A wind power short-term power prediction method and system
Patent Information
- Application Number
- CN202610957706.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-30
- Publication Date
- 2026-10-27
AI Technical Summary
风电SCADA数据具有非线性、非平稳、高噪声等固有特性，现有研究普遍缺少系统化的异常识别与缺失值修复机制
[0037]1、本发明融合SCADA数据、NWP预报数据与Open气象再分析数据，突破单一气象数据源的局限性，规避单一预报数据受地形、网格分辨率带来的固有误差，从数据源头丰富特征维度。
Smart Images
Figure CN122992660A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of wind turbine power prediction, and in particular to a method and system for short-term wind power prediction using multi-source NWP deviation correction and dual-model fusion. Background Technology
[0002] Accurate and reliable short-term wind power forecasting has become a key technological foundation for ensuring the consumption of new energy sources and supporting grid dispatching and operation.
[0003] Existing wind power prediction methods are mainly divided into three categories: physical methods, statistical methods, and deep learning methods. Physical methods rely on meteorological mechanisms and prior information about topography and geomorphology for calculation, which are highly adaptable but computationally complex. Statistical methods perform well under stable wind conditions and are suitable for scenarios with limited samples. Deep learning methods excel at uncovering nonlinear time-series patterns and have significant advantages in high-precision prediction. However, existing technologies still face two major bottlenecks in engineering applications: First, the wind speed deviation of NWP (Non-Wave Controller) data is not effectively corrected. Mainstream methods generally rely on single-source NWP data, which is limited by local topography and geomorphology and meteorological grid resolution. There is a systematic deviation between the general NWP forecast wind speed and the actual wind speed at the turbine hub. This deviation is transmitted from upstream to the power prediction stage, directly causing the output prediction to deviate from the actual operating conditions. Existing multi-source fusion research focuses on the horizontal integration of NWP products but neglects in-depth exploration of the mechanism of deviation, resulting in a bottleneck in improving prediction accuracy. Second, a data quality governance system is lacking. Wind power SCADA data has inherent characteristics such as nonlinearity, non-stationarity, and high noise. Existing research generally lacks a systematic mechanism for anomaly identification and missing value repair. Unprocessed data will severely interfere with model feature extraction, amplify prediction errors, and reduce the reliability of results. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method and system for short-term wind power prediction based on multi-source NWP deviation correction and dual-model fusion. For the NWP wind speed deviation problem, an XGBoost-BPNN multi-meteorological factor deviation correction model is constructed to eliminate systematic deviations caused by topography, using the corrected wind speed as an accurate input for power prediction. To address data quality issues, an anomaly cleaning system combining physical constraints and DBSCAN is constructed, and KNN and XGBoost are used differentially to fill missing values, ensuring the quality of input data from the source. Based on this, with the corrected wind speed as the core input, a collaborative prediction system based on an adaptive fusion XGBoost-Chronos hybrid model within the wind speed range is constructed to achieve high-precision short-term power prediction for wind turbines within 24-48 hours, supporting wind farm spot trading and stable grid operation.
[0005] The objective of this invention is achieved through the following technical solution: a method for short-term wind power prediction based on multi-source NWP bias correction and dual-model fusion, comprising the following steps:
[0006] S1. Obtain multi-source data from the wind turbine generator and perform time alignment processing on the multi-source data to obtain the basic input dataset;
[0007] S2. Extract all wind direction-related features from the basic input dataset, convert the wind direction-related features from angle features to numerical features, perform double anomaly cleaning and data imputation on the numerical features, and form a standardized dataset.
[0008] S3. Based on the standardized dataset, construct the XGBoost-BPNN wind speed correction model, and output the corrected wind speed that fits the actual operating conditions of a single wind turbine.
[0009] S4. Based on the Chronos time series model and the XGBoost regression model, construct an adaptive fusion XGBoost-Chronos hybrid model, using the corrected wind speed obtained in step S3 as input, and output the power prediction results of the dual models.
