Dynamic-LSTM hybrid models to improve seasonal drought predictions over China

Abstract

Accurate drought prediction is essential for drought resilience and water resources management. The skill of seasonal drought prediction from dynamical and statistical models has limitations. This study combines dynamical models and deep learning to construct hybrid (dynamical-statistical) models. We use the random forest to extract typical grids based on the geopotential height, sea-level pressure, and 2-m temperature. The long short-term memory (LSTM) is used to construct the statistical models, with atmospheric variables as predictors and the 3-month Standardized Precipitation Index (SPI3) as the predictand. The hindcasts of atmospheric variables from the European Centre for Medium-Range Weather Forecasts (ECMWF) SEAS5 model are processed as predictors to force the statistical models. The hybrid model, constructed using the dynamical models and the LSTM, is named dynamic-LSTM (“D-LSTM”). The results suggest that the LSTM models are of information-extraction capability and robustness. When the lead time exceeds one month, the prediction skills are significantly improved by the D-LSTM models, especially in the East, Northwest, Southwest, and Tibet. In most regions, the D-LSTM models are more skillful across all seasons for lead times exceeding 30 days and are reliable in predicting droughts in spring and summer when the ECMWF SEAS5 loses skills at the seasonal scale. Furthermore, the D-LSTM models are more accurate in drought onset prediction.