A novel hybrid DOA-PSO-enhanced LSSVM model for monthly runoff forecasting in the upper Heihe river basin

Accurately forecasting runoff is vital for balancing human water demands, mitigating flood risks, and safeguarding ecological integrity. However, the nonlinear and non-stationary characteristics of hydrological processes pose significant challenges for conventional models. To address these issues, this study proposes a novel hybrid DOA-PSO-LSSVM model, which integrates the Dream Optimization Algorithm (DOA) for global exploration, Particle Swarm Optimization (PSO) for local refinement, and Least Squares Support Vector Machine (LSSVM) for nonlinear learning. Instead of relying on signal decomposition or manual parameter tuning, the proposed framework directly optimizes LSSVM hyperparameters through a hybrid global-local search strategy, effectively reducing mode mixing, subjectivity in parameter selection, and computational burden. Using 30 years of hydro-meteorological data from the upper Heihe River Basin, China, the hybrid model was compared with LSSVM, PSO-LSSVM, and DOA-LSSVM. Results demonstrate that DOA-PSO-LSSVM consistently achieved the highest predictive accuracy, with RMSE reductions of 4-23% and correlation coefficients (r) above 0.95 across stations. Beyond accuracy, the model exhibits robustness across different meteorological input scenarios and shows strong capability in simulating peak flows and water balance. This study advances hyperparameter optimization in runoff prediction and provides practical implications for sustainable water management and disaster prevention.