LP Formulation
Purpose
Seção intitulada “Purpose”This chapter presents the complete stage subproblem LP for the Novomodelo SDDP solver: the objective function with its cost taxonomy, all constraint families, slack/penalty variables, and the Benders cut interface to the future cost function. It uses the parallel blocks formulation by default.
Reading order: System Element Modeling Overview → Equipment-Specific Formulations → this chapter → State Augmentation
For what each physical element represents and its decision variables, see System Element Modeling Overview. For variable naming conventions and index sets, see Notation Conventions.
For the entity view of this chapter, see Hydro plants, Buses and subsystems, Load and Generic constraints.
Stage LP at a Glance
Seção intitulada “Stage LP at a Glance”Given its incoming state, the stage LP minimizes the objective below subject to the rows of every family it carries, each written in its parallel-stage form with the symbols of its owning section, standing for terms that section writes in full; the in-transit bucket, plant minimum and maximum, and commitment-ring families are named in text lines. is the realized inflow of hydro and the right-hand side of cut , numbers computed before the solve (§5, §11).
The table lists the families in the order of the stage LP’s rows, with hydros and blocks at the stage.
| No. | Row family | Rows per stage | Section |
|---|---|---|---|
| 1 | Water balance | on a parallel stage, on a chronological stage | §4 |
| 2 | In-transit bucket definition | one per receiving plant and maturity lag reached at the stage | State Augmentation — Bucket definition rows |
| 3 | Load balance | , one per bus and block | §3 |
| 4 | FPHA planes | one per plane, (hydro, bus) cell and block of each FPHA hydro; none while the hydro is PreFilling or in its filling window | §6 |
| 5 | Evaporation | one per evaporating hydro on a parallel stage, one per evaporating hydro and block on a chronological stage; none while the hydro is PreFilling | §4 Evaporation Row |
| 6 | Minimum outflow | , one per hydro and block | §7 |
| 7 | Maximum outflow | , one per hydro and block | §7 |
| 8 | Minimum turbined flow | one per (hydro, bus) cell and block | §8 |
| 9 | Minimum generation | one per (hydro, bus) cell and block | §6 |
| 10 | Plant minimum and maximum | for a hydro with several cells, per block: a turbined-flow and a generation minimum row when the plant minimum binds on its cells’ total at some block of the stage, then a turbined-flow and a generation maximum row when the plant maximum does | §8 Plant Minimum and Maximum |
| 11 | Filling floor | one per filling hydro at each stage of its filling window | §8 |
| 12 | Soft dead-volume floor | one per filling hydro at each stage from its entry stage on | §8 |
| 13 | Commitment fish | one per anticipated thermal whose delivery at the stage was decided at an earlier stage or before the study | State Augmentation — Ring Rows |
| 14 | Commitment deposit | one per anticipated thermal that decides at the stage a commitment for a delivery inside the study or the declared post-study calendar at which the plant is commissioned | State Augmentation — Ring Rows |
| 15 | Commitment carry | one per ring slot holding a later delivery, decided at an earlier stage or before the study, inside the study or the declared post-study calendar | State Augmentation — Ring Rows |
| 16 | Generic constraints | one per active bound and block it covers, or one stage-level row for a block-independent bound | §10 |
| 17 | Benders cuts | one per active cut; one per resident cut under the dynamic method | §11 |
Incoming-state pinning and the lifecycle pins are column bounds, not rows: each incoming-state column is fixed at its trial value by equal bounds (State Augmentation §2), and each column a lifecycle phase freezes is fixed by its bounds (§8). The PAR(p) inflow model has no row and no column: the realized inflow and the inflow lags reach the LP through right-hand sides (§5, PAR(p) Folding).
1. Cost and Penalty Taxonomy
Seção intitulada “1. Cost and Penalty Taxonomy”Penalty System owns the cost categories, their units, their priority ordering, the ordering checks and the resolution cascade; the table lists the cost symbols of each category as this page writes them.
| Category | Cost symbols |
|---|---|
| Thermal and contract costs | , |
| Recourse slacks | , , |
| Constraint violation penalties | , , , , , , , , , |
| Regularization costs | , , , , |
Pumping carries no cost term; the power a pumping station draws enters the load balance of its bus (§3, Equipment-Specific Formulations §4).
1.1 Objective Function Structure
Seção intitulada “1.1 Objective Function Structure”The complete stage objective is:
where the per-block terms are weighted by the block duration , except the amount-basis slacks of a block-level generic constraint (§10), and the stage-level terms, namely the storage-violation slacks on end-of-stage storage, the inflow non-negativity and withdrawal slacks, the evaporation slacks of a parallel stage, the slacks of a stage-level generic constraint (§10) and the commitment cost of an anticipated thermal, apply once per stage (§2); the per-block part of each component is:
The coefficient of is the one-step discount factor of Discount Rate Formulation, equal to 1 at a zero rate.
