Heterogeneous double-stage wind power prediction method and system considering physical constraints
By constructing a heterogeneous two-order wind power prediction model and introducing physical constraints, the problems of insufficient prediction accuracy and rationality of wind power in small sample scenarios in microgrids are solved, achieving higher prediction accuracy and engineering usability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUBEI FUBIAN SPACETIME ENERGY TECH CO LTD
- Filing Date
- 2026-04-27
- Publication Date
- 2026-06-26
AI Technical Summary
In microgrid scenarios, existing wind power prediction methods have insufficient prediction accuracy and lack physical rationality in small sample scenarios, making it difficult to stably serve energy storage scheduling and multi-energy collaborative control.
A heterogeneous two-order prediction model combining XGBoost and LSTM is adopted, and physical constraints are combined. This is achieved by constructing an XGBoost master prediction model and an LSTM residual correction model, along with a 24-step multi-step direct prediction mechanism, and applying power boundary constraints and ramp rate constraints in the output stage.
It improves the prediction accuracy and stability in small sample scenarios, significantly improves the physical rationality of prediction results, eliminates abnormal prediction points that do not conform to the operating rules of wind turbines, and enhances the availability of microgrid control.
Smart Images
Figure CN122292331A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of wind power generation prediction and microgrid energy management technology, and in particular to a heterogeneous two-order wind power generation prediction method and system that takes into account physical constraints. Background Technology
[0002] In microgrid scenarios, wind turbines are typically deployed in small to medium capacity, distributed configurations. These turbines are mostly located at the end of the distribution network, and their power output is highly volatile, highly random, and poorly observable due to factors such as local micro-meteorological conditions, near-surface turbulence, and limited wind measurement capabilities. Furthermore, practical engineering projects often lack large-scale, high-quality samples with long-term stable accumulation, making it easy for end-to-end deep learning models to overfit during training. This results in predictions that are difficult to reliably serve the energy storage scheduling and multi-energy coordinated control of microgrids.
[0003] Existing wind power prediction methods can be broadly categorized into physical model methods, statistical learning methods, and deep learning methods. Physical model methods have the advantage of clear mechanisms, but they are prone to systematic biases in the low to medium wind speed range. Ensemble learning methods, represented by gradient boosting trees, have strong nonlinear fitting capabilities for structured meteorological features and are more suitable for small sample scenarios, but they typically model different prediction times independently, making it difficult to further explore the temporal correlation of residual sequences. Deep learning methods, represented by long short-term memory networks, while possessing temporal modeling capabilities, often face problems such as numerous model parameters, unstable training, and decreased generalization ability under small sample conditions.
[0004] Besides accuracy issues, existing methods also generally suffer from insufficient engineering usability. Some predictions may show negative power, power exceeding the rated power, or power fluctuations between adjacent time points exceeding the dynamic response capability of the wind turbine. Even if such results are acceptable in terms of error metrics, they do not conform to the actual operating mechanism of wind turbines and are difficult to use directly as reliable input for microgrid dispatching. Therefore, existing technologies urgently need improvement to simultaneously adapt to small sample scenarios, improve prediction accuracy, and ensure the physical rationality of the prediction results. Summary of the Invention
[0005] The purpose of this invention is to provide a heterogeneous two-order wind power prediction method and system that considers physical constraints, aiming to solve the problems of insufficient prediction accuracy and lack of physical feasibility of prediction results for small wind turbine generators under small sample and strong fluctuation conditions. By constructing a heterogeneous two-order prediction model of XGBoost and LSTM, and combining it with physical constraints, a wind power prediction method that balances accuracy and rationality is provided.
[0006] To achieve the above objectives, this invention provides a heterogeneous two-order wind power generation prediction method considering physical constraints, comprising the following steps: S1. Obtain raw data, which includes historical operating data and corresponding meteorological data of the wind turbine generator set in the target scenario; historical operating data includes at least historical wind power, and meteorological data includes at least wind speed, wind direction, ambient temperature and air humidity; S2. Preprocess the raw data, including removing outliers and filling in missing values, to obtain the preprocessed data. S3. Construct the input features of the prediction model based on the preprocessed data. The input features include the periodic encoded components of wind speed and wind direction, temperature, humidity and historical power sequence. S4. Standardize the input features and divide them into training, validation and test sets according to time order. Construct supervised learning samples based on a sliding time window. S5. Based on the validation set, the hyperparameters of the XGBoost master prediction model are determined by the Bayesian optimization method, and 24 XGBoost regression models are trained for the next 24 prediction steps to form a 24-step multi-step direct master prediction model. S6. Construct a residual sequence based on the difference between the predicted value and the true value of the XGBoost master prediction model on the training set. Use the values of the preset step size in the residual sequence as historical residual sequences to input into the LSTM residual prediction model for training. The LSTM residual prediction model is used to output the residual prediction value corresponding to the future step size. S7. The predicted values of the XGBoost master prediction model corresponding to each prediction step are superimposed with the residual prediction values of the LSTM residual prediction model to obtain the two-order prediction results. S8. Apply physical constraints to the two-order prediction results to obtain the final prediction results and output them; the physical constraints include power boundary constraints and ramp rate constraints.
