Hybrid ICEEMDAN-VMD-LSTM framework for probable maximum drought estimation along the Lower Yangtze River Basin

Study region Zhenjiang City in Jiangsu Province, China—representative riparian city in the lower Yangtze River Basin Study focus We propose an integrated model that incorporates parameter optimization through the Crested Porcupine Optimizer (CPO), dual-signal decomposition using Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) and Variational Mode Decomposition (VMD), and a Long Short-Term Memory (LSTM) network for long-term forecasting of the Standardized Precipitation Evapotranspiration Index (SPEI). During the long-term forecasting process, dynamic error compensation is incorporated to mitigate the error accumulation inherent in recursive forecasting, while preserving intrinsic stochastic characteristics of hydrological time series, and the probable maximum drought (PMD) is subsequently estimated. New hydrological insights for the region The ensemble model demonstrates better long-term prediction performance than the standalone LSTM model, preserving the statistical and volatility characteristics in the original data. Future PMD events substantially surpass historical extremes across multiple indicators, with intensity increasing by at least 17% and duration exhibiting a maximum increase of up to 282%. Although the precise timing of PMD occurrences remains highly stochastic, statistical results indicate that the occurrence of future droughts still exhibits periodicity, suggesting that the future regional climate system still maintains an underlying rhythmic structure. This study provides an effective framework for PMD estimation and offers a valuable method for regional water resource planning and drought emergency management under changing climate conditions.