NEURAL NETWORK APPROACH TO MULTIPERIOD PLANNING PROBLEMS

Abstract

A conventional way of solving large optimization problems, with all but one decoupled constraint,is to augment the objective function with the coupled constraint using the Lagrangian multiplier. Conventionally the method of finding this multiplier is long and iterative, taking up a lot of computational effort and time. For sensitivity studies, in power systems planning, this kind of problem will need to be solved very often. In this study a methodology has been proposed to replace the time consuming method of search for the optimal Lagrangian multiplier by a Feedforward Neural Network. The optimal value of the Lagrangian multiplier can be greatly help in problem decomposition, greatly reducing the number of variables in each subproblem. A Feedforward Neural Network has been trained using the Backpropaga.tion Algorithm. This network has the right hand side constraint values as its inputs and the optimal value of the Lagrangian multiplier A as its output. So it represents the complex mapping of different constraints to the corresponding X.