Adaptive models are unsupervised machine learning models that can adapt their output and training data to accommodate new data. This prevents the need to manually retrain models based on static data and increases scalability and reduces resource for tagging and sorting data.
This project aims to assess the effectiveness, practicality, and benefits of implementing the adaptive data models within our operations, focusing on its impact on system optimisation and decision-making processes. The input data model will address several identified gaps in our current capabilities, aiming to enhance forecasting, optimisation, and situational awareness in various operational scenarios.
Benefits
This feasibility study hopes to discover the time and resource savings that could be achieved when it comes to developing adaptive models of different components. There is an expectation that adaptive models will significantly improve efficiency and performance of optimisation solvers by being able to retrain themselves on updated data as it is added to the operational databases. This will add a new capability to NESO and allow for the development of scenarios while providing an envelope of outputs rather than deterministic results with only one output.
Learnings
Outcomes
The project is complete, and the outcomes include:
- The differentiation between adaptive inputs into two categories: the forecasting layer and the enrichment layer. The forecasting layer includes adaptive models, which utilise machine learning to forecast inputs required for decision making, such as generation, demand and interconnector flow. The enrichment layer includes the system requirements and the model of the transmission system, which it is suggested could be rule-based models.
- The data input assessment report identifies the necessity for adaptive input models. It highlights data streams which, if enhanced, could benefit future adaptive input models.
- Adaptive Proof of Concept models showcasing the feasibility of real-time forecasting models, providing empirical findings and recommendations for the programme
- An assessment of the feasibility and value of implementing adaptive models in NESO’s operational environment. Insight was developed on both the feasibility of model use, along with operational considerations such as retraining, update frequency and granularity. Value exploration was assessed, including an assessment of accuracy, robustness and output explainability.
- A series of recommendations for the development of adaptive models was developed, along with Volta solution design refinement considerations. An actionable roadmap and plan to deliver these recommendations was created, along with further innovation and development opportunities for consideration.
Lessons Learnt
The lessons learnt are:
- Data requests for projects of this length can be challenging if the relevant data is not publicly available. Mitigation strategies such as early data access requests, utilising data already in accessible systems, minimising data requests to that necessary for the case study / minimal viable product
- Iterative Development: Having rapid iteration and rapid feedback cycles during the feasibility phase accelerated development of the proofs of concept. Taking this approach in future phases will support efficient refinement of design and implementation.
- Collaborative working: collaboration with Control Room SMEs and the forecasting team provided grounding to ensure adaptive model design reflects the operational reality. Continued close collaboration with these SMEs and teams, along with other Volta workstreams and external experts would be beneficial for the quality and alignment of further development.