2. Objective Function
Seção intitulada “2. Objective Function”Under penalty, whose slack has no upper bound, wherever the inflow non-negativity slack enters the water balance of (§4), it supplies water to that balance at per hm³, so on a parallel stage the value of one more hm³ of water at is at most . On a chronological stage the slack enters every block row with its , and the bound holds for the duration-weighted average, with weights , of the block values, not for each block. The bound is one-sided: it caps the value of water, not a negative value of surplus water. On a parallel stage the storage cut coefficient read at the stage inherits the bound, , where enters only the water balance; an FPHA plane, the evaporation row or a generic constraint that reads adds its own terms. The truncation methods cap the slack at the truncation amount (Inflow Non-Negativity Solution Methods §1). Under truncation_with_penalty the bound holds at a solve whose realization is negative and whose slack lies strictly below its cap; at a non-negative realization the cap is and the fixed slack bounds nothing, and inherits the bound when it holds at every opening the cut combines. Under truncation, whose fixed slack bounds no water value, there is no such bound. Under both priced methods, Penalty System — Penalty Priority Ordering converts the inflow cost to $/MWh and orders it below deficit, so that the LP takes slack water before it sheds load.
Anticipated thermals (State Augmentation §5) add one stage-level commitment-cost term per plant deciding a commitment at the stage — not a per-block term: the commitment column priced at the plant’s unit cost and hours of the delivery stage and discounted from it (State Augmentation — Objective contributions). Their per-block generation carries no thermal cost at a stage where the plant has a fish row (State Augmentation — Ring Rows), which ties it to the committed rate; at any other stage it is priced like any thermal.
3. Load Balance Constraint
Seção intitulada “3. Load Balance Constraint”Each hydro plant is partitioned into one or more (hydro, bus) cells — one cell per distinct bus among the plant’s declared unit groups (see System Element Modeling Overview §5). denotes the set of buses hosting one of ‘s cells; denotes the hydros with a cell at bus (i.e. ). is the generation of hydro ‘s cell at bus , block — the quantity that actually injects at . A plant whose groups share a single bus has , and collapses to the single-cell used everywhere else in this chapter.
For each bus and block :
Dual variable: (marginal cost of energy at bus , block , in $/MW; divide by for $/MWh — see Variable Units Convention)
For the physical meaning of each element in the balance, see System Element Modeling Overview.
4. Hydro Water Balance
Seção intitulada “4. Hydro Water Balance”Every hydro has one water-balance row on a parallel stage and one per block on a chronological stage (Block Formulation Variants); every LP variable is on the left-hand side, and the realized inflow (§5), a constant of the solve, is on the right-hand side with the withdrawal target.
Parallel-Stage Row
Seção intitulada “Parallel-Stage Row”Chronological-Stage Rows
Seção intitulada “Chronological-Stage Rows”For each block , with and :
Row Terms
Seção intitulada “Row Terms”- and = outgoing and incoming storage; is pinned to the trial value by its column bounds (State Augmentation §2). On a chronological stage is the storage at the end of block
- = in-transit volume maturing at this stage on the travel-time arcs into , in hm³ and therefore outside the conversion (State Augmentation §6); a chronological stage spreads it over its blocks by the arrival density , . Absent when no travel-time arc enters
- = realized incremental inflow, a constant of the solve computed outside the LP (§5); block of a chronological stage receives the share of its volume , the innovation included
- = inflow non-negativity slack, present under every inflow method except
none, which adds water: unbounded underpenalty, capped at the truncation amount , a constant of the solve, undertruncation_with_penalty, and fixed at it undertruncation(see Inflow Non-Negativity Solution Methods) - = the plant’s own declared withdrawal target, signed and stage-level ( adds water)
- = effective withdrawal, the quantity on the right-hand side. For a non-PreFilling plant, , where is the set of PreFilling plants whose first non-PreFilling downstream plant is ; for a PreFilling plant, (PreFilling Pass-Through)
- , = stage-level under- and over-delivery slacks of the effective withdrawal; the realized withdrawal is , with the slack caps in §9
- / = signed net evaporation of a hydro with an evaporation model (a negative value is net rainfall on the lake and adds water): one stage-level column on a parallel stage of any block count, one per block on a chronological stage, each a bounded column tied to storage by its evaporation row (see Evaporation Row)
- , , = the plant’s own turbined, spilled and diverted flow in block ; every (hydro, bus) cell’s turbined column carries the same coefficient, and spillage and diversion are single per-plant columns
- Upstream release: only the turbined and spilled flow of each reaches ; a plant’s diverted flow reaches only its diversion target (the sum)
- = same-stage share of the release of on a parallel stage, with the duration of stage and the travel time of ‘s main cascade arc ( when none is declared, so the share is ); the remaining is deposited into the in-transit buckets (State Augmentation §6)
- = within-stage routing share on a chronological stage: the fraction of ‘s block- release that reaches ‘s block- row in the same stage, . Without a travel time and for ; the rest of the release is deposited into the buckets (State Augmentation §6)
- = pumped flow of station , out of its source plant’s row and into its destination plant’s row (see Equipment-Specific Formulations); on a parallel stage every block’s pumped flow enters the single stage row
- and = stage and block flow-to-volume conversions, with the duration of block in hours and its weight, so
PreFilling Pass-Through
Seção intitulada “PreFilling Pass-Through”A PreFilling hydro (see Lifecycle Phases for the phase) has the frozen identity row , and on a chronological stage per block, with right-hand side . Its inflow non-negativity slack , the releases of its upstream plants and the flows diverted into it enter the row of the first non-PreFilling plant downstream with the coefficients they would carry on the PreFilling plant’s own row ( on on a parallel stage and on each block row of a chronological stage, on every block- flow) and no travel-time share, its local inflow moves to that plant’s right-hand side, and its withdrawal target joins that plant’s effective withdrawal , so the receiving plant’s slacks and relax it at that plant’s withdrawal-slack prices. Its maturing in-transit volume enters the same row, with coefficient on a parallel stage and on block ‘s row of a chronological stage, being the PreFilling plant’s own arrival density. With no such plant downstream the water leaves the system and the withdrawal enters no row.
