Learnings

Outcomes

The delivery of the project is almost complete.  Data used for the project was sourced from the National Grid ESO open data portal.  Data on fuel consumed by power plants was requested from the DUKES team in the Department for business, energy and industrial strategy.   
The initial work package, to build models improving on the carbon intensity of a power plant compared to the current single value, has built initial models for gas, coal, and biomass, and compared them to the single values for carbon intensity currently used.  These models show reasonable initial performance given the small amount of available data and are being refined to ensure best performance and to reduce the risk of overfitting.  The models are built using features generated from the half hourly running of the power plants. Various iterations and combinations of the features are used as inputs, including considering interactions and nonlinear terms.   Due to the low data volumes and a focus on explainability, the modelling approach being used is linear regression and only a small number of features are used in the final models to reduce the risk of overfitting.  This has been concluded and delivered.    
Making use of publicly available data and the CO2 intensity analysis in the first work package, the second work package has created a simplified model of the grid, the generators, and the allocation problem. It then optimizes dispatch decisions based on this simplified model. The optimization method used is Deep Reinforcement Learning with Evolutionary Strategy finetuning.  The work package has successfully developed a proof-of-concept tool that can produce central dispatch allocation plans that are optimized based on both carbon intensity and cost. The tool can also provide estimates for the environmental benefits and financial costs of such plans. 
 
The project is currently in final report writing with work package 3. 

Lessons Learnt

The key lesson learnt from the first half of this project is that the currently publicly available data (monthly and grouped by power plants), limit the modelling techniques available for understanding the performance of the power plants.  This would be improved by collecting this data at more regular intervals, for example daily or weekly, or for single plants.  
 
The secondary lesson from the first work package is that we see model performance that suggests with the additional data we would be able to have a sufficient level of performance to model power plant performance at half hourly level.  This would give the National Grid ESO the opportunity to review the decisions made controlling the grid based on their impact on carbon intensity – for example balancing services. 
 
Key lessons learnt from work package 1: 

• Currently available data (monthly and grouped by power plants) limit the modelling techniques that can be used, and the results that can be driven. 
• The results with the limited data show improved accuracy on the current approach and also start to segment the performance of the different types of CCGT plants and the different types of operation.  This shows there is the opportunity to extend the work with additional data. 
• The calculation of carbon intensity across gas, coal and biomass in general is not standardized and future work should review this.   

Key lessons learnt from work package 2: 

• Having an accurate CO2 intensity model for every generator in the network has a big impact on the allocation optimization. However, building such models requires data on individual generators which is not available in the public domain. 
• Current model has a straight-forward energy demand forecast — the project assumes the control room  has a perfect demand forecast 16 hours ahead. Improving this forecast model through collaborations with the NG-ESO demand forecast team can enhance the allocation model and optimization.