Research On Short-term Electricity Forecasting Method Based On Deep Learning

Short-term load forecasting(STLF)is the basis of economic operation of power system.Accurate load forecasting can optimize power generation planning of power companies and improve the economic benefit and safety of power grid operation.The classic regression-based forecasting model is mainly aimed at stationary time series,but the power load is a typical non-stationary time series.The shallow neural network model is difficult to effectively capture complicated inner nonlinear pattern,while the power load presents complicated non-linear patterns due to various factors such as region,climate,and economy.Among the prediction methods based on deep neural networks,the most popular ones are based on recurrent neural network(RNN),such as long short-term memory(LSTM)networks,which has been proven to achieve the state-of-the-art forecasting performance.However,with the increase of input sequence length,the training of deep neural network becomes much more difficult.Since the power load has a large time span from day to year,LSTM cannot fully utilize the inherent correlation of power load.The prediction model based on CNN has lower computational complexity and less research at present.Based on the latest research results of deep learning,this thesis studies STLF based on real power load dataset.The main contributions of this thesis include:(1)A short-term load ensemble prediction method is designed which is based on LSTM.Based on the similar day method,a similar data selection module is built.Utilizing the powerful learning ability of LSTM in time series forecasting,the ensemble learning structure is designed by integrating multiple LSTM networks,which increases the model capacity with endurable complexity payload.Evaluation is implemented on the real power load data set,and the results show its advantage over the reference ones.(2)A CNN-based short-term load forecasting method is proposed.After the analysis of several measure metrics,a complex invariant distance-based clustering model is established,and the number of categories is then automatically adjusted.The basics of CNN are briefly introduced,and the CNN-based STLF model is designed.The experimental test is carried out based on the real load data set,and the feasibility of the prediction model is verified.