Attention-Based Dual-Source Spatiotemporal Neural Network for Lightning Forecast

Abstract

Accurate lightning forecast is significant for disaster prevention and reduction. However, the mainstream lightning forecast methods, which mainly rely on numerical simulations and parameterizations, can hardly cope with the spatiotemporal deviations. Meanwhile, the rapid and complex evolution of lightning regions go beyond the traditional extrapolation-based forecast methods. In this work, we propose a data-driven neural network model for hourly lightning forecast, which exploits both the numerical simulations and the recent historical lightning observations. The two kinds of data complement each other and play different roles at different stages of the forecast. The use of dual-source data greatly increases the amount of information available to improve the forecasting performance. To handle the variability of deviation patterns in numerical simulations, we introduce a channel-wise attention mechanism, which adaptively adjusts the proportion of each simulated parameter to maximize the useful information. The attention mechanism also enables the model to reveal the contribution of each simulated parameter for the forecast. Experimental results on a real-world dataset show that the proposed method outperforms several baseline methods. Ablation studies further demonstrate the effectiveness of our data fusion approach and attention module.