Inside the concrete dam, machines hum, running on empty. Three out of ten turbines are silent. The water level in Southeast Brazil’s largest reservoir has dropped so low that the blades are about to gulp air. The country’s 160 hydropower plants, which supply more than half its electricity, are linked into a single water network where every drop counts. Yet for years, an invisible mirage of a full reservoir lulled operators into spending water as if the rains would surely come tomorrow. This wasn’t chance: the official models were systematically overestimating future inflows.

At the heart of the management system is the stochastic dual dynamic programming (SDDP) algorithm, which calculates the 'water value': how much future cost is saved by holding back one more cubic meter in the reservoir. This value is no abstraction: it’s a mathematical shadow of tomorrow’s expenses: w = -∂Q/∂v. When the model forecasts abundant rains, the derivative of future costs with respect to volume drops, water appears cheap, and it gets flushed through the turbines today. But when inflow forecasts are systematically inflated, exactly what happened in Brazil for over a decade, the mirage of swelling rivers erodes the logic of conservation.

The researchers tested the mechanism in three stages. First, on a stylized analytical model, they showed how even a small positive bias in expected inflow weakly lowers the water value and encourages greater releases. Next, using real operational data from 2014–2026, they found a statistical link: months with overestimated forecasts coincide with overestimated planned reservoir levels. Finally, with numerical simulations, two policies (one trained on biased forecasts, the other on corrected ones) were tested on 2000 inflow realizations, modeled as a stochastic process with Markov transitions, named after Andrey Markov.

The results lay bare a drama. Under optimistic forecasts, the average reservoir level by the end of the dry season drops by 29 gigawatts — a third of the country’s average demand. Thermal generation is postponed for months, and then, at the peak of the drought, the most expensive plants are fired up. The upshot: the probability of rolling blackouts hits 10,3%, versus 0,9% under an unbiased policy. Average spot prices over five years end up 34% higher, and dry-season volatility skyrockets. Total operating costs, including blackout penalties, shoot up sevenfold.

The mirage has lifted. When forecasts are stripped of systematic error, the water in the reservoirs regains its true worth. But the transition is painful: spot prices jump even in wet seasons, and the market goes into shock. This work exposes a vicious circle familiar to any system governed by marginal storage values — from pure hydro to future hybrids with batteries and hydrogen. The next step is to embed statistical corrections directly into the SDDP algorithm, account for grid constraints, and apply Bayesian correction of forecasts, extending the analysis to other countries where water flows through turbines and mirages.