With regard to the characteristics of uncertain network-induced delay, an idea of using Generalized Predictive Control (GPC) is proposed in networked control systems. To overcome the influence of modeling mismatch and uncertainty of a system, an error predictive model is further built based on BP neural network, and error predictive value is used to compensate output predictive value, which forms a new control algorithm of GPC, and smooth filtering weighted controlled input variable is adopted. Random-delay-time characteristic of data transmission is tested in a networked control system. GPC and GPC with error compensation are compared in the system. GPC with BP error compensation is verified to have better performance in a modeling mismatch system.