Electric vehicle charging management through LSTM-based prediction of non-critical charging sessions

Source abstract

Battery electric vehicles (BEVs) reduce more GHG emissions when charging aligns with low-carbon electricity. Many real-world sessions happen during high-emission hours but are not necessary for completing daily travel. This study develops a deep learning model that helps drivers skip or shift these non-critical sessions to cleaner hours without changing travel routines. We train a Long Short-Term Memory (LSTM) network using second-by-second data from 64 BEVs. The model forecasts travel needs for the next two days and identifies non-critical charging sessions. It then shifts those sessions to lower-emission hours. Results show up to 41% GHG reduction and 49% fewer charging days. A 6kW charger performs nearly as well as a 19kW charger without requiring costly infrastructure upgrades. Tests in CAISO, PJM, and ERCOT show consistent improvements, even in grids with less daytime renewables. This approach offers a scalable path to decarbonize transport without new infrastructure or changes to driver behavior.