased for several consecutive rounds.

[0248] Model prediction process;

[0249] Preliminary predictions for S73 and GRU;

[0250] Coarse predictions are made using a GRU model trained in S3. Inputting historical normalized power and wind speed data, the GRU model outputs preliminary predictions for future periods, which serve as GRU-enhanced features for the LightGBM model.

[0251] S74, Feature Construction;

[0252] Following the feature construction method described in S4, multiple types of features are constructed for each prediction step. These features include: historical statistical features, multi-scale features, meteorological features, temporal features, interaction features, DIRMO group features, and GRU enhancement features. The multi-scale features are derived from the VMD-CEEMDAN hybrid decomposition described in S2, and the GRU enhancement features are calculated based on the preliminary GRU prediction results obtained in S1.

[0253] S75, Grouped Prediction;

[0254] For each group, features of each prediction step are extracted, and a data reconstruction method is used to set a corresponding in-group step number identifier for each prediction step, generating multiple prediction samples. Then, the LightGBM model for that group is loaded, and predictions are made on these prediction samples separately to obtain multiple prediction values. Finally, these values ​​are sorted and merged to form the final prediction result.

[0255] like Figure 4 As shown in the comparison curves of the 24-hour (short-term) forecast results for region B, the method significantly improves the forecast accuracy in region B, especially performing exceptionally well under conditions of drastic wind speed fluctuations and frequent extreme weather events. By introducing a regionally differentiated objective function, the model can more accurately capture the power variation patterns under complex wind resource characteristics. Furthermore, the adaptive differential evolution algorithm strategy for optimizing hyperparameters further enhances the model's generalization ability, enabling it to maintain stable forecast performance under different meteorological conditions. Experimental results show that compared to traditional single-objective function forecasting methods, the average absolute error of this method in region B is reduced by more than 10%. This fully verifies the effectiveness and practicality of combining the regionally differentiated objective function with the ADE algorithm.

[0256] like Figure 5 As shown in the comparison of the scatter plots of the predicted sample accuracy in regions A and B, the different performance characteristics of the prediction accuracy under the two evaluation indicators can be clearly observed. For region A, the deviation between the predicted and actual values ​​is small, reflecting high prediction accuracy and stability. This is due to the effective optimization of the weighted RMSE objective function for the evaluation indicators in region A, enabling the model to better fit the stable power change trend. The accuracy of each sample in region B is relatively low, which is due to the amplification of the prediction error of the evaluation indicators in region B during low-power periods. However, compared with the traditional method using a single objective function, its overall accuracy has been significantly improved.

[0257] Further analysis shows that in the experiment using the differentiated objective function in region A, the model captures the long-term trend component T(t) and the medium-term cycle component C(t) with particular accuracy, verifying the advantages of the VMD-CEEMDAN hybrid decomposition strategy in handling multi-scale features. When using the differentiated objective function in region B, the prediction performance of the short-term fluctuation component S(t) is most significantly improved, indicating that the combination of GRU-enhanced features and data reconstruction effectively alleviates the problem of a single model being unable to cope with complex scenarios.

[0258] Furthermore, with the hyperparameter configuration optimized by the ADE algorithm, the models in both regions exhibited stronger robustness and generalization ability. Experimental results show that the models maintained high prediction consistency regardless of whether the wind speed was high or low, fully demonstrating the potential of the adaptive differential evolution algorithm in improving model performance. These achievements provide an innovative and efficient solution for wind power prediction, and have significant practical application value.

Claims

1. A wind power prediction method based on DIRMO and a differentiated objective function, characterized in that, Comprising the following steps: S1, collecting historical power data, wind speed data and weather forecast data, and preprocessing; S2, adopting VMD-CEEMDAN adaptive hybrid decomposition method to denoise, normalize and extract features from the preprocessed power and wind speed data, and obtain multi-scale features; S3, preliminary prediction stage: Adopting GRU to model the time series of normalized power and wind speed data, capture the time trend and long-term dependence of power sequence, and generate the preliminary prediction result of 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 differentiation target function; According to different evaluation standards in different regions, design a regional differentiation custom target loss function as the training target of the LightGBM model, and derive its first-order gradient and second-order Hessian matrix; make the training target of the LightGBM model consistent with the actual evaluation index; S6, automatic optimization of hyperparameters: Adopting adaptive differential evolution algorithm to automatically optimize the hyperparameters of LightGBM model, and the optimization target is the regional differentiation 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 to make the final power prediction.

