Hydropower Production Planning

A complete hydropower seasonal planning system in Julia, from single-reservoir water value computation through multi-reservoir SDDP to seasonal planning output with P10/P50/P90 forecasts and intra-week dispatch optimisation.

Packages

PackageDescription
HydroCorePhysical model: Reservoir, Plant, Cascade, TimeSeries, Constraints
WaterValuesSingle-reservoir SDP → WaterValueTable
WeeklySchedulingSingle & multi-reservoir LP scheduling via JuMP + HiGHS
HydroSimulationMonte Carlo P10/P50/P90 validation
SeasonalPlanningMulti-cut SDDP, Markov price, intra-week dispatch, seasonal plan
HydroRunnerCase runner: YAML/CSV → SDDP → charts + CSV/JSON output

Quick Start

using HydroRunner

# Full segment strategy (segments in both SDDP passes)
result = run_case("cases/dokka")

# Nested decomposition (fast scalar LP + greedy dispatch)
result = run_case_nested("cases/dokka")

Command Line

# Default: full segment strategy
julia --project=HydroRunner HydroRunner/bin/run_case.jl cases/dokka

# Nested decomposition (experimental, faster but less accurate)
julia --project=HydroRunner HydroRunner/bin/run_case.jl cases/dokka --nested-decomp

# Skip chart generation
julia --project=HydroRunner HydroRunner/bin/run_case.jl cases/dokka --no-charts

Results on Dokka Watercourse

Comparison with the commercial ProdRisk solver and a Python SDDP implementation on the same input data (37 inflow scenarios, 220 weeks, 7 Markov price states, 42×4h intra-week price segments):

SolverRevenueRev gapVolumeVol gapRuntime
Julia SDDP (full segment)147.6M6.0%134.23.8%96 min
Python SDDP (42×4h)148.3M5.6%127.31.6%280s
ProdRisk (8-core, commercial)157.1M—129.3—890s

Architecture

HydroCore/          Physical model: Reservoir, Plant, Cascade, PQ/HV curves
WaterValues/        Single-reservoir SDP (water value tables)
WeeklyScheduling/   Weekly LP scheduling (JuMP + HiGHS)
SeasonalPlanning/   Multi-cut SDDP, Markov price, cuts, dispatch, simulation
HydroRunner/        Case loader, runner, nested decomposition, charts, CLI
HydroSimulation/    Monte Carlo simulation

Key Design Decisions

  • Multi-cut SDDP: P per-state α variables in the LP, one per Markov price state. Eliminates convergence issues from the single-α expected-cut approach.
  • Intra-week dispatch: 42×4h segmented LP with per-segment discharge and production variables. State segment profiles in the backward pass ensure Benders cuts capture dispatch flexibility value.
  • Buffer reservoirs: tiny passthrough reservoirs (< 1 Mm³) have volume fixed at min_volume, shadow prices zeroed for numerical stability.
  • Terminal water values: computed from cascade energy equivalent (sum of downstream plants' MWh/Mm³). Replaces inflated proprietary VVERD values.
  • Hot-path backward: cuts and price loaded once per (stage, state), only water balance RHS updated per scenario.

Strategies

Full Segment (default: run_case)

Segmented LP subproblem with 42×4h price segments in both SDDP forward and backward passes. The backward uses average segment profiles per Markov state. Correct water valuation but slower due to larger LP.

Nested Decomposition (run_case_nested, --nested-decomp)

Fast scalar LP for SDDP, greedy intra-week dispatch for revenue in the forward pass and simulation. The dispatch callback feeds actual scenario dispatch revenue back into the SDDP's upper bound calculation.

Experimental — the scalar backward cannot capture dispatch flexibility in cut shadow prices, leading to over-conservation.

Documentation

Build the full HTML documentation:

julia --project=docs docs/make.jl
# → docs/build/index.html