Does residual load explain German day-ahead electricity prices on a year the model never saw, and how wrong was the flat merit-order price?

research/energy-price-residual

Does residual load explain German day-ahead prices on a year the model never saw? Yes: correlation 0.87 and RMSE 9.6 EUR/MWh on 2019 against 23.0 for the flat merit-order price; the cheap-quarter mean is within EUR 4–9/MWh; the rest is the fuel-price level.

5 tests

What it found

Fits price as a non-decreasing function of residual load (load minus solar minus onshore wind) by isotonic regression on hourly German day-ahead prices 2015–2019 (OPSD/ENTSO-E), and scores it on a held-out year. Trained on 2015–2018 and tested on 2019: correlation 0.87 and RMSE 9.6 EUR/MWh, against 14.5 for an hour-of-day mean, 15.7 for the training mean and 23.0 for the flat EUR 54.6/MWh merit-order price of energy-thermal-hourly. Leave-one-year-out over all five years gives correlation 0.78–0.90 and beats every non-residual baseline in every year; the flat model is worse than the training mean in four of five. The predicted cheapest-quarter mean, which is what a storage buyer pays, is within EUR 3.7–9.2/MWh of the actual one. Most of the remaining error is the yearly price level (fuel and carbon), not shape: an oracle level cuts RMSE to 6.2–11.2. No data after 2019 (hosts unreachable from the run container); one held-out year per score, no confidence interval; realised rather than forecast load and renewables; onshore wind and solar only. All inputs are from sources.

A figure from its own output

Figure produced by research/energy-price-residual: rmse — energy-price-residual/experiments/rmse.svg, written by the project's own experiment code.

Paper, code and results

More in Energy supply and dense energy storage

Does a weather-dependent heat-pump COP change the sand-store versus heat-pump verdict?Does a weather-dependent heat-pump COP change the sand-store verdict? The table's 3.6 equals a 35 C floor-heating sink; at 45 to 55 C the store's weekly flip multiplier rises 8 to 46%, and 2018 still favours the pump.
How far must capital costs move to flip the sand-store versus heat-pump verdict, and is that more than the price year moves it?How far must capex move to flip the sand-store vs heat-pump verdict? Weekly cycling flips on a 2 to 40% store-capex change (the cost table's own vintage step is 26%); daily cycling holds to 2 to 3.5 times.
How much can the sand store's heat exchanger cost before a heat pump wins again?How much can the sand store's heat exchanger cost? Weekly cycling survives about 120 to 270 EUR per kW in four of five price years; daily cycling survives 460 to 900; 2018 never wins.
Does the sand-store versus heat-pump verdict survive a modelled electricity price instead of the real one?Does the sand-store vs heat-pump verdict survive a modelled price? Mostly: 32 of 40 cells agree with real prices (21 for the flat merit-order price), but weekly cycling, whose margin is under EUR 0.011/kWh, is wrong in sign in 5 of 5 years; as a charge scheduler the model costs under EUR 0.0011/kWh.
How much cheaper must a store's electricity be than a heat pump's for a sand heat store to deliver heat for less?How much cheaper must electricity be for a sand heat store to beat a heat pump? Sourced costs (PyPSA technology-data) on the energy-thermal model: a daily-cycled store needs the pump's electricity about 2 times dearer, weekly 5, monthly 16, seasonal 120 (house); the heat lost in standby, not the sand, rules out seasonal storage.
If a sand heat store buys electricity only in the cheap hours of a German price series, when does it beat a heat pump?When does a sand heat store beat a heat pump on hourly prices? With 189 GW solar and 137 GW wind and a modelled German price series, a daily or weekly store wins in all 5 years (EUR 0.027/kWh against 0.071); a levy of 0.10/kWh or a surplus-hour price above about 20–35 EUR/MWh removes the win.

All 39 Energy supply and dense energy storage projects →

Other research topics

Proper scoring & elicitation28 Effort & peer prediction9 Verification games & slashing19 Detection & tolerances14 Markets as routers13 Prediction-market design9 Paying contributors8 Decentralised training systems23 Digital twins & sim-to-real164 LLM chip design loops40 Robot manufacturing loops53 Ambient AI assistant46 Self-replicating manufacturing34 Simulation-first design64 Engineered clothing45 Task-specific AI accelerators40 Algorithmic chip-design loops42 Open-source purpose-built robots44 The software-engineering loop43 Adversarial human imitation41 Open-source residential construction41 Autonomous multi-angle capture34 Sensing, imaging and 3D31 Energy supply and dense energy storage39 Robot societies among people31 Agentic co-design of low-cost robots27 Laundry-hanging robots25 Machine learning on qualitative data23 Robots on the building site27 Civilizations & group psychology1 History & human sciences1

Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.