PyPSA-Earth, end-to-end

1 · What PyPSA-Earth actually is

PyPSA-Earth is a global, open-source energy-system optimisation framework built on top of PyPSA (Python for Power System Analysis). It reduces the question “how should this country decarbonise?” to a single Linear Program:

It's capacity expansion + dispatch at the same time. The optimiser doesn't just pick what to build — it co-optimises where on the grid, how often each plant runs, and how electricity / heat / gas / hydrogen flow between zones.

The four building blocks

What “sector-coupled” means here

Vanilla PyPSA-Earth was electricity-only. The sector-coupled variant adds, as additional carriers and links on the same graph:

• Heat — residential / services / industrial low-temp via heat pumps, resistive heaters, gas boilers, CHP, district heating
• Land transport — ICE oil load + EV charging links + V2G + fuel-cell vehicles
• Industry — process heat (low/high T), feedstock (naphtha, methanol), industrial CCS
• Hydrogen — electrolysis, fuel cell, pipelines (off in BD), salt caverns (none in BD)
• CO₂ accounting — atmosphere ↔ stored CO₂ buses, DAC, biomass-with-CCS
• Agriculture, biomass — crop residues, biogas potential, machinery oil

So a single LP simultaneously decides whether to electrify a steel plant, whether to build a heat pump or keep a gas boiler, and whether to import LNG or build an electrolyser.

2 · The pipeline, rule by rule

The Snakemake DAG runs in three stages. Each box below is a script in pypsa-earth/scripts/. Arrows show what feeds what.

The wildcard system (how Snakemake parameterises a run)

So the canonical filename of a final solved BD network is:

results/Case2_NetZero2050/postnetworks/elec_s_8_ec_lcopt_Co2L-24H_24h_2050_0.071_AB_0export.nc

3 · Data sources — the complete inventory

Every external dataset PyPSA-Earth touches, what it's used for, and where it lives in our copy.

Geospatial & weather

Energy / techno-economic data

Configuration files

4 · Bangladesh customisations (Paul's diff)

pdockery/pypsa-earth sits two commits ahead of upstream pypsa-meets-earth/pypsa-earth main:

fe4ee35  Add solved Bangladesh energy system network results
2f77da87 Update config.default.yaml with Bangladesh configuration

The BD-specific decisions in our team's configs/config.yaml on top:

① Geographic + temporal scope

② CO₂ trajectory (custom block, not in vanilla)

co2_budget:
  enable: true
  override_co2opt: true
  co2base_value: co2limit
  year:
    2030: 0.9    # 10% cut → 87.3 MtCO2
    2040: 0.45   # 55% cut → 43.7 MtCO2
    2050: 0.05   # 95% cut → 4.85 MtCO2

The fractions multiply electricity.co2base = 9.7 × 10⁷ tCO₂/yr (full-economy BD baseline). With override_co2opt: true, the rule rewrites the Co2L wildcard from opts using these fractions per horizon.

③ Sector flags — Phase-1 set

④ Custom industry-decarbonisation block (not vanilla — likely Paul's branch)

sector.industry_decarbonisation:
  coal_to_gas: true   # gas substitutes industrial coal
  coal_to_elec: true  # electric process heat replaces coal
  oil_to_gas: true
  oil_to_elec: true

When enabled, industrial coal/oil emissions become optimisable links rather than fixed loads — the LP gets to fuel-switch on the cost curve. This block is not in the public pdockery/pypsa-earth tree — it lives on an unpushed branch.

⑤ Existing-fleet treatment

⑥ Numerical tuning for BD's bad scaling

BD cost data spans a 6 × 10⁹ range (very small marginal costs vs. very large investments). Default Gurobi diverges. The config switches to a custom gurobi-numeric-focus profile:

NumericFocus: 3        # stability over speed
ScaleFlag:    2        # aggressive matrix scaling
BarHomogeneous: 1      # fall back to homogeneous barrier
DualReductions: 0      # don't claim "infeasible_or_unbounded" on H2 export bus
BarConvTol: 1e-4       # relaxed tolerances
FeasibilityTol: 1e-4
OptimalityTol: 1e-4

5 · Inside the Case2_NetZero2050 run

What the optimiser is actually computing, in plain language.

