Carbon-Aware Microgrid Scheduling with End-to-End Learning

The pressing challenge of climate change has accelerated global initiatives to decarbonize power systems through the adoption of renewable energy sources. While these technologies significantly reduce greenhouse gas emissions, their variability and intermittency pose challenges to grid stability, particularly in microgrids with higher renewable penetration and lower inertia. In this paper, we propose a novel end-to-end learning approach for carbon-aware microgrid scheduling that seamlessly integrates forecasting and optimization into a single neural network, aligning the entire process with the objective of minimizing operational costs. Unlike conventional microgrid scheduling methods that follow a predict-then-optimize paradigm, the proposed approach effectively mitigates cumulative errors arising from the misalignment of objectives between the forecasting and optimization stages. Our approach also embeds carbon constraints directly into the optimization process by quantifying carbon penalties and incorporating technologies such as carbon capture and storage to support prompt decarbonization. Through a case study, we demonstrate that our approach significantly outperforms conventional methods by achieving lower mean squared error (MSE) and greater accuracy in multiple test scenarios, which showcases its potential for delivering efficient, reliable, and environmentally sustainable microgrid scheduling.