National Grid is the Electricity System Operator in Britain, and as such makes sure electricity is transported safely and efficiently from where it is produced to where it is consumed. National Grid seeks to make sure that supply and demand are balanced in real-time and facilitate the connection of assets to the transmission system. To ensure that demand can and will always be met, National Grid must produce demand forecasts for every half hour of the day.

There is, however, an unknown amount of generation that feeds into the transmission network, offering supplies that are connected directly to the local distribution networks and that is therefore ‘hidden’ from National Grid. In this sense, National Grid can only see this generation as a ‘missing demand’ and cannot control it. These ‘hidden’ inputs are usually called embedded generation as the generation is embedded within the distribution systems. Examples include smaller renewable energy, like wind turbines and solar panels as well as non-renewables like biomass, gas, landfill gas and waste.

The challenge

National Grid recently acquired 4 years of historic embedded generation data, showing that in 2017 there was around 4.7 GW of non-renewable embedded generation. The Smith Institute was asked to create forecasting models predicting the contribution of the non-renewable embedded generation to the UK’s transmission network. National Grid already master forecasting tools for wind and PV generation, but previously did not have models for non-renewable embedded generation.

The solution

From the dataset provided, the Smith Institute designed and developed forecasting models as well as characteristic generation profiles for the outputs of non-renewable embedded generation. These were categorised according to features such as fuel type and weather dependency.

We created forecasting models for 1, 2 and 7-days ahead energy generation forecasts using Support Vector Regression, a machine learning technique that has been widely used in the last 15 years and is particularly effective for understanding non-linear relationships in the data. The combination of Support Vector Regression with a Gaussian kernel function (i.e. a non-linear function) in our models offered accurate fits compared to linear regression which is most often used in National Grid’s forecasting models.

Each model developed by the Smith Institute includes a range of weather variables such as effective temperature, wind speed and effective illumination and variables describing time of day and time of year, contained in National Grid’s existing demand forecasting models. The new models also consider day ahead peak electricity prices which, critically, had not been used in National Grid’s demand forecasts before but are found to be a major driver of the non-renewable embedded generation.

These innovative predictive models provide National Grid with greater visibility of type and behaviour of the embedded generation that is connected to the distribution networks and help to manage the electricity system more efficiently and securely.