Advances in the Trustworthy Machine Learning Pipeline
Abstract:
Virtual power plants (VPPs) are decentralized networks of distributed energy resources (DERs), e.g., photovoltaic panels, wind farms, electric vehicle batteries, thermostats, and other controllable grid-edge energy resources, that are collectively managed and operated as a unified entity through a central control system. Instead of generating electricity from a single location like typical power plants, VPPs mobilize and incentivize geographically dispersed generation, storage, and consumption to offer a vast range of grid services. It relies on the forecast, the optimization, and the control of these distributed assets to exploit flexibility and meet grid needs with very limited infrastructure change. The VPP is essential for a green and resilient grid and algorithms are fundamental in its making: optimization methods need complex machine learning (cML) to forecast and model the intricacy of the system, and ML models need to be integrable in control methods to exploit them while enforcing safety constraints. As much as the future success of the VPP stands on cML, the sensitivity of the application and the scale of potential failures slow down its industrial acceptability. The issue is that cML historically lacks explainability and interpretability, has limited out-of-the-box performance guarantees, can be unpredictable and thus hard to control, requires a lot of data to perform as expected, is sensitive to data corruption, and is associated with complex training procedures that do not scale well with large deployments. These flaws have labeled it as unreliable for critical applications such as the VPP. In this presentation, we overview the requirements to create a trustworthy ML pipeline for the VPP from pre-processing to exploitation and delve into some of our contributions and present work: sliced-Wasserstein-based outlier detection methods for out-of-sample data filtering, Wasserstein distributionally robust shallow convex neural networks (WaDiRo-SCNNs) for trustworthy low-stochasticity training, physics-constrained SCNNs for added guarantees when learning nonlinear dynamical systems, and safe data-driven nonlinear model predictive control.
Biography:
Julien Pallage is a Research Master's student at Polytechnique Montréal (University of Montreal) and an affiliated student researcher at GERAD and Mila. He obtained his Bachelor's in Electrical Engineering at the same institution and presently is an NSERC and FRQNT Master's scholar. His research interests comprise trustworthy machine learning, optimization, and control theory for critical applications in the energy sector.
Virtual power plants (VPPs) are decentralized networks of distributed energy resources (DERs), e.g., photovoltaic panels, wind farms, electric vehicle batteries, thermostats, and other controllable grid-edge energy resources, that are collectively managed and operated as a unified entity through a central control system. Instead of generating electricity from a single location like typical power plants, VPPs mobilize and incentivize geographically dispersed generation, storage, and consumption to offer a vast range of grid services. It relies on the forecast, the optimization, and the control of these distributed assets to exploit flexibility and meet grid needs with very limited infrastructure change. The VPP is essential for a green and resilient grid and algorithms are fundamental in its making: optimization methods need complex machine learning (cML) to forecast and model the intricacy of the system, and ML models need to be integrable in control methods to exploit them while enforcing safety constraints. As much as the future success of the VPP stands on cML, the sensitivity of the application and the scale of potential failures slow down its industrial acceptability. The issue is that cML historically lacks explainability and interpretability, has limited out-of-the-box performance guarantees, can be unpredictable and thus hard to control, requires a lot of data to perform as expected, is sensitive to data corruption, and is associated with complex training procedures that do not scale well with large deployments. These flaws have labeled it as unreliable for critical applications such as the VPP. In this presentation, we overview the requirements to create a trustworthy ML pipeline for the VPP from pre-processing to exploitation and delve into some of our contributions and present work: sliced-Wasserstein-based outlier detection methods for out-of-sample data filtering, Wasserstein distributionally robust shallow convex neural networks (WaDiRo-SCNNs) for trustworthy low-stochasticity training, physics-constrained SCNNs for added guarantees when learning nonlinear dynamical systems, and safe data-driven nonlinear model predictive control.
Biography:
Julien Pallage is a Research Master's student at Polytechnique Montréal (University of Montreal) and an affiliated student researcher at GERAD and Mila. He obtained his Bachelor's in Electrical Engineering at the same institution and presently is an NSERC and FRQNT Master's scholar. His research interests comprise trustworthy machine learning, optimization, and control theory for critical applications in the energy sector.
Practical information
- General public
- Free
Organizer
- Professor Giancarlo Ferrari Trecate