Research On Short Term Load Forecasting Method Based On Feature Selection

With the diversification of energy supply and demand,the intelligent energy system with interconnection and integration of various energy sources is of great significance for countries around the world to improve energy efficiency and achieve energy conservation and emission reduction goals.As an important part of smart energy system planning and operation,accurate short-term electricity load prediction can provide important reference for optimization of system scheduling plan and safe and economic operation,and is an inevitable requirement for promoting modern energy system planning and energy digital transformation.However,the multi-dimensional and complex characteristics composed of meteorological factors such as temperature and rainfall,as well as human factors such as power consumption habits and working hours will cause short-term electricity load fluctuations,and it is difficult for a single prediction model to fully show the variation rule of electricity load.Therefore,a prediction method based on Feature Selection(FS),Equilibrium Optimizer(EO)and Ensemble Empirical Mode Decomposition(EEMD)is proposed in this paper to optimize the single model,so as to improve the accuracy of short-term electrical load prediction.The specific research contents are as follows:Firstly,based on the actual data of a park,the characteristics and influencing factors of electricity load are qualitatively analyzed,and the cycle characteristics of short-term electricity load itself and the main influencing variables are determined.Based on the results,the characteristic variables of historical data are classified and the corresponding data are preprocessed,and the input data set containing 24-dimensional characteristics is constructed.Secondly,FS-EO combinatorial optimization method was designed to determine the optimal feature set of the model,aiming at the problem that the excessive number of features in the input data set might lead to the over-fitting and high complexity of the model.Random Forest(RF),e Xtreme Gradient Boosting(XGBoost),Least Absolute Shrinkage and Selection Operator(LASSO)and Elastic Network(EN)with high feature selection efficiency,the FS-model was constructed by its own classifier,and the optimal feature set of each model was determined by comparing the prediction accuracy of models with different feature numbers,and the two base models with better effect were selected from them.Back Propagation(BP)neural network,Long and Short Term Memory(LSTM)neural network,Extreme Learning Machine(ELM)and Convolutional Neural Network(CNN)with low feature selection efficiency,the FS-model is constructed by backward feature selection of the optimal feature set obtained above,and the optimal feature set and two optimal base models of each model are also obtained.In order to reduce the randomness of model parameter setting,the equilibrium optimizer algorithm was introduced to carry out secondary verification and parameter optimization for the four FS-base models,and further build the combined prediction model.As verified by the above data set example,FS-EO-base model has higher prediction accuracy and fewer required features compared with single item model.Finally,the FS-EO-EEMD combined optimization method was designed to solve the problem that the FS-EO-base model did not fully consider the time-frequency characteristics of electric load itself.Each electrical load modal component decomposed by EEMD algorithm was predicted and reconstructed by FS-EO-base model,so as to determine the optimal component combination mode and build the combined prediction model.The above data set examples verify that the prediction accuracy of this model is improved compared with FS-base model,FS-EO-base model and FS-EEMD-base model.At the same time,FS-EO-EEMD-base model has better overall prediction performance when the base model is extreme learning machine.