How to Measure the Impact of Electric Vehicles on Electricity Demand

Leveraging geospatial data is crucial for companies that want to bolster their digital strategy. Whether it's to win the race to build autonomous vehicles, to gain a competitive edge by locating potential customers or to optimise e-mobility solutions, the combination of Artificial Intelligence and Geographic Information Systems ("GIS") has revolutionised the way that companies make decisions. The volume of location-based data has grown exponentially in the last decade, generated by multiple sources such as smartphones, IoT sensors and drones. In our ever-growing digital world, geographic data science is indispensable for companies that want to leverage their data – scientists indicate that up to 80% of all data is currently georeferenced. Our UK Data Science team are experts in using data science methods to analyse and visualise spatial data, with particular experience in e-mobility. We use advanced data science techniques to find the optimal location to install electric vehicle chargers, given the spatial distribution of charging demand across a city or region.


As many companies and individuals turn towards clean energy solutions, the global electric vehicle ("EV") fleet will continue to expand significantly in the years to come. In its 2020 Electric Vehicle Outlook report1 , BNEF outlines that the size of the global EV fleet will rise from 8.5 million in 2020 to 116 million in 2030. The EV share of new car sales will go from 10% of global passenger vehicle sales in 2025 to 28% in 2030, and then to 58% in 2040.

In that context, how can we measure the impact of EVs on the electricity demand at scale, and satisfy the energy demand for these vehicles? Our FTI Consulting London Data Science team were tasked with the following challenges: How can we build a solution that allows us to forecast the demand for EV charging at different points in time and across Greater London? How can we make sure the energy supply is optimised given the current infrastructure? Can we suggest the best locations for new charging stations to meet the forecasted demand? Finally, can we make sure that our model is flexible enough to test a range of scenarios back-to-back? How can we ensure that the model is generalised enough to easily adapt to different geographies, while at the same time modularised enough to continually learn from new driver data?

Building a large-scale simulation of EV journeys

A highly flexible and optimised simulation model was used to simulate EV trips in London. The aim was to determine where and when drivers charge. This would facilitate a better understanding of how charging stations are being used and where new infrastructure would be required. We developed an agent-based model to achieve this goal.

At its core, agent-based modelling ("ABM") is a process by which microscopic entities are simulated to observe macroscopic effects. Agents - as individual autonomous units - operate based on a set of defined rules and processes applied to their state and environment. Some of these processes may be sampled from data-driven probability distributions using Monte Carlo methods. Taking this one step further, it is possible for the distributions to be conditionally linked. Each time an agent undergoes a process, its new state will dictate the distributions used for the next. This series of events creates a Markov chain, and the random sampling used follows Monte Carlo methods. Thus we obtain a Markov Chain Monte Carlo ("MCMC").

"Whether it's to win the race to build autonomous vehicles, to gain a competitive edge by locating potential customers, or to optimise e-mobility solutions, the combination of Artificial Intelligence and Geographic Information Systems has revolutionised the way companies make decisions."

Before the advent of ABM, geographical systems were represented by considering a population en-masse. These approaches allowed large-scale effects to be observed but were frequently limited by complexity and disequilibrium. Early attempts to separately consider all of the individual components of a system were hampered by technological limitations. Nonetheless, some applications succeeded. Economist Thomas Schelling managed to simulate the process of class and racial segreation2 . Where we now use random number generators and computer bits, Schelling used coin flips pencil and dots on graph paper.

As computational power and capabilities increased, it became increasingly feasible to separately consider substantial numbers of components of a system, leading to larger-scale simulations3 .

Simulate future scenarios for EV journeys in London

In our agent-based model for EV trips, drivers were treated as agents. The model was able to simulate a full calendar day for 10,000 drivers in a matter of a few minutes. Trip data from millions of recorded trips was analysed and condensed into a series of multidimensional probability matrices. Based on the driver's current state, these dictated factors including where they would be likely to travel next and how they would transition between different modes. These modes included taking a trip with a passenger, driving to such a trip, taking a break, charging and driving to and from home. For each agent, their home location and the time period for which they would aim to drive was also drawn from a probability distribution. The model was able to track drivers'

Our team was able to combine a variety of data sources to analyse EV journeys, including Zap-Map data (an EV charging platform showing the charge point networks), UKPN data and even weather data to see whether adverse weather would affect the patterns of EV journeys. Using our GIS data science expertise, we also supplemented these data sources with hand-picked open source data such as road networks and statistical boundaries.

Mapping London's roads

To understand road traffic in the greatest possible detail, the team were able create and manipulate a customised road network. Working with almost 400,000 sections of road, advanced graph operations needed to be applied to understand how journey data related to individual roads. The choice of network was also very important. We aimed to cover our area of interest, while minimising computation time and allowing all calculations to run. The network that we used was in the form of a directed multigraph. Each section of road had an associated direction in which traffic could move, and multiple sections could run in parallel, representing two-way streets and alternative routes.

Before data was incorporated into the system, the network was divided into strongly connected components. These are defined as sub-networks in which every node (a point on the network) is accessible from every other node.


1 BNEF, Electric Vehicle Outlook 2020,

2 Schelling, Dynamic Models of Segregation, 1971

3 Hanappi, Agent-Based Modelling History, Essence, Future, 2017; Crooks and Heppenstall, Introduction to Agent-Based Modelling, 2012

To read the full article click here

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.