[0007] Preferably, the process of removing abnormal samples in step S2 includes: removing samples with negative power values or exceeding the upper limit of the rated power of the wind turbine generator set, and removing samples with wind speeds exceeding a preset reasonable range; missing values are filled using linear interpolation.
[0008] Preferably, in step S3, the periodic coding component of the wind direction specifically involves converting the wind direction angle into a sine component and a cosine component.
[0009] Preferably, in step S4, the training set, validation set, and test set are divided according to time order as follows: the first 60% of the data is used as the training set, the middle 20% of the data is used as the validation set, and the last 20% of the data is used as the test set; the sliding time window has a window length of 24 consecutive time steps, which is used to predict the wind power in the next 24 time steps.
[0010] Preferably, in step S5, the Bayesian optimization takes the average error of 24 prediction steps as the optimization objective, and the optimized hyperparameters include at least one of the following: number of base learners, tree depth, learning rate, row sampling ratio, and column sampling ratio.
[0011] Preferably, in step S6, the preset step size is 24 steps; the LSTM residual prediction model adopts a single-layer LSTM structure with a hidden layer dimension of 32; the Adam optimizer, Huber loss function, and early stopping mechanism are used during training.
[0012] Preferably, the power boundary constraint in step S8 is specifically: limiting the wind power at any predicted time to between zero and the rated power of the wind turbine generator set; The specific method for limiting the ramp rate is as follows: the maximum allowable power change range is determined based on the statistical results of the historical power change absolute value sequence, and the predicted power value at the current moment is limited based on the historical measured power at the previous moment or the predicted power after constraint correction. The statistical results use the 95th percentile of the absolute value series of historical power changes as the maximum permissible power change range.
[0013] This invention also provides a heterogeneous two-order wind power generation prediction system considering physical constraints, comprising: The data acquisition module is used to acquire raw data from the wind turbine generator set in the target scenario. The data preprocessing module is used to preprocess the raw data by removing outliers and filling in missing values. The feature engineering module is used to construct input features based on the preprocessed data; The hyperparameter optimization module is used to determine the hyperparameters of the XGBoost master prediction module based on the validation set using Bayesian optimization. The XGBoost master prediction module is used to build a 24-step multi-step direct master prediction model and output the master prediction value. The LSTM residual correction module is used to construct the residual sequence and train it to obtain the residual prediction value; The two-order fusion module is used to superimpose the master forecast value and the residual forecast value to obtain the two-order forecast result; The physical constraint correction module is used to apply power boundary constraints and ramp rate constraints to the two-order prediction results and output the final prediction results. The results output module is used to output the final prediction results to the microgrid energy management module.
[0014] Therefore, the heterogeneous two-order wind power generation prediction method and system that considers physical constraints, as described above, have the following beneficial effects: (1) For small wind turbines with small sample scenarios, XGBoost is used as the first-order master prediction model, which can improve the fitting ability of structured meteorological features and the generalization ability of the model with low training complexity.
[0015] (2) By constructing a heterogeneous two-stage prediction structure of “XGBoost master prediction + LSTM residual correction”, the ability to model structured features and the ability to model temporal residuals are combined, and combined with the 24-step multi-step direct prediction mechanism, which helps to improve both prediction accuracy and long step length prediction stability.
[0016] (3) By introducing power boundary constraints and ramp rate constraints in the output stage, the physical rationality and engineering usability of the prediction results can be significantly improved with minimal accuracy cost. According to the results of the embodiment of the present invention, the MAE of the test set after constraint is 3.908kW, eliminating 226 physically unreasonable prediction points with less than 0.5% accuracy cost, and reducing the maximum power change rate by 57.6%.
