A Scalable and Reconfigurable Machine Learning and Deployment Platform for Power Plant

Abstract

With the rapid development of the Internet of Things(IoT) and industrial automation systems, the data accumulated by the industry is exponentially increasing. With the dynamics of the system, this platform should support different, possibly diverse types of models with different resource requirements. The life-long data collection and analysis for the complex coal-fired Power Plant requires a Machine Learning (ML) and deployment platform that is scalable and reconfigurable. This paper proposes a scalable and reconfigurable ML platform for a power plant based on docker technologies that support online model deployment, execution, and scheduling. In order to support the model retraining, a mechanism is proposed to manage the execution of the model training and the seamless transitions between models, without interrupting the online operation of model serving. This platform has been deployed in a power plant with two coal-fired units for about five months. Results of the field test prove that this system is flexible, model reconfigurable, and can achieve smooth model switches with minimal overhead.