Summing the Block Rows
Seção intitulada “Summing the Block Rows”Summing a chronological stage’s block rows over telescopes the storage terms to , returns (), and returns the parallel coefficients of , , and the right-hand side (); the per-block flows keep their . The sum has the parallel row’s form with two mode differences: in place of , and each upstream block- release carrying its own same-stage share in place of , whose duration-weighted average over is : . With one block, or no travel time on the arc, the shares coincide.
Dual variable: for the parallel row; on a chronological stage each block row has its own dual (water value — captures the marginal value of incoming storage as seen through the hydro balance, but is not used directly as a cut coefficient; the cut coefficient comes from the reduced cost of the pinned incoming-storage column, see State Augmentation §2 and Cut Management)
Evaporation Row
Seção intitulada “Evaporation Row”Every hydro with an evaporation model has an evaporation row that ties its net evaporation to storage. A parallel stage, of any block count, has one row per such hydro, on the stage’s incoming and outgoing storage:
A chronological stage has one row per block , on the block’s own storages, with and :
-
and = slope and intercept, at a reference volume, of the tangent of the evaporation flux of : the evaporation coefficient of the stage’s calendar month times the reservoir surface area from its area–volume curve, converted to a monthly-average rate in m³/s. Both are recomputed at every stage; the row’s target is the tangent evaluated at the average of the two storages the row reads. The flux, and so or , can be negative (net rainfall on the lake)
-
/ = net evaporation (Row Terms above), a bounded column, not a free one. Its box follows the sign of the stage month’s evaporation coefficient, the sign of the coefficient itself and not of or . With the magnitude of the target at the maximum storage times a fixed safety margin, recomputed at every stage:
- positive month: ;
- negative month: ;
- zero month: .
The same box applies to of each block, with the extremes taken over that block’s average storage. It enters the parallel water-balance row as and the block- row of a chronological stage as
-
, and = extremes of the row’s target over the average storages the row can read. and are the smallest and largest target, with the stage’s storage floor taken as the hydro’s minimum storage even where that floor is soft; is the largest target with the floor taken as the hard floor, which is for a filling hydro and for a PreFilling stage and the stage’s minimum storage otherwise (§8). The incoming storage ranges over the previous stage’s floors and ceiling, or, at the first stage and after a PreFilling stage, over the interval the initial storage is clamped into. The target is evaluated at both ends of each range, because the area–volume slope can take either sign
-
/ and / = evaporation above and below the target, priced at and : the row sets the evaporation to the target plus minus . On a parallel stage the one pair is priced over the stage hours , on a chronological stage each block’s pair over its duration (§9)
-
is unbounded in every month. is unbounded in a positive and in a zero month, and in a negative month it is capped, , which holds it to the amount that brings a positive target down to . In a positive month the floor of keeps it from going negative far enough for to create water, and in a negative month the ceiling keeps evaporation from removing more water than the target can
With one block the parallel and chronological forms coincide, rows and pricing alike. A PreFilling hydro has no evaporation row.
5. Realized Inflow
Seção intitulada “5. Realized Inflow”The stage LP has no column and no row for the PAR(p) inflow model. Before each solve, the realized inflow (m³/s) of each hydro is computed from the model at the incoming inflow lags and the noise of the opening:
where:
- = deterministic base (precomputed from seasonal means and AR coefficients — see PAR(p) Inflow Model §2.4)
- = original-unit AR coefficients, for a lag the model of does not use; for a hydro with the annual component each of these coefficients also carries the annual term (PAR(p) Inflow Model §7.4)
- = incoming inflow lags, the trial values of the inflow-lag state (State Augmentation §4)
- = noise innovation of the scenario or opening
enters every row that reads it as a constant on the right-hand side: the water-balance rows (§4), the rows of a PreFilling plant’s first non-PreFilling downstream plant (PreFilling Pass-Through) and the generic constraints with a hydro-inflow term (Hydro Inflow). No inflow non-negativity method alters it: keeps the raw realization, which may be negative, and the methods add their water to the water balance through instead (§4; Inflow Non-Negativity Solution Methods). The same value is the realized inflow the simulation reports, and the lag accumulation forms the outgoing inflow-lag state from it (Multi-Resolution Studies §2).
The inflow lags are not LP columns, so the cut rows evaluate their lag terms at and on the right-hand side (§11), and the lag coefficients of a new cut come from the duals by the chain rule. PAR(p) Folding derives this formulation, its equivalence with the formulation that writes the model as rows of the LP, and its limits.
