Probabilistic forecasting of residential electric vehicle charging demand for low voltage distribution network planning

Abstract

The global decarbonisation effort is accelerating transport electrification, driving increased Electric Vehicle (EV) adoption. Local clusters of EV ownership and variations in charging behaviour introduce significant spatiotemporal uncertainty in future residential charging demand on Low-Voltage (LV) distribution networks. Compounded by limited visibility at the LV level, this presents major challenges for Distribution Network Operators in planning investments and ensuring adequate network capacity. While the growing availability of open energy data presents exciting modelling opportunities, the data can be prone to inaccuracies and potential data drift. Integrating data from other sectors could enhance insights for the LV network, yet scalable and transparent methods for such integration are absent. Moreover, current EV adoption forecasts often lack the validation, spatial granularity, and uncertainty estimation needed for reliable LV network planning, and their alignment with real LV network topologies and principled risk quantification remains unresolved. This thesis presents a novel probabilistic framework for forecasting EV adoption and charging demand on the LV network, integrating diverse open datasets with advanced statistical methods to deliver uncertainty-aware forecasts that can support evidence-based investment decisions. At the framework’s core is a novel EV adoption forecasting approach that uses Bayesian inference and Gaussian process regression to integrate historical registration data with government mandates, producing accurate forecasts alongside calibrated uncertainty estimates. Validated for 115 Lower Layer Super Output Areas in Bath and North East Somerset (BANES), the framework significantly outperforms deterministic baseline models, achieving Mean Absolute Errors (MAEs) of 1.2, 0.9, and 0.4 for five-, three-, and one-year-ahead forecasts of EV market share. Another key contribution of this work is the fusion of multiple open datasets. Census data is used to calibrate vehicle registration data and estimate its uncertainty, while a novel mapping approach translates administrative-level forecasts to the LV network, accounting for spatial misalignment and uncertainty. A hierarchical Bayesian model is developed to predict probabilistic, scalable and validated EV demand profiles. Applied to 1,125 BANES LV substations, the framework reveals that 20.3% of substations could
surpass an exceedance probability of 0.1 by 2035 without mitigation. However, a 50% reduction in peak charging probability through smart charging could lower this to 14.7%. These findings highlight the critical role of open data and probabilistic risk quantification in enabling adaptive, cost-effective, and robust network planning strategies.

Date of Award	25 Jun 2025
Original language	English
Awarding Institution	University of Bath
Supervisor	Furong Li (Supervisor), Julian Padget (Supervisor) & Chris Brace (Supervisor)