Wind power short-term power prediction method based on error feedback and model dynamic optimization

Abstract

The invention discloses a wind power short-term power prediction method based on error feedback and model dynamic optimization, and the method comprises the steps: carrying out the processing of historical data and online data of a wind power cluster, constructing a core feature sample set, employing a similar scene generation model based on Kmeans-DTW, and completing the construction of an online sample set for new input data through sample expansion; providing wind power prediction based on an integrated model, predicting each scene by adopting an initial ratio weight integrated model, screening a potential to-be-optimized integrated model based on a root-mean-square error, providing a comprehensive evaluation index system, and screening according to a certain threshold to obtain a target optimization scene set; and for a target scene, performing integrated model dynamic adjustment based on a Staking algorithm, performing integrated model hyper-parameter optimization by adopting an Optuna framework, determining an optimal integrated model hyper-parameter, and realizing a dynamic optimization integrated model of a typical scene of a future input wind power sample set. According to the method, wind power online prediction is realized, and a corresponding technology is provided for scheduling system model optimization.

Description

Technical Field

[0001] The invention relates to a wind power short-term power forecasting method based on error feedback and model dynamic optimization, and belongs to the field of wind power forecasting. Background Art

[0002] The development and utilization of wind power is an important support for the balance of supply and demand. Improving my country's energy consumption structure and promoting green development and low-carbon transformation are important strategies for high-quality development. The utilization of wind power is an inevitable trend in the development of my country's energy industry. Wind power generation technology is an important component of the new power system, providing power generation sources for the power system. Due to the volatility and randomness of wind energy, the negative temperature effect of photovoltaic power generation and irradiation changes have a huge impact on the balance of the new power system. Nowadays, AI+electric energy is the goal of my country's energy transformation and development, and it is also an important way to ensure the safe balance of the power system and enhance the security of the power grid. Therefore, if AI technology can be combined to perform periodic dynamic adjustment and optimization of the model based on online data, a short-term wind power prediction method based on error feedback and model dynamic optimization can be provided to provide corresponding solutions for power grid dispatching, which is the key to the development of the safety guarantee of the new power system.

[0003] Wind power forecasting currently mainly uses offline training prediction, which cannot utilize newly added operating data in real time and dynamically optimize the model structure. Studying the weekly wind power online data collection, dividing similar feature scenarios, studying the prediction of integrated models, proposing corresponding evaluation indicators for model screening, and completing the online optimization of the model are the keys to improving the accuracy of wind power forecasting. Summary of the invention

[0004] The purpose of the present invention is to provide a wind power short-term power prediction method based on error feedback and model dynamic optimization to solve the problem that the model offline training cannot be updated and optimized in real time. The method of the present invention can not only perform online prediction, but also dynamically optimize and adjust the model through the corresponding system, which can improve the prediction accuracy of wind power, optimize the information of the model in real time, and make the prediction more reliable.

[0005] The present invention adopts the following technical solution:

[0006] The wind power short-term power forecasting method based on error feedback and model dynamic optimization is characterized by the following steps:

[0007] Step 1, constructing an online sample set of wind power clusters: collect historical numerical weather forecast NWP (Numerical Weather Prediction, NWP) data of wind power clusters and online weekly operation data to obtain sample data of wind power clusters, use data processing and feature mining to mine the sample data of wind power clusters, and obtain the core feature sample set of wind power clusters; input it into the wind power similar scene generation model based on Kmeans-DTW (Dynamic Time Warping, DTW) to obtain the sample set of each scene, and propose a wind power online sample set construction method based on the enhanced diffusion model to obtain the online sample set of wind power clusters under each weather scenario;

[0008] Step 2, wind power prediction based on ensemble learning: establish an ensemble model based on LSTM (Long Short-Term Memory, LSTM), Transformer, LightGBM, BPNN (Back Propagation Neural Network, BPNN), TCN and CNN models, input the online sample set of wind power clusters obtained in step 1 into the initial ensemble model, obtain preliminary prediction results for each scenario and cluster, screen potential optimization scenarios based on the root mean square error, and calculate relevant indicators of accuracy and pass rate; assign corresponding weights to each indicator, establish a comprehensive evaluation index, and screen out the target optimization scenario set and normal scenario set according to the corresponding threshold;

[0009] Step 3, dynamic optimization technology of wind power cluster model: based on the target optimization scenario set in step 2 as the feature set input, the integrated model based on the Staking algorithm is dynamically adjusted, and the adjusted integrated model is obtained by cross-validation; based on the Optuna framework, the hyperparameters of the adjusted integrated model are optimized, and the ideal function Trail is used to search for hyperparameters to determine the optimized integrated model. The similar typical scenarios in the new input sample set will dynamically update the initial integrated model, and finally the prediction is performed through the online optimized integrated model to obtain the prediction result.

