Dam release planning for the Arda cascade, Bulgaria

On 3 December 2017, 682 m³ of water per second poured into Studen Kladenets. The dam released 42.

Bandkeeper gives reservoir operators a release plan for the next seven days. The plan keeps room for the next flood, holds water for the dry season, and never releases more than the dam already did in 2017.

28days over the upper schedule, as operated
5days with a 3-day forecast, same peak release
Volume as operatedVolume with BandkeeperUpper schedule (flood room)Lower schedule (dry-season reserve)Inflow

Built for Track 3, topic 03

The organisers asked for hyper-local extreme-rainfall forecasts that feed two decisions. Bandkeeper takes the second one and runs it on the cascade data they shared.

“For automated urban stormwater surge controls and dam release planning.”EIT Water Hackathon Sofia 2026, Track 3 topics of interest

Predict

Reads the inflow forecast for the next one to seven days.

Prevent

Lowers the reservoir before the flood arrives, never faster than the dam has released before.

Manage

Gives the operator one release figure for today and the reason behind it.

One reservoir, two guides, one decision every morning

The operators of the Arda cascade file a daily report for Kardzhali, Studen Kladenets and Ivaylovgrad. Each report checks the reservoir against the Ministry of Environment and Water’s fill schedule, an upper and a lower volume for that time of year, and against a monthly quota on how much water the dam may release.

The schedule carries no fines. It marks where flood room runs out and where summer supply gets tight. The operators measure every day at manual and automatic stations, but the report records yesterday and shows nothing about next week’s rain.

80 daysreservoir-days above the upper schedule across the three dams in 2017
39 daysreservoir-days below the lower schedule in the same year
Both sidesKardzhali left the schedule on both sides within twelve months: 31 days too full, 34 days too empty

Counts from 1,005 daily records, February to December 2017, checked against the schedule printed in each report.

What a sector mentor told us on 24 October

Operators set daily releases from manual and automatic measurement stations. No regulation fixes a safe downstream flow; it depends on the river. The ministry sets monthly limits on how much water each dam may use, and the fill schedule carries no financial penalty. We changed the model to match.

Replay 2017 yourself

Pick a reservoir and a forecast window. The grey line shows what the operator did. The blue line shows the same reservoir following Bandkeeper’s plan on the same inflows.

Period
As operatedBandkeeperUpper scheduleLower scheduleInflow
As operatedBandkeeper

The cascade on one day

Drag through 2017. The fill shows how full each reservoir is. The frame turns red above the upper schedule and amber below the lower one.

How the plan works

  1. Read the daily report

    Volume, inflow, release and both limits for each reservoir, in the format operators already file.

  2. Take the inflow forecast

    One to seven days ahead. In this replay we use observed inflow with 5% noise.

  3. Find the strictest day

    For each day ahead, compute the release that keeps the reservoir 5 million m³ under the upper schedule. Keep the largest.

  4. Respect the river and the quota

    Cap the release at the highest flow this dam has already released, stay inside the monthly quota, and with no flood ahead, hold water above the lower schedule.

Checked against the operator’s numbers

We rebuilt each reservoir’s water balance from the reports: tomorrow’s volume equals today’s plus inflow, minus release and losses. Our balance matches the reported volumes with a correlation of 1.00 on all three dams.

Hydrologists use general reservoir models such as HEC-ResSim and RTC-Tools, which need a calibrated basin model and a specialist. Bandkeeper starts from the report the operator fills in today and runs in a browser.

What we have not proven yet

  • A real forecast misses more than 5%. Set the bias slider to −30% to see how the plan holds up.
  • No rule fixes a safe downstream flow. We cap releases at the highest flow each dam made in 2017, so the river sees nothing new. The operator can set a different figure for each river.
  • With that cap, Studen Kladenets still spends 5 days over the upper schedule in December 2017, down from 28. A higher cap clears more days, and only the operator can judge whether the river takes it.
  • The dam reports cover 2017. We have not linked rainfall to inflow yet.

The same basin dries out

Topsoil moisture in the Arda basin fell from 0.25 m³/m³ in January 2024 to 0.09 in August. A reservoir drained to dodge every winter flood leaves farms and towns short in a summer like that one.

Bandkeeper plans no release below the lower schedule unless the forecast shows the reservoir running out of flood room. When the two limits conflict, the plan shows you the trade-off and you decide.

Satellite soil moisture, monthly means, from the organisers’ Earth Engine export.

Arda basinSofia and upper Iskar

Where Bandkeeper sits, and who pays

Rain and inflow forecastNowcasting and sensor networks, such as the flash-flood services shown at this hackathon
BandkeeperTurns the forecast and the reservoir’s limits into today’s release
Reservoir operatorApproves or overrides the plan and logs the reason
Towns and farms downstreamReceive a release alert before the river rises

Reservoir owners and hydropower operators pay a yearly licence per reservoir. They keep flood room without opening the gates wider than before, keep water for summer supply and power, stay inside the monthly quota, and get a record of why they released what they did on each day.

Basin directorates get the same view for every dam they supervise. Towns downstream receive alerts at no cost. We start in Bulgaria, where one ministry sets the fill schedule and monthly quotas for large reservoirs, so one report format covers the country. Sofia’s Iskar reservoir faces the same flood and drought trade-off.

What the pilot measures

  • Days above the upper schedule, compared with earlier years.
  • Days below the lower schedule.
  • Peak release into the river, against the operator’s own record.
  • How often the operator follows the plan, and why they override it.

We set the licence price after the pilot, once we can put a value on flood room kept and water saved for summer.

What we build next

  1. Train a rainfall-to-inflow model for the Arda basin on ERA5-Land data and replace the replay forecast with it.
  2. Run a one-season pilot with one reservoir operator on live daily reports.
  3. Plan the cascade as one system, so Kardzhali’s release feeds Studen Kladenets’ forecast.
  4. Send release alerts to towns and farms downstream.

Team

We are two students from Bandung, Indonesia, competing online. We will keep working on Bandkeeper after the hackathon and want one operator to test it with us in 2027.

Dyka Hatawa

Dyka Hatawa

Data, simulator and product

Informatics, Institut Teknologi Nasional Bandung. Built the data pipeline, the simulator and this site. Trained in electrical systems and heavy-equipment maintenance.

Reisya Putri Ramadhani

Reisya Putri Ramadhani

Customers and business case

Ran the sector interview with a water-sector mentor and shaped the pricing and the pilot plan.