Under the hood of electric vehicle grid balancing

Electricity markets are complex, but one thing is sure: keeping the grid in balance is an absolute necessity. It drives the incentives and penalties that market players pay and that influence our own energy bills. Today, we take a look at these mechanisms not from the EV driver’s seat like we usually do, but from an energy market player’s perspective. We explain how operating an electric car fleet makes grid balancing easier and how machine learning plays a role in this process.

Who is who?

Let’s focus on two major market players:

• Balance Responsible Parties (BRPs). These parties oversee a portfolio of energy generators and consumers and make sure that supply meets the demand for their portfolio as a whole. If not, requests are sent by the TSO (transmission system operator, such as TenneT in the Netherlands) to solve these imbalances. Who solves the imbalances?
• Balancing Service Providers (BSPs). BSPs are market participants who balance the grid by rapidly increasing or reducing their power output. A conventional example is a gas plant working at 80% capacity and bidding the remaining 20% as a balancing capacity in this balancing market. If the need arises, this capacity will be activated, and the plant will be paid for the amount of energy it provided. The plant will be actually paid twice: once for the “promised” capacity and once for the actual volumes delivered to the energy grid. While the first cost is spread across all actors on the energy grid, the second one is paid out by a BRP and is often passed on to utilities and consumers.

In this situation, the production of Balance Responsible Party (BRP) 1 exceeds the consumption, while for BRP 2 it’s the opposite. BRP 1 can sell its excess electricity on the Intraday market to BRP 2, which can use it to meet its needs. However, there’s still a shortage because of other market players. To fix this problem, a Balancing Service Provider (BSP) is called upon to increase production or reduce the use of smart charging EVs to bring more electricity to the grid. When the grid balance is restored, the parties that caused the imbalance will have to pay a fee to cover BSP’s expenses.

Flexible demand

It may seem like BRPs, and BSPs have conflicting interests, but they both rely on flexible supply. BRPs have difficulty accurately predicting consumption and pay imbalance costs to compensate for adjusting BSP’s supply. BSPs, on the other hand, manage the complex infrastructure of reserve capacities that could be used more efficiently in terms of cost and CO2 emissions.

But what if it’s not just the supply that is flexible, but also the demand? That could create a win-win situation for both parties!

According to a 2022 report by the International Energy Agency, by 2030 electric vehicles are expected to use about 5% of total electricity demand or 110 TWh in Europe. This estimate is based on the assumption that electric vehicles will have a market share of 30% in 2030, which takes into account current government policies.

Unlike many other types of energy consumption, the power demand from electric vehicles can be flexible. That’s why at Jedlix, we are working to enable the untapped flexibility of EV charging for grid balancing for energy market players.

Under the hood of a Virtual Power Plant

Let’s take a closer look at a typical session on the Jedlix platform. Imagine a user plugs in their car in the evening during the work week and expects the battery to be fully charged by the next morning. During these hours, electricity demand is usually high, which is reflected in the prices on the day-ahead market and, as a result, the user’s tariffs. However, by delaying charging until prices are lower and scheduling charging times based on price, the user can significantly reduce the cost of the session.

The optimal schedule is determined by tariffs and user settings (for example, leaving time or the presence of solar power) and can bring up to more than 30% of cost savings for this session for users with dynamic tariffs.

This flexibility in demand can also be utilized in other energy markets that the user does not have direct access to. This is where BRPs and BSPs come into play.

When combined, the pool of electric cars becomes essentially a decentralized battery that can be charged and discharged based on market signals. We call this battery a Virtual Power Plant (VPP). The assets needed for a VPP are already present within the BRP’s perimeter or BPP’s portfolio. The only problem is… operating thousands of vehicles is much more complex than steering one industrial-sized battery.

That’s where Jedlix comes in. We offer VPPs as turn-key solutions, similar to how a real battery works. We do this by:

• Ensuring we know the capacity of our VPP
• Having visibility on how fast our VPP can charge (or discharge) at any given moment
• Having a way to interact with the car without any additional hardware

In this blog post, we’ll take a closer look at the first two points. If you’re interested in how we control car charging remotely, you can check out our developer portal.