[0010] Furthermore, the step 1 is specifically performed according to the following steps:

[0011] Step 1.1: Select wind power clusters at the same provincial level and collect numerical weather forecast (NWP) data and online weekly operation data of each wind farm and photovoltaic farm in the cluster. The numerical weather forecast (NWP) includes meteorological factors such as wind speed, wind direction, irradiance, temperature and humidity.

[0012] Step 1.2: Data processing and feature mining are used for the NWP data of the wind power cluster in step 1.1 and the online weekly operation data to construct a core sample set of the wind power cluster; data processing includes data cleaning, abnormal data removal and missing data filling, and normalization is used to smooth the data. Outliers, stray points and power-limiting points are used to delete abnormal data values, and missing data are filled by neighboring and average station filling methods; feature mining uses Pearson coefficients and SHAP values ​​for correlation analysis, and establishes high-dimensional features to construct an optimal core feature sample set;

[0013] Pearson coefficient ρ X，Y The calculation formula is:

[0014]

[0015] In the formula, cov(X,Y) represents the variance, σ represents the standard deviation, X and Y represent two different input features;

[0016] Step 1.3: Divide the core feature sample set of step 1.2 into 96h time scale, obtain sample sets of different weather scenes based on Kmeans-DTW model, use 70% of the sample sets of different weather scenes as the training set of enhanced diffusion model, use 30% of the sample sets of different weather scenes as the test set of enhanced diffusion model, input the training set into enhanced diffusion model for sample expansion, and then input the generated result into the model training result discriminator for judgment. If it is judged as no, continue to adjust the parameters of the model of the training set; if it is judged as yes, obtain the pre-trained model, input the test set into the pre-trained model for result testing, and when the data t-Distributed Stochastic Neighbor Embedding (t-sne) and principal component analysis (PCA) are consistent, output the output result of the pre-trained model, and obtain the online sample set of wind power cluster as the data basis for subsequent online prediction and optimization.

[0017] Furthermore, the step 2 is specifically performed according to the following steps:

[0018] Step 2.1: Set fixed weights for the LSTM, Transformer, LightGBM, BPNN, TCN and CNN models respectively, establish an initial integrated model, input the online sample set of wind power clusters generated in step 1.3 into the initial integrated model, and obtain preliminary prediction results for each scenario and cluster;

[0019] Step 2.2: Compare the difference of the root mean square error for the prediction results of each scenario and cluster, determine the prediction model of the potential scenario to be optimized, calculate the accuracy, pass rate, correlation coefficient between the prediction result and error, accuracy of the high wind speed segment and accuracy of the low wind speed segment under the weather scenario, and count the results of each indicator;

[0020] Step 2.3: Assign corresponding weights to the five indicator results in step 2.2, calculate the comprehensive indicator results, set the threshold to 75% to 85%, and classify the wind farm data below the threshold into the target dynamic optimization scenario set; classify the remaining wind farms into the normal scenario set.

[0021] Furthermore, step 3 is specifically performed according to the following steps:

[0022] Step 3.1: Dynamically adjust the weights of each model in the initial integrated model based on the Stacking algorithm to obtain an adjusted integrated model. The target dynamic optimization scenario set in step 2.3 is input as a feature set into the adjusted integrated model for training. The adjusted integrated model adopts cross-validation for each model. When the average accuracy of the verification reaches more than 85%, the adjusted integrated model is output;

[0023] Step 3.2: Based on the adjusted integrated model obtained in step 3.1, optimize the model hyperparameters based on the Optuna framework, define the search space, create the ideal function Trail, obtain the optimization results of the integrated model hyperparameters, complete the optimization of the integrated model, and obtain the optimized integrated model of the target scenario;

[0024] Step 3.3: For future new input sample sets, the similar typical scene part will dynamically update the initial integrated model of the normal scene set, and finally make predictions through the online optimized integrated model to obtain the final optimized prediction results.

