Wiss Utvärdering av Lastprognoser CoordiNet Uppsala (2023)

Full citation: Wiss, E. Utvärdering av lastprognoser — En undersökning om lastprognoserna skapade till flexibilitetsmarknaden CoordiNet i Uppsala (Evaluation of load forecasts). UPTEC STS 23006. Uppsala University, Civilingenjörsprogrammet i system i teknik och samhälle, February 2023. Supervisor: Nicholas Etherden; company mentor: Yvonne Ruwaida (work carried out in the Innovation & Market Outlook department at Vattenfall Eldistribution). Subject reviewer: Juan de Santiago. Examiner: Elísabet Andésdóttir.

Summary

A master’s thesis evaluating the machine-learning load forecasts used in CoordiNet‘s Uppsala flexibility market (winters 2020–2021 and 2021–2022), written from inside Vattenfall Eldistribution with a company supervisor and mentor. Data covers 1 November to 31 March of CoordiNet’s second and third winters, at hourly resolution. Its value to this wiki is not primarily the forecasting-error methodology (MAPE, and RMSE expressed as a percentage of the subscription limit, over seven candidate error factors) but a structural admission stated in the method section, not as the thesis’s own finding: CoordiNet was the first time Vattenfall Eldistribution had created load forecasts for its grid, the development and running of the forecasts was outsourced to Expektra, and outsourcing left the company without a detailed internal understanding of how its own forecasts worked.

Key claims

The outsourcing-and-understanding-gap finding

Direct (translated): “Because Vattenfall Eldistribution outsourced the development and work on the forecasts to Expektra, there was a lack within the company of a detailed understanding of how the forecasts function and are built.” The thesis gives this as the reason its first phase was information-gathering: interviews with Expektra, Vattenfall Eldistribution and Vattenfall R&D staff (plus E.ON on the platform and CoordiNet’s project manager and operators), and an interactive presentation to operations in Trollhättan (27 October 2022), to find where the information gaps were. A separate two-hour workshop (21 December 2022, with the Market Outlook & Innovation department) was used to discuss the candidate error factors that the interviews had raised, not the understanding gap itself.

First-ever grid load forecast

CoordiNet is described as the project in which Vattenfall Eldistribution, for the first time, created load forecasts for the grid (“för första gången skapa lastprognoser för elnätet”, popular summary; the closing chapter says for Vattenfall Eldistribution’s control centre). The thesis does not elaborate on what forecasting existed before.

Two named internal information gaps

Interviews (with Expektra and with Vattenfall staff) surfaced two large information gaps: (1) how long before the operating hour the day-ahead forecast (DAP) and the intraday forecast (IDP) are created, and (2) how the models for the two forecasts differ from each other. The thesis later describes both as produced by the same algorithm, with parameters weighted differently.

Forecast error drivers

Of seven candidate error factors examined (weather-forecast uncertainty, load variation, significant grid users/SGUs, extreme weather, planned switching operations, electricity-price variation, and flexibility activations), two dominate: SGUs deviating from their submitted production plan (in practice the electric boiler, about 20 percent of Uppsala’s subscribed capacity; a deviation above 10 percent of the subscription limit had a large effect), and flexibility activations (called flexibility shows up as a forecast deviation even when the forecast served its purpose). Unexpectedly, forecast error generally decreased at high load: with higher electricity prices, in extreme cold, and at flexibility activations when the error was measured against estimated load without flex (i.e. when the subscription limit might have been exceeded). Measured against the metered load, flexibility activations increased the error. The thesis notes this pattern is what benefits the flexibility market, but its analysis is descriptive and cannot establish statistical significance, so all conclusions need further study with more data; some factors also rest on few cases (only five planned switching operations were analysed).

E.ON’s Switchmarket as the CoordiNet Uppsala platform

The thesis describes Switchmarket as the market platform used for CoordiNet, created and designed by E.ON with requirements from Vattenfall Eldistribution, and interviews E.ON (David Bjarup, 17 January 2023) about it alongside Expektra and Vattenfall staff. The resulting split is DSO (Vattenfall Eldistribution), platform (E.ON) and forecasting provider (Expektra). Switchmarket is the same platform name as today’s SWITCH marketplace; the thesis itself does not describe it as a predecessor, so that link is this wiki’s inference.

Relevance to the wiki