The ESO, the electricity system operator for Great Britain, ensures that the GB power grid is always secure and operating efficiently. One of the many ways they do this is to keep the grid ‘balanced’ – informally meaning that supply always meets demand. Increased renewable generation and an ever-changing electricity grid makes this a significant challenge, coupled with the economic challenge presented by rising energy prices.

Problem

Fundamentally, balancing the grid involves ensuring generation of power must match the demand from consumers. Predictions are made of future demand, to ensure the ESO can plan ahead and operate as economically as possible. Similarly, generators forecast their expected outputs, and the maximum and minimum limits in the future. Forecasts – in this context ranging from around 6 hours ahead to 2 days ahead – are never perfect, and each of these has an associated error of margin. One way to protect against these errors is to secure ‘spare capacity’ – also called reserve – that can be called upon when needed to cover instances where forecasts do not match the actual outputs from generators or true demand on the system. Without reserve, instances where actual behaviour deviates from these forecasts could jeopardise the grid security, leaving insufficient generation to meet demand and ultimately resulting in an unstable grid frequency.

Reserve is therefore vital to balance the grid, and every day it helps keep the energy grid in Britain stable. However, securing reserve is not easy (generators can play many roles to balance the grid, and holding one for reserve limits options in other services), and purchasing it incurs a cost. This cost not only impacts the ESO, but ultimately the consumer (the bill payer) due to a knock-on effect through energy bills. Choosing the amount of reserve to secure is therefore a balance between risk and cost. Secure too little and the cost is reduced, but the risk to the grid is increased. Secure too much and the risk to the grid is minimised, but the cost increases, and potentially the amount of carbon emissions generated increase dependent on the type of reserve held. The ESO’s current approach to reserve setting, in contrast to the one developed during this project, is static – it updates at clock-change to reflect differences in summer and winter but does not take into account changing dynamic aspects such as daily weather conditions. An ideal alternative would incorporate this dynamic data to produce half-hour, ultra-responsive reserve recommendations that truly represent the current state of the system.

Solution

The Smith Institute worked with the ESO on a Network Innovation Allowance funded project to develop a data-driven, probabilistic and explainable machine learning methodology for determining the optimal amount of reserve to hold that balances this trade-off. Smith Institute worked closely with domain experts at the ESO to understand the operational considerations when securing reserve, and developed a data-processing pipeline, model and dashboard tailored to these practical considerations. The data-pipelines extract and process forecasts and actual values of weather, system flows, generator parameters, breakdowns and various other relevant aspects, unifying them and ensuring they are appropriate for modelling. Using features derived from these, we trained a probabilistic and explainable machine learning model, that recommends reserves at a provided risk appetite. The modelling approach adheres to some key principles throughout:

• The risk appetite is adjustable – rather than setting a fixed risk appetite, it can be varied by experts based on a given situation, allowing for real-time understanding of the cost and risk trade-off.
• The model is explainable – rather than a black-box that simply gives a value of reserve to hold, the model must detail to control room engineers exactly what is contributing to this recommendation and by how much. This not only builds confidence between the expert and the model, but it gives a finer grain understanding of the current state of the system.
• The model responds to current conditions. The amount of reserve required to secure the system depends on the forecast errors one might get at a given time. These in-turn depend on the weather conditions, system flows and time of year. The model should dynamically update reserve values as these features vary, accounting for the fundamental drivers of reserve.

After establishing the pipelines for data-processing and the model to generate reserve recommendations, a dashboard was built to ensure results were shown to experts in an intuitive, clear and simple way. In any given moment, balancing engineers have an enormous amount of information to process, so ensuring a clear and quick display of information was vital for practical usage of the solution. The dashboard dynamically updates through time, showcasing the risk being operated at, explanations for the model outputs, and also allows users to adjust the reserve recommendations and understand how doing so alters the risk they are exposing themselves to. Coupled together, the data-pipelines, model and dashboard create a system with which the setting of reserves can be transformed and brought into the future, with explainable machine learning at the heart of it.

Results

The solution developed by Smith Institute could save over 300 megawatts of reserve each settlement period (30 minutes) compared to the existing ESO’s approach once implemented into the live environment.

This is akin to roughly the same amount of energy produced by half a nuclear powerplant.

When energy prices are high, every megawatt counts, so this could represent a real, practical saving of millions of pounds a week to the ESO, and therefore consumers too. The benefits go beyond purely cost. The robust modelling approach also allows the risk that is being taken when securing reserve to be better understood by operators and allows them to alter the risk at any time, based on their expert judgement. Since the approach used has explainability as a fundamental aspect, control room operators can see exactly why a particular amount of reserve is being recommended, and how the weather, current time, as-well as forecast and historical state of the system contribute to this recommendation. The solution fundamentally enables operators to act more efficiently, with more confidence, and save significant amounts of money.