BSUoS is the Balancing Services Use of System charge, paid by transmission connected generation and demand to cover the cost of balancing the electricity system. To set the tariff, an accurate forecast of the costs and the variability is required, but these balancing costs are highly volatile and difficult to forecast accurately.
This project is looking to improve existing short term (<12 month) forecasts by applying machine learning and cutting-edge forecasting methods. Additionally, the project seeks to increase the temporal granularity to weekly or daily.
Accurate BSUoS forecasting benefits all consumers by enabling better business planning and risk management by NGESO and its customers. It may also bring opportunities for the control room and planners to reduce spend by taking more cost-efficient actions.
Benefits
The energy system is going through rapid and extensive change, with changes in supply and demand. This project will facilitate NGESO's role though the energy system transition by:
Financial
Accurate BSUoS forecasting benefits the customers by enabling better business planning and risk management. Ultimately, this benefits all consumers as more reliable risk estimation around balancing costs supports better risk provisioning.
Customer/stakeholders & Financial
BSUoS forecasts are published publicly, and are used by many industry participants, including suppliers and generators. Any improvement to forecasts will therefore allow industry participants to have more confidence in their financial planning.
System Security
With the ability to forecast balancing spend accurately in the short term and to a higher granularity comes opportunities for the control room and planners to reduce that spend by taking different actions (i.e. delaying outages). Model development might provide insights on the drivers and improve NGESO understanding of how costs might be reduced.
Learnings
Outcomes
The project improved the understanding of the drivers behind the ESO’s current model, and after thorough investigation of many different approaches identified some avenues for further investigation whilst ruling out many others:
- The variables with the best predictive power are the ones used in the live model (renewable generation as a proportion of demand and wholesale electricity prices).
- The best modelling technique was the ‘Prophet’ modelling package. This gave a modest improvement in accuracy compared to the existing model. Over the testing period (Apr-21 to Nov-22), the Prophet model gave a mean absolute error of £73 million, compared to £78 million (per month) using the existing model. However, this is based on using actual (rather than forecast) values for the input variables. When using forecasts instead, the accuracy difference between the models is negligible (£208 million with Prophet, and £209 million with the existing model).
- GARCH modelling showed some promise for simulating wholesale prices, as it allows for volatility that changes over time. This is an area that needs more research to determine if this method gives an improvement over the existing one.
- Daily granularity forecasts were produced but were not able to capture the volatility in daily costs enough to provide useful insight. Aggregating these forecasts up to monthly level did not give increased accuracy over just producing a monthly forecast directly.
More details on the project outcomes can be found in the final report published on the Smarter Networks Portal.
Lessons Learnt
In terms of how the work packages were planned, the differences between Work Packages 2 and 4 were not as clearly defined as first thought. Machine learning is a wide field and it made sense to apply these techniques within Work Package 2 alongside other approaches. Work Package 4 has therefore focused on neural networks as a sub-category of machine learning. Potential overlap has been managed through collaboration between the two teams working on these packages and ensuring there are enough people reviewing to spot any repeated work.
In terms of modelling approaches, the “Prophet” package explored in WP2 is an approach that could apply to other time series forecasting projects within the ESO.
In terms of the variety of modelling approaches, performing initial comparisons with existing models and evaluating performance at an earlier stage facilitates a quicker transition to alternative approaches.
