Looking for the live map? It now lives here
Imagine you’re building a system that can decide the optimal time to charge electric vehicles or the optimal time to run hydrogen electrolyzers. If your goal is to decarbonize the grid, which signal do you base your decision on?
As we’ve previously written, there are two types of CO2 signals: marginal and average. But which one should you use in practice when making scheduling decisions? This article attempts to answer this question based on our experience developing both signals as part of the Electricity Maps API.
In order to reason through this problem, let’s imagine an overly simplified electricity grid with 1 solar power plant and 1 fossil fuel gas turbine. For simplicity we’ll assume that this grid is never congested, and that the gas turbine is large enough to cover the electricity demand at any time.
A simplified grid with 1 solar power plant and 1 fossil fuel gas turbine.
As a grid consumer, the average carbon intensity is obtained by assessing the emissions of the mix of electricity consumed. This means that every power plant contributes to the average signal, which will thus be lower at times where the solar plant produces electricity.
On the other hand, the marginal carbon intensity is obtained by identifying which power plant provides the additional electricity requested as a result of the decision to consume at a specific time. Assuming that our simplified grid is never congested, the only power plant that has the ability to increase its production to deliver additional electricity is the gas turbine (one can’t demand that the solar panel produces more, as one can’t ask the sun to shine more strongly). Therefore, the marginal intensity doesn’t change over time because the marginal power plant is always the gas turbine.
On this simplified grid, the average signal is lower during sunny hours. The marginal signal is constant, as the marginal power plant is always the gas turbine.
If you were to decide the charging time of an electric vehicle, the average signal would incentivize scheduling it at noon, where the sun shines the strongest. On the other hand, with the marginal signal, all charging times are equivalent, because regardless of when you decide to charge the vehicle, the impact will always be to adjust the production of the gas turbine to accommodate that request. So which decision is the best? Charging at any time or charging when the sun shines?
The key to answering this question is to consider the time scale of the impact considered. The marginal signal accurately depicts the immediate impact that a decision has on the grid. But what if that vehicle gets charged at the same time every day for 10 years? What if that vehicle is part of a larger fleet that also gets charged at the same time? How might the grid adjust?
Let’s revisit our simplified grid by now considering the situation where enough electric vehicles are systematically charged during sunny hours. The persistent additional demand during these sunny hours would create incentives to install new solar panels, as they are probably the cheapest (and cleanest) option to expand the grid to deliver electricity at these times. If the additional demand wouldn’t have happened during sunny hours, chances are that the grid would need to be expanded with new fossil fuels, or with storage systems. In both cases, these will cause more emissions than if solar electricity was consumed at times when it can be produced. Intuitively, in the long run, charging during sunny hours reduces the need for fossil fuels.
On this simplified grid, the marginal signal, by focussing exclusively on the immediate impact, misses the long-term benefits of incentivising consumption at times when renewables can be abundant. Thus, it might lead to less renewables (and more fossil power plants) being installed 🤯. But wait, isn’t the marginal signal supposed to be the optimal signal for decision making? Correct, but it is only the optimal signal in the short run: as we’ve seen in this example, optimizing for immediate impact can make things worse in the long-run.
It should be noted that research is currently underway to assess whether long-term effects can be incorporated into a marginal signal by investigating how the future grid would change if demand systematically was allocated to specific times. However, at present, no method currently exists for estimating future structural changes in real-time. So what do we do in the meantime?
It turns out that on this simplified grid, the average signal, by incentivizing electricity use during sunny hours, incentivizes the proper long-term behavior. This is not a coincidence. A fundamental aspect of the marginal signal is that it doesn’t change when the wind blows or the sun shines, unless wind or solar are being curtailed. Rather, it bases itself on the power plant that would change its production (often a fossil fuel power plant) in reaction to additional demand.
On the other hand, the average signal incorporates all power plants of the grid, and thus incentivizes using electricity during times where the grid is clean. Thus, it is a signal that motivates the alignment of consumption and low-carbon production. Because this is one of the key challenges to build the low-carbon grid of the future, the average signal creates better conditions for decarbonisation investments to happen.
There are some cases where the average signal doesn’t lead to long-term impact. For example, if there are no existing renewables on the grid, then the signal won’t incentivize using electricity during hours where renewables could supply clean electricity in the future. There are also cases where the marginal signal does lead to long-term impact, when e.g. there’s an oversupply of renewables to the point where they get curtailed and thus appear in the marginal signal. However, in most situations, the average signal is a better proxy for long-term impact (see e.g. this study).
As consumers demand more sustainability insights from products, a challenge appears when the insights that they see conflict with their intuition. For example, in the US, marginal emissions have increased although the grid has become cleaner. Wouldn’t EV drivers expect their drive to become cleaner as the grid gets cleaner?
It is quite common for California to have the majority of their electricity coming from solar power during the day. This means that intuitively, Tesla drivers charging their car at these times will expect their electricity to come mostly from solar power. However, in these situations, the marginal power plant will probably not be a solar panel and will almost certainly be a fossil fuel power plant. If the charging time is based on a marginal signal, Tesla owners will be confused to see a display that shows that their car is powered by fossil fuels, whereas they know the grid is currently dominated by solar power.
The Californian grid is relatively clean during sunny hours. Source: Electricity Maps app.
This confusing situation also happens in France, where users will be confused to see that their electricity doesn’t predominantly come from nuclear power, although the French power system is well known for this particular feature.
Based on our experience crafting user interfaces and collaborating with product teams seeking to build such features, we have found that the use of marginal signal conflicts with common intuition and ends up confusing end-users.
Let’s imagine you just plugged in your electric vehicle for charging, and let’s furthermore assume that the marginal signal tells you that a gas turbine provides that additional electricity. How do we verify that the marginal signal is correct?
The only way to verify is to observe the identity of the marginal power plant, which means comparing the production of each power plant between a situation where you plugged in the EV compared to an alternative world where you didn’t. The power plant that changed its production between these two situations would be the undisputed marginal power plant. However, it has so far proven impossible to rewind time to put the grid back in the exact same conditions to run the experiment where the EV wasn’t connected. The parallel world where you didn’t make the decision to plug your EV is unknown to us, and therefore, the real identity of the marginal power plant is unobservable1, and can thus be disputed.
Complex models therefore have to be devised to describe this parallel world. The lack of a time machine or parallel world explorer means the best option for verifying these models is to perform randomized experiments, which are costly and would only provide partial confidence. In practice, it is difficult to assess the extent to which a marginal signal is correct. Therefore, it is also difficult to assess whether a certain marginal model is better than another.
On the other hand, the electricity production of each power plant is information that is known and physically measured. It is therefore simple to verify that the underpinnings of the average signal reflect the physical reality it represents. Its simplicity also explains why the average signal is and will be more readily available in countries around the world.
Making automated real-time decisions based on a marginal signal generally doesn’t lead to the best long-term outcomes and causes confusion when presented to end-users. Even if it were leading to long-term impact and could be understood by users, there is unfortunately no objective way to assess whether a marginal signal is correct, and thus, consensus among various methodologies is difficult to reach.
On the other hand, the average signal, due to its simplicity, is intuitive for users and provides a useful heuristic that in most cases optimizes for long-term impact. Its simplicity makes it readily available for many countries and allows for scrutiny and verification. Although the average signal is not perfect, our experience working with both signals shows that it currently represents the best suited signal for use in automated real-time decisions.