Research On Short-Term Prediction Model Of Wind Power System Based On Multi-Level Data-Driven Approach

The modern power system with high shares of renewable energy and power electronics is facing low-carbon,digital,and intelligent transformation.The construction of wind power generation units,stations and clusters has a rapid momentum.Wind speed and wind power have multi-scale spatio-temporal coupling and random fluctuation,which seriously affects the safe,stable,efficient,and economic operation of power grid.How to improve the multi-level prediction accuracy of wind power system represented by wind speed and wind power is the core problem to ensure the friendly access of wind power.Taking "point prediction layer probability prediction layer cluster prediction layer" as the research path,this paper aims to study the prediction model of multi-level wind power system based on data-driven.The main work is described as follows:(1)Point Prediction Layer.A point prediction model based on Stacking fusion is proposed to significantly improve the accuracy and generalization of wind speed short-term prediction model.First,a kernel ridge regression model based on different kernel functions is established.Then key parameters are selected using an improved firefly algorithm.Improve the global search ability and convergence speed of the algorithm by introducing adaptive parameters,global search and Levy flight.Finally,t he independent models are fused by the Stacking algorithm to enhance the generalization of the combined model,and the cross-validation is used to further improve the prediction accuracy.The measured results of different wind fields and in different seasons are selected to simulate the prediction effect of the proposed model,and its prediction accuracy and generalization ability are verified through comparative analysis.(2)Point Prediction Layer.A point prediction model based on dual Q-learning is proposed,which fully considers the wind speed fluctuation characteristics and related attributes in each period.The model proposed in this paper realizes the adaptive model screening in the set based on data-driven and the error correction through dual Q-learning,which significantly improves the accuracy and autonomy of the prediction model.First,build a wind speed Q learning model set consisting of 5 basic prediction algorithms,fully consider wind speed fluctuations and attribute factors,select the best prediction model for each time period through the Q learning,and get the preliminary wind speed prediction results.Calculate the prediction error based on the wind speed prediction result,construct the second-stage error Q learning model library,screen the best model in the model library to correct the preliminary prediction value,obtain the final prediction result.Finally,the effectiveness of the proposed method is verified by predicting the wind speed of the actual wind field in different seasons.(3)Probability Prediction Layer.A probability prediction model based on improved Natural Gradient Boosting is proposed,aiming at providing more uncertain information,focusing on improving the coverage of the prediction area and the proportion of the average width of the prediction interval of the model,so as to provide a more accurate reference for building an efficient and intelligent new energy power system.First,the initial data set is preprocessed by outlier detection and feature variable screening.Then for the defects of the ordinary gradient to solve the probability prediction problem,a natural gradient is proposed and a general solution is given.Thereby the Natural Gradient Boosting is established and the Blending Fusion is applied to enhance the learning effect.Finally,the model is validated and validated with the help of measured data from Dalian Tuoshan Wind Farm.The results show that the model proposed in this paper can provide complete wind power probability prediction information at a specified confidence level,reduce the forecast area coverage probability while ensuring higher proportion of average width of prediction interval,and can be used to build efficient and intelligent new energy power system.(4)Cluster Prediction Layer: A wind power cluster prediction model based on cluster division is proposed to improve the speed of cluster division and avoid the impact of data noise,so as to realize the rapid and accurate prediction of wind power clusters.First,blending fusion xgboost is used to reduce the dimension and screen the clustering index.Then,the clustering division of wind power clusters is realized by using the spatial clustering algorithm based on density with the help of dynami c time integration optimization.Finally,the Rhododendron algorithm is used to improve the regularized limit learning machine to realize the short-term power prediction of wind power cluster.Based on the measured data of a regional wind farm,the prediction effect of the proposed model is analyzed and verified.The results show that the proposed method has fast cluster division speed,can effectively avoid the influence of noisy data,and can realize the rapid and accurate prediction of wind power clusters.The short-term prediction model of multi-level data-driven wind power system proposed in this paper has been successfully applied to practical examples of several wind farms.The experimental results show that the models proposed in this paper achieve the expected objectives.