As is customary, our latest blog is related to our recent energy seminar. At this week’s talk, Adriaan Hilbers from the Department of Mathematics presented his work on how we plan a future electricity system with uncertainties in weather and consumer demand. He has kindly summarised the talk for us here and you can download his slides as a PDF.

The energy transition is expensive. Infrastructure for decarbonised transport and heating is forecast to cost £240bn in the UK alone over the next 30 years. It’s hence important to get things right; spending even 5% more than necessary means billions wasted. At the same time, the decarbonisation process is full of complicated choices.

Consider the decision whether to build new nuclear plants, which has seemingly split the energy community almost exactly in half between those who view it as the technology for a secure and sustainable energy future and those who discard it for its high cost.

Decision-making in electricity systems (e.g. should we build a new wind farm, gas plant or battery?) is particularly complex for two main reasons.

The first is that they are fiendishly complicated, resembling the London Tube: complicated networks throughout which supply must match demand. This makes the effect of a new wind farm on the system as-a-whole difficult to estimate.

The second is the uncertainty such decisions are made under; a wind farm has a lifetime of around 20 years, during which future grid developments, energy policy, electricity prices or even wind speeds are unknown.

To aid in their choices, decision-makers often employ power system models: computer representations simulating how electricity is generated and distributed throughout the grid.

These models allow informed electricity strategy; one can, for example, add a candidate wind farm into the model to predict its effect on electricity security, price and emissions. They use demand & weather data, such as hourly demand levels and wind speeds at different locations throughout the grid, as inputs. A model is run using, for example, the demand and wind data from the year 2018.

Using this data leads to uncertainties with important implications.  Suppose demand & weather data from the year 2018 indicates the cheapest way to decarbonise is by installing two new wind farms and one new gas plant. Would this also have been the case if 2017 data had been used?

These questions are important, since choosing the “wrong year” may lead to suboptimal energy strategy, wasting billions in unnecessary costs. It’s also lead to fiery academic debates, including one in which a Stanford professor took a colleague to court over criticism of his modelling methods, partially revolving around the choice of demand & weather data.

Electricity strategy becomes a statistical problem: given some (uncertain) demand & weather data, how sure can one be that the recommended policy is optimal in the long run?

An obvious way to understand this “demand & weather uncertainty” is to use multiple samples, e.g. each year from 2000 to 2018. This gives a range of outputs and allows a user to pick strategy that is optimal in the largest number of years, or a middle ground to hedge one’s bets against uncertain outcomes, for example Setting A and Setting B.

Consider two different scenarios Model A and Model B. Both indicate the optimal mix of baseload (e.g. nuclear), peaking (e.g. gas) and wind technologies to meet demand in a large European region. The dots and error bars indicate the point estimate and uncertainty bounds respectively.

Model A is based on a 10-year demand & weather sample, while Model B uses a 1-year sample. In both settings, the recommended strategy is to install around 90GW baseload, 100GW peaking and 80GW wind.

However, the demand & weather uncertainty ranges differ considerably in size. For Model A, a decision-maker can be fairly confident that building 90GW of baseload, 100GW of peaking and 80GW of wind is optimal. For Model B, this is much less clear, with optimal wind capacity anywhere between 50GW and 110GW. In this setting, a decision-maker might opt to obtain more information before committing.

What are the implications of such analysis? The first is that both demand and weather uncertainty should be considered when forming energy policy; a range of possible outcomes provides a significant information advantage over a point estimate without any indication of the degree of certainty.

The second is that demand and weather samples should be as long as possible. The uncertainty on analysis conducted on 10 years of data is smaller than that on a single year since a decade-long sample already contains encapsulates a higher proportion of the possible scenarios.

While simple enough in principle, in practice such “statistically-informed electricity policy” becomes difficult due to the complexity of the grids. It may not be possible to run models using long samples, especially not multiple times. An active area of research involves using shorter demand & weather samples to conduct statistical analysis. Creating such samples, and accounting for the reduction in sample length, requires robust statistical arguments.

Since energy strategy is rife with uncertainty and complexity, statistical considerations are essential to robust decision-making. The money involved in decarbonisation is enormous, and mistakes are costly. Dealing with (demand & weather) uncertainty in electricity system planning ensures an energy future that is secure, affordable and sustainable.