Power plant start-up scheduling: a reinforcement learning approach combined with evolutionary computation

Article type: Research Article

Authors: Kamiya, Akimoto | Kimura, Hajime | Yamamura, Masayuki | Kobayashi, Shigenobu

Affiliations: Toshiba Corporation, 1-1, Shibaura 1-Chome, Minato-ku, Tokyo 105-8001, Japan. E-mail: akimoto.kamiya@toshiba.co.jp | Tokyo Institute of Technology, 4259, Nagatsuta, Midori-ku, Yokohama 226, Japan. E-mail: gen@fe.dis.titech.ac.jp, my@dis.titech.ac.jp, kobayasi@dis.titech.ac.jp

Abstract: Power plant start-up scheduling is aimed at minimizing the start-up time while limiting turbine rotor stresses to acceptable values. In order to increase on-line performance of searching an optimal or near-optimal start-up schedule during power plant operation, we propose to integrate neural network-based reinforcement learning with evolutionary computation implemented by means of Genetic Algorithms (GA). GA guides reinforcement learning to learn optimal schedules with respect to a number of representative sets of stress limits prior to the start-up process. During start-up, GA combined with reinforcement learning will search an optimal or near-optimal start-up schedule at a given set of stress limits. This approach significantly reduces the time needed for learning and searching. On a SPARC station 20, experiments show that it can search an optimal or near-optimal schedule within tens of seconds of CPU time, a time range which should be acceptable in power plant operations.

Journal: Journal of Intelligent and Fuzzy Systems, vol. 6, no. 1, pp. 99-115, 1998