What we simulate, and how precisely.

Every number on the platform belongs to exactly one of three tiers, and the product interface must never let a T3 number look like a T1 number:

The computable grid model is derived deterministically from the OSM data: every recomputation returns identical IDs and identical results for identical inputs.

Scope and voltage classes

• Three voltage classes: 380 kV, 220 kV, 110 kV. The highest voltage tag decides the assignment: ≥300 kV → 380 kV class, ≥175 kV → 220 kV, ≥60 kV → 110 kV (150 kV lines are added to the 110 kV class). Anything below 60 kV or above 800 kV lies outside the model.
• Spatial extent: Germany plus a 20 km border buffer. Foreign grid meshes without load and generation data would form artificial parallel paths; interconnectors end as boundary nodes. Transit ring flows (AT/PL/CZ) are therefore deliberately outside the model (see section 12).
• Metric: all distance calculations run in EPSG:25832 (ETRS89/UTM 32N, true metres).

Node formation (clustering)

• Substation geometries are clustered with a union-find grid method at a 500 m radius (PyPSA-Eur convention). One grid node = (location cluster × voltage class); the centroid is the mean of the members, name and operator take the most frequent value.
• Line ends without a matching substation within 2,000 m become junction nodes (400 m raster bundling), so that lines meeting mid-span share a node. Lines are never dropped.
• Parallel circuits are combined into one edge with circuits=N: R and X per circuit ÷ N, thermal rating × N.
• Transformers connect adjacent voltage classes of the same cluster as separate edges; HVDC links (frequency tag 0) are typed as dc_link and excluded from the AC screening (controlled setpoint flows).
• Degenerate edges (zero length <1 m, self-loops) are removed and counted; minimum counts (≥1,000 substations, ≥5,000 lines) prevent a broken OSM import from emptying the model (fail-closed).

Resulting size: around 7,950 screened edges after clustering. Quality metrics (island shares, parametrisation level per class, connected components) are logged on every run.

Generation

• Plant fleet: all MaStR units marked "in operation" (status code 35) with coordinates and rated capacity, assigned to the grid model through the unit-to-substation mapping (section 10), otherwise to the nearest node that meets the same hard voltage floor.
• Wind capacity factor from the ERA5 100 m wind, with a cubic power curve between cut-in and rated wind speed:

CF_wind(v) = 0 for v < 3 m/s or v > 25 m/s
           = (v³ − v_ci³)/(v_r³ − v_ci³) for v_ci ≤ v < v_r
           = 1 from rated wind speed upwards
Fleet split by coastal AGS prefix {01,02,03,04,13}:
  high_wind v_r=13.5 m/s, otherwise low_wind v_r=10.0 m/s

• PV capacity factor from global irradiance: CF_pv = (G/1000 W/m²) · PR with a performance ratio of 0.9, clipped to [0,1] (v1 without orientation or tilt; an optional temperature derating of −0.4%/K exists but is off by default).
• Scaling to reality: the capacity factor model supplies the spatial pattern; the national hourly sum for wind and PV is scaled multiplicatively onto the SMARD actual generation (unscaled, the wind model ran 31% high, which would have overstated the north-south flows).
• Conventional plants, biomass, hydro and pumped storage: SMARD actual generation per carrier group, distributed pro rata by installed capacity per node (time-invariant weights, national sum preserved exactly). No unit commitment model in v1 (see section 12).

Load

• eXtremOS supplies hourly sectoral load profiles per NUTS-3 region (402 regions; weather year 2012, leap day removed, so 8,760 hours map 1:1 onto 2025).
• Regionalisation onto 110 kV nodes uses ENLAPA's own land use data: industrial load by industrial and commercial areas, household plus commerce and services load by settlement areas. Three energy-preserving corrections: (1) nodes with more than 1 TWh/a of assigned industrial energy are moved to the nearest ≥220 kV node (≤25 km), a direct-connection heuristic motivated by BASF Ludwigshafen (around 6.35 TWh/a); (2) every weight polygon is distributed inverse-quadratically across the 3 nearest 110 kV nodes; (3) capping: at most 50% of a (region × sector) energy on a single node, with the excess redistributed pro rata.
• SMARD anchor: the national hourly sum is scaled onto the realised grid load (SMARD filter 410). Fail-closed gates: annual energy 400–600 TWh, hourly mean 40–90 GW.

