German Energy Shift

Architecture

How the system is shaped. Full implementation stays private.

1. Step 1

   Ingest

   SMARD day-ahead prices (filter 4169) and When2Heat DE heat demand plus COP, both CC BY 4.0 public series.

2. Step 2

   Price model

   Offline train on 2020 to 2021. Publish ridge MAE on the hard 2022 holdout year. Freeze metrics.json.

3. Step 3

   Heat stress

   Normalized DE space-heat demand blended with inverted ASHP COP into a unitless stress index.

4. Step 4

   Shift score

   score = w1 * norm_price + w2 * norm_stress. Top-K run hours and avoid hours. Worker serves frozen scenarios only.

Holdout day-ahead MAE (2022)

SMARD DE-LU filter 4169, hourly 2020 to 2021 train and 2022 holdout. The table publishes MAE in EUR/MWh as the primary score. Floored MAPE was also computed with |price| at least 5 EUR/MWh. It stays secondary because near-zero and negative prices explode raw MAPE.

Algorithm

Hybrid price-stress shift ranking

Normalize day-ahead price and heat-pump stress, blend with recruiter-tunable weights, then pick the cheapest low-stress windows.

price_n = minmax(prices_eur_mwh)
stress_n = minmax(stress_index)
score = w_price * price_n + w_stress * stress_n
run_hours = argsort(score)[:top_k]
avoid_hours = argsort(score, descending)[:top_k]

Key logic

Compact illustrative snippet (Deterministic shift blend (illustrative)). Not the full codebase.

def build_shift_recommendation(prices, stress, price_weight=0.6, stress_weight=0.4, top_k=24):
    w1, w2 = clamp01(price_weight), clamp01(stress_weight)
    score = w1 * minmax(prices) + w2 * minmax(stress)
    order = np.argsort(score)
    return {"run_hours": order[:top_k].tolist(), "avoid_hours": order[::-1][:top_k].tolist()}

Stack