Feed-forward control of thermal power plants using neural networks

In thermal power plants, it is an important theme to improve the control performance of main steam pressure and temperature etc. during load up/down. This paper focuses on temperature control that is the most difficult problem due to the non-linearity and long dead times of power plants. Model Reference Adaptive Control (MRAC) is applicable to the feed-forward control of power plants, but there are some problems. The most serious problem is that persistently exciting (PE) condition is not satisfied, and so it is difficult to estimate plant parameters using the well-known recursive least squares method. It is proposed in this paper that Jacobians of the neural networks (NN) are applied to identify the above mentioned plant parameters and control law is obtained by two methods, that is, one is the method to use the Jacobians of the NN plant model which is obtained by off line forward model learning, the other is the method to utilize the Hessian of the cost function. This method is evaluated by a detailed simulator that represents accurately the dynamics of power plants, and usefulness and effectiveness of the proposed method is proved.