Machine Learning and Its Application at Nooksack Falls Hydroelectric Station

Abstract

Copyright © 2005, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. The objective of this project is to control water delivery and distribution at Nooksack Falls Hydroelectric Station (NFHS) in order to maximize efficiency of the system, thereby increasing energy generation. Two machine learning algorithms will be applied. (1) Q-learning a reinforcement learning approach to obtain a set of roughly optimal configurations. (2) Recurrent Neural Network (RNN) rough configurations gathered by the Q-learning agent will be used to train the RNN. The RNN will refine these configurations as well as enabling lifelong learning. The significance of this project is to demonstrate the practical utility of the machine learning techniques described above when applied to realworld processes such as NFHS. Utility process control has advanced towards increasing productivity and reliability, while decreasing operating costs by replacing humans with automated systems. When NFHS began operation in 1906, it required 24-hour manned operation, a job shared by several men with families living on site. In the 1970s a system of electromechanical relays, DC positioning motors, and low pressure hydraulics were added, decreasing the need for operator attendance. In 2003 a more sophisticated set of controls were added including: high pressure hydraulics, a programmable logic controller (PLC), and a host of relay and analog sensors providing information on plant state. The evolved control configuration outlined above is typical of modern utility process control. Sensor inputs, along with a control scheme devised by plant operators and engineers are programmed into the PLC, which operates the plant according to adjustable set-points. This works well for static processes, those without a great variance of operating conditions. However, NFHS is a dynamic process. Fluctuating river flow conditions, waterborne debris, equipment wear, and mechanical problems are variables impacting NFHS's energy production. For the process to remain optimal, set-points should be adjusted accordingly.