Hybrid Particle Swarm Optimized Models for Rainfall Prediction: A Case Study in India

Abstract

Predicting rainfall is crucial across multiple sectors and activities, impacting agriculture, water management and disaster preparedness. In this study, the Particle Swarm Optimization (PSO) algorithm is used to optimize the hyperparameters of hybrid deep learning and machine learning models such as Bidirectional Long Short-Term Memory (BiLSTM), Long Short-Term Memory (LSTM), Recurrent Neural Network (RNN), Artificial Neural Network (ANN) and Support Vector Regression (SVR). The performances of the PSO-optimized models are compared using the monthly rainfall dataset of Aizawl Weather Station and the all-India monthly average rainfall dataset. For the all-India rainfall datasets, the results of the PSO models are also compared with models from previous studies. The results show that, for the all-India rainfall dataset, the hybrid model PSO-BiLSTM IV achieved an RMSE of 225.12 and outperformed an existing RNN model by 14% and an existing single-cell LSTM, Vanilla LSTM and stacked LSTM by 11%, 10% and 8% respectively. In the Aizawl Weather Station dataset, the hybrid model PSO-BiLSTM II achieved the best result with an RMSE of 76.6, a benchmark result for this dataset. Overall, the hybrid PSO-BiLSTM models have the lowest RMSE score and the SVR models achieve the lowest performance for both datasets.