Tariffs and prices

A German electricity bill is a stack, and every layer moves for its own reasons: five day-ahead sources, § 51 EEG, and an advisor for the § 14a Module.

Since 1 October 2025 the German day-ahead auction clears in quarter hours, and since § 41a EnWG every supplier has to offer a tariff that follows it. That makes “what does a kilowatt-hour cost at 17:15 on Thursday?” a question with an answer — and the answer is the input the planner spends most of its objective on.

hems-tariff is that answer. Like every domain crate it does no I/O: the parsers are &str in, a series out, and the fetching lives in tariffd.

The bill is a stack

Every layer moves for its own reasons and on its own clock.

The layers stay separate all the way through and are summed only at the edge. The optimiser wants one number per direction per slot; the household wants to see the stack; and the Modul advisor has to take it apart again to answer “would Modul 2 be cheaper for us?”. A design that adds them up early can serve exactly one of those three.

flowchart LR
  subgraph pub["published, fetched by tariffd"]
    E["day-ahead energy
ct/kWh, per quarter hour"]
  end
  subgraph conf["configured from the price sheet"]
    N["network working price
+ Modul 3 level"]
    L["levies and taxes"]
  end
  E --> S["SlotPrice"]
  N --> S
  L --> S
  S --> IMP["import_ct, gross
what the household pays"]
  S --> EXP["export_ct
what it earns — zero when
the spot price was negative"]
  S --> SH["shared_import_ct
a § 42c allocation"]
  IMP --> P["the planner"]
  EXP --> P
  SH --> P

Money is exact, control is not

Prices and costs are rust_decimal::Decimal in cents per kilowatt-hour. Optimisation runs on f64, because a solver has no use for exact arithmetic — but anything a household is shown or billed comes back through Decimal, and the conversion happens once, at SlotPrice::import_f64. A settlement that went through a double is one nobody can reproduce.

The five sources, and what is a decision in each

hems-tariff::source parses what ENTSO-E, SMARD, aWATTar, Tibber and Energy-Charts actually publish. Each is a pure function tested against a captured response, on a machine with no network, including the two days a year with 92 and 100 quarter hours.

Three things in them are decisions rather than plumbing:

• The factor of ten lives in one place. €/MWh and ct/kWh differ by ten, and a conversion written at five call sites is wrong at one of them within a year.
• An hourly price expands, it is never averaged. A source that still publishes hours becomes four identical quarter hours. Averaging four quarter hours into one destroys exactly the structure the planner exists to exploit.
• A position is resolved as an instant, not an index. “Position 27 of the day” is not a time until somebody says which day, and on the March and October days it is not the same quarter hour as position 27 of an ordinary one.

And one that costs real money: Tibber publishes a gross consumer price. PriceBasis refuses to let it be used as a wholesale one, because adding the levies on top of a price that already contains them roughly doubles a household’s modelled bill — and a planner optimising against a doubled bill makes confident, expensive, wrong decisions.

§ 51 EEG: negative prices earn nothing

§ 51 Abs. 1 reduces the anzulegender Wert to zero in a quarter hour with a negative spot price. Both remuneration schemes are computed from it — § 53 Abs. 1 for the Einspeisevergütung, Anlage 1 zu § 23a Nr. 1 for the Marktprämie — so the rule zeroes the tariff and the premium.

The rule turns on the sign of the market price, not on what the household pays, so the raw spot price is kept beside the retail one. A household in Direktvermarktung priced at spot + Prämie that still books the premium in a negative hour is being paid in exactly the hours the statute is paying it to stop.

Whether the rule reaches a plant at all is a date, and it is deliberately not the same date that lifts the § 9 EEG cap — see the grid rules.

Modul 3: time-variable network charges

A network charge that changes with the clock is the second dynamic layer, and it moves for entirely different reasons from the first: the market prices energy, Modul 3 prices the network at times it is busy. The two are routinely out of phase, and a planner that only sees one of them will move load into an hour that is cheap on the wrong axis.

