The Planning Hierarchy
Hydropower planning is decomposed into a hierarchy of nested models, each covering a different time horizon at a different resolution. This is not just an engineering convention — it reflects a fundamental truth about computational tractability.
Why a hierarchy?
Imagine trying to solve one model that:
- Covers 5 years of operation
- Uses 15-minute resolution
- Models 50 reservoirs with stochastic inflow
The number of variables would be in the billions. No solver can handle this directly.
The solution is decomposition: solve approximate models at each level, and pass summary information (boundary conditions) between levels. Each level is feasible to solve; the hierarchy approximates the full problem.
The four levels
┌─────────────────────────────────────────────────────────────────┐
│ LANGTIDSPLANLEGGING (Long-term planning) │
│ Horizon: 3–10 years │ Resolution: weekly │
│ Method: Stochastic DP → Water value table WV(v, week) │
└──────────────────────────────┬──────────────────────────────────┘
│ Water value table
▼
┌─────────────────────────────────────────────────────────────────┐
│ SESONGPLANLEGGING (Seasonal planning) │
│ Horizon: 1–2 years │ Resolution: weekly │
│ Method: LP with water values as end-value │
└──────────────────────────────┬──────────────────────────────────┘
│ End-of-week reservoir targets
▼
┌─────────────────────────────────────────────────────────────────┐
│ KORTTIDSPLANLEGGING (Short-term planning) │
│ Horizon: 1–2 weeks │ Resolution: hourly │
│ Method: LP or MIP (unit commitment) │
└──────────────────────────────┬──────────────────────────────────┘
│ Hourly generation schedule
▼
┌─────────────────────────────────────────────────────────────────┐
│ SIMULERING (Simulation / validation) │
│ Horizon: any │ Resolution: depends on question │
│ Method: Monte Carlo simulation using water value tables │
└─────────────────────────────────────────────────────────────────┘Level 1: Long-term planning (Langtidsplanlegging)
Question: How should we manage reservoirs over the next several years to maximise long-run revenue, given uncertain future inflow?
Method: Stochastic Dynamic Programming (SDP) over a discretised state space.
The key output is the water value table: for each reservoir, for each week of the year, for each possible filling level — what is the marginal value of one additional Mm³ of water?
WV(reservoir, week, filling_level) → [NOK/MWh]This table encodes the answer to "is it better to generate now or save for later?" into a single lookup.
In practice (Norwegian tools): EMPS (EFI's Multi-area Power-market Simulator), Vansimtap. Both closed-source.
Level 2: Seasonal planning (Sesongplanlegging)
Question: Given where we are now, what is the optimal production plan for the next 1–2 years, week by week?
Method: Linear Programming. The water value table from Level 1 is used as the end-value function — instead of optimising to infinity, we optimise over a finite horizon where the value of water left in storage at the end is given by the table.
This is the classical "reservoir management" LP:
Maximise Σ_t price(t) × production(t) + WV(v_T)
Subject to:
v(t+1) = v(t) + inflow(t) - release(t) - spill(t) [water balance]
v_min ≤ v(t) ≤ v_max [storage limits]
q_min ≤ release(t) ≤ q_max [discharge limits]Key insight: The water value table is what makes this LP well-posed. Without it, the LP would drain the reservoir at the end of the horizon (no reason to save water if the horizon is finite).
Level 3: Short-term planning (Korttidsplanlegging)
Question: What is the hour-by-hour generation schedule for the next 1–2 weeks?
Method: LP or Mixed Integer Programming (MIP).
At this resolution, features that were too detailed for higher levels become important:
- Unit commitment (which turbines to start/stop — binary decisions → MIP)
- Start/stop costs
- Ramping constraints
- Intra-day price variation (peak/off-peak)
- Transmission constraints (network model)
The boundary condition from Level 2 (weekly reservoir targets) constrains the LP.
Level 4: Simulation
Not an optimisation level — rather a validation and calibration tool.
Given a water value table and historical or synthetic inflow scenarios, simulate the operation year by year. The simulation tells us:
- How often does the reservoir hit the minimum (risk of shortage)?
- How often is there unnecessary spill?
- What is the expected annual revenue?
Simulation results are used to calibrate the water value table: if the simulated reservoir levels are systematically too low (frequent near-empty), the water values are underestimated and should be revised upward.
Information flow between levels
The hierarchy works by passing boundary conditions downward and validation signals upward:
Long-term SDP → water value table
↓
Seasonal LP → weekly volume targets
↓
Short-term MIP → hourly schedule
↓
Simulation → validation statistics
↑
Calibration ← adjust water valuesThis coupling is the most practically important concept in the compendium. A poorly calibrated water value table will lead to systematically bad decisions at every lower level, regardless of how good the short-term model is.
What we model in this project
We build from the bottom up:
- HydroCore — the shared physical data model (this is what Chapter 02 covers)
- WaterValues — the SDP for Level 1 (single reservoir first, then multi-reservoir)
- ShortTermPlanning — the LP for Level 3 (using JuMP)
- HydroSimulation — Level 4 simulation for calibration
Level 2 (seasonal LP) uses the same LP machinery as Level 3, just at weekly resolution — so it comes naturally once Level 3 is built.