Research On Time Series Neural Network Model For Load Forecasting

With the rapid development of the global economy and the increasing demand for energy in various social activities,the production and dispatch of electric energy will be greatly improved with the help of accurate load forecasting,which will also save a lot of unnecessary energy waste from the perspective of environmental protection.Load forecasting is one of the most important tasks in today’s power system management,however,today’s society is increasingly rich in electricity consumption patterns and traditional load forecasting models may not be able to cope with such complex situations.In this paper,we first briefly explain the concepts related to load forecasting,analyze the nature of electric load variation based on the unstable characteristics and periodicity of load,and list several types of commonly used classical load forecasting methods.Based on this,this paper proposes a TEL(Transformer Enhanced LSTM)model with Transformer enhanced LSTM semantic extraction capability,and an integrated forecasting model WTEL(WaveletTEL)combined with wavelet transform strategy.For the matrix representation initially extracted by the LSTM in the load data,the self-attentive mechanism and residual connectivity within Transformer can further abstract the original power consumption pattern features,i.e.,the high-level semantic information.This high-level semantic information is used in the prediction process to initialize the weight matrix of the LSTM network and enhance the learning capability of the model for dynamic features of load sequences.In addition,it is difficult to simultaneously solve the periodic and nonlinear problems often found in electric loads due to traditional models.A deep learning forecasting method TLSTMSA(Temporal LSTM with Self-attention)based on the modeling of time-series periodic residuals is proposed to address this problem.In this paper,the load forecasting process is divided into two parts: one part forecasts the large-scale trend of electric load changes,which directly encodes and models the overall trend of load changes over time based only on the weekly,daily,and hourly information of load;the other part models the amount of residuals between the original load and the overall trend,in which an LSTM network with a self-attentive mechanism is designed as the encoder and decoder.In this paper,the public data set is simulated by using the actual load used by the Texas Electric Reliability Council and the 9th "China Electrical Engineering Society Cup" National Student Electrician Mathematical Modeling Competition.First,the effectiveness of the proposed TEL model in single-step prediction is verified with different data sets,and the accuracy of the proposed TEL model is improved to a certain extent by combining the wavelet transform with the effective decomposition of the sequences and then recombining them after individual prediction.Second,the experimental results show the improvement of the TEL model’s prediction ability on the compared statistical learning methods,machine learning methods,and neural network methods.In the TLSTMSA model,an alternative approach to feature selection is used,i.e.,prediction is modeled directly based on the date information of the load instead of inputting the historical load into the model.The results of the ablation experiments surface the effectiveness of each module within the TLSTMSA forecasting method.Both TEL and TLSTMSA models have higher prediction accuracy than other comparison models in the multi-timescale prediction task,and the error performance is stable and generalized in different regions of the prediction task.