SMART-Rain: iTransformer for Climate-Aware Multivariate Rainfall Forecasting

Abstract

Accurate rainfall forecasting is essential for climate monitoring, disaster preparedness, and agricultural planning, particularly in data-scarce regions such as Yemen where rainfall variability significantly affects environmental and socioeconomic stability. This paper proposes the SMART-Rain framework for multivariate rainfall forecasting. The paper utilizes around four decades (1982–2025) of subnational rainfall observations and integrates temporal, anomaly, cyclical, and spatial embedding features to capture regional variability. The proposed SMART-Rain framework integrates spatio-temporal feature engineering, including lag-based rainfall variables, cyclical temporal encoding, and regional spatial embeddings (PCODE), within a unified deep learning forecasting pipeline. Historical rainfall data spanning from 1982 to 2025 covering 318 subnational regions in Yemen were utilized to predict future dekadal rainfall (rfh). To improve learning stability and reduce rainfall distribution skewness, logarithmic transformation and MinMax normalization were applied during preprocessing. Three sequence models, namely LSTM, Transformer Encoder, and iTransformer, are evaluated under a unified experimental setting with Bayesian hyperparameter optimization. The results of the SMART-Rain framework show that Transformer-based models outperform recurrent architectures. The iTransformer achieves the best performance with an R2 score of 0.7703, demonstrating its capability to learn long-range temporal dependencies and multivariate feature interactions. The SMART-Rain framework captures seasonal patterns and rainfall trends, but predicting extreme rainfall events remains challenging due to data imbalance and distribution skewness. The SMART-Rain findings confirm the effectiveness of attention-based architectures for multivariate rainfall forecasting and provide useful insights for climate-related time-series prediction.