MSACN-LSTM: A multivariate time series prediction hybrid network model for extracting spatial features at multiple time scales

In the current era of big data, which constantly generates massive data, all walks of life are producing a large number of multivariate time series data sets. The ability to predict these multivariate time series data sets is essential for effective decision making and management in industry. Considering that the relationship between variables of multivariable time series on different time scales is not the same, this paper proposes a multi-time scale prediction hybrid network (MSACN-LSTM) for multi-time scale multivariable time series prediction. Firstly, in order to explore the relationship between variables at multiple scales, we construct MSACN module. Specifically, we introduce CBAM attention mechanism based on TCN structure and add residual connection to form 4 branches with different convolution kernel sizes, and use these 4 branches to mine the relationship between variables of the down-sampled multivariate sequences. Secondly, in order to mine the time dependence of elements within variables at the same scale, we construct a two-layer LSTM module to extract the time features within the data series. In order to verify the validity of the model, we conducted ablation experiments on several public data sets, and the results show that the use of MSACN and LSTM modules can improve the prediction accuracy, because they can mine the temporal and spatial characteristics of multivariate serial data at different scales. In order to verify the superiority of this model over several different deep learning models, we also conducted comparative experiments. The experimental results show that the new model has good prediction effect and high prediction accuracy in single and multi-step prediction of multivariable sequences.