An incremental constructive layer neural network based power system stabiliser

In this paper, a systematic approach for a neural network based Power System Stabiliser (PSS) design using an incremental constructive layer algorithm is presented. The proposed design is essentially based on incremental training and network pruning with an intermediate phase of cross validation. The algorithm starts training the network with a single node in the hidden layer with one data point. After training the first data point with required number of hidden nodes, successive data points are trained incrementally with the architecture growing automatically as needed. After completing the training of the network with the one data set, it is validated against the remaining data sets and the resultant network is finally pruned using a measure of goodness factor of each node. The design has the advantage in that, if the operating region changes, the network can be incrementally trained with little training. The proposed design has been implemented both for identification and control of the power plant. Satisfactory closed loop time responses are obtained at the nominal operating point as well as within ±20% variation with the implementation of the proposed NN based PSS.