Time Series Forecasting Based On Deep Learning

Time series data is widely used in many fields such as meteorology,economy,industry,etc.The prediction of time series data is of great significance to actual production and life.Traditional methods of forecasting time series are data-critical and often have no way to learn longer historical values.This paper uses the more cutting-edge deep learning technology in the field of artificial intelligence research as a theoretical framework to process and predict two different types of multivariate time series data,and improve some existing models for processing time series data.The specific work is as follows :1)A combined model(AR?CLSTM)based on encoder-decoder structure and linear regression is proposed for multivariate time series with univariate and multiple sources.AR?CLSTM uses a convolutional neural network to learn the feature information of the input time series data.Then,the extracted features are input into the encoder-decoder network structure based on time-step attention to learn the dependencies between time series data.In addition,an attention mechanism is introduced between the encoding network and the decoding network,which can more accurately select the relevant encoded hidden states of all time steps for prediction,thereby improving the model's ability to express multivariate time series.At the same time,a linear autoregressive module is added to the original model to realize the combined model of linear model and nonlinear model.The fault tolerance of the model is increased,and the problem that the original model is not sensitive to the learning of linear features of the data is improved.This paper validates the effectiveness of the model by testing it on four univariate multi-source multivariate time series datasets.2)Considering that the multivariate single source multivariate time series data is not only affected by its own historical value but also affected by other variables,this paper refactors its characteristics,In addition to the other influencing factors collected,the mean and variance of the data at different frequencies,which are the characteristics of the data itself,are constructed to make up for the model's learning ability of the characteristics of the data.On the basis of feature construction,PCA is used to reduce the dimension,and while reducing the dimension,as much data information as possible is saved.It eliminates redundant data to a certain extent,improves prediction accuracy,and shortens training time.Then,the improved temporal convolutional network DATCN model is used to improve the accuracy of electric load prediction.The DATCN model uses a stacked deep temporal convolutional network(TCN)to learn the dependency information of longer time steps and then uses the self-attention mechanism to capture the internal correlation of the data to further improve the prediction accuracy of the network.In this paper,the model is tested under the real power load data of two different areas.