Weather radar echo extrapolation is an important approach for convective nowcasting, which predicts the evolution of convective systems in a short term. In recent years, radar echo extrapolation approaches based on deep learning have made significant progress and have been widely applied for radar echo extrapolation. However, existing methods only use a composite reflectivity product as the input data for echo extrapolation, which contains limited echo information and tends to have noise and data missing, resulting in low accuracy and short effective time of radar echo extrapolation. To alleviate the mentioned limitations and leverage multiple radar products, an extrapolation method based on multiple products fusion is proposed in this paper. In our method, three types of radar products including compositive reflectivity (CR), vertical integrated water (VIL) and echo top height (TOP) are taken as input, features ofeachproductare extracted by residual multi-scale extractor, then the concatenated features are fused in channel dimension and space dimension by parallel convolutional block attention module（PCBAM）. The model is trained with the fused feature sequences and the hierarchical loss function is used to deal with the uneven data distribution. Experiments shows that, the effective time and accuracy of our proposed extrapolation model have improved.
Described in the txt in ZIP file.