⚠️ Pre-1.0 development release. API surface may still change before 1.0. Code and analyses are functional and tested. Pin to a specific Git tag if you rely on reproducible results. SeeCHANGELOG.mdfor the current cut. Issues and PRs welcome.
Six paths, three decades, one open model. Every assumption documented, every source cited, every claim reproducible in five minutes.
Each panel shows one path's capacity build-up over 2026–2055 against
the dark-doldrum peak-demand line (LOLE-P95 reserve standard).
Red-hatched zones mark residual demand the backup architecture cannot
cover. Reproduce: python examples/generate_chart_stress_rampup.py.
The German energy debate argues about the wrong axis. Most discussions ask: renewable or nuclear? But the actual decision is two-dimensional — base-load × backup. This repository contains the quantitative model that walks through both axes for six realistic paths from 2026 to 2055, then asks: which path wins under any plausible set of assumptions?
| Want to … | Then look at … |
|---|---|
| Try the interactive app | Live demo on Streamlit Community Cloud — anchor camp left, variant camp right, six charts side by side |
| Try the model without installing anything | Open the Quickstart notebook in Colab — 5 minutes, no setup |
| Use the model programmatically | Quickstart § — six lines of Python |
| Audit the assumptions | docs/SOURCES.md — every default with primary-source tag |
| See the formulas | docs/FORMULAS.md — derivation and worked examples |
| Challenge a finding | Open an issue with model parameters and your counter-evidence |
A Python model that compares six pathways for the German energy transition, arranged along two independent axes (base-load × backup):
| Backup: Gas | Backup: H₂ | |
|---|---|---|
| Status quo / inaction | WEITER-SO | (n/a) |
| Existing-fleet emphasis | BESTAND | (n/a) |
| EE-only | EE-GAS | EE-H2 |
| EE + KKW | KKW-GAS | KKW-H2 |
The model has explicit treatment of:
- Forward costs (no sunk costs in investment decisions)
- Time-path dynamics (2026 → 2055, with realistic build-times)
- Asymmetric uncertainty (different assumption-sets for renewables vs. nuclear advocates)
- Sector coupling efficiency (heat pumps and EVs as 2:1 primary energy multipliers)
- Grid stability requirements (inertia, black-start, frequency control)
- Winter dunkelflaute stress test (10-day cold dark calm periods)
The model traces every default parameter to a primary source (see
docs/SOURCES.md) and exposes every
formula transparently (see docs/FORMULAS.md).
The German energy-modelling landscape is rich — PyPSA-Eur, REMIND, MESSAGE, oemof, AnyMOD, calliope, Agora and Fraunhofer studies, IEA scenarios. This model is not a replacement for any of them; it has a much narrower lens, focused on one question: which path is robustly cheapest as a forward decision, across the assumption substrates the opposing camps would actually defend?
- Not a high-resolution dispatch model. No hourly grid simulation, no unit-commitment, no network flows. For that: PyPSA-Eur, oemof, calliope.
- Not an integrated assessment / macro-economy model. No CGE feedback loops, no inter-sectoral capital allocation, no global emissions pathways. For that: REMIND, MESSAGE, IEA ETP.
- Not a policy roadmap or a strategy paper. For sector-coupling pathways with policy detail: Agora Energiewende, Fraunhofer ISE studies, BMBF-Ariadne project.
- Not a forecast. The model answers "given these assumptions, what follows?" — not "what will Germany do?".
If your question fits one of the categories above, those tools are the right starting point. The four properties below describe what this model is doing inside its narrower lens.
Five properties — each verifiable in the code, not just claims. Full
discussion in docs/methodik/methodology.md.
-
Forward-cost framing, sunk costs excluded by construction. KKW decommissioning fund, EEG-Altlast, Endlager fund sit in dedicated context fields and never enter the LCOE arithmetic. Most policy discussions silently mix the two. (
core/path_model.py,core/path_inputs.py) -
Two-dimensional path space (base-load × backup). Six paths, not the usual one-axis "renewables vs nuclear" framing. KKW-GAS and KKW-H₂ exist as separate paths precisely because the backup vector is independent of the base-load source.
