Developing meta-heuristic optimization based ensemble machine learning algorithms for hydraulic efficiency assessment of storm water grate inlets

This research explored the performance comparison of machine learning algorithms, i.e. random forest (RF) and Adaboost, each coupled with the genetic algorithm (GA) and particle swarm optimization (PSO), in intercepted discharge calculations. Thus, six different storm water grate inlets were evaluated through laboratory experiments, and the acquired data was used to construct the integrated prediction framework. Consequently, the RF-based models outperformed the models constructed with Adaboost. Overall, the PSO-RF was found as the best strategy with Nash-Sutcliffe Efficiency index and determination coefficient values of 0.8896 and 0.8990, respectively. In addition, the game-theoretical SHAP analysis demonstrated that the approach flow depth is the most influential variable for hydraulic efficiency evaluations, followed by the grate inlet width and transversal slope of the road. This study further concluded that the lower the grate inlet width, the lower the intercepted discharge capacity, while the increase in transversal slope results in an increase in hydraulic efficiency.