Heimdall: Only the Safe Shall Pass

Model card for the Heimdall thesis (BSc Software Engineering, SDU Sønderborg, with Danfoss A/S as industrial partner) by Phongsakon Mark Konrad and Tim Lukas Adam. This repository holds the forecaster checkpoints and serves as the project landing page.

The power grid keeps one promise every second, that supply matches demand, and the balancing market is where that promise is settled in real time. As machine learning models and large language models start to take part in this market, they bring failure modes of their own. A forecasting model hides its own uncertainty, and a language model can hallucinate. Heimdall is the safety guard that sits between an autonomous bidder and the grid. It uses conformal prediction to turn an uncertain price forecast into a promise the operator can act on, that an accepted bid stays above their loss limit with the coverage they asked for.

We run three kinds of bidder behind the guard, a deterministic chooser, a single language model, and a multi-persona society. We report honestly that the language model does not beat the deterministic baseline on profit. What it adds is auditability, because every accepted and rejected bid carries its reasoning, and steerability, because the society answers to plain-language instructions. The guard holds its coverage to within a fraction of a percent of the 90 percent target, including straight through the March 2025 rule change that reshaped the market. Every number reproduces from real DK1 panels and real model traces.

Where things live

Artifact	Location
Source code	https://github.com/phomarkon/heimdall-src
Forecaster checkpoints	this repository
Data bundle (CC-BY-4.0)	DVC, reproducible from public sources

Forecaster checkpoints

The weights predict 15-minute imbalance prices for the Nordic mFRR EAM in bidding zone DK1 under the post-2025-03-04 single-price regime. The reproducible validation and test metrics, predictions, leaderboard, and training logs live in the source repository (phomarkon/heimdall-src).

Per-model layout

<F-id>/seed-<n>/
    model.pt | model.npz       # weights
    stats.pkl                  # train-set normalization (mean/std)
    config.json                # architecture + training config

Forecaster zoo

ID	Family	Notes
F0	Seasonal AR(24)	deterministic baseline
F1	Quantile LightGBM	gradient-boosted quantile trees
F2	Bayesian Linear Regression	closed-form posterior
F3	Deep Ensemble (5x F7)	aggregation over F7 seeds (ADR-0006)
F3-Lite	LSTM DeepAR	appendix, original day-3 implementation
F4	MC-Dropout transformer	K=30 over F7 backbones
F7	patchTST + quantile + split-CP	univariate workhorse
F8	patchTST + online ACI	regime-shift-aware, default in production
F9	TimesFM-2.0 zero-shot	foundation model, deterministic
F10	Chronos-Bolt zero-shot	foundation model, appendix
F11	PriceFM-shaped surrogate	8-layer x 192-d patchTST, no public PriceFM weights

Reproduce

git clone https://github.com/phomarkon/heimdall-src
cd heimdall-src
uv sync --frozen
uv run dvc pull

Forecaster checkpoints hydrate from this repository on first use. The full runbook is in docs/REPRODUCE.md.

License

Apache 2.0.

Citation

@misc{heimdall2026,
  title={Heimdall: Only the Safe Shall Pass},
  author={Konrad, Phongsakon and Adam, Tim Lukas},
  year={2026},
  note={BSc thesis, SDU S{\o}nderborg; industrial partner Danfoss A/S.}
}