Using neural networks to forecast marginal emissions factors: A CAISO case study

Abstract

Marginal Emissions Factors (MEFs) quantify the time-dependent changes in CO2 emissions resulting from changes in electricity consumption. Accurate MEFs are critical for calculating the emissions impact of demand-side management (DSM) activities and programs, but current methods of calculating MEFs are limited by their temporal resolution, accuracy (particularly in grids with high penetrations of variable renewable energy), and ability to predict MEFs ahead of time, reducing their utility for DSM. We improve upon existing techniques by introducing a novel multi-layer perceptron to linear composite model that uses publicly available grid data to calculate historical MEFs and predict day-ahead MEFs. We test our model on publicly-available data from the California Independent System Operator over the period of 2019–2021, a grid with high daytime VRE generation. Results indicate that our model produces more accurate and more granular demand-based MEF estimations than comparable regression techniques and maintains high accuracy when use to forecast future MEFs. Our MEF framework can be applied to other regional grids to evaluate and design DSM strategies that leverage CO2 emissions-reductions as motivation for altering electricity consuming behaviors.