Deployment and Management of Time Series Forecasts in Ocean Industry

Abstract

Machine learning has not achieved the same degree of success in environmental applications as in other industries. Challenges around data sparsity, quality, and consistency have limited the impact of deep neural network approaches and restricted the focus to research applications. An alternative approach - that is more amenable to the characteristics of data coming from disparate IoT devices deployed at different times and locations in the ocean - is to develop many lightweight models that can be readily scaled up or down based on the number of devices available at any time. This paper presents a serverless framework that naturally marries a single IoT sensor device with a forecasting model. Aspects related to data ingestion, data processing, model training and deployment are described. The framework is applied to a fish farm site in Atlantic Canada.