Multi-period interaction networks for time series forecasting

Abstract

Forecasting time series with complex and overlapping periodic patterns remains a major challenge, especially in long-term prediction tasks where both local and cross-period dependencies must be modeled. In this work, we propose Multi-Period Interaction Networks, a fully multilayer perceptron based architecture designed to capture temporal dynamics across multiple periodic components. The core of the model is a Period-Frequency Interaction module, which enables dynamic modeling of multi-periodic structures in time series data. We evaluate MPINet on the task of lithium-ion battery state of health estimation, where accurate long-term prediction is essential for ensuring system reliability and safety. Extensive experiments on real-world battery datasets demonstrate that MPINet achieves state-of-the-art forecasting accuracy while maintaining high computational efficiency, highlighting its effectiveness for both battery health monitoring and broader time series forecasting applications.