Decision variables

Constraints (the LP wall)

• Nodal energy balance at every bus, every snapshot: generation + storage discharge − storage charge − load + net imports = 0
• Kirchhoff for AC lines (PTDF formulation)
• Capacity factors: p ≤ p_max_pu(t) × p_nom (RE) or p ≤ availability × p_nom (thermal)
• Storage state evolution: SoC(t) = η · SoC(t-1) + p_charge − p_discharge
• Global CO₂ cap: Σ (carrier emission factor × dispatch) ≤ co2limit × fraction(year)
• Line-volume cap: Σ length × s_nom ≤ line_volume_cap
• Sector links: heat-pump COP timeseries, electrolyser efficiency, EV-charger availability windows, biomass-transport costs

Objective

min  Σ_techs ( annuity(CAPEX) × p_nom + FOM × p_nom )
   +  Σ_techs,t ( marginal_cost(t) × p(t) )
   +  Σ_lines  ( line CAPEX × s_nom )
   +  Σ_co2    ( emission_price × emissions )

“Annuity” converts CAPEX into a per-year cost using the discount rate (0.071) and tech lifetime. So we're minimising annual system cost in 2030 / 2040 / 2050, separately, with each horizon's solution feeding the next as fixed brownfield.

Myopic foresight, in pictures

2030 horizon:
  - existing fleet (~38 GW, year ≥2022)            → fixed
  - extendable: solar, onwind, OCGT, CCGT, batteries, H2…
  - CO2 cap = 0.9 × 9.7e7 = 87.3 Mt
  - SOLVE → p_nom_opt[2030]

2040 horizon:
  - p_nom_opt[2030] is now BROWNFIELD (fixed)      ← add_brownfield
  - retire what aged out
  - new build optimised on top
  - CO2 cap = 0.45 × 9.7e7 = 43.7 Mt
  - SOLVE → p_nom_opt[2040]

2050 horizon:
  - p_nom_opt[2030] + p_nom_opt[2040] BROWNFIELD
  - CO2 cap = 0.05 × 9.7e7 = 4.85 Mt
  - SOLVE → p_nom_opt[2050]

Myopic ≠ optimal-overall. The 2030 LP doesn't know 2050's cap is 95 % cut — it only sees its own 10 % cap. So you can get short-sighted gas build-out in 2030 that becomes stranded in 2050. That's by design; it mimics how real planners decide. Perfect-foresight (Case3, future) would avoid it.

6 · Outputs — what comes out, how to read it

The headline files per horizon

How to read a solved network in Python

import pypsa
n = pypsa.Network("results/Case2_NetZero2050/postnetworks/...2050_0.071_AB_0export.nc")

# Capacities
n.generators.groupby("carrier").p_nom_opt.sum() / 1e3   # GW per tech, year 2050
n.links.groupby("carrier").p_nom_opt.sum() / 1e3        # heat pumps, electrolysers…
n.stores.groupby("carrier").e_nom_opt.sum() / 1e6       # TWh of H2 / battery storage

# Dispatch
n.generators_t.p.sum() * n.snapshot_weightings.objective.iloc[0] / 1e6  # TWh/year
n.statistics.energy_balance().round()                   # full sectoral balance

# Costs
n.statistics.capex().round()                            # €/yr CAPEX by tech
n.statistics.opex().round()                             # €/yr OPEX by tech
n.statistics.installed_capex().round()                  # one-off €

# Emissions
n.global_constraints                                    # CO2 cap + shadow price
n.generators_t.p.mul(n.generators.carrier.map(n.carriers.co2_emissions))

The 8 buses

With alternative_clustering: true + gadm_layer_id: 1, the 8 buses are the GADM-1 divisions:

7 · Karim outputs — first results off the network

Run on pypsa-bd/networks/elec_s_10*.nc (Paul's pre-built BD network, 10 buses, 2013 hourly). Source scripts in karim_outputs/scripts/.

Demand fundamentals

Transmission stress — pandapower load-flow snapshot

Equity — demand per capita by division

Energy security — fleet composition & concentration

8 · Data caveats & open issues

Generated for the FalgunXBangladesh team, May 2026. Source repo: /Users/karimarnous/FalgunXBangladesh. PyPSA-Earth fork: github.com/pdockery/pypsa-earth. PyPSA-BD analysis bundle: github.com/pdockery/pypsa-bd. Karim outputs: /Users/karimarnous/FalgunXBangladesh/karim_outputs/.