Development and evaluation of a novel DTIB-LSSVM model for efficient and accurate runoff forecasting

• First use of LSSVM as initializer for BKA, improving convergence speed and global search capability. • TVFEMD decomposes runoff data into multi-scale sub-sequences, enriching feature information for model training. • IBKA optimized with intelligent initialization enhances global optimization ability and LSSVM parameter tuning. • Integration of DCS module enables parallel task decomposition, improving computational efficiency for large-scale data. • Model maintains forecasting accuracy while achieving significant improvement the computational efficiency. Efficient and accurate runoff forecasting is essential for flood prevention and mitigation, water resources management, and ecological protection. However, the runoff process exhibits nonlinearity, non-stationarity, and multi-scale characteristics, which challenge accurate and efficient runoff forecasting. To address these challenges, we propose a novel combined forecasting model, DTIB-LSSVM, which integrates the time-varying filtered empirical mode decomposition (TVFEMD)-intelligent black kite algorithm (IBKA), least-squares support vector machine (LSSVM), and distributed computational server (DCS) module. This model constructs an efficient parallel combinatorial optimization forecasting structure by embedding the DCS module. Firstly, the TVFEMD technique is employed to decompose non-smooth data into multiple Intrinsic Mode Functions (IMFs), which exhibit relatively smooth properties, thereby effectively reducing data complexity. Secondly, the regularization and kernel function parameters of LSSVM are optimized by IBKA, which significantly improves the model's nonlinear fitting ability and generalization performance. To address the time-consuming computation in model parameter optimization, this study further embeds the DCS module, which significantly enhances the model's computational efficiency through parallel task allocation. Finally, accurate and efficient daily runoff forecasting is achieved by superposing the forecasting results of each mode. Moreover, the article proposes an intelligent initialization strategy that innovatively uses a machine learning model as the initializer of swarm intelligence optimization algorithms, thereby enhancing the global search capability of the optimization algorithm. To verify the model's effectiveness, this study takes the Shebu Hydrological Station, Quinebaug River, and Housatonic River basins as case studies and assesses the performance using four evaluation indexes and nine comparative models. The results demonstrate that the Nash-Sutcliffe efficiency coefficients of the proposed model are 77.4 %, 23.8 %, and 187.9 % higher than those of the original model, with all other indicators outperforming the comparison models. After embedding the distributed computing module, the computational efficiency of runoff forecasting is improved by 24, 26, and 22 times under three hydrological conditions, respectively. Overall, the model shows excellent potential for efficient and accurate runoff forecasting in different basins.