Water level predictive control in cascaded canals based on the variation pattern analysis of integrator-delay model parameters

The Beijing-Shijiazhuang section of the Middle Route of the South-to-North Water Diversion Project (MRP), China. The integrator-delay (ID) model exhibits time-varying parameters, whose relationships with system state variables remain unexplored. However, traditional prediction relies on fixed parameters, resulting in a model–plant mismatch that generates parameter uncertainties and the accumulation of prediction errors, which degrades the performance of real-time control of water levels. This study investigates the variation patterns of the parameters by identifying system state-related influencing factors and quantifying their individual functional relationships via polynomial regression and correlation analysis, guiding key factors screening. Although these single-factor analyses provide insights into individual relationships, multi-factor interactions can alter such relationships. Hence, a dynamic parameter identification method using feedforward neural networks is proposed to address this issue; based on this, an adaptive-parameter model predictive controller is further developed. Evaluation on a simulation model of the MRP indicates that using FFNNs for dynamic parameter identification reduces accumulated prediction errors by 43.15 %, improving the predictive accuracy of ID. Compared with traditional fixed-parameter MPC, control actions are reduced by 19.49 % (0.2 % increase under Disturbance 2) and 29.69 % in magnitude and frequency on average, and water level control effect is not significantly improved (enhancement: 0.81 % on average; paired t-test, P > 0.05, 95 % CI [-2.90 %, 4.45 %]). The proposed method reduces scheduling pressures in canal pools. • The variation patterns of backwater area A d with influencing factors are investigated. • Based on correlation and polynomial regression analysis, Key influencing factors are screened out. • A dynamic identification method is proposed based on Feed Forward Neural Networks. • An adaptive parameter model predictive controller is further developed.