ConvLSTM: Introducing Convolutional Structure into LSTM to Address Spatiotemporal Sequence Prediction for Precipitation Nowcasting

Precipitation nowcasting aims to predict rainfall intensity in localized areas over short timeframes (e.g., 0–6 hours), which is critical for generating emergency alerts and airport weather guidance. Traditional methods primarily rely on numerical weather prediction or radar echo extrapolation based on optical flow, but the latter is limited by separated steps and difficult-to-determine parameters. This paper proposes treating precipitation nowcasting as a spatiotemporal sequence prediction problem and designs a Convolutional Long Short-Term Memory network (ConvLSTM). The model introduces convolutional structures in both input-to-state and state-to-state transitions, thereby effectively capturing spatial correlations. Experiments show that ConvLSTM outperforms fully connected LSTM (FC-LSTM) and the ROVER algorithm developed by the Hong Kong Observatory on both synthetic datasets and real radar data, enabling more accurate prediction of future frames. For researchers engaged in meteorological AI, video prediction, or spatiotemporal data mining, it is recommended to start with the introduction to understand the problem background, focusing on Section 3 for model architecture design and Section 4 for experimental comparison results. Readers are encouraged to consult "Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting" (2015) to deeply understand the construction logic and empirical effects of this classic spatiotemporal prediction model.