Methods

Complexity follows evidence.

The study uses the simplest transparent model that the available public evidence can support, then adds complexity only when the next question requires it.

Progressive evidence ladder

From source record to bounded inference.

AuditSource dates, definitions, missingness and accounting boundaries.
ReconstructModel only quantities that are not directly observed.
ConstrainPreserve observed totals and published planning anchors.
Stress-testVary transfer, hydro timing and candidate resources.
Cross-checkCompare implementations and independent planning evidence.
GateStop before claims that require unavailable physical evidence.

Hourly demand reconstruction

Observed daily energy; reconstructed intraday shape.

Continuous public Kerala interval-load telemetry was not available in the admitted evidence. The model therefore fixes what is known and reconstructs only what is missing.

For observed day d:
Σₕ Load(d,h) = observed SLDC daily energy(d)
354 FY2024–25 daily totals are source-observed. Eleven missing daily totals are explicitly model-only in the 8,760-hour chronology.

Held-out intraday-extrema performance

Fixed shapeRMSE 477.9 MW
Sparse static162.0 MW
ERA5-sensitive155.2 MW

January–March extrema were withheld from fitting. These errors measure intraday-shape reconstruction conditional on known daily energy, not day-ahead forecasting.

Chronological adequacy

Every hour is solved in sequence.

The statewide screen asks whether generation, storage and imports can meet demand in each of 8,760 hours.

Demand scenarios

Official future energy/peak trajectories from different studies remain separate rather than being averaged into a synthetic consensus.

Transfer scenarios

CEA's 4,455 MW ATC is the dated reference. 80% and 60% cases are transparent stress sensitivities.

Candidate resources

Renewable and storage ceilings are research scenario envelopes, not legal buildable capacity.

Stage 1: minimize Σ unserved energyₜ
Stage 2: preserve Stage-1 adequacy, then minimize candidate investment + partial import cost

The lexicographic objective avoids inventing a Kerala-specific Value of Lost Load. It is an adequacy-first screen, not a total-system welfare optimum.

Hydropower hierarchy

Separate timing value from reservoir physics.

v1.1 · same-day timing

Daily hydro MWh fixed; hourly timing flexible within power limits.

→
v1.2 · interday windows

Hydro energy conserved over 1/3/15/30-day windows.

→
v1.3 · stateful pilot

Add storage state with reconstructed net-water forcing.

→
v1.4/v1.5 · source gates

Require direct physical inflow/storage evidence before stronger claims.

v1.3 reconstructed forcing: Bₜ = Sₜ₊₁ − Sₜ + Gₜ/c
State replay: Sₜ₊₁ = Sₜ + Bₜ − Gₜ/c

Because B is constructed from storage change and generation, historical storage closure is algebraic. It verifies implementation and date alignment, not hydrology.

Public transmission screen

A source-backed topology with explicit proxy layers.

542buses
561lines in reconstructed topology
233source-backed primary lines screened
8boundary-injection allocations tested

Spatial inputs

  • Public KSEBL topology and equipment evidence.
  • Distribution-interface transformer MVA as a load-location proxy.
  • Mapped generation where source location is resolved.
  • Generic voltage-class reactance where source X is unavailable.

Interpretation boundary

Outputs are candidate stressed corridors and robustness to import-injection location. The model does not assess reactive power, voltage, transient behavior or N-1 security.

Verification & validation

Different checks answer different questions.

CheckPurposeWhat it establishes
Source provenanceTrace values to a named agency/document/date.Evidence identity and accounting boundary.
Conservation / invariant checksVerify energy/accounting transformations.Arithmetic correctness, not physical truth.
Unit/integration testsCheck software guards and equations.Implementation correctness.
PyPSA ↔ SciPy ↔ OSeMOSYSReproduce common mathematical formulations independently.Reduced framework-specific implementation risk.
Held-out extremaEvaluate demand-shape reconstruction on withheld observations.Out-of-sample shape evidence.
CEA transmission plan comparisonCompare network-screen geography with independent planning evidence.External corroboration of selected hotspot locations.

Method selection

Why more complex methods were not used yet.

MethodWhy it was not the primary approachEvidence needed to upgrade
Deep-learning hourly forecastingNo dense measured hourly target series.Multi-year interval load + weather observations.
Full unit commitmentPlant-level ramping, startup, minimum-output, heat-rate and outage data incomplete.Operator-grade fleet parameters.
Probabilistic LOLP / ELCCForced-outage, multi-weather-year and hydro uncertainty distributions unavailable.Reliability datasets and stochastic scenarios.
AC / N-1 network studyMeasured bus injections, complete impedances and reactive-power data unavailable.Utility-grade electrical model and operating states.
Single least-cost 2040 optimizationWould collapse unresolved demand, transfer, spatial, ecological and cost uncertainty into one precise-looking answer.Planning-grade physical, siting and economic inputs.