Integrating seasonal autoregressive integrated moving average and long short-term memory for multi-scale standardized precipitation evapotranspiration index forecasting

ABSTRACT

Workflow of SARIMA-LSTM model for SPEI forecasting over Tibetan Plateau, integrating SARIMA linear components, LSTM spatial attention, and dynamic residual weighting to improve multi-scale prediction performance.

Drought evolution exhibits pronounced multi-temporal-scale variability and coupled linear–nonlinear dynamics. Existing statistical models or deep learning models alone are often unable to achieve satisfactory prediction accuracy across different time scales, while spatial correlation information in regions with complex terrain remains insufficiently exploited. To address these challenges, this study proposes a hybrid forecasting framework integrating the Seasonal Autoregressive Integrated Moving Average (SARIMA) model and the Long Short-Term Memory (LSTM) network to improve the accuracy of multi-scale Standardized Precipitation Evapotranspiration Index (SPEI) forecasting. The proposed model adaptively integrates linear and nonlinear prediction components through a dynamic residual-weighting mechanism and incorporates a lightweight spatial attention module to enhance the representation of spatially correlated features. Using grid-based multi-scale SPEI data collected over the Qinghai–Tibet Plateau during 1981–2022, the proposed framework was trained, validated and tested. Model performance was evaluated across multiple time scales (SPEI-1, SPEI-3, SPEI-6, and SPEI-12) and compared with several benchmark models. Results indicate that the proposed approach consistently outperforms the benchmark models across all considered time scales, with the introduction of the spatial attention module and dynamic residual-weighting mechanism improving the Nash–Sutcliffe Efficiency (NSE) from 0.88 to 0.91. Overall, the proposed framework provides an effective approach for multi-scale drought forecasting in regions with complex terrain.