-
Camp-symmetric defaults — no built-in bias. Every contested parameter carries EE-/atom-/bestand-optimistic alternatives alongside the neutral default. In the point estimate each camp wins its preferred path (atom_optimistic → KKW-GAS, three others → EE-GAS). The recommendation (EE-GAS) follows from min-max-regret across the four camps, not from "EE wins everywhere". (
core/camp_ranges.py,core/regret_decision_tree.py) -
Parameter-substrate robustness check. Swap the entire enesys assumption substrate for the PyPSA-Tech-Data substrate that feeds PyPSA-DE / BMBF-Ariadne; structural path ordering survives. The name is deliberately not "cross-validation" — that term has a specific statistical meaning that does not apply here. (
tests/consistency/test_ariadne_convergence.py,docs/PARAM_SETS.md) -
Nuclear start-year robustness check — regret is not a timing artifact.
nuclear_start_year_regret_analysisoverrides the camp-derived KKW start years (2036/2046/2050) with a uniform value X across all camps. At default parameters,kkw_regret_crossover_year((2020, 2055))returnsNone— KKW max-regret stays ~3 ct/kWh above EE-GAS regardless of when nuclear first delivers. The recommendation survives the counterfactual where KKW arrives in 2028. (core/regret_decision_tree.py)
| Model | Strategy | Key feature |
|---|---|---|
| WEITER-SO | Status quo continuation | Baseline: political inaction, Kohle until 2038, Erdgas wachsend |
| BESTAND | Existing-fleet emphasis, dampened EE-Zubau | Shows what »keep what we have« actually costs |
| EE-GAS | Renewables + Storage + Gas-Backup with green ramp-up | Pragmatic optimum, robust to H₂ uncertainty |
| EE-H2 | Renewables + Storage + H₂-Backup | Pure energy transition, H₂ wager |
| KKW-GAS | Renewables + Nuclear (post-2042) + Bridge-Gas | Nuclear renaissance with realistic build-times |
| KKW-H2 | Renewables + Nuclear + H₂-Backup | Reveals structural independence of backup choice |
Headline finding (30-year average 2026-2055, neutral_default camp;
CO₂ on a system boundary — power-sector emissions plus external
sector-coupling emissions from heating and mobility that remain fossil
in WEITER-SO/BESTAND):
Deterministic baseline — the six paths at their neutral mid-point assumptions:
| Path | Cost (LCOE) | Cumulative CO₂ (system boundary) | vs WEITER-SO |
|---|---|---|---|
| EE-GAS | 16.79 ct/kWh | 2,937 Mt | saves 1,495 Mt (−34 %) |
| WEITER-SO | 16.97 ct/kWh | 4,432 Mt | (baseline) |
| EE-H2 | 17.26 ct/kWh | 2,689 Mt | saves 1,743 Mt (−39 %) |
| BESTAND | 17.33 ct/kWh | 4,870 Mt | adds 438 Mt |
| KKW-GAS | 17.57 ct/kWh | 3,312 Mt | saves 1,120 Mt (−25 %) |
| KKW-H2 | 17.82 ct/kWh | 3,022 Mt | saves 1,410 Mt (−32 %) |
Reading this table requires care. The cost spread across the six paths is 1.03 ct/kWh — smaller than the Monte-Carlo P5-P95 spread within any single path (1.7-3.6 ct/kWh across 2,000 runs sampling uniformly over the camp ranges). Reading off ranks 1-3 from the cost column is reading noise. The robust claims sit elsewhere.
Cost robustness — Monte-Carlo, n = 2,000 over camp ranges:
| Pair comparison | P(left cheaper than right) |
|---|---|
| EE-GAS vs KKW-GAS | 100 % |
| EE-GAS vs KKW-H2 | 100 % |
| EE-GAS vs EE-H2 | 97 % |
| EE-GAS vs BESTAND | 53 % (coin flip) |
| EE-GAS vs WEITER-SO | 24 % (WEITER-SO cheaper in 76 % of runs) |
What survives the noise:
- Among the four active programmatic paths (EE-GAS, EE-H2, KKW-GAS, KKW-H2), EE-GAS is the robust cost choice. Every other active path is more expensive in ≥ 97 % of runs; nuclear paths sit at the bottom of the cost distribution with certainty.
