PhysAttNet Enhances Time Series Forecasting with Physics-Informed Attention

Accurate time series forecasting is vital across various physical processes, from manufacturing to astrophysics. While convolutional neural networks (CNNs) are popular for these tasks due to their efficiency, their learned temporal representations can sometimes be unstable or physically inconsistent, hindering robustness and interpretability.PhysAttNet addresses this by integrating physics-informed attention into a lightweight CNN forecaster. It applies three complementary regularization constraints during training: alignment regularization to encourage attention to follow smooth, peak-centered structures; smoothness regularization for continuous temporal evolution; and sparsity regularization to focus on informative intervals.These differentiable terms embed physics-guided inductive bias without needing explicit annotations. Experiments in predicting milling cutting forces and blazar flare forecasting demonstrate that PhysAttNet significantly improves forecasting accuracy, generalization, and performance on critical events, making predictions more reliable and physically consistent.