Linear Variable Convolution and Cross Attention LSTM Method for Radar Echo Extrapolation

Radar echo extrapolation is critical in short-term weather forecasting. To enhance the accuracy and adaptability of radar echo prediction, this paper proposes a novel method that integrates a linear variable convolution (LVC) module and a cross-attention (CA) mechanism into the spatiotemporal long short-term memory (ST-LSTM) framework, named LVC-LSTM. The LVC module enables dynamic adjustment of the convolutional sampling shape, allowing the network to capture the irregular and evolving structures of radar echoes more accurately, thereby offering improved flexibility and representation capability compared with traditional and deformable convolutions. The CA mechanism introduces a pyramid-based CA structure that incorporates cross-scale embeddings and hierarchical attention blocks, effectively capturing multi-scale features and long- and short-range spatial dependencies inherent in radar echo dynamics. Experimental results using a real-world radar echo dataset demonstrated that the LVC-LSTM model outperformed comparative models across multiple evaluation metrics, including the critical success index, heidke skill score, root mean square error, mean absolute error, and structural similarity index, indicating its strong potential for operational radar echo extrapolation.