First Author: WANG Jingyang
Corresponding Author: BAIYIN Baoligao
Journal: Ecological Modelling

Abstract

Because of the disconnection between hydraulic design of fishways and fish behavioral theory, the operational efficiency of fishways is still significantly lower than that of natural rivers. Fish behavior models, as emerging research tools, can accurately capture the dynamic interaction between fish and their hydrodynamic environment, thereby improving the prediction of fish behavior during passage and providing essential technical support for fishway design.

This study innovatively constructed a fish behavior prediction framework (ML-IBM) that integrates machine learning (ML) with individual-based modeling (IBM). Specifically, the framework integrates the random forest (RF) algorithm with the Eulerian–Lagrangian Agent Method (ELAM), thereby significantly improving prediction accuracy of fish behavior.

Using upstream monitoring data of Schizothorax prenanti in a vertical-slot fishway, the model validation demonstrated that ML-IBM achieved an accuracy of 83.4% for swimming behavior classification. The R2 for swimming velocity prediction reached 0.77, with a root-mean-square error (RMSE) of 7.35 and mean absolute error (MAE) of 6.26, both lower than those of existing models.

Feature importance analysis revealed that hydrodynamic factors such as turbulence intensity, flow velocity gradient, and total force response were the dominant drivers of fish swimming behavior. The results highlight that ML-IBM enhances model interpretability, allowing fish swimming behavior to be explained more mechanistically from a hydraulic perspective.

This research provides an effective approach to developing fish behavior models driven by both hydraulics and individual traits, offering important practical value for optimizing fishway design and supporting eco-hydraulic engineering.

Keywords: Fish behavior model; Individual-based model (IBM); Machine learning; Fishway