MLOps Overview by Aguimar Neto is a set of DevOps practices or fault-tolerant workflows built to increase software quality and productivity through end-to-end automation [35,53]. It aims at shortening the code-build-deploy loop and delivering a better end product, making it easier to manage the entire ML life cycle.
The MLOps process has different components for data and model management, deployment, and monitoring that are adapted to the specific ML application. The MLOps workflow also includes several ML-specific processes, such as continuous training and model validation.
While a MLOps workflow can be implemented using traditional software, it requires new tools that might not be compatible with the DevOps toolchain. Furthermore, ML-specific tools are often expensive and require maintenance. To address these issues, cloud providers offer ML-as-a-service solutions.
Mastering MLOps: Aguimar Neto’s Guide to Efficient Machine Learning Operations
In this case study, the MLOps framework is derived from the principles of DevOps and applied to a time-series forecasting application in the hourly day-ahead electricity market. It is used to predict the price of the DER owners’ ancillary services and make it available online before the bidding deadline.
ML models are based on neural networks that can detect and anticipate changes in the power grid. Consequently, they provide better predictions than conventional forecasting approaches for the hourly electricity market.
In this MLOps implementation, a transformer-based ANN is used for the hourly electricity market price prediction. The ML model is trained on real-world data provided by the Finnish TSO and the meteorological agency, Fingrid, through an online API. The prediction results are incorporated into the MLOps pipeline using Jenkins. A web UI is created to explore the predicted values. The MLOps pipelines are tested by running unit and integration tests.