Power generation scheduling by neural network

A new method for solving a power generation scheduling problem in an electric power system is presented. The objective is to determine the hourly start-up/ shut-down schedules of all generators so that forecasted hourly power demands per day may be met and total operating costs, the sum of setup and fuel costs for a given day, may be minimized. The problem may be formulated as a large-scale combinatorial optimization problem which includes 0-1 variables representing the start-up/shut-down of generators and continuous variables representing the power outputs. Determination of an optimalsolution within practical time limits is consequently difficult. Until now, the lagrangian relaxation method has been studied as it appeared to be the most practical method for obtaining an approximate solution to the problem. The efficiency of this method, however, depends on how the Lagrange multipliers are determined. Here, it is proposed that the Lagrange multipliers be estimated by utilizing the neural network and results determined from examination of the possibility of applying the backpropagation algorithm to pattern recognitions which presume the relationship between power demand pattern and Lagrange multipliers are reported. Through numerical experiments, it was established that the Lagrange multipliers, estimated by the neural network, are applicable to the problem.