Cascade LP — Multi-Reservoir Scheduling

This document explains the multi-reservoir cascade LP in WeeklyScheduling. It extends the single-reservoir LP (Phase 2) to K reservoirs connected by the Cascade topology from HydroCore. This model becomes the LP subproblem that SDDP calls at every forward/backward pass stage in Phase 6.

Reference: Kompendium 3, §4 (Vassdragstopologier, Optimaliseringsstruktur)


The problem

Given a T-stage planning horizon and a cascade with K reservoirs and J plants:

max  Σ_t  price[t] × Σ_j prod[t,j] × Δt  +  Σ_k end_val_k(v[T,k])

subject to:
    v[t,k] = v[t-1,k] + inflow[t,k]                    (water balance)
             − Σ_{out edges} outflow[t,edge]
             − spill[t,k]
             + Σ_{in edges} inflow_from[t−delay, edge]
    v_min_k ≤ v[t,k] ≤ v_max_k                         (storage bounds)
    0       ≤ r[t,j] ≤ r_max_j                         (discharge bounds)
    spill[t,k] ≥ 0                                     (residual spill)
    prod[t,j] ≤ slope_s × r[t,j] + intercept_s         (PQ curve per plant)
    end_val_k ≤ vf[i] + wv[i] × (v[T,k] − grid[i])    (end-value cuts)

Topology and water balance

The cascade is a directed acyclic graph (DAG) of nodes and edges. Each reservoir's water balance constraint accounts for:

  1. Natural inflow — external inflow to this reservoir [Mm³/stage]
  2. Outgoing plant discharge — water flowing through turbines
  3. Outgoing waterway flow — spillway, bypass, river reach flows
  4. Residual spill — emergency slack (exits system, ensures feasibility)
  5. Incoming plant discharge — water arriving from an upstream reservoir
  6. Incoming waterway flow — with routing delay applied

Routing delays

When a waterway has delay_h > 0, the flow from stage t only arrives at the downstream node at stage t + delay_stages, where:

delay_stages = floor(delay_h / step_hours(resolution))

For stages before the horizon (t − delay < 1), the warm-start flow is assumed to be zero.

Residual spill

Each reservoir has a spill[t,k] ≥ 0 variable that exits the system entirely (not routed downstream). This ensures the LP is always feasible even when all physical outflow paths are saturated. The spill has zero revenue — the LP naturally minimises it because releasing through turbines earns revenue.

This follows the penalty/slack approach from Kompendium 3, §7.2.


Shadow prices (water values)

After solving, the dual variable of each water balance constraint gives:

shadow_price[t,k] = marginal value of 1 extra Mm³ of inflow
                    at reservoir k, stage t  [NOK/Mm³]

These are the β⃗ gradient components used in SDDP Benders cuts (Phase 6):

α(k+1) ≥ β_i + μᵀ_i × x(k)

where μ_i = vector of K shadow prices at stage t

Constraints from HydroCore

The cascade LP translates all Cascade.constraints to JuMP:

HydroCore typeLP translation
FlowConstraintTightens bounds on r[t,j] or q_way[t,w]
VolumeConstraintTightens bounds on v[t,k]
RampConstraintr[t,j] − r[t-1,j] ≤ ramp_up per stage
MinEnvFlowRaises lower bound on waterway flow

Using the model

using HydroCore, WeeklyScheduling

# Build cascade (see HydroCore docs)
cascade = Cascade("ALTA", "Alta System")
add_node!(cascade, r_upper); add_node!(cascade, r_lower); add_node!(cascade, sea)
add_edge!(cascade, plant_A); add_edge!(cascade, plant_B); add_edge!(cascade, spill_A)

# Natural inflow per reservoir [Mm³/stage]
inflows = Dict(
    "R_upper" => fill(3.0, 52),
    "R_lower" => fill(0.5, 52)
)

# Electricity price [NOK/MWh]
price = fill(400.0, 52)

# End-value (optional, per reservoir)
end_vals = Dict{String, EndValueFunction}()  # defaults to ZeroEndValue

# Solve
result = optimize_cascade(cascade, inflows, price, end_vals;
                          resolution=WEEKLY)

result.volume         # T × K matrix [Mm³]
result.discharge      # T × J matrix [Mm³/stage]
result.spill          # T × K matrix [Mm³/stage]
result.production     # T-vector [MWh/stage]
result.shadow_price   # T × K matrix [NOK/Mm³] — water values for SDDP
result.revenue_NOK    # total revenue [NOK]
result.reservoir_ids  # column labels for K dimension
result.plant_ids      # column labels for J dimension

Known limitations

  1. Fixed reference head — PQ curves are evaluated at reference head. Variable head correction (Kompendium 2, §5.4.6; Kompendium 3, §3.4) is not implemented. For cascades with small lower reservoirs where head varies significantly, this introduces error.

  2. Fixed routing delays — Waterway.delay_h is a constant. In reality, delay depends on current flow rate and channel fullness (Kompendium 3, §2.5.1).

  3. No unit commitment — The LP relaxes min-discharge to zero. Start/stop decisions and associated costs (Kompendium 3, §3.1) are not modelled.

  4. No intra-stage price variation — A single price per stage. Price segments within a week (Kompendium 2, §3.4.4) are not modelled.


Connection to the planning hierarchy

WaterValues (Phase 1 SDP)
    ↓  produces WaterValueTable (single reservoir)
WeeklyScheduling — single-reservoir LP (Phase 2)
    ↓
WeeklyScheduling — cascade LP (Phase 4)  ← this document
    ↓  produces CascadeScheduleResult with shadow_price
    ↓  (Phase 6: embedded in SDDP as subproblem)
SeasonalPlanning (Phase 6–7)
    ↓  produces SeasonalPlan with Benders cuts