[0025] Compared with the prior art, the present invention has the following beneficial effects: the present invention provides an online prediction method for short-term wind power based on comprehensive evaluation indicators, which can perform online prediction and model optimization of wind power and photovoltaic clusters, and construct a set of online samples, utilize dynamic adjustment of the model and hyperparameter optimization to provide dynamic updates for similar scenarios in the future, effectively improve prediction accuracy, select typical scenarios for optimization, and increase calculation rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the overall process structure flowchart of the present invention. DETAILED DESCRIPTION

[0027] The present invention is further described below in conjunction with the accompanying drawings and examples. The purpose of the present invention is to provide a method for short-term wind power prediction based on error feedback and model dynamic optimization.

[0028] This example uses a wind power cluster from the Ningxia Hui Autonomous Region in northwest China. The data set contains wind power data from 2015 to 2017. The training set and test set are used to simulate the input of historical data, online data, and real-time data. The wind power cluster contains 16 wind farms, and there are certain differences between different farms. The time window of the sample set is 15 minutes, and the prediction time scale is 96 hours. The root mean square error, average error, accuracy, and qualified rate are selected as the basis for judging the prediction effect.

[0029] like Figure 1 As shown, the present invention provides a method for short-term wind power prediction based on error feedback and model dynamic optimization, which is characterized by the following steps:

[0030] Step 1, constructing an online sample set of wind power clusters: Collect historical numerical weather forecast NWP (Numerical Weather Prediction, NWP) data of wind power clusters and online weekly operation data to obtain sample data of wind power clusters, use data processing and feature mining to mine the sample data of wind power clusters, and obtain the core feature sample set of wind power clusters; input it into the wind power similar scene generation model based on Kmeans-DTW (Dynamic Time Warping, DTW) to obtain the sample set of each scene, and propose a wind power online sample set construction method based on the enhanced diffusion model to obtain the online sample set of wind power clusters under each weather scene; specifically, follow the following steps:

[0031] Step 1.1: Select wind power clusters at the same provincial level, collect NWP data and online weekly operation data of the wind farms and photovoltaic stations in the cluster. NWP includes meteorological factors such as wind speed, wind direction, irradiance, temperature and humidity.

[0032] Step 1.2: Data processing and feature mining are used for the NWP data of the wind power cluster in step 1.1 and the online weekly operation data to construct a core sample set of the wind power cluster; data processing includes data cleaning, abnormal data removal and missing data filling, and normalization is used to smooth the data. Outliers, stray points and power-limiting points are used to delete abnormal data values, and missing data are filled by neighboring and average station filling methods; feature mining uses Pearson coefficients and SHAP values ​​for correlation analysis, and establishes high-dimensional features to construct an optimal core feature sample set;

[0033] Pearson coefficient ρ X，Y The calculation formula is:

[0034]

[0035] In the formula, cov(X,Y) represents the variance, σ represents the standard deviation, X and Y represent two different input features;

[0036] Step 1.3: Divide the core feature sample set of step 1.2 into 96h time scale, obtain sample sets of different weather scenes based on Kmeans-DTW model, use 70% of the sample sets of different weather scenes as the training set of enhanced diffusion model, use 30% of the sample sets of different weather scenes as the test set of enhanced diffusion model, input the training set into enhanced diffusion model for sample expansion, and then input the generated result into the model training result discriminator for judgment. If it is judged as no, continue to adjust the parameters of the model of the training set; if it is judged as yes, obtain the pre-trained model, input the test set into the pre-trained model for result testing, and when the data t-Distributed Stochastic Neighbor Embedding (t-sne) and principal component analysis (PCA) are consistent, output the output result of the pre-trained model, and obtain the online sample set of wind power cluster as the data basis for subsequent online prediction and optimization.