6. Hydro Generation Constraints
Seção intitulada “6. Hydro Generation Constraints”Novomodelo supports two production models, in increasing order of complexity. The model can vary by stage or season per hydro.
Both models are evaluated per cell (§3) rather than per plant; for a single-cell plant the per-cell constraint is the plant-level constraint.
Constant Productivity Model (for each cell of hydro , block ):
Constant-productivity hydros carry no separate generation column: is this direct multiple of the cell’s own turbined-flow column, so a cell’s generation bound is enforced by folding it into that column’s bound (§8) rather than by a bound on itself.
FPHA Model (for each plane , cell of hydro , block ):
where (MW) is the FPHA plane slack, one per cell and block , and is the average storage during the stage, with being the incoming storage LP variable (State Augmentation §2) and the end-of-stage storage on a parallel stage; on a chronological stage the row of block reads that block’s own average (Block Formulation Variants §2.4) — a single plant-level quantity shared by every cell, since storage is not partitioned. is cell ‘s apportionment share of plant ‘s declared maximum turbined flow,
(the ratio of the cell’s own unit groups’ declared maximum turbined flow to the plant’s total, when the plant’s total is ), so that for every plant with a positive declared maximum turbined flow. Only the plane’s flow-independent part — the intercept , the storage term, and the spillage term — is apportioned by ; the flow coefficient stays on the cell’s own unscaled, because it alone is homogeneous in the cell partition (summing the per-cell rows at fixed , , and recovers the plant-level bound this replaces). A single-cell plant with a positive declared maximum turbined flow has exactly, so its row is the plant-level row.
The slack enters the row of every plane of its cell and block with coefficient , so generation may exceed every plane of the cell by the slack. It sits only in those plane rows: in no water-balance, load-balance or generic-constraint row, so the water balance and the inflow coupling carry no term for it. It is priced at ($/MWh) times the block duration (§9). The plane displays that follow hold at .
For a plant with a positive declared maximum turbined flow the shares sum to , so the per-cell rows of each plane sum to that plane’s plant-level row (above). A sum of minima is at most the minimum of the sums, so at the same , and total turbined flow the cap the per-cell rows put on the plant’s generation , on the left, is at most the plant-level envelope, the minimum over the planes of the plant-level rows, on the right:
Equality holds when the same plane attains the minimum in every cell, in particular when the cells turbine in proportion to their shares . For the minimum of cell is times the plant-level envelope at the flow , so when every share is positive the display is Jensen’s inequality for the concave envelope with the weights . A cell with keeps only its term, and the display still holds by the sum-of-minima argument.
At a block where cell has no turbine capacity (Hydro Production Function Models §2.9), its plane rows are free and is fixed at , while a cell with capacity keeps its rows at its share , which is not re-apportioned. An FPHA plant without turbine capacity at every block of the stage has no generation column, no plane rows and no plane slack at that stage.
Generation Bounds (per cell , block — the FPHA generation column’s own bound; a constant-productivity cell has no such column, so its generation cap is folded into the turbined-flow bound below instead):
Both sums run over the cell’s unit groups with turbine capacity at block , and when the plant’s own limits have none.
Generation bounds are user-defined (declared per unit group, not derived from turbined flow). The lower bound is soft, with one slack per cell — priced at the plant’s own penalty at full magnitude on every cell of a split plant, never divided by cell count; a constant-productivity cell’s floor couples this same slack to rather than to a generation column. The upper bound’s plain sum over the cell’s own unit groups closes against the plant’s own resolved maximum — a bounds override may never raise a cell above it. The plant’s own minimum and, on a plant with several cells, its maximum also bind on the total of its cells (§8 Plant Minimum and Maximum). See System Element Modeling Overview §5. Matches the per-cell defined in Notation Conventions §4.3.
For details on the FPHA construction and production function model variants, see Hydro Production Function Models.
7. Outflow Constraints
Seção intitulada “7. Outflow Constraints”Outflow Definition (per hydro , block ):
Outflow Bounds (with slacks for soft enforcement):
8. Variable Bounds and Minimum Constraints
Seção intitulada “8. Variable Bounds and Minimum Constraints”Storage Bounds (per hydro )
Seção intitulada “Storage Bounds (per hydro hhh)”The dead volume is a hard lower bound for every hydro except two cases: a hydro with a filling configuration has a floor of in every phase, the per-stage filling floor below taking its place while it fills, and from its entry stage on the dead volume returns as the soft floor , the only storage-below-minimum slack of the LP, priced above deficit; a hydro without a filling configuration has a floor of while PreFilling, where its frozen identity (§4) holds the storage at its incoming value. The upper bound is hard; excess water leaves through spillage. While a hydro is PreFilling its upper bound is the larger of the stage’s and the hydro’s own maximum storage, so a stage maximum below the hydro’s own maximum never cuts off a held storage within the hydro’s own storage range. On a chronological stage the block-end storages carry the same column bounds, and the soft floor applies to the end-of-stage storage only.
