Complexity-efficiency dynamics of metaheuristic-optimized recurrent neural network models for drought forecasting in hyper-arid Kuwait

Kuwait, Arabian Peninsula, Western Asia This study develops an optimization-based drought-forecasting framework in which LSTM hyperparameters are optimized using Bat Algorithm (BA), Ant Colony Optimization (ACO), and Grey Wolf Optimization (GWO) to predict the Precipitation Index at 12- and 24-month scales (PI12 and PI24). To benchmark the computational efficiency of the proposed optimization framework, a non-optimized baseline LSTM and a Gated Recurrent Unit (GRU) model were also implemented under identical data partitions and training configurations. Unlike conventional models relying on probabilistically normalized indices, the framework utilizes the deterministic and distribution-free PI index, which is well-suited for zero-inflated, data-scarce conditions in hyper-arid regions. This study benchmarks multiple metaheuristically optimized LSTM configurations for drought forecasting in hyper-arid Kuwait using the Precipitation Index. Model training relied on long-term monthly precipitation data, with performance evaluated using RMSE, MAE, and R ², and computational complexity quantified by the number of trainable parameters. At the PI24 scale, optimized LSTMs achieve accuracy comparable to that of the baseline model, indicating the dominance of long-term precipitation accumulation in regional drought dynamics. Compact architectures, such as the GRU, further demonstrate the efficiency–accuracy trade-off relevant for operational drought monitoring. • Dual-scale PI12–PI24 drought forecasting with LSTM and GRU. • Bounded metaheuristic tuning of LSTM (ACO, BA, GWO). • Complexity–efficiency analysis using parameters and epoch time. • Compact models matched larger networks in accuracy. • PI24 demonstrated improved stability and reduced error variability.