Precipitation Nowcasting

The Challenge

Predicting local weather patterns (1–6 hours ahead) is computationally expensive and difficult with traditional numerical models, especially at local spatial resolution.

The Solution

Developed a ConvLSTM deep learning model to predict the next 6–60 frames of radar echo images based on historical sequences from the CIKM radar dataset.

Key Results

• Reduced forecasting error to RMSE = 11 dBZ.
• Provided reliable short-term prediction for climate-sensitive use cases.