An Interpretable Multivariate Time Series Prediction Method Based on VSN-GRU

Multivariate time series prediction is crucial in various domains such as electricity, where the interpretability of predictive models plays a key role in decision-making. Traditional black-box models, such as LSTM and Transformer, are constrained by their inability to interpret variable importance and capture temporal dynamics. In this paper, we propose an interpretable, end-to-end prediction framework that integrates a Variable Selection Network (VSN), Gated Recurrent Unit (GRU), and Feedforward Neural Network (FFN). This framework is designed to address the challenges of interpretability while preserving high predictive accuracy. Specifically, the VSN facilitates instance-level variable importance analysis, the GRU models local temporal dependencies, and the FFN generates multi-step predictions. Additionally, an adaptive gating mechanism is incorporated to dynamically regulate the flow of information. Experimental results on multiple real-world datasets show that the proposed method not only achieves competitive accuracy in comparison with machine learning methods but also offers valuable insights through variable selection and weight analysis.