Abstract: The matrix format of multi-input multi-output Hammerstein model is deduced, and a kind of the new improvedneural dynamics algorithm is proposed. The algorithm can be used to identify many groups of unknown Hammertein modelparameters, which improves accuracy and convergence rate. Firstly, the parameters’convergence of the new improved neuraldynamics algorithm is analyzed. Then a new hybrid model based on Hammerstein model is deduced to build an error modelbetween actual system and mechanism system. The hybrid model has good compensating effect. Since the new neural dynamicsmethod can adjust Hammerstein model parameters online, the hybrid model can be used to simulate dynamic behaviorof complex processes in a large scale exactly. The rationality and efficiency of the presented method are demonstrated bysimulation experiment.