[0037] Step 2, wind power prediction based on ensemble learning: establish an ensemble model based on LSTM, Transformer, LightGBM, BPNN, TCN and CNN models, input the online sample set of wind power clusters obtained in step 1 into the initial ensemble model, obtain preliminary prediction results for each scenario and cluster, screen potential optimization scenarios based on the root mean square error, and calculate relevant indicators of accuracy and pass rate; assign corresponding weights to each indicator, establish comprehensive evaluation indicators, and screen out the target optimization scenario set and normal scenario set based on the corresponding thresholds; proceed as follows:

[0038] Step 2.1: Set fixed weights for the LSTM, Transformer, LightGBM, BPNN, TCN and CNN models respectively, establish an initial integrated model, input the online sample set of wind power clusters generated in step 1.3 into the initial integrated model, and obtain preliminary prediction results for each scenario and cluster;

[0039] Step 2.2: Compare the difference of the root mean square error for the prediction results of each scenario and cluster, determine the prediction model of the potential scenario to be optimized, calculate the accuracy, pass rate, correlation coefficient between the prediction result and error, accuracy of the high wind speed segment and accuracy of the low wind speed segment under the weather scenario, and count the results of each indicator;

[0040] Step 2.3: Assign corresponding weights to the five indicator results in step 2.2, calculate the comprehensive indicator results, set the threshold to 80%, of course, the threshold can be set between 75% and 85%. In this embodiment, the threshold is set to 80%, and the wind farm data below the threshold is classified as the target dynamic optimization scenario set; the remaining wind farms are classified as the normal scenario set.

[0041] Step 3, dynamic optimization technology of wind power cluster model: based on the target optimization scenario set in step 2 as the feature set input, the integrated model based on the Staking algorithm is dynamically adjusted, and the adjusted integrated model is obtained by cross-validation; the hyperparameter optimization of the adjusted integrated model is performed based on the Optuna framework, and the hyperparameter search is performed using the ideal function Trail to determine the optimized integrated model. The similar typical scenario part in the new input sample set will dynamically update the initial integrated model, and finally the prediction is performed through the online optimized integrated model to obtain the prediction result; specifically, follow the following steps:

[0042] Step 3.1: Dynamically adjust the weights of each model in the initial integrated model based on the Stacking algorithm to obtain an adjusted integrated model. The target dynamic optimization scenario set in step 2.3 is input as a feature set into the adjusted integrated model for training. The adjusted integrated model adopts cross-validation for each model. When the average accuracy of the verification reaches more than 85%, the adjusted integrated model is output;

[0043] Step 3.2: Based on the adjusted integrated model obtained in step 3.1, optimize the model hyperparameters based on the Optuna framework, define the search space, create the ideal function Trail, obtain the optimization results of the integrated model hyperparameters, complete the optimization of the integrated model, and obtain the optimized integrated model of the target scenario;

[0044] Step 3.3: For future new input sample sets, the similar typical scene part will dynamically update the initial integrated model of the normal scene set, and finally make predictions through the online optimized integrated model to obtain the final optimized prediction results.

Claims

1. A wind power short-term power forecasting method based on error feedback and model dynamic optimization, characterized by Follow these steps: Step 1, constructing an online sample set of wind power clusters: collecting historical numerical weather forecast NWP data of wind power clusters and online weekly operation data to obtain sample data of wind power clusters, and using data processing and feature mining to mine the sample data of wind power clusters to obtain a core feature sample set of wind power clusters; The data is input into the wind power similarity scenario generation model based on Kmeans-DTW to obtain the sample set of each scenario. A method for constructing a wind power online sample set based on the enhanced diffusion model is proposed to obtain the online sample set of wind power clusters under various weather scenarios. Step 2, wind power prediction based on ensemble learning: establish an ensemble model based on LSTM, Transformer, LightGBM, BPNN, TCN and CNN models, input the online sample set of wind power clusters obtained in step 1 into the initial ensemble model, obtain preliminary prediction results for each scenario and cluster, screen potential optimization scenarios based on the root mean square error, and calculate relevant indicators of accuracy and pass rate; assign corresponding weights to each indicator, establish comprehensive evaluation indicators, and screen out the target optimization scenario set and normal scenario set based on the corresponding thresholds; Step 3, dynamic optimization technology of wind power cluster model: based on the target optimization scenario set in step 2 as the feature set input, the integrated model based on the Staking algorithm is dynamically adjusted, and the adjusted integrated model is obtained by cross-validation; based on the Optuna framework, the hyperparameters of the adjusted integrated model are optimized, and the ideal function Trail is used to search for hyperparameters to determine the optimized integrated model. The similar typical scenarios in the new input sample set will dynamically update the initial integrated model, and finally the prediction is performed through the online optimized integrated model to obtain the prediction result.

