Below the intake of a small hydropower plant there is a stretch of river that lives on what the plant lets past. The licence says how much that must be, the minimum flow. The plant has to be able to show that it was released, day after day, and in practice that showing is a record: a gauge, a meter, a level over a weir.
1,487 plantsbuilt hydropower stations under 10 MW in NVE's register, expected to give 13.2 TWh a year (NVE, NLOD)
In one decision NVE fined a plant NOK 500,000 after a physical control measurement found it releasing 13.7 litres a second where its licence requires 20; the operator, in its reply quoted in the decision, called the condition absolute (NVE, Stølselva kraftverk). NVE's guidance from April 2020, Slipp, måling og dokumentasjon av minstevannføring, is about how the release is made, measured and documented.
What we could test, and on what
We found no open record from a Norwegian small hydropower plant. NVE's gauges are open data, but the interface needs a key of one's own. So we took the strongest open record we found: five U.S. Army Corps of Engineers plants in Oregon, of 15 to 35 MW, with their hourly output, turbine flow, spill and levels (USACE, a U.S. federal work) and a USGS gauge below each (USGS, public domain). They are larger than Norwegian småkraft, and the minimum flow there is the whole release of a dam, not the water left in a bypass reach. The physics is the same.
Turbine flow from what the generator says
4% to 8%median error of turbine flow worked out from output and head, five years after the efficiency was fitted, at all 5 plants
One efficiency per plant, fitted on 2015 to 2019, gave the turbine flow of 2020 to 2024 within that range. Over all plant-days it beat a constant flow per megawatt, 5.5% against 9.1%. The head helped where a reservoir rises and falls, and made no difference at two dams with a steady head. These were the plants' own figures, which the Corps most likely works out from its turbine ratings, so this says the estimate agrees with the plant, not with a meter in the pipe.
The negative that matters most
Output and levels were not enough to close the balance. Only 2 of 5 plants agreed with their gauge within 10% on at least 80% of days when the turbine flow was worked out instead of recorded. Water leaves by routes the generator does not see: a fish ladder, an outlet, a day the powerhouse is off. In Norway the minimum flow itself does not pass through the turbine at all. A pilot needs a record of the release, and the product will not pretend otherwise.
Two records that agree, and two that do not
Where the release was recorded, it agreed with the gauge below within 10% on at least 90% of days at 4 of 5 plants, and on up to 99.6%. At Foster it agreed on 73%, because the gauge there sees more water than the plant releases.
That gap decides what gets reported. At Foster the plant's record put 63 more days below 95% of the requirement than the gauge did. Which of the two is right is a question for a person who knows where each one sits. The console puts the two side by side and does not answer it for them.
A meter that reads 10% high
58% against 22%of 500 faults planted in the plants' records, found by the balance against the gauge and by fixed limits, at the same false-alarm budget
Fixed limits on one record found 99% of the frozen records and almost nothing else: 1% of the records that read 10% high, against 42% for the balance. A meter that reads high is the one that hides a low release. The balance found up to 94 of 100 faults at a plant where the two records agree closely, and only 20 of 100 where they disagreed for weeks in the calibration years. It fell short of the 80% we set in advance, and it stayed within its alarm budget at 3 of 5 plants. A second record is only as useful as its agreement with the first.
What this means for a plant on shift
The operator does not need a new model of the river. They need to see, for each day, what was released, what the gauge below saw, what the licence asked for, and which of those numbers is measured, which is worked out and which is unknown. When the two records part, someone should be told before the report is sent, not after a control measurement.
Every figure here is from the results files of experiments P67 and P68, preregistered before the held-out years were scored. Both came out PARTIAL, and the reports say why.