The current methodology for calculating carbon intensity from fossil plants is based upon a simplistic calculation which does not capture variability between different generators of the same type, or within individual generators over time. National Grid ESO’s ability to improve the carbon intensity calculation is hampered by lack of available data on the fuel consumption of individual generators, which is considered commercially sensitive by the data owners.
Using available data and relevant knowledge from scientific literature, this project will research and develop a refined model that will improve the accuracy of carbon intensity for power generation. This data is important in tracking the progress towards de-carbonising the electricity system, and in future could also be used to optimise the dispatch of power based on carbon intensity.
Benefits
This project will develop our understanding of how operating carbon emitting plants in different ways impacts the carbon intensity released, improving the representation of progress to net zero, and enabling future options for carbon optimisation.
Learnings
Outcomes
The initial work package developed models that improved upon the current single-value carbon intensity of a power plant, and produced models for gas, coal, and biomass. These models were compared against the single values currently used for carbon intensity.
The models demonstrated reasonable initial performance given the limited available data and were refined to optimise performance and reduce the risk of overfitting. They were built using features derived from the half-hourly operation of power plants, with various iterations and combinations of these features used as inputs—including interactions and nonlinear terms.
Due to the low volume of data and the emphasis on explainability, a linear regression approach was used. Only a small number of features were included in the final models to further mitigate the risk of overfitting.
The project suggested several activities as follow-on activities to further the impact of the project:
- NESO needs more access to more granular data on power plant inputs.
- With additional data, additional modelling techniques should be explored to improve the performance of the machine learning.
- Explore more nonlinear relationships between the variables and correlations between the variables
- The best performing model for CCGT plants is a Bayesian regression model using time spent on and time spent in ramp up as features.
- If data is available, look at the biomass plants separately to separate the behaviour of Drax from the other plants.
- Linear regression, using only the time the plant was on, is the best-performing model for biomass plants.
- The most predictive model for coal is that without an intercept that has the variables time spent on and % of time constant.
- Replace the current values in the grid carbon intensity calculation with the modelled ones for the different gas generators.
- Do not change the value used for coal in the grid carbon intensity calculation.
- Explore a project conducting full lifecycle analysis of carbon emissions for all the main fuel types.
- Calculate improved carbon intensity calculations for the non-wood pellet biomass plants
Lessons Learnt
Lessons learnt from this project indicate that the data, which was collected monthly and for groups of power plants, restricts the modelling techniques available for assessing the performance of these plants. This issue could be alleviated by collecting data at more frequent intervals, such as daily or weekly, or by focusing on individual plants. Future projects should extend this work to get predictions of carbon intensity per BM Unit per half hour with more granular data.
With improved modelling, detailed analysis of the impact of dispatch and flexibility services should be undertaken.
