Reserve is essential in securing the electricity grid. Choosing how much reserve to procure is a balance between risk and cost. This problem has been tackled in a previously successful project named Probabilistic Machine Learning Solution for Dynamic Reserve Setting (DRS) NIA2_NGESO003; however, this project considered reserve at the national level and in practice, different amounts of reserve are required in different locations across Great Britain. This project will further the DRS work by building explainable, risk-based dynamic models for reserve that generate predictions at finer spatial resolutions. Using these models, NESO will have access to accurate, risk-based predictions of reserve requirements at different locations and can then make more informed decisions to maximise its usage and minimise costs.
Benefits
Setting the reserve regionally could unlock further savings by allowing NESO to procure reserve where it is required, lowering transmission losses, ensuring reserve purchased is not inaccessible due to constraints, further lowering overall reserve setting costs by allowing for offsetting of reserve in neighbouring regions. Also, the benefits to consumers would be lower energy costs and carbon emissions. Better modelling reserve requirements at desired risk appetites would maintain the trajectory that NESO is on to move to a carbon free electricity grid by 2035 and to do so in a way that maintains security of supply while optimising the balancing cost.
Learnings
Outcomes
The existing data has been meticulously cleaned, processed, and validated, with missing data identified from both internal and external sources. Initial feature selection for each regional model has been completed, resulting in feature sets that will be used for training each regional model.
The next phase involves training the models to generate initial regional reserve setting results, followed by fine-tuning the models and developing a geographical dashboard to display reserve predictions in real-time.
Lessons Learnt
This project builds upon the highly successful National Dynamic Reserve Setting model, that has delivered savings of hundreds of megawatts per reserve settlement period compared to the previous national methodology. For the initial part of this project, locational data was provided for Grid Supply Points (GSPs), grid nodes, and grid regional boundaries, providing the geographical framework for the aggregation of model features and target variables at the regional level. To achieve alignment of datasets, the project developed mapping methods that standardise the integration of regional information across heterogeneous data sources. This process provided validation of the classifications made when assigning regions to both model features and target data, eliminating inconsistencies that could compromise model performance or interpretability. Strategic Spatial Energy Planning (SSEP) and Regional Energy Strategic Planning (RESP) are integral to NESO’s comprehensive energy strategies. The mapping methods and data sources utilised in this project would be relevant to the RESP process as well as related applications and future projects.
In the feature selection stage of model training, both national-level features and region-specific predictors were included as initial candidate features for each given regional model. National features were included as they can capture broader system dynamics that could influence local conditions, meaning they may retain power for predicting reserve at the regional level. The initial pool of candidate features was refined using rigorous selection techniques, including correlation analysis and feature importance plots, to determine the optimal set of predictors for estimating reserve requirements in each region. Preliminary results indicate that national-level features remain significant drivers of regional reserve needs. These are effectively complemented by region-specific variables such as demand forecasts and errors, embedded wind forecasts, and air temperature. The relevance of these features has also been highlighted in related regional modelling efforts, including constraint forecasting. This underscores the potential for improved forecast accuracy to enhance the performance of various regional forecasting applications, particularly in the context of reserve setting as explored in this project.