[0010] S5. Divide the wind speed into different intervals, assign corresponding weights to the dual-model prediction results, and obtain the final power prediction result through weighted calculation. At the same time, perform physical limiting on the final power prediction result to ensure that it conforms to the actual operating range of the wind turbine.
[0011] Furthermore, the method includes the following steps:
[0012] S6. Calculate the error index between the final power prediction result and the measured power value of the wind turbine, and verify the prediction effectiveness of the adaptive fusion XGBoost-Chronos hybrid model.
[0013] Furthermore, step S1 includes:
[0014] Acquire SCADA operation data, NWP data, and Open meteorological reanalysis data of the target wind turbine. Resample the data using a preset time granularity, align the time series across all dimensions using timestamps as the sole reference, delete duplicate and empty timestamp records, and unify field names and data units to form a basic input dataset that is time-continuous, formatted, and dimensionally complete.
[0015] Furthermore, the time granularity is set to 15 minutes.
[0016] Furthermore, step S2 includes:
[0017] Extract all wind direction-related features from the basic input dataset, perform sine and cosine trigonometric function transformations on the wind direction-related features to convert them from ring-shaped angle features to continuous numerical features, and delete the original wind direction angle field.
[0018] Based on the operating mechanism of wind turbines, unreasonable data such as non-zero power at low wind speeds and abnormal power at high wind speeds are marked. Outlier noise data is identified by combining the DBSCAN clustering algorithm, and abnormal data is uniformly cleared. For the missing wind speed and power values that appear after cleaning, the KNN model is used to fill in the missing wind speed values, and the XGBoost model is used to fit and fill in the missing power values according to the rated power, forming a standardized dataset without discontinuities.
[0019] Furthermore, step S4 includes:
[0020] An adaptive fusion XGBoost-Chronos hybrid model is constructed based on the Chronos time series model and the XGBoost regression model. The Chronos time series model takes corrected wind speed and pre-acquired historical power data as input and outputs predicted power values including quantiles. The XGBoost regression model takes corrected wind speed, wind speed at multiple altitude levels, and gust intensity as input and learns the nonlinear mapping relationship between wind speed and power. Both the Chronos time series model and the XGBoost regression model use a unified time window to perform predictions and output day-ahead power predictions.
[0021] Furthermore, step S6 includes:
[0022] The predictive effectiveness of the adaptive fusion XGBoost-Chronos hybrid model was verified by calculating the error indices between predicted and measured values. These error indices included mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination. The formulas for each error index are as follows:
[0023] ;
[0024] ;
[0025] ;
[0026] in, This is the measured power value. This is the predicted power value. The mean of the measured power sequence is given. This represents the total number of power samples.
[0027] A wind power short-term power prediction system based on multi-source NWP bias correction and dual-model fusion, used to implement the aforementioned wind power short-term power prediction method based on multi-source NWP bias correction and dual-model fusion, includes:
[0028] The multi-source data reading and fusion module is used to read SCADA operation data, NWP data, and Open meteorological reanalysis data, and completes the time alignment, format unification, and feature standardization of the data, providing a unified basic input dataset;
[0029] The data cleaning and missing value imputation module combines the physical operation mechanism of wind turbine units with the DBSCAN clustering algorithm to perform dual identification and removal of abnormal data. The KNN model and XGBoost model are used to complete the imputation and repair of wind speed and power data respectively, and a standardized dataset is constructed.
[0030] The wind speed correction module uses a standardized dataset as input to build an XGBoost-BPNN wind speed deviation correction model and outputs the corrected wind speed.
[0031] The dual-model fusion prediction module is used to construct an adaptive fusion XGBoost-Chronos hybrid model. It takes the corrected wind speed as input, outputs the dual-model power prediction results, and sets differentiated weights based on different wind speed ranges to perform weighted fusion of the dual-model power prediction results.
[0032] The result verification module is used to calculate the error index between the power prediction result and the measured power value, and to verify the power prediction result.
[0033] The visualization module is used to visually display the power prediction results and the verification of error indicators.