Filling floors (for filling hydros, at every filling stage , from the filling start stage up to, not including, the entry stage):
is the dead volume in force at stage (a stage may override it), is the last filling stage, the stage before the entry stage, and the minimum accumulation rate of stage , which converts into hm³. The floor at stage is the dead volume of stage minus the accumulation the schedule still owes after , never above the dead volume of stage itself; it reaches at . Every filling stage carries its floor. The slack is priced at , which Novomodelo expects below deficit (Penalty System — Penalty Ordering Validation checks it as given). See Penalty System §6.
Turbined Flow Bounds (per cell , block )
Seção intitulada “Turbined Flow Bounds (per cell (h,b)(h, b)(h,b), block kkk)”The lower bound is soft, with one slack per cell, priced at the plant’s own penalty at full magnitude on every cell — never divided by cell count; it is the plain sum of the resolved minimum turbined flow of the cell’s own unit groups with turbine capacity at the block, , except on a single-cell plant whose own minimum is tighter (Plant Minimum and Maximum). The upper bound is hard and closes against the plant’s own resolved maximum, never raised above it:
where for an FPHA hydro (turbined flow and generation are independent columns there) and for a constant-productivity hydro, and for a unit group without turbine capacity at the block, in every model; the closing term is when the plant itself has none, so there. Each group’s own flow cap and MW-implied flow cap must be folded before summing across the cell, since does not distribute over a sum of groups that bind on different sides. This is the mechanism that enforces a constant-productivity cell’s generation cap (§6): there is no separate generation column to bound directly. Matches the per-cell defined in Notation Conventions §4.3.
Plant Minimum and Maximum
Seção intitulada “Plant Minimum and Maximum”A plant’s own resolved minimum turbined flow and minimum generation bind on the total of its cells where they are tighter than the cells’ floors. Take the relative tolerance . The plant minimum binds at a block where , the sum running over every unit group of the plant, those without turbine capacity included; elsewhere the cell floors already enforce it. Where it binds, the plant enforces , and alike. A single-cell plant takes as its cell floor, and a plant with several cells gets one residual row on top of its cell rows:
is the cell’s generation: for an FPHA cell, for a constant-productivity cell, and nothing for an FPHA plant without a generation column, which is PreFilling, in its filling window, or without turbine capacity at every block of the stage. The residual slacks are priced at the plant’s own penalties, on and on , at full magnitude. Because each residual row also carries the cell slacks, the cheapest slacks charge for the turbined flow and the same form for the generation, so a shortfall is charged once. A plant with a residual row at some block of the stage has the row at every block of the stage; at a block where its cells cover the plant minimum the row is free and the residual slack is fixed at . Where the plant cannot produce at a block (it is PreFilling or in its filling window, or it has no turbine capacity at the block), every turbined-flow and generation minimum it has there is dropped: each cell floor, a single cell’s plant floor included, is with its slack fixed at , and the residual row is free. The minimum outflow is not dropped (Lifecycle Phases).
A plant minimum below its cells’ floors does not loosen them: the cell floors still bind. A plant with several cells also gets one hard row per block for its maximum turbined flow and one for its maximum generation, without a slack:
Each binds at a block where the plant maximum is below the sum of its cells’ caps by more than of it, the cells’ caps being for the turbined flow, and (constant productivity) or (FPHA) for the generation; at any other block of the stage the row is free. A cell without turbine capacity at the block has cap , so the row is free where no cell can turbine. A single-cell plant has no such row, since its cell bound already closes against and , and neither has a plant that is PreFilling or in its filling window, whose turbined flow is fixed at .
Diversion Flow Bounds (per hydro , block )
Seção intitulada “Diversion Flow Bounds (per hydro hhh, block kkk)”is the channel capacity, or a maximum diversion flow set for the stage or for the block, which replaces it. A channel declared without a capacity has , and a hydro without a diversion channel has . The lower bound is unless a minimum diversion flow is set for the stage or for the block, which replaces it. Both bounds are hard. Diversion cost is a regularization term (see §1), not a violation penalty. Neither this bound nor the spillage bound below depends on storage: Novomodelo models no sill below which the diversion channel or the spillway closes.
Spillage Bounds (per hydro , block )
Seção intitulada “Spillage Bounds (per hydro hhh, block kkk)”Spillage is bounded below by and above by unless a minimum or maximum spillage is set for the stage or for the block, which replaces the bound. Both bounds are hard. While the hydro is PreFilling, spillage is fixed at (Lifecycle Phases).
Pumping Flow Bounds (per station , block )
Seção intitulada “Pumping Flow Bounds (per station yyy, block kkk)”Both bounds are hard.
