Power plant operator assistant: An industrial application of factored MDPs

Abstract

Markov decision processes (MDPs) provide a powerful framework for solving planning problems under uncertainty. However, it is difficult to apply them to real world domains due to complexity and representation problems: (i) the state space grows exponentially with the number of variables; (ii) a reward function must be specified for each state-action pair. In this work we tackle both problems and apply MDPs for a complex real world domain -combined cycle power plant operation. For reducing the state space complexity we use a factored representation based on a two-stage dynamic Bayesian network. The reward function is represented based on the recommended optimal operation curve for the power plant. The model has been implemented and tested with a power plant simulator with promising results.