Research On Time Series Load Forecasting Method Based On Deep Learning

As an important direction of time series forecasting,power load forecasting needs to quickly obtain accurate forecasting of power load in a limited time to avoid system paralysis caused by excessive load.However,the traditional time series prediction method is difficult to achieve long-term memory function for complex waveforms,and it is difficult to filter invalid information interference.In view of this challenge,this paper studies the task of power load forecasting,and fully explores the relevant characteristics of power load,the factors affecting power load and the relevant algorithms of time series forecasting,In order to use the relevant time series information for fast and accurate power load forecasting,a power load forecasting method based on a variety of neural network hybrid framework and a power load forecasting method based on time convolution network and multi head attention mechanism encoder are proposed.Firstly,in view of the characteristics that the long-term and short-term characteristics in the power load forecasting task can not be taken into account at the same time,the ShuffleNet V2-GRU module and SSTCN module are constructed.The ShuffleNet V2-GRU module can effectively alleviate the long-term dependence in the time series forecasting task on the premise of improving the training speed through the long-term feature extraction method of weighted channel group convolution gated cycle unit and time series convolution network.Secondly,in view of the feature discreteness and feature uncertainty in power load forecasting,the feedforward neural network is used to realize the preliminary cleaning and effective expansion of features,and the automatic selection of feature information is realized by feature aggregation.Then,the encoder based on multi head attention mechanism is used to further refine the ability of the model to focus on different dimensions of information,so as to enhance the robustness of the model.Thirdly,aiming at the exponential growth of computing resources and model complexity caused by the increase of model in time convolution network,SSTCN structure is proposed.By integrating the deep learning idea of separable group convolution,channel shuffling and feature relearning,the network is decomposed into multiple small time convolution networks that can run in parallel,so as to reduce the complexity of the model.Finally,for the SV2 GRUE model and the SSTCNE model,sufficient experiments are carried out on the ASHRAE dataset and the ARCHIVE dataset,respectively,and a comprehensive analysis of the experimental results,model parameters,and model training time is carried out.