A flat neural network is designed for the on line state prediction of engine. To reduce the computational cost of weight matrix, a fast recursive algorithm is derived according to the pseudoinverse formula of a partition matrix. Furthermore, the forgetting factor approach is introduced to improve predictive accuracy and robustness of the model. The experiment results indicate that the improved neural network is of good accuracy and strong robustness in prediction, and can apply for the on line prediction of nonlinear multi input multi output systems like vehicle engines. To reduce the computational cost of weight matrix, a fast recursive algorithm is derived according to the pseudoinverse formula of a partition matrix. Furthermore, the forgetting factor approach is introduced to improve predictive accuracy and robustness of the model. The experiment results indicate that that improved neural network is of good accuracy and strong robustness in prediction, and can apply for the on line prediction of nonlinear multi input multi output systems like vehicle engines.