[0034] A non-transitory computer-readable medium storing instructions, characterized in that, when the instructions are executed by a processor, the steps of the wind power short-term power prediction method according to the above-described multi-source NWP deviation correction and dual-model fusion method are performed.
[0035] A computing device includes a processor and a memory for storing processor-executable programs. When the processor executes the program stored in the memory, it implements the above-described method for short-term wind power prediction with multi-source NWP deviation correction and dual-model fusion.
[0036] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0037] 1. This invention integrates SCADA data, NWP forecast data, and Open meteorological reanalysis data, breaking through the limitations of a single meteorological data source and avoiding the inherent errors caused by terrain and grid resolution in single forecast data, thus enriching the feature dimensions from the data source.
[0038] 2. A dual anomaly identification method combining wind turbine operation mechanism rules and DBSCAN density clustering is adopted to accurately identify abnormal samples that deviate from the operation rules and data distribution. At the same time, KNN and XGBoost algorithms are used to fill in the missing wind speed and power data differentially, effectively eliminating noise interference and filling in the time series data breakpoints, significantly improving the overall quality and engineering usability of the modeling dataset.
[0039] 3. To address the discrepancy between the wind speed forecast by the general NWP and the actual local wind conditions of a single wind turbine, the XGBoost-BPNN combined model is introduced to perform wind speed correction. This effectively corrects the forecast system deviation caused by topography and local environment, and outputs a corrected wind speed that closely matches the actual wind conditions at the wind turbine hub. This provides a highly reliable front-end input for subsequent power prediction and suppresses the propagation of prediction errors from the source.
[0040] 4. The Chronos time series model and XGBoost regression model are used for parallel modeling and prediction, giving full play to the dual advantages of Chronos in capturing the evolution of time series and XGBoost in fitting the nonlinear relationship between wind speed and power. At the same time, the model weights are configured differently according to different wind speed ranges to achieve adaptive fusion of multi-model results. It can flexibly adapt to various and complex wind conditions. Compared with a single prediction model, the overall prediction accuracy and robustness under operating conditions are significantly improved. Attached Figure Description
[0041] Figure 1 This is a flowchart of a method for short-term wind power forecasting.
[0042] Figure 2 This is a heat map showing the correlation between wind power and meteorological factors.
[0043] Figure 3 This is a schematic diagram illustrating the results of identifying and filling in missing values for abnormal wind power data.
[0044] Figure 4 This is a schematic diagram of the wind speed correction results.
[0045] Figure 5 A schematic diagram of the dual-model power prediction results for adaptively fusing the XGBoost-Chronos hybrid model. Detailed Implementation
[0046] The present invention will be further described below with reference to specific embodiments.
[0047] Example 1
[0048] See Figure 1As shown, this embodiment presents a method for short-term wind power prediction based on multi-source NWP bias correction and dual-model fusion, applied to the short-term power prediction of actual operating wind turbine units in a wind farm. The method includes the following steps:
[0049] S1. Acquire multi-source data from wind turbine generators and perform time alignment processing on the multi-source data to obtain the basic input dataset. This includes acquiring SCADA operation data of the target wind turbine generators, including time, measured wind speed, and active power; acquiring NWP data, including forecast wind speed, forecast wind direction, forecast temperature, and forecast humidity; and acquiring Open meteorological reanalysis data, including multi-level wind speed and direction at 80m, 120m, and 180m, as well as gust intensity at 10m height. All data are resampled at a uniform 15-minute time granularity, and time-series alignment is completed using timestamps as the basis.
[0050] S2. Extract all wind direction-related features from the basic input dataset, convert the wind direction-related features from angular features to numerical features, perform double anomaly cleaning and data imputation on the numerical features, and form a standardized dataset. The wind direction-related features are shown in Table 1 below:
[0051]
[0052] Table 1. Wind Direction Related Characteristics
[0053] See Figure 2 The image shows a heatmap illustrating the correlation of characteristics related to each wind direction.
