Learnings

Outcomes

Workstream 1 has achieved the following: 

• Constructed a data pipeline to read, extract, organise and store relevant data fields from a custom NGESO control room system
• Identified a range of forecasts accessible within these datasets, key forecasts of interest identified by SMEs
• Forecasting performance measured at a more granular level than is otherwise available from other internal data sources
• Individual generator forecasting and time dimensions explored in forecasting errors
• Dispatch Algorithm utilised in novel way to cost model forecasting error against alternative dispatch solutions
• Situational awareness visualisations and generator forecast performance rankings prototyped
• Implementation test running from historical snapshot of data successful on internal business supported system, demonstrating future deployment techniques

A sample of available screens and data are displayed on the Addendum sheets on the Smarter Networks portal https://smarter.energynetworks.org/projects/nia_ngso0032/

Workstream 2 has achieved the following outcomes:

• The production of skillful and statistically valid forecasts of BMU wind power generation and embedded solar power, using best-in-class data-driven forecasting algorithms.
• Skillful and statistically valid forecasts of demand and embedded solar.
• An example approach for using probabilistic forecasting to support rescheduling and redispatch.
• An example approach for using probabilistic forecasting to support day ahead reserves recommendations.
• An example approach for using probabilistic forecasting to support margins analysis.

Lessons Learnt

Workstream 1: Smith Institute

One learning from the project was that accessing historic data from the control room systems involves manual input and is time consuming. The time to extract historic backup data from control room systems should be considered when planning future projects. In future, it should be possible to build a live link into control room systems to access this data on a real-time basis and a selection of this data could be stored in a database (similar to the static database created in this project).

The control room backup data was also used to analyze the error across different forecasts for the same target variable. For example, the forecast error for individual wind balancing mechanism units was analyzed for two different forecasts: an NGESO forecast, and forecasts submitted by the wind units themselves in their final physical notifications. It was previously thought that NGESO forecasts provided greater accuracy than the final physical notifications for all wind BMUs. However, for wind units we learnt, for the period of study, for which times of day, and for which units, the final physical notifications provide a better forecast.

The proof-of-concept dashboard explored many ways to quantify and visualize error and uncertainty in power systems operations for the control room engineers. We learnt that it is possible to use the control room system data to produce new visualizations which could aid the control room engineers in their real-time operations. The dashboard will allow control room engineers to easily interpret historic trends in the forecast data, this will be increasingly useful as growing renewable penetration will increase the complexity for managing the system. 

A power dispatch model (MDA) was used to explore costs associated with forecast errors. The MDA was computed once using forecast data and once using measured data and the difference in costs between these two solutions was used as a proxy for the forecast error cost. We learned that for errors in the National Generation Requirement, the main features of the distribution of cost gradients match the main features of the distribution of marginal costs and that during Nov-20:

• £60/MWh was a typical marginal cost under conditions where increasing generation is required to close the imbalance, but the amount of the increase is forecast incorrectly (38% of periods at 15 mins forecast leadtime).
• <£30/MWh was the typical marginal cost under conditions where reducing generation is required to close the imbalance, but the amount of the reduction is forecast incorrectly (41% of periods at 15 mins forecast leadtime).

We also applied this technique to estimate the costs associated with errors in forecasts for group constraints, interconnector flows, and wind generation. We learnt that this approach for historically calculating the costs associated with forecast errors provides only a little more information than the (simpler to calculate) marginal costs and is not a good proxy for the true incurred costs. This is because incurred costs are more complicated to calculate as they relate to opportunity costs which have incurred directly as a result of, and in the settlement-periods following, the forecast error and consequently, are unavoidably included in any historical analysis of system costs. For calculating costs associated with forecast errors in constraints (including National Generation Requirement and individual group thermal constraints), the cost gradient does provide more information than the marginal costs. However, there is not sufficient added value to recommend its implementation and the cost gradient approach is not suitable for estimating the costs of forecast errors in wind generation and ICs. Finally, we have proposed a new online approach for estimating incurred cost for which we have also identified a plausible deployment route, through synergies with work on MDA.

Workstream 2: TNEI
Workstream 2 explored different methods for developing probabilistic forecasts for demand and renewable power generation. Learnings from the project demonstrated that it is possible to produce univariate probabilistic forecasts e.g. of demand, wind, solar and that these univariate probabilistic forecasts can be successfully combined into a multivariate probabilistic forecast using gaussian copulas.

It was also learned how probabilistic forecasts can be applied across three different use cases within NGESO: 

• Rescheduling and redispatch at between four hours ahead and real-time
• Day-ahead operational reserves advice
• Margins forecasting

Additional lessons learned include how probabilistic forecasting can be used to support: rescheduling and redispatch, day-ahead operational reserves recommendations, and margins analysis. 

Review of benefits case 

In workstream 1, insights into which forecasts provide more accurate information in the data snapshot provided was demonstrated. The recommended next steps are to link these visualisations into a live data stream to validate on continuous operational data, so the comparative accuracy can be seen in live operation.

In workstream 2, all individual probabilistic forecasts were delivered and three example use-cases for these probabilistic forecasts have been shown. 

Next steps 

For workstream 1, the next step is to establish a live link into the control room data so that the dashboard tool can be deployed for use by control room staff in real time and the control room data can be stored for easier offline analysis.

For workstream 2, the next steps are for an internal dissemination of the probabilistic forecasting methods developed in the project and for the relevant subject matter experts to explore using probabilistic forecasts as described in the three use cases identified in the project. 

Dissemination 

Control REACT was presented at the ENIC conference in Autumn 2021 and the project was included in the Innovation Annual Summary https://www.nationalgrideso.com/future-energy/projects/nia-annual-summary-report.