Research On Short Term Load Forecasting Methods Based On Deep Learning

With the deepening of power market reform,more and more attention has been paid to the optimal dispatch of power system and the dynamic balance between generation and load.With the development of information intelligence in power industry,the data acquisition system of smart grid provided massive data sources for load forecasting.It was of great significance to explore the information behind the power data.Accurate short-term load forecasting could not only ensure the safe and reliable operation of the system,but also improved the utilization rate of power generation and avoided the waste of power resources.According to the basic principle of power system short-term load,this thesis determined the main influencing factors.Based on the analysis of the uncertainty,conditionality and timeliness of short-term load,three short-term load forecasting methods were proposed.The research contents of this thesis included:1.The basic principle of power system short-term load was introduced in detail.the characteristics and main influencing factors of power system short-term load were analyzed and verified through experiments.the characteristics of power system short-term load was uncertainty,conditionality and timeliness,and the main reference characteristics of load forecasting research was determined.This thesis introduced the main work flow of load forecasting and data preprocessing,including abnormal data detection and cleaning and data normalization method.2.Aiming at the instability of recurrent neural network in the analysis of long-term dependence,a method combining dual attention mechanism and gated recurrent unit(GRU)was proposed and applied to short-term load forecasting.Using feature attention mechanism to analyze the impact of different input parameters on load,the problem that environmental factors and residents’ own subjective factors affect the real-time change of short-term load forecasting was solved,and the potential associated information was mined.Through the time series attention mechanism,the importance of the load at each historical time to the load at the forecast time was analyzed,and the key time point data was selected to improve the prediction accuracy.Data analysis results showed that the average absolute percentage error of short-term load forecasting was reduced to 3.82%.3.In order to realize the fast extraction of power short-term load time series characteristics,the time convolution network with simple network structure and fast learning was adopted,and the self attention mechanism was combined to extract the relevant information within the network,so as to learn the dependence of power short-term load position for a long time.Although this method could enhance the real-time performance,the prediction accuracy was decreased.Therefore,this thesis proposed a multi-scale time attention convolution network to mine power data,which effectively improved the accuracy and timeliness of short-term load forecasting.Data analysis results showed that the average absolute percentage error of short-term load forecasting was reduced to 2.36%.4.Considering that the time convolution network was a single variable time series input,the influence of weather,electricity price,holidays and other external factors were ignored.In this thesis,a method of fusing time convolution and GRU network was proposed.By introducing time convolution network to extract the temporal features of historical load data,and fusing the features with external features,the feature vector of load in projection space was obtained.The feature vectors were input into the combined network of GRU and jump GRU respectively,and the output of the combined network was obtained by linear superposition.The data analysis results showed that the accuracy of short-term load forecasting was 97.26%,which could realize the effective short-term load forecasting.