STANDARDIZED PRECIPITATION INDEX FORECASTING IN NORTH-CENTRAL MEXICO USING TRANSFORMER MODELS

Abstract

Drought prediction is crucial for water resource management, agriculture, and climate adaptation in arid and semi-arid regions such as Zacatecas, Mexico. This study evaluates the advanced neural network architectures Long Short-Term Memory (LSTM), Vanilla Transformer, and Informer, for forecasting the Standardized Precipitation Index (SPI) using monthly precipitation data from 1964–2020 collected at 31 weather stations. SPI series were clustered into four regional climate zones. Models were implemented using the Nixtla NeuralForecast framework, and performance was assessed with Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Diebold-Mariano significance tests. The Informer model achieved the highest predictive accuracy, reducing average MSE by approximately 15 % relative to LSTM and consistently outper-forming the Vanilla Transformer in most regions. Statistical testing confirmed regional differences in model performance, suggesting that an adaptive, region-specific modeling approach is optimal for drought forecasting. These results demonstrate the robustness, efficiency, and transferability of Transformer-based models, particularly Informer, for operational drought monitoring under variable climatic conditions.