A Transformer Time-series Architecture with Multi-metric Enhanced Marine Predator Algorithm (MMEMPA) for Meteorological Drought Forecasting

Accurate drought forecasting is critical for building climate resilience and ensuring sustainable water resource management, particularly in vulnerable regions like India.While deep learning models such as LSTMs are widely used, they often struggle to capture long-term temporal dependencies and require intensive, manual hyperparameter tuning.This study proposes a novel hybrid framework-a Transformer-based Time Series model optimized with a Multi-Metric Enhanced Marine Predator Algorithm (TST-MMEMPA)-for multi-scale forecasting of the Standardized Precipitation Evapotranspiration Index (SPEI1-SPEI12).Applied to five meteorological stations in Gujarat, India, using daily climate data, the model leverages the Transformer's self-attention mechanism to capture complex, long-range dependencies in climate series, while the MMEMPA metaheuristic automates and refines hyperparameter optimization to enhance generalization and predictive accuracy.Experimental results demonstrate the superior performance of TST-MMEMPA, achieving low error metrics (MSE: 0.41-0.46,RMSE: 0.63-0.67)and high explanatory power (R²: 0.86-0.91)across all stations and SPEI timescales.This represents a significant improvement over traditional models, including LSTM, BiLSTM, N-BEATS, Autoformer, and an unoptimized Transformer, as statistically validated by paired t-tests (p < 10⁻⁶).The proposed framework proves to be a robust and accurate tool for meteorological drought early warning across climatically diverse stations within Gujarat, with transferability subject to retraining on local data.