CSPP-RNN: A Precipitation Nowcasting Approach That Couples Similar Precipitation Processes with Sequence-to-Sequence RNNs

Accurate precipitation nowcasting is critical for many aspects of human life. A recurrent neural network (RNN) has demonstrated strong and relatively mature performance in machine learning approaches for precipitation nowcasting. However, their inherent recursive prediction structure leads to error accumulation, causing progressively blurred outputs and limiting practical applicability. To address this issue, we propose CSPP-RNN (Coupling-Similar-Precipitation-Processes RNN), a net that couples similar precipitation processes with a sequence-to-sequence RNN. For each prediction timestep, similar precipitation processes are retrieved, and their segments are then input into the encoder to obtain the corresponding hidden states. These hidden states replace the ones influenced by earlier predicted results in the recursive structure. Based on radar data from Beijing Daxing station, the comparison experiments of CSPP-RNN and ConvLSTM indicate that: (1) Over the 36–60 min lead time across the 0.1, 5.0, and 20.0 mm/h thresholds, the POD and CSI improved by 0.0334, 0.0170 on average, respectively, whereas the FAR degraded by 0.0586; (2) error accumulation was mitigated, retaining richer fine-scale structures in the predicted images; (3) the extra computational cost of coupling was controlled within an acceptable range. In conclusion, CSPP-RNN mitigates the error accumulation problem in RNN by coupling similar precipitation processes as part of the modification of the recursive prediction structure. This provides a potential new direction for optimizing the application of RNN in precipitation nowcasting.