[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the overall framework of the heterogeneous two-order wind power generation prediction method considering physical constraints of this invention. Figure 2 This invention presents a scatter plot of wind turbine wind speed and measured power, and a corresponding deviation box plot, in an embodiment of a heterogeneous two-order wind power generation prediction method and system considering physical constraints. Figure 3 This is a convergence curve of the Bayesian hyperparameter optimization process in the embodiment of the heterogeneous two-order wind power generation prediction method and system considering physical constraints of the present invention. Figure 4 This is a verification diagram of the power change distribution and the effect of the ramp rate constraint on the heterogeneous two-order wind power generation power prediction method and system embodiment that considers physical constraints of the present invention. Figure 5 This is a comparison chart of wind power prediction results before and after physical constraint correction during typical fluctuation periods in the implementation of the heterogeneous two-order wind power prediction method and system considering physical constraints of the present invention. Figure 6 This is a diagram showing the module connection relationship of the heterogeneous two-stage wind power generation prediction system that takes into account physical constraints, as described in this invention. Detailed Implementation
[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0021] Example 1 This embodiment provides a heterogeneous two-order wind power generation prediction method that considers physical constraints, such as... Figure 1 As shown. In this embodiment, "heterogeneous two-stage" refers to the use of different types of machine learning models for the first-stage master prediction model and the second-stage residual correction model. Specifically, the first stage uses an XGBoost ensemble learning model, and the second stage uses an LSTM deep learning model. The method specifically includes the following steps: S1. Obtain raw data: Obtain historical operating data and corresponding meteorological data of the wind turbine generator set in the target scenario. The historical operating data should include at least historical wind power output, and the meteorological data should include at least wind speed, wind direction, ambient temperature, and air humidity.
[0022] S2. Data Preprocessing: The raw data is preprocessed to remove outliers and fill in missing values, resulting in preprocessed data. Specifically, samples with negative power values or exceeding the rated power limit of the wind turbine generator are removed, as are outlier samples with wind speeds exceeding a preset reasonable range (e.g., greater than 25 m / s). Missing data is filled using linear interpolation between adjacent time points.
[0023] S3. Constructing Input Features: Based on the preprocessed data, construct the input features for the prediction model. Input features include periodic encoded components of wind speed and wind direction, temperature, humidity, and historical power sequences. The periodic encoded component of wind direction is achieved by converting the wind direction angle into sine and cosine components to eliminate the feature discontinuity problem between 0° and 360°.
[0024] S4. Standardization and Sample Construction: The input features are standardized (e.g., Z-score standardization), and the data is divided into training, validation, and test sets according to time sequence. In this embodiment, the first 60% of the data is used as the training set, the middle 20% as the validation set, and the last 20% as the test set. Supervised learning samples are constructed based on a sliding time window, with a window length of 24 consecutive time steps, to predict wind power output for the next 24 time steps.
[0025] S5. XGBoost Master Prediction Model: Based on the validation set, the hyperparameters of the XGBoost master prediction model are determined using Bayesian optimization. Bayesian optimization aims to optimize the average error over 24 prediction steps. The optimized hyperparameters include at least one of the following: number of base learners (n_estimators), tree depth (max_depth), learning rate (learning_rate), subsample ratio (row sampling ratio), and colsample_bytree ratio (column sampling ratio). Subsequently, 24 XGBoost regression models are trained for each of the next 24 prediction steps, forming a 24-step multi-step direct master prediction model.
[0026] S6. LSTM Residual Correction Model: A residual sequence is constructed based on the difference between the predicted values and the true values of the XGBoost master prediction model on the training set. The values of past preset step sizes (24 steps in this example) in this residual sequence are used as historical residual sequences and input into the LSTM residual prediction model for training. The LSTM residual prediction model adopts a single-layer LSTM structure with a hidden layer dimension of 32. During training, the Adam optimizer, Huber loss function, and early stopping mechanism are used. The trained model is used to output the residual prediction values for future step sizes.
[0027] S7. Fusion of two-order prediction results: The predicted values of the XGBoost master prediction model corresponding to each prediction step are superimposed with the residual prediction values of the LSTM residual prediction model to obtain two-order prediction results.
[0028] S8. Apply Physical Constraints: Apply physical constraints to the above two-order prediction results to obtain and output the final prediction results. Physical constraints include power boundary constraints and ramp rate constraints. Specifically, the power boundary constraint limits the wind power at any prediction time to between zero and the rated power of the wind turbine generator. The ramp rate constraint determines the maximum allowable power change range based on the statistical results of the historical absolute power change sequence, and uses the historical measured power at the previous moment or the predicted power after constraint correction as a benchmark to limit the predicted power value at the current moment; the 95th percentile of the historical absolute power change sequence is used as the maximum allowable power change range.
