Solving the problem of water level fluctuations in small hydropower station systems is challenging under traditional industrial control methods. This difficulty arises from the system’s high nonlinearity and the complexities involved in mechanism modeling. To address this, an improved neuro-fuzzy approach is proposed. In which, the multi-head attention mechanism based long short-term memory network is used to describe complex water level change patterns, and the fuzzy controller is introduced to dynamically adjust the control parameters to reduce water level fluctuation. Simulation-based on real hydropower station system data is carried out, and the superiority of the improved model under complex dynamic conditions is verified by comparing the prediction accuracy of different neural network methods and the effects of fuzzy controller and traditional PID control.