Research On Multivariate Time Series Forecasting Method Based On Deep Neural Networks

Time series forecasting is an important task in engineering applications,and it has a wide range of applications in many fields such as engineering,economics and finance,meteorology,and energy science.With the rapid development of the Internet,the scale of time series data has grown rapidly,and time series data has also been converted from univariate time series to multiple time series.Compared with univariate time series,multivariate time series have more complex and high-dimensional characteristics,and the nonlinear trend and noise of multivariate time series lead to the unstable interdependence and time periodicity of multivariate time series,so it is difficult to accurately analyze and predict the data.Most of the existing time series prediction methods focus on capturing the long-term dependence of time series data,while ignoring the local details of the time series data,resulting in insufficient detection capabilities of the model for outliers.The research of this thesis focuses on accurately capturing the long-term dependence relationship between time series data and improving the model's ability to extract local features of multiple time series data.Two deep neural network model based on the encoder-decoder framework are proposed for non-attribute multiple time series prediction tasks and multiple time series with attribute prediction tasks.For the task of non-attribute multivariate time series prediction,LFE-Transformer is proposed in this thesis.Specifically,LFE-Transformer first uses Causal Convolutions to extract and enhance the local features of the time series data,and then introduces the enhanced time series features into the self-Attention.The self-attention mechanism can quickly obtain the global information and long-term memory of the data,so it can effectively enhance the time series prediction model's ability to predict long-term dependencies and outliers.Experiments show that the performance of LFE-Transformer on the commonly used time series prediction datasets is better than other non-local enhancement feature methods,which indicates that the LFE-Transformer model proposed in this thesis can effectively enhance local key information in multiple time series and capture long-term dependence.For the task of multivariate time series prediction with attributes,this thesis proposes the FAT-TACNet model.The coding stage of FAT-TACNet consists of attribute feature extraction network: Attrext Net and time series feature extraction network:Timext Net.Attrext Net uses 1D convolution to extract features in both the positive and reverse order of attribute values.Timext Net uses causal convolution to extract the characteristics of time series data,and integrates self-Attention mechanism into the convolution to capture long-term information.The features extracted by above two networks are fused and then decoded to obtain the predicted time series.Experiments on two time series datasets with attributes show that the performance of FAT-TACNet is better than other methods,which indicates that the FAT-TACNet model proposed in this thesis can effectively extract attribute features and time series features and can obtain accurate prediction series.