[0054] Furthermore, dual anomaly cleaning is performed based on physical mechanisms and DBSCAN:
[0055] Physical mechanism screening: Data that violates the operating characteristics of the wind turbine, such as non-zero power below the cut-in wind speed, significantly low power above the rated wind speed, negative active power, and power exceeding the rated upper limit, are marked as abnormal.
[0056] Clustering screening: The wind speed and power data are standardized, and the DBSCAN algorithm is used to identify low-density points in the wind speed-power scatter distribution that are far from the main distribution area as outliers.
[0057] Both types of abnormal data were uniformly set to null values to complete the anomaly cleaning. The cleaning effect is shown in the left figure of Figure 3.
[0058] Furthermore, missing value imputation based on KNN and XGBoost is performed.
[0059] For missing data items resulting from data cleaning, a differentiated strategy is used to fill them in:
[0060] Wind speed missing data filling: Using multi-source meteorological features as input, a KNN regression model is used to fit and fill in the missing wind speed data;
[0061] Power missing value imputation: In the high wind speed range, the rated power of the unit was directly used as the input, while in other ranges, the effective measured wind speed was used as the input. The XGBoost regression model was used to complete the imputation of missing power values. After imputation, the time series data became continuous and complete. The imputation effect is shown in the right figure of Figure 3.
[0062] S3. Based on the standardized dataset, construct an XGBoost-BPNN wind speed correction model. The XGBoost-BPNN wind speed correction model outputs a corrected wind speed that fits the actual operating conditions of a single wind turbine. The correction effect is shown in [reference]. Figure 4 As shown;
[0063] S4. Based on the Chronos time series model and the XGBoost regression model, construct an adaptive fusion XGBoost-Chronos hybrid model, using the corrected wind speed obtained in step S3 as input, and output the power prediction results of the dual models.
[0064] Chronos time series model prediction: Using historical active power, measured wind speed and corrected wind speed as context input, it completes the day-ahead short-term power prediction and outputs the prediction interval results including the median, 0.1 quantile and 0.9 quantile.
[0065] XGBoost regression model prediction: Select highly correlated features as core inputs, including corrected wind speed, wind speed at 120m height, and gust intensity at 10m height. Construct an XGBoost regression model to learn the nonlinear mapping relationship between wind speed and power. Complete the prediction using the same time window and output the prediction results.
[0066] S5. Divide the wind speed into different intervals, assign corresponding weights to the dual-model prediction results, and obtain the final power prediction result through weighted calculation. At the same time, perform physical limiting on the final power prediction result to ensure that it conforms to the actual operating range of the wind turbine.
[0067] Weighted fusion by wind speed range: The operation range is divided into multiple levels according to the magnitude of the corrected wind speed. In different ranges, the Chronos time series model and the XGBoost regression model are assigned weight ratios. The two sets of prediction results are weighted and fused to obtain the final short-term power prediction value. The fusion prediction effect is shown in Figure 5.
[0068] S6. Calculate the error index between the final power prediction result and the measured power value of the wind turbine, and verify the prediction effectiveness of the adaptive fusion XGBoost-Chronos hybrid model.
[0069] Mean absolute error, root mean square error, and coefficient of determination were selected as quantitative evaluation indicators. The proposed method was compared with two comparative models, XGBoost and TXMixerx, which directly rely on the original NWP data for modeling. The error indicators of each model are shown in Table 2. The corresponding evaluation indicators for the proposed method are MAE=0.1868, RMSE=0.2805, and RMSE=0.2805. =0.8431.
[0070]
[0071] Table 2 Short-term power prediction error indices for each model
[0072] As can be seen from the comparison, the average error and root mean square error of this method are significantly lower, and the coefficient of determination is closer to 1. Compared with the traditional single model, it has higher prediction accuracy and data fitting effect, and can well meet the reliability and accuracy requirements of wind power short-term power prediction in engineering scenarios such as grid optimization scheduling and electricity spot trading.
