Replace real-export corroboration with a wallbox-aware battery projection (v2.39.0)

v2.37 required the grid meter to show a real feed-in before a device could
switch on via the modelled-surplus path, to catch cases where remaining_
surplus read positive only because the wallbox reservation was modelled
too low relative to a gross-PV upswing. In practice it also blocked
genuinely safe starts: confirmed live 2026-09-17, the pool pump's window
opened at 09:00 and modelled surplus cleared its need by ~09:20-09:30 on
a clear, EV-away morning, but the real-export gate held it off until a
forced price slot rescued it at 10:48 - the house battery (32% SOC at
sunrise) was legitimately absorbing the "surplus" the model saw, and the
gate had no way to know that was fine given how much margin the battery
still had before its own deadline.

_battery_would_still_reach_full already answers the question the export
gate was approximating, more precisely: it explicitly models the
worst case (the device's whole draw comes straight out of the battery's
charge rate, not some separate "free" pool) and re-evaluates every cycle
against the battery's own reach-full-by-deadline math. The one gap: it
never accounted for the wallbox's reservation, only for other managed
devices' committed draw. Fixed by seeding cumulative_battery_competing_
draw with wallbox_reserved_kw, so every device's projection treats a
car that still wants power as a prior claim on the battery's charge rate
- whether or not the wallbox is drawing that reservation yet.

With the wallbox now counted, would_delay_battery_full is a strictly
more precise substitute for what the export gate was a blunt proxy for,
so v2.37's export_uncorroborated check (and its EXPORT_CORROBORATION_
MARGIN_KW / cumulative_export_committed bookkeeping) is removed outright
rather than kept alongside it. The separate battery-affordability path
(battery_would_last / export_gate_open, v2.35/36) is untouched.

Backtested against real sensor history from both 2026-09-16 (EV home,
wallbox reserving) and 2026-09-17 (EV away): on the EV-home day the new
rule's first qualifying cycle for the pool pump landed at 11:15, the same
10-minute window the forced price slot actually used that day - no
earlier, because the wallbox's reservation correctly held it back through
three prior hours of weak/contested surplus. On the EV-away day it would
have started ~75 minutes earlier than the forced fallback that actually
fired.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Author: xchillxx

custom_components/surplus_load_switch/const.py

@@ -525,21 +525,6 @@ def _minutes_to_cycles(minutes: float) -> int:
 EXPORT_GATE_MEDIAN_WINDOW = 3
 EXPORT_GATE_SOC_OVERRIDE = 99.0
 
-# v2.37: a device may only be *switched on* via the modelled-surplus path
-# (`remaining_surplus > need`) while the meter also shows the household is
-# really exporting at least that device's own draw (minus what devices
-# already granted this cycle have claimed from the same reading). The
-# modelled surplus can briefly read strongly positive when the wallbox
-# reservation lags a gross-PV upswing — confirmed live 2026-09-10, a
-# broken-cloud day: "Überschuss" spiking to +7 kW while the Tibber meter
-# showed ~0.5 kW real feed-in (the car was ramping up and taking the
-# rest), the whole low-priority cascade switching fully on, then off
-# again a few minutes later. Only enforced while battery_full_projection_
-# applies (daytime) and an export sensor is configured; an already-on
-# device is never shed by this (its own draw is already in the meter
-# reading) — it only blocks a fresh turn-on.
-EXPORT_CORROBORATION_MARGIN_KW = 0.0
-
 # Default monthly solar offsets (hours after sunrise until PV is useful)
 DEFAULT_SOLAR_OFFSETS = [3.5, 3.0, 2.5, 2.0, 2.0, 2.2, 2.2, 2.0, 2.5, 3.0, 3.5, 4.0]
 

custom_components/surplus_load_switch/coordinator.py

@@ -72,7 +72,6 @@
     DEFAULT_SOLAR_OFFSETS,
     DISCHARGE_SMOOTHING_SAMPLES,
     DOMAIN,
-    EXPORT_CORROBORATION_MARGIN_KW,
     EXPORT_GATE_MEDIAN_WINDOW,
     EXPORT_GATE_MIN_KW,
     EXPORT_GATE_RELEASE_KW,
@@ -2836,15 +2835,13 @@ async def _evaluate_devices(self, data: CoordinatorData) -> None:
         # cumulative_committed above; kept separate since this one only
         # ever grows from surplus-path grants (battery_would_last is
         # already covered by the reactive battery_eligible_ids gate, and
-        # force_runtime is exempt everywhere).
-        cumulative_battery_competing_draw = 0.0
-        # v2.37: predicted draw of every device turned on via the modelled-
-        # surplus path *this cycle that wasn't already on* — subtracted
-        # from the export-meter reading before the next device's turn-on
-        # is corroborated against it, so a single positive meter reading
-        # isn't spent by several devices at once. Already-on devices are
-        # left out: their draw is already reflected in the meter.
-        cumulative_export_committed = 0.0
+        # force_runtime is exempt everywhere). Seeded with the wallbox's
+        # own reservation (v2.39), not 0 — a car that still wants power is
+        # a prior claim on the battery's charge rate too, worst-case, the
+        # same way _battery_would_still_reach_full already treats every
+        # device's draw: whether or not the wallbox is drawing that
+        # reservation *yet*, the projection assumes it will.
+        cumulative_battery_competing_draw = data.wallbox_reserved_kw
         now_dt = dt_util.utcnow()
         horizon_end = now_dt + timedelta(hours=data.effective_h_to_solar + BATT_OK_BUFFER_H)
 