[0029] To verify the effectiveness of the above technical solution, in one specific embodiment, the target wind turbine is a 50kW small wind turbine in a certain area of Hubei Province. The data covers the entire year from 2022 to 2023, with a time resolution of 1 hour, totaling 17,520 sample points. The raw data includes historical wind power, wind speed, wind direction, ambient temperature, and air humidity.
[0030] Preprocessing is performed according to step S2: samples with negative power values or exceeding 50kW, as well as samples with significantly abnormal wind speeds (e.g., greater than 25m / s), are removed; missing data is filled in using linear interpolation. In this embodiment, the proportion of abnormal samples removed is approximately 0.78%, and the proportion of missing data is less than 1%.
[0031] Feature engineering is performed according to step S3: Sine and cosine encoding of the wind direction angle is used, combined with wind speed, temperature, humidity, and historical power sequences to form input features. For example... Figure 2 As shown, a scatter plot and a box plot of the wind speed and measured power are presented. It can be seen that the data distribution is reasonable and there are no obvious anomalies.
[0032] Standardization and sample splitting were performed according to step S4: the first 60% of the data (approximately 10,512 samples) served as the training set, the middle 20% (approximately 3,504 samples) served as the validation set, and the last 20% (approximately 3,504 samples) served as the test set. A sliding time window was used to predict wind power output for the next 24 hours based on historical data from 24 consecutive hours.
[0033] XGBoost main model training was performed according to step S5: the optimal hyperparameters obtained through Bayesian optimization were: n_estimators=300, max_depth=4, learning_rate=0.024, subsample=0.9, colsample_bytree=0.9. The convergence curve of the optimization process is shown below. Figure 3 As shown.
[0034] Train the LSTM residual correction model according to step S6: use the predicted residual sequence (past 24 steps) of the main model on the training set as input, a single-layer LSTM with a hidden layer dimension of 32, employing the Adam optimizer and Huber loss function, and setting early stopping. After training, output the residual prediction values for the next 24 steps.
[0035] Following step S7, the master prediction value and the residual prediction value are superimposed to obtain the two-order prediction result.
[0036] Apply physical constraints according to step S8: First, limit the power to between 0 and 50 kW; then, determine the maximum allowable power change range (climbing rate constraint threshold) as 12.0 kW based on the 95th percentile of the historical power change absolute value sequence. Gradually limit the predicted sequence starting from the initial time: use the historical measured power of the previous hour as a reference for the first time, and use the corrected power of the previous time as a reference for each subsequent time, truncating any change exceeding 12.0 kW. Figure 4 A verification diagram of the power change distribution and the effect of the climbing rate constraint is given. In the figure, ΔP(t) represents the absolute value of the power difference between adjacent time points, and the red box shows the comparison before and after the constraint. Figure 5 The comparison of wind power prediction results before and after physical constraint correction during typical fluctuation periods is shown. It can be seen that the power curve after constraint is smoother and conforms to the operating law of wind turbines.
[0037] The implementation results show that, compared with baseline models such as LSTM, CNN-LSTM, XGBoost, and homogeneous two-order models, the heterogeneous two-order prediction structure of this invention maintains superior prediction performance in small-sample scenarios. After applying physical constraints, the MAE of the test set is 3.908kW. 226 physically unreasonable prediction points are eliminated at the cost of less than 0.5% accuracy, and the maximum power change rate is reduced from 28.3kW to 12.0kW, a reduction of 57.6%, significantly improving the direct applicability of the prediction output in microgrid control scenarios. The final prediction results can be output to the microgrid energy management module for energy storage scheduling, biomass unit coordinated control, and power purchase plan generation.
[0038] like Figure 6 As shown, the heterogeneous two-order wind power prediction system considering physical constraints corresponding to the above method includes: a data acquisition module, a data preprocessing module, a feature engineering module, a hyperparameter optimization module, an XGBoost master prediction module, an LSTM residual correction module, a two-order fusion module, a physical constraint correction module, and a result output module. Each module works together to execute the above method.
[0039] Therefore, this invention employs the aforementioned heterogeneous two-order wind power prediction method and system that considers physical constraints. By constructing a heterogeneous two-order prediction architecture combining XGBoost and LSTM, it effectively balances structured feature fitting and temporal residual modeling capabilities, improving prediction accuracy and long-step prediction stability while adapting to small-sample scenarios. Furthermore, by introducing power boundary constraints and ramp rate constraints, the physical rationality of the prediction results is significantly improved with minimal accuracy sacrifice, eliminating abnormal prediction points that do not conform to the wind turbine's operating characteristics, and ensuring that output power changes align with the unit's dynamic response capabilities.
