Energiforsk 2018-537 Dataanalys och Avancerade Algoritmer (2018)

Full citation: Persson, M. (RISE), Sandels, C. & Nilsson, A. Dataanalys och avancerade algoritmer — Möjligheter med utökad mätinfrastruktur. Energiforsk Report 2018:537. Stockholm: Energiforsk AB, October 2018. Program: Smarta Elnät. ISBN 978-91-7673-537-4.

Open access: Published by Energiforsk. Project led by RISE (Mattias Persson, project manager); the reference group was Anders Lindskog (RISE), David Westerlund (Rejlers Embriq) and Lars Olsson (Seniorit). Same Smarta Elnät programme board as Source - Energiforsk 2018-540 Framtidens Nätstation (2018) (which includes Vattenfall Eldistribution, Ellevio and Jämtkraft); Ellevio, Vattenfall Eldistribution and Jämtkraft Elnät are among the listed project stakeholders.

Summary

A RISE study for Energiforsk’s Smarta Elnät programme on what richer metering — more parameters per meter, higher sampling frequency, nätstation-level measurement — actually enables for a DSO, using non-technical loss (ITF, icke-tekniska förluster: mis-documented networks, theft, other unmetered consumption) detection and localization as the concrete worked case. The report combines an interview survey of seven DSOs with an evaluation of correlation analysis, power-flow analysis, and machine learning for detecting and localizing losses.

Key claims

The regulatory backdrop and scale of the problem

Ei has proposed an efficiency incentive for network companies. As of the report’s data, 54 of 159 Swedish network companies had annual distribution losses above 4%. If those 54 companies had reduced losses to 4%, the aggregate reduction would be 143 GWh/year; at an assumed cost of 50 öre/kWh, that is equivalent to 72 million SEK/year (the report’s own back-of-envelope exercise: 4% is its assumed ceiling, against a national average of 3.6%; 143 GWh is 3.9% of total Swedish network losses; the 50 öre/kWh comes from Ei R2015:07) — either unnecessary cost paid at the connection point, or lost revenue for the network owner, depending on where the loss actually sits.

Storage cost is not the barrier

The report’s summary states that: “Compared to the cost of the new generation meters, storage costs for additional metrics and increased measurement resolution will not be the major cost driver.” The DSO interviews found wide variation in how network companies manage losses and identify/localize non-technical losses specifically — practice is inconsistent across the sector, not just immature everywhere equally.

Localization methods evaluated

Correlation analysis, power-flow analysis, and machine learning are each evaluated for detecting and localizing different types of non-technical loss. In the report’s simulated networks, machine learning can localize losses without perfect knowledge of network structure, provided the losses follow a consumption pattern and loss-free periods exist to train the algorithm on. For random (non-pattern-following) losses, three specific methods were developed and evaluated — SiM, K:SE, and K:V — using energy-meter voltage values, with promising localization and detection results in simulation; the report notes that a real implementation of all three is still needed, and that K:SE was not promising in the larger simulated network. ML localization fails for uncorrelated (random) losses (0-5% accuracy), which is why the voltage-based methods were developed. The report also studied how metering resolution affects localization accuracy, and specified requirements for next-generation meters needed to support the proposed analysis tools; its conclusion is that DSOs should set their own meter requirements ahead of the next roll-out, with voltage values a wise and relatively cheap one.

Relevance to the wiki