Availability is narrow and worth knowing before building on it: only together with Modul 1, only at a location without registrierende Leistungsmessung, and only with an intelligent metering system. Windows and price levels are fixed per calendar year for the whole network area, must be billed in at least two quarters, and must appear in the preliminary price sheet by 15 October of the preceding year.

There is no machine-readable national format for any of it — a PDF and an Excel sheet per network operator — so the installer transcribes the operator’s calendar into [tariff.modul3] and hemsd run --check refuses a box whose calendar breaks the Anwendungshilfe. Window membership is decided on local wall-clock time, because that is how the price sheet is written.

At fleet scale the same transcription happens once per Netzgebiet instead of once per household: tariffd keeps a curated catalogue — one entry per network operator and year, in the identical shape a box uses, validated against the same Anwendungshilfe when the daemon starts — and serves it openly at /v1/modul3/{netzbetreiber}. A calendar that breaks a rule refuses the whole daemon rather than warning, because a fleet serving windows nobody may sell prices a whole Netzgebiet against a tariff nobody may be billed on.

It is worth planning against for a reason the wholesale curve cannot match: the windows are fixed a year in advance. A household learns tomorrow’s spot price each afternoon and learns its Hochtarif in the preliminary price sheet of the preceding October, so the one flexibility it can schedule around with certainty is this one.

The Modul advisor

Modul 1, 2 and 3 are three different ways of paying the network, and a household may only pick one. Nobody selling a tariff can be relied on to run that comparison in the customer’s favour. A household’s own energy manager can, and it is holding the only data that answers it.

compare_moduls prices a household’s own quarter-hourly history under all three:

  Modul 2 pays above                     2417 kWh/a
  …on this day it would have         -3.97 € on the energy

It reports a threshold in kilowatt-hours a year, not a projected saving, and that is the whole design. One day is not a year: multiplying a January Thursday with a car on the cable by 365 tells every household with an electric vehicle that it is losing four figures. A threshold is a statement the data supports — above this annual consumption, Modul 2 is the cheaper contract — and the household already knows which side of it they are on.

§ 42c: the neighbours’ roof is a cheap block, not a free one

A household in an energy-sharing community pays the community’s energy component for the kilowatt-hours the Aufteilungsschlüssel allocated, and its supplier’s for the rest. SlotPrice::shared_import_ct is the same stack with the community’s energy price substituted, so the two are directly comparable.

The result is less dramatic than the phrase “free solar from the neighbours” suggests, and the arithmetic is the point: a community selling at 12 ct/kWh net still delivers a 32,5 ct kilowatt-hour against the supplier’s 47,9 — a third off. Network charges, the Stromsteuer, the Konzessionsabgabe and 19 % VAT do not care where the electron came from, because it crossed the public grid to get here.

SlotPrice::sharing_discount_f64 floors at zero, and never goes negative. That keeps the planner’s sharing term convex; the reasoning is in the planner.

Two queries a household actually asks

Most of what a household wants from a price curve is not an optimisation.

let stack = PriceStack::resolve(&tariff, horizon);

// "When should the car charge?" — answerable without a solver.
let cheapest = stack.cheapest_slots(8);

// "When does feeding in earn nothing?" (§ 51 EEG)
let free = stack.zero_export_slots();

// The number that decides whether grid-charging a battery pays at all:
// the spread has to cover the round trip and the wear before anything is left.
let spread = stack.spread_ct();

Where the prices come from at runtime

tariffd fetches the published curves, reconciles a curve that arrives twice under a written trust order, and holds two days each way so a WAN outage never costs a plan. That is on the fleet page; what matters here is that none of it is in this crate, and that a price stack which merely stops short of the horizon is still allowed — a horizon can outrun the last published auction, and being indifferent about when to act out there is exactly the state of knowledge.