SED internal workshop

Day-ahead forecasting of net active and reactive power in the Marly low-voltage grid

Data, pipeline, resultsNovember 2026

power to the customers solar power back into the grid

Contents

01Forecasting modulewhat goes in, what comes out
02The datastructure, grid, meters
03Cleaningthe meter as the unit, and what the model trains on
04The pipelinestage by stage
05Resultsthe forecasts delivered for the grid
06SED Data Explorera short tour
07Next stepsand what would help

01Forecasting module

Tomorrow, quarter hour by quarter hour, with a range

The data

Structure, grid, meters

02The data

One database, three kinds of table

02The data

The grid is a tree

★18 kV feed-in transformer switch consumer prosumer both

nodes, 195 of them customers

transformers, 18 kV to 0.4 kV

switches, all closed

hops on the deepest path (19 without switches)

02The data

Most nodes have one meter, a few have very many

02The data

Three properties that shape every decision

> 4years

between the first and the last meter of one node

Meters of the same node arrive years apart, so a building's history is a staircase.

0days

median history of the 922 meters reporting in September 2026

Two commissioning waves, 2024 and 2025: one to two years of history per meter.

meters at the 13 transformers

Nothing is measured at the transformer. A station's load is the sum of the meters below it.

Cleaning

The meter as the unit, and what the model trains on

03Cleaning

The meter is the unit of cleaning

One clock

Daylight-saving fold resolved, timestamps snapped to the 15-minute grid, duplicates removed, regular index.

Live span

Before commissioning or after a meter stops, a slot is out of span: a different fact from missing, never repaired.

Quality checks

EN 50160 voltage bounds, frozen signals (8 identical slots or more), a −0.05 kW floor for consumers only.

Repair

Gaps of 1 to 2 slots interpolated, across midnight too. Longer gaps from the same weekday 7 to 28 days earlier, past only, level-matched, capped at 25 % of a day (40 % for prosumers).

Cleaning works meter by meter, in four steps. [click] One clock: the daylight-saving fold is resolved, timestamps are snapped to the fifteen-minute grid and duplicates are removed. This part is shared by all the meters of a node. [click] Each meter then gets its live span. Before it was commissioned, or after it stopped, a slot is out of span. That is a different fact from missing, and it is never repaired. [click] Quality checks: voltage outside the EN 50160 bounds, frozen signals of eight identical slots or more, and a small negative floor for consumers only, since prosumers export. [click] Repair: gaps of one or two slots are interpolated, also across midnight. Longer gaps are filled from the same weekday seven to twenty-eight days earlier, from the past only, level-matched, and capped at a quarter of a day for consumers and forty percent for prosumers. Every value keeps a code saying which of these it is: measured, interpolated, copied from a reference day, or missing.

03Cleaning

What the model is trained on

A day trains or scores only if all 96 of its quarter hours are measured or interpolated. Reference-filled values never train and are never scored.

The pipeline

Stage by stage

04The pipeline

Eight stages, one station at a time

04The pipeline

Scope: which meters belong to a station

Edge direction is unreliable. 32 of 282 edges are stored child to parent: walk the graph undirected, orient it from the feed-in.
A customer node can have children. 19 of 195 do (a looped service): keep walking through them.
cn_fk is the stable key, node ids are not.
Declared is not reporting. In scope only with 3 months of active power; stopped for more than 14 days means history, not forecast.

Counts measured on the release the models were trained on.

Scope answers a simple question: which meters belong to a station? We walk the tree from the transformer down; here, the station with the most customers. Four rules, each of which silently gives a wrong scope if ignored. [click] Edge direction is unreliable: 32 of 282 edges are stored child to parent, so the graph is treated as undirected and oriented from the feed-in. [click] A customer node can have children: 19 of 195 do, a looped service, so the walk must continue through them. [click] cn_fk is the stable key; node ids are not. [click] And declared is not reporting: a meter is in scope only if it sent active power over at least three months, and a meter that stopped more than fourteen days before the end of the data is history, not something to forecast. These counts were measured on the release the models were trained on.

