AI and Machine Learning are undoubtedly soon to become pervasive technologies across society. The range of applications and the power of deployable solutions is proliferating. With this comes the growing need for replicable, robust, monitorable, maintainable, and scalable solutions.
This set of properties forms the basis of what we can call Industrial-Grade Machine Learning. Apart from simple one-of-a-kind or proof-of-concept applications, handcrafted models deployed, for example, as dedicated web services are insufficient when the value to be extracted spans a large number of assets, each requiring a personalized model.
The energy and power utilities sectors are characterized by high multiplicities that require scalable and replicable solutions. A representative example is load monitoring and forecasting at secondary substations. A typical electrical power distribution utility has from a few hundred to several thousand substations, each one having a unique load profile and requiring a tailor-made prediction model. Manually developing, deploying, integrating, operating, monitoring, and maintaining thousands of prediction models is neither feasible nor sustainable. Similar examples from the same industry are production forecasts for residential solar power installations, load forecasts for EV charging stations, flexibility forecasts for prosumers, and automated analysis of inspection data, such as imagery.
Industrial-Grade Machine Learning entails, in our view, a set of basic requirements, illustrated in the following example of predictive analytics services, though the same principles apply equally well to inspection and other analytics services.
Standardization. Prediction services require a unified interface (API) to facilitate standardization, maintainability, and interchangeability between services. This also allows “consumers” to query the forecasting capability of each prediction service.
Logging and monitoring. The activity and performance of each prediction service should be logged and monitored to ensure traceability and early detection of performance degradation which might trigger a prediction model update or re-training. Supporting tools for large-scale monitoring such as graphical dashboards provide additional support.
Automated Routing. A prediction request should be automatically routed to the best available prediction service that can handle the request. Alternative services can potentially be available that differ in performance profile and data requirements, and a mechanism to select the best available service given the current conditions and constraints makes the system significantly more robust, for example in the case of missing data.
Automated Re-training. Re-training prediction service instances, either on a schedule, on demand, or triggered by observed performance degradation, must be possible.
Continuous Integration. Changes in prediction service code, trigger automated unit tests, code quality checks, configurations, and deployment.
As machine learning and AI applications continue to mature, there will be an increasing need for mass-produced intelligent systems, replacing handcrafting with automation, just as conveyor belts and assembly lines replaced artisan workshops and revolutionized the manufacturing industry.
At eSmart Systems, we have a continuous focus on Industrial-Grade Machine Learning. It spans from large-scale cloud deployment of deep learning models for automated image analysis to our newest Robust Predictive Analytics Architecture (RPAA) for deploying prediction models at scale. We firmly believe this is a fundamental strategic advantage for delivering economy of scale value to the energy and power utilities sector.