[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A heterogeneous two-order wind power generation prediction method considering physical constraints, characterized in that, Includes the following steps: S1. Obtain raw data, which includes historical operating data and corresponding meteorological data of the wind turbine generator set in the target scenario; historical operating data includes at least historical wind power, and meteorological data includes at least wind speed, wind direction, ambient temperature and air humidity; S2. Preprocess the raw data, including removing outliers and filling in missing values, to obtain the preprocessed data. S3. Construct the input features of the prediction model based on the preprocessed data. The input features include the periodic encoded components of wind speed and wind direction, temperature, humidity and historical power sequence. S4. Standardize the input features and divide them into training, validation and test sets according to time order. Construct supervised learning samples based on a sliding time window. S5. Based on the validation set, the hyperparameters of the XGBoost master prediction model are determined by the Bayesian optimization method, and 24 XGBoost regression models are trained for the next 24 prediction steps to form a 24-step multi-step direct master prediction model. S6. Construct a residual sequence based on the difference between the predicted value and the true value of the XGBoost master prediction model on the training set. Use the values of the preset step size in the residual sequence as historical residual sequences to input into the LSTM residual prediction model for training. The LSTM residual prediction model is used to output the residual prediction value corresponding to the future step size. S7. The predicted values of the XGBoost master prediction model corresponding to each prediction step are superimposed with the residual prediction values of the LSTM residual prediction model to obtain the two-order prediction results. S8. Apply physical constraints to the two-order prediction results to obtain the final prediction results and output them; the physical constraints include power boundary constraints and ramp rate constraints.
2. The heterogeneous two-order wind power generation prediction method considering physical constraints according to claim 1, characterized in that, The process of removing abnormal samples in step S2 includes removing samples with negative power values or exceeding the upper limit of the rated power of the wind turbine generator set, as well as removing samples with wind speeds exceeding a preset reasonable range; missing values are filled using linear interpolation.
3. The heterogeneous two-order wind power generation prediction method considering physical constraints according to claim 1, characterized in that, In step S3, the periodic coding component of wind direction is specifically: converting the wind direction angle into sine and cosine components.
4. The heterogeneous two-order wind power generation prediction method considering physical constraints according to claim 1, characterized in that, In step S4, the training set, validation set, and test set are divided according to time order as follows: the first 60% of the data is used as the training set, the middle 20% of the data is used as the validation set, and the last 20% of the data is used as the test set; the sliding time window has a window length of 24 consecutive time steps, which is used to predict the wind power in the next 24 time steps.
5. The heterogeneous two-order wind power generation prediction method considering physical constraints according to claim 1, characterized in that, In step S5, Bayesian optimization takes the average error of 24 prediction steps as the optimization objective, and the optimized hyperparameters include at least one of the following: number of base learners, tree depth, learning rate, row sampling ratio, and column sampling ratio.
6. The heterogeneous two-order wind power generation prediction method considering physical constraints according to claim 1, characterized in that, In step S6, the preset step size was 24 steps; the LSTM residual prediction model adopts a single-layer LSTM structure with a hidden layer dimension of 32; the Adam optimizer, Huber loss function and early stopping mechanism are used during training.
7. The heterogeneous two-order wind power generation prediction method considering physical constraints according to claim 1, characterized in that, The power boundary constraint in step S8 is specifically to limit the wind power at any predicted time to between zero and the rated power of the wind turbine generator set. The specific method for limiting the ramp rate is as follows: the maximum allowable power change range is determined based on the statistical results of the historical power change absolute value sequence, and the predicted power value at the current moment is limited based on the historical measured power at the previous moment or the predicted power after constraint correction. The statistical results use the 95th percentile of the absolute value series of historical power changes as the maximum permissible power change range.
8. A heterogeneous two-order wind power generation prediction system considering physical constraints, used to execute the heterogeneous two-order wind power generation prediction method considering physical constraints as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire raw data from the wind turbine generator set in the target scenario. The data preprocessing module is used to preprocess the raw data by removing outliers and filling in missing values. The feature engineering module is used to construct input features based on the preprocessed data; The hyperparameter optimization module is used to determine the hyperparameters of the XGBoost master prediction module based on the validation set using Bayesian optimization. The XGBoost master prediction module is used to build a 24-step multi-step direct master prediction model and output the master prediction value. The LSTM residual correction module is used to construct the residual sequence and train it to obtain the residual prediction value; The two-order fusion module is used to superimpose the master forecast value and the residual forecast value to obtain the two-order forecast result; The physical constraint correction module is used to apply power boundary constraints and ramp rate constraints to the two-order prediction results and output the final prediction results. The results output module is used to output the final prediction results to the microgrid energy management module.