04The pipeline

Features: what is known on the evening of day D

Features are everything known on the evening of day D, when we forecast day D plus one. Each training row is one meter, one day and one quarter hour; here the target is 13:00 tomorrow. [click] The largest family is the meter's own load, 21 features: yesterday at this slot and two slots either side, the same weekday last week at this slot and one either side, the past week at this slot, and statistics of yesterday. [click] Then tomorrow's calendar and weather: slot, weekday, weekend, Fribourg holidays and time of year; temperature, humidity and dew point, plus solar radiation and clear-sky irradiance for prosumers. [click] Plus the meter's size: 38 features for consumers, 41 for prosumers. A test perturbs every load value from day D plus one onward and checks that no feature moves, so nothing from tomorrow can leak in.

04The pipeline

One XGBoost per class, all 19 quantiles at once

Pooled over meters, made fair

Each meter scaled by its own mean |P|; each meter weighs the same in total; its size goes back in as a feature.

One model, 19 quantiles

5 % to 95 % for 96 quarter hours, sorted so they never cross, then corrected per quarter of the day on the validation month.

Station 220, final run

	Consumers	Prosumers
Meters	37	14
Features	38	41
Training rows	1.09 M	1.02 M
Trees kept	597	1,107
Fit on Baobab	48 min	76 min

Depth 6, learning rate 0.05, at most 2,000 trees, early stopping after 50 rounds.

The model: one XGBoost per class, consumers and prosumers, each pooled over all the meters of the station. [click] Pooling is only fair with three precautions. Each meter is scaled by its own mean absolute power over the training period; each meter carries the same total weight, however long its history; and its size goes back in as a feature, so the model can still tell a large consumer from a small one. [click] One model outputs all 19 quantiles, from 5 to 95 %, for every quarter hour. The quantiles are sorted so they never cross, then shifted per quarter of the day on the validation month, because the raw intervals were too narrow. [click] For station 220: 37 consumer and 14 prosumer meters, about a million training rows each, 597 and 1,107 trees kept by early stopping, 48 and 76 minutes on Baobab. Two data rules on top: meters that read zero throughout training are forecast as zero, and a 2021 to 2024 block that moved between meters across releases is left out of training.

04The pipeline

Walk-forward, with a sealed final month

Two references: same as yesterday, and same day last week, each with a range from its own errors over the previous 56 days.

Scores: CRPS, MAE, nMAE, 90 % coverage, and Diebold-Mariano tests against both references.

Caveat: training and evaluation use the measured weather of tomorrow, so the scores assume a perfect weather forecast.

04The pipeline

What the model relies on

04The pipeline

Running it reproducibly on Baobab

One commit

The code is frozen in git; nothing uncommitted can run.

One bundle

The commit, its file list and a checksum for every file.

Slurm jobs

One job per station and step on the public CPU partition.

Results, stamped

Forecasts, scores and logs, each traceable to the commit that made them.

Results

The day-ahead forecasts delivered for the grid

05Results

What was delivered

customer nodes, the whole grid

P & Q

active and reactive power

quarter hours a day, median and 90 % interval

05Results

Against two simple references

05Results

A prosumer, Saturday 26 September

Active power

Reactive power

05Results

A consumer, Sunday 20 September

Active power

Reactive power

SED Data Explorer

A short tour

06SED Data Explorer

NodesFilter to prosumer and mixed nodes, sort by length of history, open one.
Node detailNode 241: P, Q, voltage, current; gaps and partial sums flagged; one week; its PV meter alone.
NetworkThe grid as a graph, by transformer station.
New NetworkThe October topology with its switches, linked to the current records.
ForecastsStation 6, node 107, the sealed test month: forecast, 90 % range and measured, day by day.

New Network shows the October release; the other views serve the August 2026 release.

The Nodes view: every node with its equipment and how much data it has. Here filtered to prosumer and mixed nodes, sorted by length of history, and node 241 opened. [click] Node 241 in detail: active and reactive power, voltage and current. Gaps are left empty, and stretches where only some of its meters report are shaded. Then one week at fifteen-minute resolution, and its PV meter alone, exporting at midday. [click] The Network view: the grid as a graph, grouped by transformer station. Hovering a station highlights its customers. [click] New Network shows the October topology, with its switches dashed, and every node links back to its current record. [click] And the Forecasts view: the XGBoost station forecasts. Here station 6, node 107, a prosumer, in the sealed test month: forecast, 90 % range and measured values, day by day, with the day's scores.

07Next steps

Next steps

Consistent levels

Make station and grid forecasts agree with the node forecasts beneath them (reconciliation).

Would help most

One transformer measurement

Even for a single station: today a station's load is only the sum of its meters, with nothing to check it against.

SED internal workshop

Thank you

Questions

November 2026