Design of a Multi-Step Predictive Control Alum Dosing System Based on GRU Neural Networks^*

This paper proposes a predictive control alum dosing system based on a data-driven approach. Considering the highly nonlinear, uncertain, and time-varying parameters in the coagulation and sedimentation process of water treatment plants, a multi-step predictive control algorithm (GRU-MPC) is designed. This algorithm uses current and previous alum dosing amounts and effluent turbidity data to predict effluent turbidity over multiple future time steps. The approach includes: constructing a multi-step predictive neural network model, designing an objective function to minimize the difference between effluent turbidity and the target value while reducing the fluctuation of alum dosing amounts, optimizing the objective function using a sequential quadratic programming algorithm, and applying the optimal alum dosing amounts for real-time control. Simulation results show that the GRU-based multi-step predictive control method can effectively enhance the stability and resource utilization efficiency of the water quality control system.