Improving Meteorological Drought Forecasting Through a CPO ‐Tuned VMD ‐Liquid Neural Network

Abstract

ABSTRACT This paper introduces a novel hybrid model for one‐month‐ahead meteorological drought forecasting. The proposed approach employs the Crested Porcupine Optimizer (CPO) to adaptively tune the key parameters of Variational Mode Decomposition (VMD). This optimization step effectively decomposes and denoises the raw Standardized Precipitation‐Evapotranspiration Index (SPEI) time series, yielding cleaner input features for a liquid neural network (LNN) regression model. To develop parsimonious CPO‐VMD‐LNN models, the most informative inputs are selected via mutual information between SPEI and lagged decomposed vectors. The effectiveness of the model was evaluated through two case studies in the eastern and western regions of the Urmia Lake Basin, Iran. Comparative analysis against three benchmark models, namely SARIMA, LSTM, and LNN, demonstrated that the proposed hybrid model consistently outperformed the alternatives in both regions. Notably, the CPO‐VMD‐LNN achieved root mean square errors below 0.25 across the study areas, highlighting its strong potential for operational drought forecasting tasks.