Learnings

Outcomes

The Volta: Real Time Prediction project successfully delivered a validated, high-temporal resolution machine learning–based demand forecasting approach for NESO’s Control Room. The project demonstrated that modern deep learning architectures, when appropriately tuned and operationally optimised, can materially improve the accuracy, transparency and risk-awareness of minute-by-minute electricity demand prediction. 
 
A range of statistical, regression and deep learning models were developed and systematically assessed. Through structured cross-validation, feature analysis and hyperparameter tuning, the Temporal Fusion Transformer (TFT) architecture was identified as the most robust and highest performing approach for both single-point and probabilistic forecasting. The final single-point model achieved performance within NESO’s defined accuracy targets across short-term and full 24-hour horizons as shown in the table below, with significant improvement in the first four hours ahead — the timeframe most critical to control room decision-making. 
 
Accuracy metrics for single-point forecasts in Appendix 1 
 
 
Beyond deterministic forecasting, the project successfully implemented probabilistic forecasting at 1-minute granularity over a full 24-hour horizon. By applying quantile regression and conformal calibration, the solution produced interpretable confidence intervals aligned with empirical coverage. This ensured that forecast bands (e.g. 10–90% or 1–99%) could be directly translated into risk thresholds, supporting more informed balancing and reserve decisions. This represents a step change from traditional single-point forecasts toward risk-informed operational forecasting. 
 
Accuracy metrics for probabilistic forecasts in Appendix 2 
 
 
Significant optimisation of the deep learning training pipeline reduced model training times from multiple days to a matter of hours without degrading forecast accuracy. This improvement materially enhanced the practicality of retraining schedules and future operational deployment. The project also established a robust quarterly cross-validation framework to simulate live retraining and ensure that performance was validated across seasons and demand conditions. 
 
Further details of project methodology and model development can be found in the 4 core deliverables:  D2: ML Model for Single Point Forecasting; D4 Assessment of Different ML and Deep Learning Methods for Single Point Prediction; D5: Assessment of Different ML and Deep Learning Methods for Probabilistic Prediction; D6: Top performing ML model for Probabilistic Prediction. 

Lessons Learnt

The project explored the suitability of different machine learning and deep learning model architectures for demand forecasting at high temporal granularity. The top performing single point and probabilistic models were based on a Temporal Fusion Transformer architecture. While the project was successful in meeting the accuracy targets provided by NESO when evaluated on historic data demonstrating theoretical improvements in forecasting accuracy, to meet the next TRL level, the prediction models need to be integrated into the real control room systems, including piping of real data into the models.  The ways these predictions work (including probabilistic outputs) is very different to the existing real-time forecasting approach which involved frequent manual manipulation of historic demand curves to match real-time monitoring. Once integrated, user training and testing should be undertaken to establish the prediction’s usefulness for the OEM regarding usability and resulting balancing mechanism dispatch efficiency. 
 
The UI work was critical to bring the AI-driven demand prediction model to life for OEMs and other stakeholders. The architecture work was essential to illustrate how an Innovation project can be adopted into plans for operations.  
 
The level of accuracy demonstrated with the TFT model architecture contributed to the implementation of same models in other project within NESO, whose scope spans from National Demand forecast to forecasts on Interconnectors flows.