Using AI-powered modelling for drought forecasting in the Democratic People’s Republic of Korea

Context The Democratic People’s Republic of Korea (DPRK) is vulnerable to drought due to monsoon-driven variable precipitation that affects agricultural production and water security. Reliable multi-timescale drought forecasting aids risk mitigation; however, nonlinear non-stationary drought signals are hard to capture indata-scarce regions. Aims This study developed and validated a Complete Ensemble Empirical Mode Decomposition with Adaptive Noise-Long Short-Term Memory (CEEMDAN-LSTM) hybrid Standardised Precipitation Evapotranspiration Index (SPEI) drought forecasting model. Its performance is systematically compared against a standalone LSTM for multiple SPEI timescales under DPRK climatic conditions. Methods Monthly meteorological records from 10 stations collected over a period of 60 years were used to derive SPEI time series. CEEMDAN decomposed SPEI into intrinsic mode functions for noise reduction and trend extraction as LSTM inputs; hyperparameters were optimised by Bayesian optimisation. Model performance was assessed with Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). Key results CEEMDAN-LSTM generally outperformed standalone LSTM with timescale-dependent skill. Short-term SPEI predictions had inherent accuracy limits (NSE < 0.5) yet reduced relative errors. Medium- to long-term SPEI yielded large accuracy gains (max NSE = 0.95 for SPEI-12), with residual-distribution analysis showing effective suppression of extreme prediction errors. Spatial heterogeneity existed; complex terrain at stations such as Chongjin limited decomposition-driven improvements. Conclusions CEEMDAN-LSTM improves forecast skill by processing non-stationary drought signals. Without complex external predictors, this hybrid method enables reliable medium-long-term drought monitoring and extreme-event detection. Implications This single-variable time-series method enables data-efficient drought early warning for data-scarce regions, supporting drought response and adaptive water management without large-scale remote-sensing or global datasets.