Chlorine dosing is a complicated system with time delay, time-varying, non-linear and coupling. In this paper, multivariable adaptive generalized predictive controller based on Smith predictor is proposed. Instead of the optimal predictor, the Smith predictor with adaptive identifying parameters can increase the robustness of the MIMO system. Simulation and application in water-works at Suzhou (China) shows that the algorithm can overcome time-varying, time delay and disturbance.