Hybrid Attention-TCN With Multimodal Radar Features for Atmospheric Electric Field Prediction

Traditional single-architecture models struggle to meet the performance requirements for atmospheric electric field (AEF) prediction. This paper proposes a novel hybrid architecture that integrates multimodal radar features for AEF prediction. Specifically, given the strong correlation between real-time radar charts and AEF signals, this study determines the usable regions of radar charts based on the detection range of AEF apparatus, then designs handcrafted radar chart features and concatenates them with convolutional neural network-extracted features to form a multimodal feature set. Leveraging a cross-modal attention module, it temporally aligns the asynchronous low-frequency radar chart features and high-frequency AEF signals. Finally, the aligned data are input to a temporal convolutional network (TCN) to predict AEF signals. Experimental results demonstrate that our method achieves the most competitive performance, reducing the root mean square error by 42% compared to long short-term memory-based method. The proposed method provides a feasible solution for asynchronous data fusion in meteorological environment monitoring.