- The CO₂ separation is structural, not parameter-driven. Active
paths emit ~1,500-2,200 Mt less cumulative CO₂ than WEITER-SO/BESTAND
over 30 years — the gap follows from fossil heating + mobility staying
in WEITER-SO/BESTAND vs. being electrified in the active paths, not
from cost assumptions. A power-sector-only CO₂ view would understate
WEITER-SO and BESTAND by roughly a factor of two (their external
emissions live in
r.co2_external_mt_per_year, which the system-boundary total aggregates). KKW paths also emit more system-boundary CO₂ than their EE counterparts within the 30-year window (KKW-GAS 3,312 vs EE-GAS 2,937 Mt) — bridge-phase fossil coverage outweighs later operational gains.
What does not survive the noise — and is therefore stated honestly as a tied cluster, not a ranking:
- The cost ordering between WEITER-SO, EE-GAS, BESTAND and EE-H2 sits inside Monte-Carlo overlap. The cost-only race between WEITER-SO and EE-GAS is a 76/24 split, not a 0.18-ct/kWh "near tie" — that's the "humility thesis" quantified sharply: on cost alone, political inaction is the most likely winner; the case against WEITER-SO is the +1,495 Mt CO₂, not the price tag.
Other camps yield other deterministic top paths (see distinctive property 3 above); the recommendation comes from min-max-regret across camps, not from any single deterministic table.
Numbers reproducible with python -c "from enesys import compute_path; ..."
and python -c "from enesys import monte_carlo_all_paths; ..." (see
Quickstart below).
Zero-setup path — Google Colab. Open
notebooks/01_quickstart.ipynb in
Colab
and run all cells. The first cell %pip installs enesys from this repo;
the rest walks through the forward-cost snapshot, the four-camp lager-
symmetry, the 30-year trajectory, the tornado sensitivity, and a small
Monte-Carlo — five minutes end-to-end.
Local install:
git clone https://github.com/berndhardung/enesys.git
cd enesys
pip install -e .Easiest setup — VS Code + Docker (no Python on the host needed): after git clone open the folder in VS Code (code .) and accept the "Reopen in Container" prompt. The included .devcontainer/ provisions Python 3.12, uv, all dependencies and editor extensions; the venv is built during the container build so the environment is ready as soon as VS Code attaches.
Bare-metal setup with uv: install uv once, then make venv builds the full dev environment deterministically from uv.lock in a few seconds.
curl -LsSf https://astral.sh/uv/install.sh | sh # one-time, ~5 s
make venv # ~5 s, uses uv.lockOr in GitHub Codespaces: click Code → Codespaces → Create codespace. Same .devcontainer/ setup, runs in the cloud.
The compare view (anchor camp ↔ variant camp, six core charts side by side, sliders for any of 19 model parameters) ships as a Streamlit app. The hosted demo lives at enesys.streamlit.app; to run locally:
pip install -e . # streamlit + matplotlib + pandas + plotly all included
streamlit run app/streamlit_app.pyThree pages reachable via the sidebar navigation:
- 🏠 Overview — the headline finding (cost and CO₂ vs BESTAND), the structure of the six paths, the five camps explained.
- 📊 Charts — pick an anchor camp on the left and a variant
camp on the right; the six core charts render side by side.
