NEURAL NETWORK APPROACH TO UNIT COMMITMENT PROBLEM

Abstract

As generating facilities are expanding, power system operations have become complex. Daily load patterns in many power houses exhibit extreme variation between the peak and off-peak demand. So to meet the load, the power generation should be adjusted throughout the scheduling horizon. It should be decided which unit is to taken on line and for how much time. The Unit Commitment is an important optimization task in the daily operation for planning of power system. The task involves the optimum scheduling of the units , such that the production cost including fuel cost, start-up and shut-down cost is minimum. The problem should be fast and flexible enough to handle numerous operating restrictions as well as conditions. In the present work a methodology using neural network has been developed to solve unit commitment problem. This method has two steps in the first step the unit commitments are determined using a trained BackPropagation network and in. the second step the power generation by committed units are determined using simulated annealing method. A method is also presented for generating the training pattern. , The 3 layer feedforward neural network has been used. This network is trained by a set of load profiles and their corresponding commitment schedule by using gradient algorithm. The simulated annealing method works iteratively in two phases, in first phase the production cost is evaluated- at a number of feasible points and in the second phase, these points are replaced by local searches to obtain global minima. The annealing concept is used to jump out of the local optimum solutions to global solution. The generator scheduling problem involving 10 units and 24 hour time period is tested on this method.