Particle Swarm Optimized Deep Learning Models for Rainfall Prediction: A Case Study in Aizawl, Mizoram

Rainfall is vital to all life on Earth, and rainfall prediction is essential for various sectors and aspects of human society. Hilly areas such as the state of Mizoram in India have suffered from landslides during the rainy season. This study compares twelve hybrid deep learning and machine learning models to predict daily rainfall using meteorological variables such as maximum humidity, minimum temperature, maximum temperature, and rainfall. The compared models include Particle Swarm Optimization (PSO)-Artificial Neural Network (PSO-ANN I), PSO with stacked ANN (PSO-ANN II), PSO-Bidirectional Long Short-Term Memory (PSO-BiLSTM), PSO-BiLSTM-ANN without Dropout Layer (PSO-BiLSTM-ANN I), PSO-BiLSTM-ANN with Dropout Layer (PSO-BiLSTM-ANN II), Stacked BiLSTM with ANN (PSO-BiLSTM-ANN III), PSO-Long Short-Term Memory (PSO-LSTM), PSO-LSTM-ANN without Dropout Layer (PSO-LSTM-ANN I), PSO-LSTM-ANN with Dropout Layer (PSO-LSTM-ANN II), Stacked LSTM with ANN (PSO-LSTM-ANN III), PSO-Recurrent Neural Network with ANN (PSO-RNN-ANN), and PSO-Support Vector Regression with Linear Kernel (PSO-SVR). We trained and tested the models using 12,418 days of meteorological data from 1985 to 2018 collected by the Aizawl Weather Station in Mizoram, India. The study used Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and coefficient of determination ( \text{R}^{2} ) to evaluate the performance of the models. It is observed that the PSO-BiLSTM-ANN II model, which is a stack of BiLSTM, ANN and Dropout layer, achieved the best performance and outperformed the PSO-SVR model by 6.4%. The PSO-BiLSTM-ANN II model also requires fewer cells in the hidden layer than other models and converges with the lowest epochs. The results show the advantage of adding the ANN layer in the RNN, LSTM, and BiLSTM models, and this study provides a benchmark model for predicting rainfall in the study area.