StreamDL: Deep Learning Serving Platform for AMI Stream Forecasting

Abstract

Advanced Metering Infrastructures (AMIs) facilitate individual load forecasting. The individual load forecasting not only improves the accuracy of aggregated load forecasting but is a fundamental component of various power applications. With the highlight of deep learning (DL) in the individual load forecasting, a serving platform specialized in deep learning is required to forecast with AMI stream data. However, the existing serving platforms for DL models do not consider stream data as an input but usually support image or text data through RESTful API. To solve this problem, we propose StreamDL that is a serving framework providing deep learning inference with AMI stream data. It leverages Apache Kafka to support stream data and Kubernetes to support the cloud environment. StreamDL considers the specific requirements for stream data, which supports stream parsing to fit any DL model especially recurrent network and continual training to alleviate accuracy degradation by the change of stream distribution. In this paper, we introduce the detail of the StreamDL platform and its use-cases using real AMI data.