2. The method of wind power prediction 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 is to merge the three data sources by inner join according to the timestamp, ensuring that there are complete power, wind speed and weather data at each time point; The data time resolution unification is to unify all data according to the preset time resolution, ensuring that all data are consistent with the time granularity of power prediction; The data structure definition forms a preprocessing data set containing timestamp, power value, wind speed value and multiple weather features from the data with unified time resolution.

3. The method of wind power prediction 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, 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 by CEEMDAN, realizing the separation of trend and fluctuation; S23, adopting hierarchical VMD decomposition strategy for different frequency components, including long-term trend component, medium-term periodic component and short-term fluctuation component, and setting different mode numbers, penalty factors and denoising strategies; S24, adopting energy entropy guided mode number determination method, setting stop criterion according to 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.

4. The method of claim 1, wherein the method is based on DIRMO and a differentiated objective function. In S3, the specific process of time series modeling includes: S31, constructing a multi-layer GRU deep neural network; S32, using mean square error as loss function to optimize prediction accuracy, and using early stopping strategy on validation set to prevent overfitting; S33, after training, the GRU model outputs the preliminary prediction result of the future specified steps.

5. The method of claim 1, wherein, In S4, the specific steps of the DIRMO grouping strategy include: S41, divide the total prediction steps F into G groups, each group containing S steps, satisfying F=G×S; build an independent LightGBM model for each group, the input features of each group model are the same, and distinguish different groups and step sequences in each group by adding group features; S42, build multi-class features; extract features at different time scales to improve prediction accuracy; S43, use data reconstruction method to flatten the multi-step output of each group into a one-dimensional vector, and copy the features for each prediction step and add step identification features, so that the LightGBM model can handle the multi-step prediction task in a single output form.

6. The method of wind power prediction 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: S61, initialize the population, each individual represents a set of hyperparameter configurations; S62, calculate the population diversity and perform normalization processing 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 differentiation 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.

7. The method of wind power prediction 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 prediction steps, set the optimization range of all hyperparameters for each group of LightGBM models, 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, use the optimal hyperparameter configuration obtained by the ADE algorithm to train a LightGBM model for each group of the DIRMO strategy; S73, use the GRU model to perform coarse prediction on the normalized power and wind speed data, and use the preliminary prediction result of the future period output by the GRU model as the GRU enhanced feature of the LightGBM model; S74, build multi-class features for each prediction step, including historical statistical features, multi-scale features, weather features, time features, interaction features, DIRMO group features, and GRU enhanced features; S75, for each group, extract the features of each prediction step of the group, set the corresponding group step number identifier for each prediction step using data reconstruction method, generate a plurality of prediction samples; then load the LightGBM model of the group, predict these prediction samples respectively to obtain a plurality of prediction values, and finally sort and merge to form the final prediction result.

8. A wind power prediction system for implementing the wind power prediction method based on DIRMOs and a differentiated objective function according to any one of claims 1-7, characterized in that, The system 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 configured to acquire historical power data, wind speed data, and weather forecast data. The data preprocessing module is configured to merge and unify the time resolution of the acquired data, and form a preprocessed data set from the preprocessed data. The preliminary prediction module is configured to perform time series modeling on the 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 configured to divide the multi-step prediction task into a plurality of groups, construct a LightGBM model for each group, and use data reconstruction method to convert the multi-output problem into a single-output problem for training and prediction. The regional differentiation module is configured to design a regionally differentiated custom target loss function as the training target of the LightGBM model according to different evaluation criteria for 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 automatic hyperparameter optimization module is configured to automatically optimize the hyperparameters of the LightGBM model, and the optimization target is the regional differentiation 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 multi-scale features.

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

10. The wind power prediction system based on the method of wind power prediction based on DIRMO and differentiated objective function according to claim 8, characterized in that, The refined prediction module comprises 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, and construct an independent LightGBM model for each group, each LightGBM model being responsible for predicting the output of each step of the group. The multi-scale feature construction module is configured to construct multi-scale features, which 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.

Images