[0073] Example 2
[0074] This embodiment discloses a wind power short-term power prediction system based on multi-source NWP deviation correction and dual-model fusion, used to implement the wind power short-term power prediction method based on multi-source NWP deviation correction and dual-model fusion described in Embodiment 1, including:
[0075] The multi-source data reading and fusion module is used to read SCADA operation data, NWP data, and Open meteorological reanalysis data, and completes the time alignment, format unification, and feature standardization of the data, providing a unified basic input dataset;
[0076] The data cleaning and missing value imputation module combines the physical operation mechanism of wind turbine units with the DBSCAN clustering algorithm to perform dual identification and removal of abnormal data. The KNN model and XGBoost model are used to complete the imputation and repair of wind speed and power data respectively, and a standardized dataset is constructed.
[0077] The wind speed correction module uses a standardized dataset as input to build an XGBoost-BPNN wind speed bias correction model and outputs a corrected wind speed, eliminating the systematic bias in NWP forecasts caused by topography from the meteorological driving source.
[0078] The dual-model fusion prediction module is used to construct an adaptive fusion XGBoost-Chronos hybrid model. Taking corrected wind speed as input, it outputs dual-model power prediction results. The XGBoost regression model learns the nonlinear mapping relationship between wind speed and power based on corrected wind speed and meteorological characteristics at multiple altitudes; the Chronos time-series model focuses on capturing time-series trends. Both models perform predictions simultaneously to ensure the comprehensiveness and reliability of the prediction results. Differential weights are set based on different wind speed ranges to perform weighted fusion of the dual-model power prediction results.
[0079] The result verification module is used to calculate the error index between the power prediction result and the measured power value, and to verify the power prediction result.
[0080] The visualization module is used to visually display the power prediction results and the verification of error indicators.
[0081] Example 3
[0082] This embodiment discloses a non-transitory computer-readable medium storing instructions that, when executed by a processor, perform the steps of the wind power short-term power prediction method with multi-source NWP deviation correction and dual-model fusion as described in Embodiment 1.
[0083] In this embodiment, the non-transitory computer-readable medium can be a disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), USB flash drive, portable hard drive, etc.
[0084] Example 4
[0085] This embodiment discloses a computing device, including a processor and a memory for storing processor-executable programs. When the processor executes the program stored in the memory, it implements the wind power short-term power prediction method of multi-source NWP deviation correction and dual-model fusion described in Embodiment 1.
[0086] The computing device described in this embodiment may be a desktop computer, laptop computer, smartphone, PDA handheld terminal, tablet computer, programmable logic controller (PLC), or other terminal device with processor functionality.
[0087] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.
Claims
1. A method for short-term wind power prediction using multi-source NWP bias correction and dual-model fusion, characterized in that, Includes the following steps: S1. Obtain multi-source data from the wind turbine generator and perform time alignment processing on the multi-source data to obtain the basic input dataset; S2. Extract all wind direction-related features from the basic input dataset, convert the wind direction-related features from angle features to numerical features, perform double anomaly cleaning and data imputation on the numerical features, and form a standardized dataset. S3. Based on the standardized dataset, construct the XGBoost-BPNN wind speed correction model, and output the corrected wind speed that fits the actual operating conditions of a single wind turbine. S4. Based on the Chronos time series model and the XGBoost regression model, construct an adaptive fusion XGBoost-Chronos hybrid model, using the corrected wind speed obtained in step S3 as input, and output the power prediction results of the dual models. S5. Divide the wind speed into different intervals, assign corresponding weights to the dual-model prediction results, and obtain the final power prediction result through weighted calculation. At the same time, perform physical limiting on the final power prediction result to ensure that it conforms to the actual operating range of the wind turbine.
2. The method for short-term wind power prediction based on multi-source NWP deviation correction and dual-model fusion as described in claim 1, characterized in that, Includes the following steps: S6. Calculate the error index between the final power prediction result and the measured power value of the wind turbine, and verify the prediction effectiveness of the adaptive fusion XGBoost-Chronos hybrid model.
