Monsoon rainfall prediction using a BiLSTM model with an attention mechanism: a case study of Odisha state

Effective water resource management and disaster mitigation require accurate rainfall prediction for planning purposes. Forecasting of rainfall is a difficult time series problem as it is highly non-linear and non-stationary, and dependent on season. This study suggests a hybrid deep learning framework that integrates Convolutional Neural Network (CNN), Bidirectional Long Short Term Memory (BiLSTM) with Transformer Architecture and Attention, as well as the ensemble based on meta-learning. This framework employs CNN layers to gather temporal local patterns, BiLSTM networks to shape sequential dependencies, and transformer blocks to embrace the self-attention mechanism to formulate long-range patterns. An attention mechanism is used to dynamically weigh relevant time steps to achieve higher accuracy and interpretability in predictions. Moreover, various forecasts from LSTM, CNN–BiLSTM–Attention, CNN–Transformer–Attention, and XGBoost are fused together using a Random Forest-based meta-learner for enhanced robustness and generalization in the ensemble model. The proposed method utilizes Monte Carlo dropout for uncertainty estimation, enhancing the trustworthiness of the aforementioned approach by providing confidence intervals. The model undergoes assessment through a variety of performance metrics such as RMSE, MAE, R², MAPE, and NRMSE. Experimental results show that the ensemble framework consistently outperforms individual models, achieving better predictive accuracy and greater stability than individual models for a variety of rainfall, including extremes. The proposed model combines attention-based interpretability, transformer-based global learning, and ensemble meta-learning for rainfall forecasting, which is robust and scalable.