@@ -3371,42 +3368,32 @@ async def _evaluate_devices(self, data: CoordinatorData) -> None:
             # surplus path only (see should_on/should_off below) —
             # battery_would_last is already fully covered by the
             # reactive gate.
+            # Worst-case, not an exact simulation (see
+            # _battery_would_still_reach_full): assumes this device's whole
+            # draw, on top of the wallbox's reservation and whatever
+            # higher-priority devices already claimed this cycle
+            # (cumulative_battery_competing_draw, seeded with
+            # wallbox_reserved_kw above), comes straight out of the
+            # battery's own charge rate — never out of some separate pool
+            # of "free" surplus that doesn't affect the battery either way.
+            # That makes this the sole gate the surplus path needs. v2.37
+            # had additionally required the grid meter to show a real
+            # feed-in before a fresh turn-on, to catch cases where
+            # `remaining_surplus` read positive only because the wallbox's
+            # reservation was modelled too low, not because power was
+            # genuinely spare — v2.39 removes that on top of this check:
+            # with the wallbox's reservation now counted here directly, a
+            # device that passes this check isn't taking anything the
+            # wallbox needed or the battery couldn't spare, independent of
+            # what the grid meter shows at this exact moment.
             would_delay_battery_full = not self._battery_would_still_reach_full(
                 data, cumulative_battery_competing_draw + predicted_power
             )
-            # v2.37: while an export sensor is configured and the sun is up,
-            # a device may only be *switched on* via the modelled-surplus
-            # path if the meter also shows a real feed-in that covers its
-            # own draw (minus what devices turned on earlier this cycle
-            # have already claimed from the same reading). The modelled
-            # `remaining_surplus` can read strongly positive for a few
-            # cycles when the wallbox reservation lags a gross-PV upswing,
-            # while almost nothing is actually leaving the house — see
-            # EXPORT_CORROBORATION_MARGIN_KW. Never sheds an already-on
-            # device (its draw is already in the meter reading); only
-            # blocks a fresh turn-on. At night / no export sensor this is
-            # always False. Same SOC override as the battery-path export
-            # gate above (EXPORT_GATE_SOC_OVERRIDE): once the house battery
-            # is effectively full it can't be the thing silently absorbing
-            # the modelled surplus, so the meter reading and the model can
-            # no longer disagree for that reason — the corroboration
-            # requirement has nothing left to protect against.
-            export_uncorroborated = (
-                data.export_sensor_configured
-                and data.battery_full_projection_applies
-                and data.soc < EXPORT_GATE_SOC_OVERRIDE
-                and not is_on
-                and (
-                    data.export_smoothed_kw - cumulative_export_committed
-                    < predicted_power + EXPORT_CORROBORATION_MARGIN_KW
-                )
-            )
             should_on = (
                 force_runtime
                 or (
                     remaining_surplus > predicted_power + SURPLUS_ON_THRESHOLD
                     and not would_delay_battery_full
-                    and not export_uncorroborated
                 )
                 or battery_would_last
             )
@@ -3491,17 +3478,6 @@ async def _evaluate_devices(self, data: CoordinatorData) -> None:
                         f"{predicted_power:.2f} kW benötigt, Akku würde nicht bis Solar-Start reichen)"
                     )
                 await self._log_decision(dev, False, decision_titel, decision_reason)
-            elif export_uncorroborated and remaining_surplus > predicted_power + SURPLUS_ON_THRESHOLD:
-                # Modelled surplus would switch this on, but the meter
-                # doesn't back it up yet — held, not switched, so the log
-                # doesn't read as a plain hysteresis hold.
-                await self._log_decision(
-                    dev, False, "Wartet — keine echte Einspeisung",
-                    f"rechnerischer Überschuss reichte ({remaining_surplus:.2f} kW), aber es "
-                    f"wird nur {max(data.export_smoothed_kw - cumulative_export_committed, 0.0):.2f} "
-                    f"kW wirklich eingespeist ({predicted_power:.2f} kW nötig) — noch nicht "
-                    f"eingeschaltet",
-                )
             else:
                 stable_titel = "Bleibt an — Hysterese" if is_on else "Bleibt aus — Hysterese"
                 decision_reason = (
@@ -3527,13 +3503,6 @@ async def _evaluate_devices(self, data: CoordinatorData) -> None:
                 # what's already been committed ahead of it, not just its
                 # own isolated draw in a vacuum.
                 cumulative_battery_competing_draw += predicted_power
-                # A device turned on this cycle that wasn't already on will
-                # start drawing from real export the moment it switches —
-                # so the next device's export-corroboration check must see
-                # that reduced headroom, not the raw meter reading (which
-                # still reflects a moment before this one came on).
-                if not is_on:
-                    cumulative_export_committed += predicted_power
 
             if should_on and not is_on:
                 # Capped rather than incremented without bound: once

custom_components/surplus_load_switch/manifest.json

@@ -9,5 +9,5 @@
   "iot_class": "local_polling",
   "issue_tracker": "https://github.com/xchillxx/surplus-load-switch/issues",
   "requirements": [],
-  "version": "2.38.0"
+  "version": "2.39.0"
 }