HoST: integrating Heuristic knowledge with attention-based LSTM networks for Thunderstorm prediction.

Abstract

Accurate forecasting of severe convective events is vital for meteorologists, as it directly supports their efforts to understand atmospheric risk patterns and enable effective early-warning systems. This paper presents an Integrating Heuristic Knowledge with Attention-based LSTM Networks for Thunderstorm Prediction (HoST) for probabilistic thunderstorm prediction. The proposed framework integrates attention-enhanced recurrent modelling with physically informed heuristic constraints, allowing the model to capture complex nonlinear atmospheric dynamics while maintaining meteorological consistency. The model is evaluated on a real-world observational dataset, where it demonstrates strong predictive capability in capturing spatiotemporal convective patterns. Furthermore, the framework is assessed in a quasi-operational forecasting setting, exhibiting low-latency inference, computational efficiency, and stable predictive performance across multiple forecast lead times (5-60 minutes). To further validate the robustness of the approach, controlled experiments are conducted using synthetically generated atmospheric scenarios that emulate key thermodynamic and kinematic relationships. Results show improved classification stability, enhanced probabilistic calibration, and superior performance compared to Random Forest, SVM, Improved Decision Support, Deep Neural Network, SALAMA, BLSTM-GRU, MetNet, FourCastNet, GraphCast, HRRR, and AROME models. Overall, the findings highlight the effectiveness of integrating heuristic knowledge with data-driven learning, demonstrating the potential of HoST as a physically consistent and operationally viable framework for short-term thunderstorm forecasting.