2. The method for short-term wind power forecasting based on error feedback and model dynamic optimization according to claim 1 is characterized in that: The step 1 is specifically performed according to the following steps: Step 1.1: Select wind power clusters at the same provincial level and collect numerical weather forecast (NWP) data and online weekly operation data of each wind farm and photovoltaic farm in the cluster. The numerical weather forecast (NWP) includes meteorological factors such as wind speed, wind direction, irradiance, temperature and humidity. Step 1.2: Data processing and feature mining are used for the NWP data of the wind power cluster in step 1.1 and the online weekly operation data to construct a core sample set of the wind power cluster; data processing includes data cleaning, abnormal data removal and missing data filling, and normalization is used to smooth the data. Outliers, stray points and power-limiting points are used to delete abnormal data values, and missing data are filled by neighboring and average station filling methods; feature mining uses Pearson coefficients and SHAP values ​​for correlation analysis, and establishes high-dimensional features to construct an optimal core feature sample set; Pearson coefficient ρ X，Y The calculation formula is: In the formula, cov(X,Y) represents the variance, σ represents the standard deviation, X and Y represent two different input features; Step 1.3: Divide the core feature sample set of step 1.2 into 96h time scale, obtain sample sets of different weather scenes based on Kmeans-DTW model, use 70% of the sample sets of different weather scenes as the training set of enhanced diffusion model, use 30% of the sample sets of different weather scenes as the test set of enhanced diffusion model, input the training set into enhanced diffusion model for sample expansion, and then input the generated result into the model training result discriminator for judgment. If it is judged as no, continue to adjust the parameters of the model of the training set; If the answer is yes, a pre-trained model is obtained, and the test set is input into the pre-trained model for result testing. When the data t-distribution random neighborhood embedding and principal component analysis are consistent, the output result of the pre-trained model is output, and an online sample set of the wind power cluster is obtained as the data basis for subsequent online prediction and optimization.

3. The method for short-term wind power forecasting based on error feedback and model dynamic optimization according to claim 2 is characterized in that: The step 2 is specifically performed according to the following steps: Step 2.1: Set fixed weights for the LSTM, Transformer, LightGBM, BPNN, TCN and CNN models respectively, establish an initial integrated model, input the online sample set of wind power clusters generated in step 1.3 into the initial integrated model, and obtain preliminary prediction results for each scenario and cluster; Step 2.2: Compare the difference of the root mean square error for the prediction results of each scenario and cluster, determine the prediction model of the potential scenario to be optimized, calculate the accuracy, pass rate, correlation coefficient between the prediction result and error, accuracy of the high wind speed segment and accuracy of the low wind speed segment under the weather scenario, and count the results of each indicator; Step 2.3: assign corresponding weights to the five index results in step 2.2, calculate the comprehensive index results, set the threshold to 75% to 85%, and classify the wind farm data below the threshold into the target dynamic optimization scenario set; The remaining wind farms are classified as the normal scenario set.

4. The method for short-term wind power forecasting based on error feedback and model dynamic optimization according to claim 3 is characterized in that: The step 3 is specifically performed according to the following steps: Step 3.1: Dynamically adjust the weights of each model in the initial integrated model based on the Stacking algorithm to obtain an adjusted integrated model. The target dynamic optimization scenario set in step 2.3 is input as a feature set into the adjusted integrated model for training. The adjusted integrated model adopts cross-validation for each model. When the average accuracy of the verification reaches more than 85%, the adjusted integrated model is output; Step 3.2: Based on the adjusted integrated model obtained in step 3.1, optimize the model hyperparameters based on the Optuna framework, define the search space, create the ideal function Trail, obtain the optimization results of the integrated model hyperparameters, complete the optimization of the integrated model, and obtain the optimized integrated model of the target scenario; Step 3.3: For future new input sample sets, the similar typical scene part will dynamically update the initial integrated model of the normal scene set, and finally make predictions through the online optimized integrated model to obtain the final optimized prediction results.

Images