Where the river loses draft
The whole convoy loads for the worst pass on the route, and the decisions that matter get made weeks before loading, when the forecast only says how much it will rain.
The river restricts navigation at short stretches where the riverbed rises and draft drops, and the model we propose maps them by crossing two signals:
- Public hydrometric data (ANNP): how much water each stretch has.
- AIS: the positions that vessels themselves transmit (picked up by satellite or coastal stations), showing where convoys slow down, wait, or transfer cargo.
The second chart shows what is at stake: the Brazil-Paraguay-Bolivia stretch moved 19.1 million tons in 2020, with soy and its derivatives in front (BCR, Feb 2022), and low water has a price: in July, the Paraguayan fleet estimated it would miss between USD 120 and 150 million across all of 2021 (La Nación PY, Jul 2021). Every centimeter of draft lost is cargo left sitting at the port.
The model we propose, and the decision it enables
This is how we design it: hydrometric data, AIS and the global weather models ECMWF/GFS feed an LSTM, a neural network built for time series, with two outputs:
- Expected level per stretch, with an uncertainty band, 7 to 30 days ahead.
- Projected state of each critical pass: when it crosses the threshold and for how long.
The data: DHN and ANNP level series and AIS traces, all of them public. The freight and demurrage costs and the dredging cost per stretch are data we do not have today: they get collected with each client at the start, with source and frequency agreed on.
The technique: the LSTM forecasts the level; on top of it, mixed integer linear programming (MILP) builds the dredging plan under a finite budget.
The decision: which stretches to dredge, in what order and with which dredge; and for you as operator, when to sail and how much to load.
Which stretch to dredge first. Dredging the fourth stretch helps little if the third stays restricted; that is why the MILP optimizes the plan over the whole network. The analysis is strictly technical: it does not evaluate contracts, bids or contractors.
This week’s decision (illustrative example)
| What the forecast says | Decision |
|---|---|
| The critical pass crosses the threshold in ~2 weeks | Sail earlier, full load |
| Stretch already restricted, level rising | Wait for the window |
| Chronic restriction with heavy traffic | Dredge that stretch first |
Example scenarios, not a real recommendation: the rule gets validated with the operator’s data.
Other options on the same window: scheduled lightering · draft adjustment · rescheduled departure.
The validation
The validation will put the model up against real corridor data, and for that we are looking for operators: the paid pilot is one operator, one stretch, one season and metrics agreed on before starting. The result gets documented, whether it works or not.
The business case. Two numbers almost nobody calculates, and they come straight from your own shipments:
- What the last low-water period cost you and how much was avoidable with two to four weeks of notice.
- Which pass costs you draft and which dredged stretch gives you back the most cargo per trip.
How to measure it: forecast error against the observed level per stretch, and days of notice gained over the reference bulletin. It rests on a digital twin of the critical stretch: your own water-level and sedimentation sensors on top of the Navy and ANNP bulletins. A sail-and-load rule learned from your operation comes later, on top of that twin.
This map of critical passes stays open for debate across the corridor’s whole chain: cargo owners, shipowners, ports, exporters. If your operation reads the draft differently, or your records tell a different story, write to us and we will check the map against the data.