Towards Accurate Rainfall Volume Prediction: An Initial Approach with Deep Learning, Advanced Feature Selection, Parameter Optimisation, and Ensemble Techniques for Time-Series Forecasting

Accurate rainfall forecasting is crucial in sectors such as agriculture, transportation, and disaster prevention. This study introduces an initial approach that combines deep forecasting techniques, advanced feature selection, parameter optimisation, and ensemble method to enhance the accuracy of rainfall volume prediction. The proposed methodology is evaluated using a historical weather dataset from Bath, United Kingdom, spanning from January 1, 2000, to April 21, 2020. To address challenges related to generalisation, uncertainty, reliability, and inappropriate predictors, a hybrid mechanism is created by combining various LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) models with a Fuzzy Inference System. The resulting ensemble system comprises five individual hybrid models. Through baseline experiments and comparisons with benchmarks, the effectiveness of the methodology is demonstrated, revealing significant performance improvements over previous studies, across a range of performance indices. Overall, the proposed ensemble approach exhibits better generalisation compared to benchmarks. This research has the potential to revolutionise rainfall volume predictions by leveraging deep learning, advanced feature selection, parameter optimisation and ensemble techniques, overcoming many limitations of the existing approaches.