Lifecycle Phases
Seção intitulada “Lifecycle Phases”At each stage a hydro is in exactly one lifecycle phase, set by its commissioning window and, for a filling hydro (a hydro with a filling configuration), by its filling start; every other entity is either in service (inside its commissioning window) or out of service.
| Phase | Applies at | Storage row and floor | Turbined flow and generation | Spillage | Diversion | Inflow, upstream releases and withdrawal | Operational floors |
|---|---|---|---|---|---|---|---|
| PreFilling | A hydro without a filling configuration: every stage outside its commissioning window, before its entry stage or from its exit stage on. A filling hydro: every stage before its filling start | Frozen identity , per block on a chronological stage, with right-hand side (§4); floor , and a ceiling never below the hydro’s own maximum storage (§8); no evaporation row | fixed at on every cell; no FPHA rows and no generation column | fixed at , whatever its bounds | fixed at | The local inflow with its inflow non-negativity slack, the releases of the upstream plants, the flows diverted in and the in-transit volume maturing into the plant enter the row of the first non-PreFilling plant downstream, and the withdrawal target moves to that plant’s right-hand side; with no such plant the water leaves the system | The minimum-outflow row is present, and with turbined flow, spillage and generation at , a positive minimum outflow falls wholly on its slack, at its penalty (§9). The minimum turbined flow and minimum generation are dropped (Plant Minimum and Maximum) |
| Filling | A filling hydro: every stage from its filling start up to, but not including, its entry stage | Water balance of §4, with the evaporation row of an evaporating hydro; floor and the per-stage filling floor of §8 Storage Bounds | fixed at on every cell; no FPHA rows and no generation column | Within its bounds | fixed at | On the plant’s own row (§4) | Spillage can meet a positive minimum outflow, and the outflow slack covers any shortfall. The minimum turbined flow and minimum generation are dropped (Plant Minimum and Maximum) |
| Operating | A hydro without a filling configuration: every stage inside its commissioning window, so every stage when it declares none. A filling hydro: every stage from its entry stage on; it has no exit stage | Water balance of §4, with the evaporation row of an evaporating hydro; the hard dead-volume floor for a hydro without a filling configuration, and for a filling hydro floor with the soft dead-volume floor | Turbined flow within the Turbined Flow Bounds above; generation by the production model of §6 | Within its bounds | Within the Diversion Flow Bounds above for a hydro without a filling configuration; fixed at for a filling hydro | On the plant’s own row (§4) | The same rows are present; the minimum turbined flow is met by the cell’s turbined flow, the minimum generation by its generation, and the minimum outflow by turbined flow plus spillage, with the slack covering any shortfall at its penalty (§9); at a block where the plant has no turbine capacity, its turbined-flow and generation minima are dropped (Plant Minimum and Maximum) |
Outside its commissioning window a thermal’s generation, a line’s flow in each direction, a non-controllable source’s generation, a pumping station’s pumped flow and a contract’s power are fixed at in every block; an anticipated thermal’s commitment is gated by its delivery stage in addition, and is opened only when the plant is in service at that delivery stage, whatever its status at the decision stage (State Augmentation §5).
9. Constraint Violation Penalty Terms
Seção intitulada “9. Constraint Violation Penalty Terms”The per-block constraint violation penalties in the objective (referenced from §2) are:
The FPHA plane slack is per cell, like the generation-minimum slack, and is priced at the plant’s at full magnitude on every cell; it exists wherever the FPHA plane rows of §6 exist and is absent for every other hydro. Penalty System lists it with the recourse slacks because it keeps the stage LP feasible when a generation floor or a constraint forces generation above the planes.
The turbined- and generation-minimum slacks are per cell — each of a split plant’s cells carries its own slack column and its own row, priced at the plant’s penalty at full magnitude, never divided across cells. The residual slacks and exist only for a plant with a residual row (§8 Plant Minimum and Maximum) and are for every other plant. The outflow slacks stay per plant: outflow has no per-cell column to attribute a floor to. The evaporation slacks are per plant too, stage-level on a parallel stage and per block on a chronological stage (Evaporation Row).
where is the total stage duration in hours, and the evaporation sums run over the hydros that model evaporation. Withdrawal violation slacks (, ) are stage-level (not per-block) and bidirectional: penalizes under-delivery (the realized withdrawal falls short of the effective withdrawal), and penalizes over-delivery. The effective withdrawal is signed (§4), and the slack bounds ensure the realized withdrawal cannot flip sign relative to it:
- (scheduled removal): (under-delivery slack capped at the effective withdrawal; floors ), unbounded.
- (scheduled inter-basin return/addition): (over-delivery slack capped at ; caps ), .
- : both slacks are pinned to zero.
This cap guards a degenerate case: an unbounded under-delivery slack would let a run-of-river plant “un-withdraw” past its effective withdrawal and inject phantom water into the reservoir.
10. Generic Constraints
Seção intitulada “10. Generic Constraints”User-defined linear constraints (per constraint ) take a two-sided interval form:
where is a quantity of the stage LP that the generic-constraint variable catalog exposes — a column such as storage, a flow, a generation or a deficit, or a fixed combination of columns such as the inflow of a hydro (below).
A term on the turbined flow or the generation of a plant split across buses may address one (hydro, bus) cell or all of the plant’s cells at once (§3).
A term may also address a plant’s useful volume — its storage above the physical minimum . It is realised on the storage variable itself by shifting both endpoints by the same amount,
so it adds no LP variable; an absent endpoint stays absent, and a negative lowers both endpoints.
A term may also address a stage volume in hm³, a flow summed over the blocks of the stage with the block conversions of §4: the released volume of hydro , its diverted volume, or the pumped volume of pumping station ,
A constraint with a stage-volume term combines it only with other stage volumes and with the storage or the useful volume at the start or end of the stage, and no coefficient or endpoint of it varies by block, so it is a single stage-level row (row materialization below).
