The ESOs ability to forecast electricity demand has reduced as a consequence of the COVID-19 pandemic. Historically, the underlying demand profile was driven primarily by day of week and time of day. That demand has become harder to forecast partly because working patterns have not returned to pre-pandemic levels and also because we are seeing new technologies such as Electric Vehicles (EVs) come onto the system. This project seeks to utilise mass mobility data (anonymised telematics vehicle monitoring data) to generate new features for electricity demand forecasting models. It will investigate potential features of value supporting electricity demand estimation and will use historical data to correlate and evidence potential predictive value from said features.
Benefits
This project has the potential to bring new insights into normal (and temporally abnormal) behaviour patterns concerning citizens in transit, home, workplace and state transitions across the population. The regionally resolved data will also provide features in regional forecasting that can support constraint management. In addition, the use of telematics derived features will provide complete spatial coverage over all of GB with minimal sampling bias due to its diverse geographic sampling. Though the project focus is not EV usage, it may be able to explore features related to EV charging behaviour and its impact on demand.
Learnings
Outcomes
The project is currently 75% completed.
Work packages 1 and 2 have provided a shortlist of traffic data features to be assessed in the demand forecast model as well as their requirements.
Work packages 3 and 4 have recently been completed, the project supplying several batches of data to be tested. A baseline demand forecasting model has been developed. A Jupyter notebook has been created to test features in demand forecast model variants and visualize results.
Currently the project is working on work packages 5 and 6, refining and assessing the value of the traffic data features. Initial results show that when integrating any of the two most promising traffic data features, an improvement of 5% in model accuracy has been observed compared to the baseline model. The improvement was much higher when considering specific “worst performing” dates for demand forecasting such as dates during the Christmas festive period.
Preparation for the last work package – final report write-up of project results and gap analysis for deployment in NESO operations – has also started.
Lessons Learnt
Interim Project Outcomes
A register of traffic data features has been created, from which a list of potential features to be assessed in the demand forecasting model has been selected.
A benchmark demand forecasting model has been developed based on energy and time variables such as weather and holidays. A model framework has been developed to easily add or change variables to the model. Data relating to the shortlisted traffic data features has been gathered and is being tested in demand forecasting model variants against the baseline model.
A more focused analysis has been performed on the worst performing dates for demand forecasting.
Review of benefits case
Two model variants integrating individual features showed an improvement in overall model performance when compared to the baseline demand forecasting model.
Another benefit is the improvement to demand forecasting on difficult dates such as the Christmas festive period, which currently incur a high balancing cost.
Next steps
- Further evaluation of features to determine value for demand forecasting.
- Visualisation of data: baseline and model variants integrating new features.
- Test of candidate traffic data features on clusters of small geographical areas classified as having similar characteristics.
- Gap analysis to usage in NESO operations: steps needed to productionise and deploy features within NESO demand forecast model.
