Hybrid Temporal-Spatial Attention Network for Rainfall Prediction in Urban Microclimates

Climate change has led to more frequent occurrences of extreme weather conditions and irregular rainfall patterns, making accurate prediction of rainfall amounts essential. In this paper, the performance of a Hybrid Temporal-Spatial Attention Network (HTSAN) model in predicting rainfall for the next day is evaluated using 12 meteorological predictors from the ERA5 reanalysis dataset. The geographical area covered by this research spans a 17 × 17 grid with 0.25° × 0.25° nominal resolution around Bangalore. HTSAN’s architecture involved a convolutional feature extractor, recurrent temporal representation, a transformer encoder, and a regression layer. HTSAN's performance was benchmarked against recurrent, convolutional-recurrent, and attention-based approaches using chronological test data. In this, benchmarking, the TFT approach resulted in the best RMSE value of 7.1474, while HTSAN had an RMSE of 7.5911, an MAE of 3.6779, and an R² of 0.4259. Extreme-rainfall benchmarking also revealed that HTSAN’s critical success index values were lower than those of the comparison approaches at the selected grid-cell rainfall thresholds. Thus, it can be inferred that the current version of HTSAN did not perform better than the selected benchmarks.