Either endpoint, or , may be absent — never both — with an absent side read as the corresponding infinity. This single interval subsumes every shape a constraint can take: a present alone is a floor, a present alone is a cap, is an equality, and both present together bound a two-sided band. A constraint’s shape — floor, cap, equality or band — follows from which endpoints are present and, when both are, whether they coincide.
Coefficients and Endpoints
Seção intitulada “Coefficients and Endpoints”Each coefficient is an arithmetic expression over literal numbers and named scalar parameters — sums, products, quotients and groupings of them — and the constraint as written must be linear in the variables: a product of two variables, or a division by an expression that holds a variable, is refused. Terms on the same variable are combined into one coefficient, and a term whose coefficient folds to exactly zero contributes nothing to the row.
Each endpoint composes independently from up to two pieces, summed together: a numeric base that may itself vary by stage (and, for the finest-grained parameter kind below, by block within the stage), plus an optional affine remainder layered on top of it — a constant plus a weighted sum of named scalar parameters. An endpoint carrying neither piece is simply absent; a base alone, a remainder alone, or their sum are all valid, so a single endpoint can combine a scheduled floor or cap that already varies by stage with a further parameter-driven adjustment on top.
A named scalar parameter takes one of six kinds: one value for every stage; one per stage; one per season; a quantity computed from a plant’s geometry and energy conversion, §4 of Hydro Production Function Models; one per stage and block; or a formula over other parameters, which varies wherever the parameters it reads vary.
Resolution happens once at LP-build time, and every resolved coefficient and endpoint must be finite, so the LP coefficients and endpoints are still numeric at solve time — the parameter mechanism does not introduce LP-variable coupling between constraints. Methodology relevance: it lets the corpus express ramping limits, capacity caps, and operator-imposed quotas that vary by stage, season, or block without authoring a separate constraint per stage. Coefficient and endpoint values can therefore be stage- and block-varying constants, not just literal numbers.
Two-Sided Slack
Seção intitulada “Two-Sided Slack”A constraint may optionally carry a slack, penalized per unit of violation, so the row relaxes at a cost instead of forcing infeasibility. The penalty is declared on one of two bases. For a row whose quantity is a rate (MW, m³/s), is a price per unit and hour, and a slack on a row covering a period of hours costs : on a per-block row and on a stage-level row. For a row whose quantity is an amount (MWh, hm³), is a price per unit, and the slack costs on every row, independent of the stage or block duration, so on a per-block row it carries no block-duration weight (§1.1). A one-sided row — only or only present — carries a single slack column relaxing that one endpoint. A two-sided row — both and present — carries two independent slack columns, relaxing the floor upward and relaxing the cap downward: a single column cannot represent both “how far below the floor” and “how far above the cap” without conflating the two directions.
The reported violation is the signed net : positive when the row sits below its floor, negative when it sits above its cap, matching the sign of the underlying deviation rather than reading as an unsigned magnitude. The objective charges both columns, , which coincides with the net’s magnitude whenever only one direction is active — the case at any optimum, since paying for both directions on the same row at once is strictly dominated by paying for neither of the excess.
Row materialization: a constraint bound declared for every block of the stage over a block-independent expression — one whose every term references a stock variable (incoming storage , outgoing storage , evaporation (the stage-level on a parallel stage, a named block’s on a chronological stage), or an anticipated-thermal commitment) or a stage volume, which reads the columns of every block itself — is materialized as a single stage-level row, since per-block rows would be identical, provided its coefficients and endpoints are block-independent too: a coefficient or an endpoint drawn from the per-(stage, block) parameter kind above, or from a formula that reads one, varies within the stage, which forces the per-block row set even when the expression alone would otherwise qualify for the collapse. A term whose coefficient folds to zero is dropped before this test, so it neither prevents nor causes the collapse. The slack of a stage-level row is priced over the stage, on the rate basis and on the amount basis. A bound declared for every block on a block-level expression, or a bound declared for one block, still produces one row per relevant block. This is an LP row-count optimization that is cost- and parity-neutral.
See Generic Constraints for the authoring grammar, the activation grid, and the per-file field tables this formulation implements.
Hydro Inflow
Seção intitulada “Hydro Inflow”A hydro-inflow term reads the inflow of hydro in block , a rate in m³/s built from LP columns and the realized inflows of the solve:
Each upstream release is credited to block by the travel time of its arc:
- , = realized inflows of the solve (§5), constants: a term with coefficient on subtracts from each finite bound of its row, and for each
- = the inflow non-negativity slack of (§4), which carries the truncation correction of the local inflow under the truncation methods (Inflow Non-Negativity Solution Methods §2); under
truncationandtruncation_with_penaltyand underpenalty, and undernonethe column does not exist - = turbined plus spilled outflow of (§7), credited with the same-stage share or the within-stage shares of §4
- = arrival density of State Augmentation §6 on a chronological stage and on a parallel stage: the share of the maturing in-transit volume that arrives in block , which the division by turns into a rate; the maturing term is absent when no travel-time arc enters
- = PreFilling plants at stage whose first non-PreFilling downstream plant is (see PreFilling Pass-Through); the local inflow of each with its slack under the truncation methods, the flows diverted into it and the releases of its upstream plants enter whole, with no travel-time share, and its maturing in-transit volume enters as the rate , by its own arrival density
The term excludes pumping, the inflow non-negativity slack under penalty, the plant’s own outflows, evaporation and withdrawal.
