Deep Learning-based Time Series Forecasting And Its Application In Household Electricity Forecasting

The importance of time series forecasting is becoming increasingly significant,and it is widely used in many fields such as new energy power generation,quantitative investment,household electricity demand side response,and traffic forecasting.In the smart grid,it is of great practical significance for the power system’s stable operation to use to mine internal laws from the historical household electricity consumption data,predict the household electricity consumption for a period of time in the future,and cooperate with the power grid to respond to the demand side.However,time series data has the characteristics of coexistence of multiple periodic features and coexistence of multiple time series features.Existing time series forecasting methods are difficult to fully and accurately extract complex time series features,resulting in inaccurate modeling and unsatisfactory forecasting effects.With the development of artificial intelligence-related theories,deep learning models provide an effective solution for improving the accuracy of time series forecasting.In this essay,taking the household electricity consumption scenario as an example,based on the analysis of the characteristics of time series data,a feature-guided prediction model is constructed based on deep learning to try to solve the problem that complex time series features are difficult to fully and accurately extract.1)A univariate time series prediction model(PSTNet)based on feature fusion and deep learning is proposed.For univariate time series,most of the existing time series forecasting methods cannot fully and accurately extract the time series features implicit in the data.Taking the electricity consumption data of a single household as an example,this paper analyzes in detail the multiple time series features implied by univariate time series,and proposes a univariate time series prediction model based on feature fusion and deep learning,which mainly uses CNN to extract multi-period features,Seq2Seq encoder to extract short-term dependent features,Embedding to extract time-related features.Besides,the Attention mechanism uses the time features in Embedding to dynamically weight multiple periodicity,making the modeling more accurate.Finally,in the Seq2Seq decoding layer,a variety of time series features are deeply fused,and the predicted values of the required future time steps are obtained in turn.The experimental results show that PSTNet model can fully considers time series features,and is superior to the current mainstream prediction models in effectiveness and accuracy,and has great practical value.2)A multivariate time series forecasting model(MFDL)based on multi-task learning and deep learning is proposed.For multivariate time series,existing time series forecasting methods are limited for global correlation feature extraction.Taking the electricity consumption data of multiple adjacent households in the same community as an example,this paper proposes a multivariate time series prediction model based on multi-task learning and deep learning,which mainly uses multi-task learning to jointly model the multi-variable household electricity consumption time series.Then,the global correlation is extracted,and CNN and LSTM are embedded in the parameter sharing layer of the multi-task learning framework,so that multiple types of time series features can be extracted at the same time,and then the time series features of the prediction output layer are further fused,and finally all target variables are realized at the same time.The MFDL model proposed in this paper is applied to real datasets,and experiments show that the MFDL model can effectively improve the accuracy of time series prediction due to the comprehensive consideration of multiple time series features.