Rough unedited code used in data science contest - CIKM AnalytiCup 2017 challenge - to predict short-term rainfall.

Animation of radar maps over 1 hour, first images in the sequence are input and the final image is the predicted radar map 15 minutes into the future. Although the prediction is blurry, the general speed and direction of movement have been estimated accurately.

This solution used a convolutional recurrent neural network (convLSTM) analysing sequential radar maps, and predicting the future state of the clouds. From these projections, a further model was built linking the predicted radar map with the actual rainfall. Maps were provided at 4 different altitudes, and a separate model was built for each, with predictions averaged.

Solution was inspired by the paper Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting

Final score (RMSE) was 14.8 (top 100), relative to top 5 accuracies ranging 11.0 to 13.3.

Static illustration - two of the radar maps input and the system prediction for the state 15 minutes into the future, compared to the truth.

Data was provided in text file format. The first part of the program processes and stores this as .h5 file systems for ease of loading and analysing (data was too large to fit into memory).

Missing data points were simply assigned the median value of that image.

Full details here

"Short-term precipitation forecasting such as rainfall prediction is a task to predict a short-term rainfall amount based on current observations. It has been a very important problem in the field of meteorological service. An accurate weather prediction service can support casual usages such as outdoor activity and even provide early warnings of floods or traffic accidents. To predict short-term rainfall amount, we usually make use of radar data, rain gauge data and numerical weather outputs. In this challenge, we focus on the first type of data - radar data, more specifically radar echo extrapolation data. Our target is to build a rainfall prediction model by solely using the radar data."