Optimization-in-the-Loop AI for Energy and Climate - IPAM at UCLA

Overview

Explore optimization-in-the-loop AI techniques for addressing climate change and energy challenges in this comprehensive lecture. Delve into the integration of artificial intelligence and machine learning methods with physics-based constraints and complex decision-making processes. Learn how this framework can be applied to design learning-based controllers that enforce stability criteria and operational constraints in low-carbon power grids and energy-efficient buildings. Discover task-based learning procedures that consider downstream decision-making processes, significantly improving performance and preventing critical failures. Gain insights into differentiable optimization, deep reinforcement learning, robust control, and decision-cognizant approaches for demand forecasting and power system optimization. Understand the potential of these innovative techniques in unlocking AI and ML capabilities for high-impact climate action problems.

Syllabus

Climate change warrants rapid action
Climate & energy problems involve physics, hard constraints, and decision-making
Machine learning methods struggle with physics, hard constraints, and decision-making
Optimization-in-the-loop ML
Talk outline
Overview: Differentiable optimization
Background: Deep learning
Differentiating through optimization problems
Follow-on work in differentiable optimization
Overview: Enforcing hard control constraints
Deep reinforcement learning vs. robust control
Differentiable projection onto stabilizing actions
Details: Finding a set of stabilizing actions Insight: Find a set of actions that are guaranteed to satisfy relevant Lyapunov stability criteria at a given state, even under worst-case conditions
Illustrative results: Synthetic NLDI system
Energy-efficient heating and cooling
Differentiable projection onto feasible actions
Results on realistic-scale building simulator
Summary: Enforcing hard control constraints
Overview: Incorporating downstream decision-making
Decision-cognizant demand forecasting
Decision-cognizant approach can dramatically improve generation scheduling outcomes
Approximating AC optimal power flow
Approximate robust power system optimization
Summary: Incorporating downstream decision-making