Achieving reliable rainfall forecasting through ensemble deep learning, fuzzy systems, and advanced feature selection

Classical Machine Learning (ML) models, like Random Forests, struggle with weather variability. Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models, though designed for time-dependent predictions, also face challenges with unseen data and extended rainfall forecasts. To address these, the Fuzzy-Integrated Deep Ensemble for Rainfall Forecasting with advanced Feature Selection (FIDERFFS) is developed, combining LSTM and GRU with Fuzzy Inference Systems to create Fuzzified LSTM and GRU hybrids, integrating them with a Transformer model, and leveraging Fuzzy Rough Feature Selection to improve accuracy, generalisation, and interpretability. Tested on meteorological datasets (Aberystwyth and Bath, UK), the Enhanced Ensemble Model outperforms benchmarks and the Foundational Ensemble Model by 12.26% (MAE), 3.21% (RMSE), and 8.77% (RMSLE), in Aberystwyth’s complex terrain. Results show how topography and maritime proximity influence forecasting and position FIDERFFS as a superior alternative to Classical ML while complementing Numerical Weather Prediction, by reducing computational demands and time lags.