Water value and future cost
The water value is the opportunity cost of using water now instead of storing it. It is not an input: it emerges from the optimization. Novomodelo summarises the future in a future cost function, which SDDP builds step by step from Benders cuts.
Why it matters
Seção intitulada “Why it matters”The future cost function is the expected cost from the current stage to the end of the horizon, as a function of the initial state: reservoir levels and inflow lags. When the subproblems are LPs, it is convex and piecewise linear.
SDDP never builds the whole function. It approximates it from below with a growing collection of Benders cuts, each a valid underestimate. During training, no active cut leaves the lower-bound problem, so that bound does not decrease.
Each cut has, for each reservoir, a storage coefficient, the marginal value of water. Its sign is generally negative or zero: more storage usually means lower future cost.
Cuts draw the future cost function
Seção intitulada “Cuts draw the future cost function”What the model decides
Seção intitulada “What the model decides”- Approximate future cost of each stage $
- Water value of each plant, per stage and block $/hm³
The model is not given the water value: it results from the optimization. In each stage, the future cost variable is bounded by all current cuts, and each plant's water balance supplies the stage's water value.
What you provide
Seção intitulada “What you provide”There is no water value to provide. What you control is how training builds the function and where it starts from.
- The training stopping criteria: an iteration or time limit, a stalled lower bound or, with an exact upper bound, a gap within tolerance.
- The policy mode: start from scratch, with an empty future cost function, or start from a previous checkpoint.
- The condition at the end of the horizon: by default, the future cost after the last stage is zero, but a study can import it from a terminal boundary (boundary cuts).
- Optionally, cut selection and a limit on active cuts, to control the growth of the cut pool.
How the function is built
Seção intitulada “How the function is built”Each SDDP iteration has two steps.
- Forward simulation. A batch of scenario trajectories is simulated under the current policy, and the state each trajectory reaches at each stage is recorded: the visited states.
- Backward simulation. From the last stage to the second, each stage’s subproblem is solved at each visited state, for all of the stage’s openings. The cut coefficients come from the reduced costs of the pinned incoming-state columns and, for the inflow history, from the row duals by the chain rule, and one new cut per visited state is aggregated for the previous stage.
After these steps, the lower bound is evaluated by solving the first stage over its openings with the current cuts. The cycle repeats until the stopping criteria are met, and the final cuts define the policy.
Two similar numbers that are not the same
Seção intitulada “Two similar numbers that are not the same”In the methodology, the cut’s storage coefficient is called the marginal value of water. The water value reported in the results is a different quantity: the dual variable of the plant’s water-balance constraint, in $/hm³. They differ because the cut coefficient is the reduced cost of the pinned initial-storage column, which combines the water-balance dual with the duals of the other constraints that read the initial storage, such as those of the production function and of evaporation.
How to read the water value in the results
Seção intitulada “How to read the water value in the results”- The
water_value_per_hm3column ofsimulation/hydros/holds each plant’s water value. In a chronological stage, each block carries its own block’s value; in a parallel stage, every block carries the stage row’s value. - To aggregate a chronological stage, weight the blocks by their hours.
- The value is zero or negative whenever one extra hm³ reduces the cost, because the water displaces thermal generation or deficit, now or later. A larger magnitude means scarcer water, which the solver preserves for later stages; a value near zero means an abundant reservoir.
- A positive value appears when the extra water can only leave the reservoir at a cost, for example through a spillage penalty.
The per-stage cost breakdown reports the immediate cost and the future cost separately.