Probabilistic forecasting of net active and reactive power in the Marly
low-voltage grid
Yverdon WorkshopNovember 2026
HES-SO
power to the customerssolar power back into the grid
Contents
01The questionwhat a day-ahead forecast is for
02The districtMarly, and what we can see of it
03What makes it hardthree obstacles in the data
04Our approachthree levels, and an interval rather than a number
05Where we are, and what we needfirst results, and the one thing missing
01
The question
What a day-ahead forecast is for
01The question
What a grid operator would like to know
What will each part of this district draw from the grid, or push back into it,
tomorrow, quarter hour by quarter hour?
01The question
Measured today, forecast tomorrow
The measured day is the real total of the district's 184 customer nodes. The forecast day is the sum of
the node forecasts; its band is approximate, drawn to show the shape of the output.
01The question
Why an interval and not a number
Decisions taken today are about capacity
How much flexibility to reserve, which feeder to watch.
A single predicted curve hides how wrong it could be
A range does not.
02
The district
Marly, and what we can see of it
02The district
Marly: a real district, fully metered at the customer
Meters in the same node arrive years apart, so a building's history is a staircase
rather than a record of its behavior.
0
meters at the 13 transformer stations
Nothing is measured at the transformer. A station's load can only be rebuilt by adding up
the meters below it.
03What makes it hard
The fleet is young
923
meters reporting in August 2026 5 in 2022
04
Our approach
Three levels, and an interval rather than a number
04Our approach
Our approach
Forecast every level, and keep them consistent
A forecast for a station should equal the sum of the forecasts beneath it. That consistency is what makes the numbers usable.
Give a range, not a line
Each forecast carries an interval, so an operator can size a margin instead of trusting a single value.
Hierarchical reconciliation is the direction being adopted; it is not yet implemented.
04Our approach
Pipeline
Hierarchical reconciliation is the direction being adopted; it is not yet implemented.
05
Where we are, and what we need
First results, and the one thing missing
05Where we are
A real day-ahead forecast
Node 123, a prosumer Tuesday 29 September 2026
net active power, every 15 minutes
measuredforecast, median90 % interval
0kW
mean absolute error on this day
0%
lower error than repeating yesterday, over 30 test days
0%
of measured values inside the 90 % interval, over 30 test days
05Where we are
The same day, reactive power
Node 123, a prosumer Tuesday 29 September 2026
net reactive power, every 15 minutes
measuredforecast, median90 % interval
0kvar
mean absolute error on this day
0%
lower error than repeating yesterday, over 30 test days
0%
of measured values inside the 90 % interval, over 30 test days
05Where we are
Where we are
Done
Day-ahead forecasts with 90 % intervals now exist for every customer node in the grid:
184 nodes, active and reactive power, across all 13 transformer stations.
Next
Making the station and grid levels consistent with them is the next step.
05What we need
The one thing missing
A measurement at the transformer, even for a single station.
Today a station's load can only be estimated by summing the meters below it, which leaves out public
lighting, station auxiliaries and losses, and gives us nothing to check the estimate against.