3. The method for short-term wind power prediction based on multi-source NWP bias correction and dual-model fusion as described in claim 1, characterized in that, Step S1 includes: Acquire SCADA operation data, NWP data, and Open meteorological reanalysis data of the target wind turbine. Resample the data using a preset time granularity, align the time series across all dimensions using timestamps as the sole reference, delete duplicate and empty timestamp records, and unify field names and data units to form a basic input dataset that is time-continuous, formatted, and dimensionally complete.
4. The method for short-term wind power prediction based on multi-source NWP deviation correction and dual-model fusion as described in claim 3, characterized in that, The time granularity is set to 15 minutes.
5. The method for short-term wind power prediction based on multi-source NWP bias correction and dual-model fusion according to claim 4, characterized in that, Step S2 includes: Extract all wind direction-related features from the basic input dataset, perform sine and cosine trigonometric function transformations on the wind direction-related features to convert them from ring-shaped angle features to continuous numerical features, and delete the original wind direction angle field. Based on the operating mechanism of wind turbines, unreasonable data such as non-zero power at low wind speeds and abnormal power at high wind speeds are marked. Outlier noise data is identified by combining the DBSCAN clustering algorithm, and abnormal data is uniformly cleared. For the missing wind speed and power values that appear after cleaning, the KNN model is used to fill in the missing wind speed values, and the XGBoost model is used to fit and fill in the missing power values according to the rated power, forming a standardized dataset without discontinuities.
6. The method for short-term wind power prediction based on multi-source NWP bias correction and dual-model fusion according to claim 5, characterized in that, Step S4 includes: An adaptive fusion XGBoost-Chronos hybrid model is constructed based on the Chronos time series model and the XGBoost regression model. The Chronos time series model takes corrected wind speed and pre-acquired historical power data as input and outputs predicted power values including quantiles. The XGBoost regression model takes corrected wind speed, wind speed at multiple altitude levels, and gust intensity as input and learns the nonlinear mapping relationship between wind speed and power. Both the Chronos time series model and the XGBoost regression model use a unified time window to perform predictions and output day-ahead power predictions.
7. The method for short-term wind power prediction based on multi-source NWP bias correction and dual-model fusion according to claim 2, characterized in that, Step S6 includes: The predictive effectiveness of the adaptive fusion XGBoost-Chronos hybrid model was verified by calculating the error indices between predicted and measured values. These error indices included mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination. The formulas for each error index are as follows: ； ； ； in, This is the measured power value. This is the predicted power value. The mean of the measured power sequence is given. This represents the total number of power samples.
8. A short-term wind power prediction system with multi-source NWP bias correction and dual-model fusion, characterized in that, The method for implementing the multi-source NWP deviation correction and dual-model fusion wind power short-term power prediction according to any one of claims 1-7 includes: The multi-source data reading and fusion module is used to read SCADA operation data, NWP data, and Open meteorological reanalysis data, and completes the time alignment, format unification, and feature standardization of the data, providing a unified basic input dataset; The data cleaning and missing value imputation module combines the physical operation mechanism of wind turbine units with the DBSCAN clustering algorithm to perform dual identification and removal of abnormal data. The KNN model and XGBoost model are used to complete the imputation and repair of wind speed and power data respectively, and a standardized dataset is constructed. The wind speed correction module uses a standardized dataset as input to build an XGBoost-BPNN wind speed deviation correction model and outputs the corrected wind speed. The dual-model fusion prediction module is used to construct an adaptive fusion XGBoost-Chronos hybrid model. It takes the corrected wind speed as input, outputs the dual-model power prediction results, and sets differentiated weights based on different wind speed ranges to perform weighted fusion of the dual-model power prediction results. The result verification module is used to calculate the error index between the power prediction result and the measured power value, and to verify the power prediction result. The visualization module is used to visually display the power prediction results and the verification of error indicators.
9. A non-transitory computer-readable medium storing instructions, characterized in that, When the instruction is executed by the processor, the steps of the wind power short-term power prediction method with multi-source NWP deviation correction and dual-model fusion as described in any one of claims 1-7 are performed.
10. A computing device, comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the wind power short-term power prediction method of multi-source NWP deviation correction and dual-model fusion as described in any one of claims 1-7.