Decentralized Importance Sampling in Variational Autoencoders to generate Industrial Carbon Emissions Scenarios

Decarbonization of energy-intensive industries like chemical process heating is expected to rely heavily on electrification through existing, renewable-driven power system infrastructure. Operations planning for decarbonized industrial and power system processes requires data-driven solutions that account for uncertainties across both sets of stakeholders. Stochastic optimization (SO) solutions for decision-making under such uncertainties require data-driven scenarios that capture spatiotemporal interdependencies the fidelity of which can severely impact SO solution quality. Additionally, the distribution of generated scenarios must also be estimated to align closely with target distributions contained in stakeholder datasets. Finally, these scenarios must also be generated from siloed data across diverse industrial stakeholders regarding data residency and privacy requirements. In this talk, we explore the potential of variational auto-encoders (VAEs) and their decentralized variants in importance sampling - a method for computing mathematical expectations with respect to a target distribution - on scenario distribution modeling to address these challenges effectively. Using real-world carbon emissions data, we demonstrate how these models capture spatial and temporal interdependencies across multiple chemical plants to generate high-fidelity scenarios. Our results also highlight how these generated scenarios, along with their associated importance weights obtained from decentralized training and inference, are utilized to drive reliable and efficient SO solutions when compared with other state-of-the-art methods.