For a hydro that is not PreFilling at stage , the term mirrors the inflow terms of its water balance (§4) listed below. On a chronological stage equals the inflow terms of block ‘s water-balance row: the local inflow with its slack under the truncation methods, the flows diverted in, the credited releases, the maturing transit volume and the PreFilling pass-through; under penalty the row’s slack is left out. On a parallel stage equals the same terms of the stage row, and only the duration-weighted stage total matches the balance; the split across blocks is a convention.
For a PreFilling hydro the term reads the same columns, with empty, and they match no row of , whose row is the frozen identity (PreFilling Pass-Through): its local inflow with its slack under the truncation methods, the flows diverted into it, the releases of its upstream plants and its maturing transit volume enter the row of the first non-PreFilling plant downstream, or leave the system when there is none (Delayed-arrival water-balance entry).
reads per-block columns, so a bound without a block expands to one row per block (see row materialization in §10).
11. Benders Cuts
Seção intitulada “11. Benders Cuts”For each cut in the stage LP, that is each active cut or, under the dynamic method, each cut resident in the solve (Cut Management §8):
where:
- = cut intercept
- = coefficient for storage state variable, present when the cut keeps the storage dimensions
- = coefficient for the inflow lag of the next stage’s incoming state, which is the realized inflow for and the incoming lag for , both numbers of the solve, so each lag term joins the right-hand side ; present when the cut keeps the inflow-lag dimensions
- = the terms on the outgoing commitment-ring slot and in-transit bucket columns (below)
The display is the form at a stage that completes a lag period that begins at it. Where a lag period spans several stages, the lag terms evaluate the outgoing lags of Multi-Resolution Studies §2.1: at a stage that leaves its period open they are , with no re-indexing, and at a stage that completes a period begun earlier lag 1 is the period’s duration-weighted mean in place of .
The cut row has no entry on the inflow lags, which are not LP columns (§5). splits into a part fixed by the trial point and a part fixed by the opening, and the lag coefficients of a new cut come from the duals of the solve by the chain rule (PAR(p) Folding §3, §4).
When anticipated thermals are present, the cut carries one coefficient per commitment-ring slot (State Augmentation §5), read, like the storage coefficients, from the reduced costs of the pinned incoming columns; when travel-time arcs are present, it likewise carries one coefficient per in-transit bucket dimension (State Augmentation §6). Unlike the storage and lag coefficients, the ring-slot and bucket coefficients are always part of the cut projection (State Augmentation §7).
The future-cost variable has lower bound at every stage, which presumes . Stage objectives can be negative: export contracts earn revenue (§2), and the curtailment term of §2 is a reward: it is never positive, and it is constant for a must-run source. Where is negative the bound lies above it, so the cut approximation need not lie below and the lower bound is not guaranteed. At the last stage of the finite horizon, with no terminal boundary loaded, is (); with a terminal boundary loaded, keeps the floor and the imported cuts bound it from below, so the terminal function is the larger of and the imported cuts.
Cuts live in an append-only pool at stable slot indices: every cut ever generated is retained for the lifetime of the run, and only the active subset is baked into each iteration’s stage template. Deactivation excludes a cut from each iteration’s stage-template rebake rather than mutating any row; the persistent lower-bound LP is append-only (its rows are never removed, so the lower bound stays monotone). Slot indices stay stable, so reactivation — re-baking the cut into the template at the same slot — is exact. The dynamic method deactivates no cut itself: each solve loads a resident subset of the active cuts and grows it with the omitted candidate cuts its solution violates. A per-stage cap on the number of active cuts the run generated, when set, deactivates the excess among the cuts of earlier iterations, with or without Level 1 selection; it is not admitted with the dynamic method (Cut Management §5).
For cut coefficient derivation, aggregation, and selection strategies, see Cut Management.
Cross-References
Seção intitulada “Cross-References”- Notation Conventions — index sets, parameters, decision variable naming
- System Element Modeling Overview — physical meaning of each element, decision variables, Variable Units Convention
- Penalty System — three-category taxonomy, penalty names, priority ordering, cascade resolution
- SDDP Algorithm — iterative structure that solves this LP at each stage
- State Augmentation — the state vector, its pinning by column bounds, the outgoing state and the cut row, and the inflow-lag, commitment-ring and in-transit state families
- LP Layout and Scaling — the column and row layout of this LP, its cost scaling and prescaling, and the unscaling of its duals and reduced costs
- PAR(p) Inflow Model — complete AR inflow model specification
- PAR(p) Folding — the realized inflow outside the LP, the cut right-hand side and the chain rule for the inflow-lag coefficients
- Hydro Production Function Models — constant and FPHA model details
- Cut Management — dual extraction, cut coefficients, aggregation, and selection
- Equipment-Specific Formulations — per-equipment constraint derivations, pumping details