Comparative performance study of QN and LM algorithms in predictive control for NNARX-identified model of hydro-power plant

摘要

In this paper input-constrained predictive control strategy for NNARX (neural network non-linear auto-regression with exogenous signal) model of hydro-turbine is presented. The input (gate position) and output (turbine power) data are generated by means of dynamic plant model. The collected data are utilized to develop the NNARX model of the plant. Then NN-based predictive control (NNPC) scheme is applied to control the turbine power. The control cost function (CCF) includes the squared difference between the model predicted output and desired response and a weighted squared change in the control signal. The CCF is minimized with both Quasi-Newton and Levenberg— Marquardt iterative algorithms. To demonstrate the suitability of the strategy, the plant has been simulated on two different reference signals.