Large-scale analytics for sustainability under data and model variability

Département Information Systems et Operations Management

Intervenant: Vassilis Digalakis (MIT)

Salle T004

Abstract:

Analytics provide unique opportunities to effectively combat major societal issues, such as climate change. Modern problems present new challenges, including large volumes of data or decisions, and (e.g., temporal) variability in the data or the underlying model. I will present my work in developing prescriptive and predictive methodologies that address the aforementioned challenges in impactful, real-world applications in the sustainability and energy space.
In the first half of the talk, I will describe a machine learning- and robust optimization-based methodology for solar capacity expansion, which we develop in collaboration with OCP, one of the world's largest fertilizer producers, to guide their 1 billion USD investment in solar energy. Our model forecasts to reduce OCP's carbon emissions by over 60% and to save 2.5 billion USD in operational costs over the next 20 years.
In the second half of the talk, I will focus on an energy consumption prediction application, and introduce the novel framework of slowly varying machine learning, whereby the underlying model varies smoothly under some graph-based temporal or spatial structure. For sparse regression models, I will present new theoretical advances for the underlying mixed integer optimization problem; for decision trees, I will show how we address long standing challenges regarding such models' stability.