面向多元时间序列的线性和非线性加权融合预测

Multivariate time series forecasting has been widely used in finance, environment, transportation and other fields. However, traditional statistical prediction models usually assume that the time series conforms to a certain distribution or functional form, and cannot capture the complex nonlinear relationships. Although neural network based algorithms have powerful learning abilities, they usually ignore the linear features in time series. By weighted and fused both Linear and Nonlinear Predictions, this paper proposes a novel WFLNNet, where the linear prediction module is designed based on an autoregressive model while the nonlinear prediction module is designed based on the neural network and consists of a feature extraction encoder, an interactive attention network, and a fully connected layer to capture the most effective features in temporal and spatial correlations, as well as a mutual influence among multivariate time series. We have done experiments using 4 real datasets by comparing them with 6 baseline algorithms. The experimental results demonstrate that WFLNNet outperforms the 6 baseline algorithms with more accurate prediction.