Capacity forecast

A major difference between a decentralized battery on wheels and its conventional stationary colleague is variable capacity. We simply don’t know what would be the number of cars plugged in at every moment of the future (and how many of them we can reschedule and still respect the user settings), which makes it hard to bid this capacity on the balancing market. However, we can forecast aggregated flexibility based on historical data, as it exhibits seasonality.

Available flexibility is highly dependent on the hour and day of the week. For example, more users stay connected over the night which increases the total available flexibility.

This seasonality is driven by electricity prices (think of peak/off-peak tariffs), consumer lifestyle (think of plugging the car in the evening after work) and even weather patterns. We capture those features separately for different fleets, as market prices and even timezones differ across the markets where Jedlix is active. In addition to that, we take into account the growing size of our virtual power plants by forecasting a scaled target that we rescale back at a later stage.

One of our go-to approaches for this is gradient boosting. It is a powerful machine-learning technique that combines the predictions of multiple decision trees to make more accurate forecasts and can handle both numerical and categorical features.

Currently, we are also experimenting with deep learning models and multivariate forecasting. As our models require historical data to train on, it can be challenging to create a forecast for a new fleet. “Reusing” the information from our other fleets will be especially helpful in this context.

Charge power estimations

So far we have talked about aggregated flexibility, however, it is the charging process of the individual assets we schedule to charge at the most appropriate time. To create such optimal schedule, we need to know how fast every asset in our pool can consume energy from the grid, or simply the charge power.

Since we mentioned charge power, many wonders, why this is unknown if the car manufacturers often list it in the specifications. In fact, charging power depends more on the charging setup rather than the car model. Let’s take a detour here.

A power line can have 1 or 3 phases. The current (or the flow of energy, that we denote I) in the cable is defined in Amps: most households have 16 Amps available at home. The “pressure” on the line is defined in Voltage (V), which in Europe mainland is mostly 220–240v. This brings us to the four most common charge powers possible while charging at home.

Charge power distribution from a subset of sessions recorded on the Jedlix platform in the years 2021–2022.

However, if we take a look at charge power distribution at the Jedlix platform, we would notice, that it’s not just those four most frequently occurring values, but the whole range between them. Also, you can see the peaks close to the ones we computed before, but shifted to the left. This is due to the losses, individual properties of the car, dependency on the battery state of charge, and many other things.

In the graph below, we depicted the charge powers for the exact same vehicle that had sessions in different locations. Here we used averaged charge powers computed after-the-fact, however, in reality, these estimations can vary in the session as well.

Charge power computed after the session varies from one charge location to another, but also can be different at the same location.

That is why at Jedlix, to schedule cars optimally and provide an accurate estimate of available flexibility, we have to determine the charge power based on the car’s previous activity on the platform and update our estimates during the session itself.

By the way, we have recently updated our charge power estimation model tailored specifically for session start. Our previous algorithm was designed to push the car into calibration mode even by the slightest variance in estimated charge powers from previous sessions. This led to a brief charging period at the start of each session, which was often during expensive hours for our users and reduced flexibility for our VPP.

Model step-by-step: unlike deterministic model, the output is a distribution that we can use to incorporate uncertainty in our estimates.

The new probabilistic model allows us to optimize the charging process for EVs with unstable previous charge powers sooner, resulting in less need for charge power calibration. This helps us maintain the same level of confidence that the battery is filled to the desired state of charge as before while also delivering flexibility to the grid.

To handle the low-latency and high-throughput requirements of our charge power estimation model, we use the NVIDIA Triton inference server. This allows us to scale our forecasting service to provide charge power estimates as our platform grows!

Conclusions

To sum up, we walked you through some of the challenges we solve with machine learning to make operating the Virtual Power Plan under uncertainty a reality. This includes (but is not limited to) available flexibility forecast and real-time charge power estimations. Want to know more about how we do this at scale? From smart charging APIs to complete Virtual Power Plants, get support for any case you need to accelerate your e-mobility offering by heading to our website or contacting us directly.