evoDLA: A Semantic Learning System for Predictive Cost Minimization in Power Generation System

Abstract

In this paper, we propose evolutionary deviant learning system (evoDLA) for optimal power generating systems. It is a temporally structured learning system for the predictive cost minimization task at the next time step expressed in a semantic way. Cost minimization in power generation is an assignment problem that requires the determination of the best set of operating parameters to enable generators perform optimally. Presented as an alternative to the conventional unit commitment strategies, the predictive system is designed to operate in a semi-supervised manner allowing the dynamic superimposition of linguistic memory units into the mixed-integer power system generation data. A set of representative linguistic biased integers are programmed into an equivalent fuzzy-like integer representation. These are transformed through a series of temporal deviant states and then re-transformed back into true linguistic form. Simulation studies are performed in comparison with the Long Short-Term Memory (LSTM), a proven artificial neural network (ANN) method for sequential learning tasks. The results show competitive performance with the present system and the unique capability of the novel system in inferring the optimal set of output generation parameters.