Multivariate Time Series Forecasting based on Time-Frequency Decomposition

Multivariate time series forecasting is a current research hot spot in academic and industrial fields; however, it is still challenging to effectively mine the underlying patterns in multivariate time series data and achieve accurate long-term predictions. To cope with this problem, a model based on timefrequency decomposition is proposed. First, the multivariate time-series data are decomposed into different frequency components by using frequency domain decomposition methods, combined by their contributions in energy, and then the combined different components are reduced back to the time domain, and the multi-scale time windows are used to extract the global and local attention of different components in the time series. Secondly, for the prediction task, a balance of prediction accuracy and stability is achieved by adaptively selecting two loss functions, mean square error (MSE) and mean absolute error (MAE), and adjusting the loss weights according to the prediction steps. There is an overall improvement in the performance of the model on six standard datasets and two real datasets. The results show that the model can effectively mine frequency patterns in multivariate time-series data and achieve accurate long-term forecasts.