Slider overrides snap the variant to Individuell with the
deviation shown in the caption above each chart. Deep-links:
?lang=de/en,?anchor=…,?variant=…,?mobile=1for the portrait layout. - 📚 Sources — every default in the model is backed by a primary citation; slider tooltips deep-link to the matching tag.
from enesys import compute_path, baseline_all_paths
# Forward-cost trajectory for the EE-GAS path, 2030-2055, in the
# neutral_default camp:
results = compute_path("ee_gas", years=[2030, 2040, 2050, 2055], camp="neutral_default")
for r in results:
print(f"{r.year}: LCOE = {r.lcoe_ct_kwh:.2f} ct/kWh, CO2 = {r.co2_mt:.1f} Mt")
# Compare all six paths at one year:
prices = baseline_all_paths(year=2045, camp="neutral_default")
for path, lcoe in sorted(prices.items(), key=lambda x: x[1]):
print(f" {path:<10} {lcoe:6.2f} ct/kWh")pytest tests/ -vCovers path-model tests (Demand, Forward LCOE, 30-year integration, six-path invariants, winter-stress test), sensitivity tests (tornado, Monte-Carlo, camp presets), source-traceability tests, and convergence tests against the external parameter substrate (PyPSA-Technology-Data).
enesys/
├── README.md, LICENSE, CHANGELOG.md, AUTHORS.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md
├── pyproject.toml, uv.lock, Makefile, VERSION
│
├── src/enesys/ Model library
│ ├── core/ Data structures, path model, sensitivity, WACC
│ ├── extensions/ land use, consumer, winter stress, profile costs
│ └── viz/ Chart building blocks (matplotlib + plotly theme)
│
├── tests/ pytest suite
└── docs/ Methodology, formulas, sources, architecture
docs/QUICKSTART.md— five-minute orientationdocs/HOWTO_RUN.md— running locally and in Dockerdocs/FORMULAS.md— every formula with derivation and worked examplesdocs/SOURCES.md— every default parameter with source, range, and camp-specific assumptionsdocs/VERSIONING.md— version scheme and how the version is pinned across model packagesdocs/methodik/modell_architektur.md— high-level architecture overviewdocs/methodik/— further methodology deep dives (steady-state parameter consistency, bridge phase parameters, English methodology overview)
The model is calibrated against:
- Fraunhofer ISE — Stromgestehungskosten erneuerbarer Energien 2024 (PDF)
- BloombergNEF — Energy Storage Cost Survey 2025 (10 Dec 2025); LCOE Report 2026 (18 Feb 2026)
- EWI Köln — Hydrogen storage requirements
- Cour des Comptes (FR) — Flamanville EPR audit (17 years, €23.7 bn)
- EDF — Hinkley Point C status reports
- KENFO — Geschäftsbericht 2024
- KernD — Bewertung der Fraunhofer ISE Studie (pro-nuclear critique)
- Modo Energy — Inertia in Europe (Nov 2025)
- DE-TSOs — Joint inertia procurement (22 Jan 2026)
Full traceability is in docs/SOURCES.md.
- Pinned dependencies in
uv.lock(76 packages with hashes) - Deterministic Monte-Carlo with fixed seeds in tests
- Python 3.10+ with explicit version constraints
- CI on every commit via GitHub Actions
- All defaults annotated with
[SRC: ...]tags pointing todocs/SOURCES.md
If you can't reproduce a result, please open an issue with the parameter set you used and the command you ran.
If you use this model in academic or policy work, please cite:
@software{enesys_2026,
author = {Hardung, Bernd},
title = {enesys: A transparent cost-robustness analysis
for the German energy transition},
year = {2026},
publisher = {GitHub},
url = {https://github.com/berndhardung/enesys},
note = {See CITATION.cff for current version}
}This project welcomes:
- Bug reports for incorrect formulas, source mismatches, or test failures
- Source updates when newer studies (BNEF, ISE, BNetzA) are published
- Translations of the model into other regulatory contexts (FR, PL, etc.)
- Critical reviews of methodology — issues that say "your assumption X is wrong because Y" are especially welcome
What this project does not want:
- Pull requests that hardcode advocacy positions (pro/anti renewables, pro/anti nuclear) — the model must remain neutral and parameter-driven
- Deletion of "inconvenient" parameters or sources
See CONTRIBUTING.md for details.
Source code (src/, tests/, top-level configuration) is licensed
under the MIT License — see LICENSE.
Methodology documentation under docs/ is distributed under MIT alongside
the source code; a separate license tier (CC0 or CC-BY-4.0) may be applied
at a later date.