Research On Short-term Wind Power Forecasting Based On Local Learning

In order to achieve the goal of "carbon peak by 2030 and carbon neutral by 2060",my country is speeding up the adjustment and optimization of energy structure and vigorously developing renewable energy.Among them,clean,pollution-free,and widely distributed wind energy has become one of the ideal energy sources that have attracted much attention in the field of renewable energy.At present,wind energy is mainly consumed through the integration of wind power,but due to the strong randomness and volatility of wind energy,the integration of wind power will have a non-negligible impact on the safety and stable operation of the power system.Wind power prediction is one of the effective means to solve this problem,but the prediction performance of the global model is easily affected by the randomness of wind energy,and the adaptive ability is poor.For this reason,based on the local learning strategy,this paper conducts research on the short-term prediction of wind power.The main research contents are as follows:Aiming at the problem of the weak adaptability of offline global models to wind power data,a wind power generation based on the combination of Just in Time Learning(JITL)strategy and Back Propagation Neural Network(BPNN)is proposed.Forecasting algorithm.First,use the quartile method to clean the abnormal wind power data,and use the mean value of the adjacent points before and after to fill in the missing data,so as to obtain an effective data set.Secondly,considering the strong nonlinearity of the wind power data and the insignificant overall characteristics,it is proposed to use Euclidean distance to measure the similarity between the sample to be estimated and the historical data,and to select historical data with high similarity to construct a local BPNN model.Finally,the model proposed in this paper is compared with the global model of back-propagation neural network,long-short-term memory neural network,and Gaussian process regression.The results show that the local modeling method proposed in this paper can effectively improve the prediction accuracy of the model.Aiming at the problem that static samples cannot reflect the local dynamic characteristics of wind power data,a wind power prediction method combining DBSCAN(Density-based spatial clustering of application with noise)clustering and real-time learning is proposed.This method is based on the continuity and trend similarity of the time series,and uses the comprehensive distance combining Euclidean distance and cosine distance as the clustering metric,and constructs the continuous onehour wind speed as a clustering feature,thereby dividing the samples into different dynamic characteristics Clusters.Secondly,by calculating the similarity between the sample to be estimated and the center of each cluster,the cluster with the most similar dynamic characteristics is selected.Then the real-time learning strategy is adopted in the cluster,and modeling samples are selected according to Euclidean distance.Finally,use back-propagation neural network to model predictions.The simulation results show that the proposed prediction algorithm can better capture the local dynamic characteristics of wind power data and further improve the prediction accuracy of the model.Aiming at the problem that online prediction requires high model training speed,a wind power adaptive local modeling algorithm based on dynamic knowledge distillation is proposed.First,use DBSCAN clustered data to construct a BPNN model for each cluster,and define the model as a teacher model.Then,initialize the BPNN model with a smaller structure as the student model.When the sample to be estimated selects the modeling sample through the DBSCAN-JITL method,the soft loss between the teacher model and the student model and the hard loss between the student model and the real value are combined,which is used as the loss function of the student model to guide student model training.The simulation results show that the proposed method can effectively improve the modeling speed of online local models and maintain high accuracy.In summary,a short-term wind power prediction algorithm based on local learning is proposed,which effectively improves the adaptive ability and prediction speed of the wind power prediction model,and provides new ideas for wind power prediction.