Headroom answers one question: how much additional feed-in ΔP can node n absorb in hour t before any line exceeds its limit? The additional feed-in is taken up by the distributed slack; the slack-adjusted sensitivity of line l to node n is:

S_ln = PTDF_ln − Σ_k w_k·PTDF_lk
headroom_n(t) = min_l ( cap_l − |f_l(t)| ) / |S_ln|
   across all l with |S_ln| > 0.05 and known capacity

• Sensitivity threshold 5%: the conventional "significantly affected element" limit. (A threshold of 10⁻⁴ proved far too permissive: a distant line with a sensitivity of 0.001 and a negative margin zeroed out every node.)
• Positive-margin semantics: lines that are already at or beyond their limit in a given hour (margin ≤ 0) drop out of that hour's minimum. Base-case violations are TSO redispatch reality, not a marginal constraint on additional feed-in. Without these semantics, headroom would be identically zero across Germany (at around 10 TWh/a of redispatch, some EHV element is violated in almost every hour).
• Constraint set = EHV (≥220 kV), NEP scope. 110 kV edges are reported in full but are not used as a headroom constraint (the NUTS-3 load clustering produced around 500 chronic artefact overloads there). Edges with suspect data (own p95 loading > 2.0) are excluded and counted.
• Percentile criterion: the 8,760-hour series per node yields p100 (the minimum), p99 and p95 = the feed-in that would be possible without overload in at least 95% of the hours of the year, and that is the default display of the map layer. Capped at 100,000 MW (no binding element); nodes with no constraint at all are shown as "not assessed" (grey). Each node also carries the most frequently binding line (limiting_edge).
• Direction: the value is a feed-in headroom (additional generation). For withdrawal (battery charging, electrolysis) the sign of the grid effect reverses; a dedicated withdrawal headroom is on the roadmap. Until then, curtailment hotspots and §13k relief regions are the reliable siting signals for flexible load.

N-1 calculation (LODF): implemented, disabled for the customer layer

f_l^post(k) = f_l + LODF_lk·f_k
LODF_lk = (PTDF_l,f(k) − PTDF_l,t(k)) / (1 − (PTDF_k,f(k) − PTDF_k,t(k)))

Contingency set = EHV; bridges (island formers, |denominator| < 10⁻⁶) are excluded and counted (their N-1 event is loss of load, not redistribution). The N-1-aware headroom adds, for each monitored line, its 5 worst contingencies as pseudo-lines with sensitivity PTDF[l,·] + LODF[l,k]·PTDF[k,·], and is by construction ≤ N-0. A deliberate decision: the customer layer shows N-0 headroom. At today's parametrisation level, the strict N-1 headroom (median p95 ≈ 10 MW; 2,053 of 4,217 substations below 10 MW) is parameter noise rather than signal. It will be released only once the corridor impedance coverage turns the N-1 corridor rank correlation positive (section 8). The N-1 loading and calibration metrics keep running in the background.

TSO side (measured): redispatch measures from all four TSOs including the measured energy (MWh). The free-text unit names are geocoded against the MaStR: normalisation (remove everything outside [a-zA-Z0-9], lower-case), preferring an exact match, otherwise a substring search over the first non-stopword token of at least 5 letters, with the highest rated capacity as the tiebreak. Coverage is around 68.9% of the name references.

DSO side (E.ON group): measured facts, modelled energy. The feed supplies, per measure × unit × time slice: start and end, duration, curtailment step and the EEG unit key (joined onto the MaStR; 46.4% of the keys match a unit of at least 30 kW), but no energy volume. That volume is modelled in two stages:

v1 (upper bound):  E = P_rated · (step%/100, clamped to [0,1]) · (duration_min/60)
CF-refined:        E = P_rated · step% · ∫ CF_ERA5(t) dt   over the measure window

The integral sums, per hour, overlap(window, hour) × CF(hour, grid cell); wind uses the two-class power curve, PV the global irradiance (0.25° cell of the unit). Windows longer than 168 h are extrapolated with the window mean after the first exactly integrated week; functional tests are excluded. Both the v1 and the CF value are stored, so the difference stays visible. Why an upper bound: v1 assumes full load throughout the curtailment window; the CF weighting replaces that with the actual weather resource, but it remains a model estimate (not a measurement of lost energy, and structurally impossible to validate, because DSOs do not publish unit-level lost energy).

Consistency rules in the display: measures are counted everywhere as distinct events (not as time slices); measured facts (count, hours, weekly profile) carry "measured", energy volumes carry "model".

All thresholds are configurable as documented environment variables; changes run through the monthly calibration loop with its regression guard.

Every value carries its reliability tier in the product (measured / model). Companion document to the grid intelligence page.