The base layer (loss, heat, bias) showed values only, so a zone with weeks of
samples looked equally trustworthy whether or not the data ever varied its
inputs. Each base coefficient now reports a confidence from its own posterior
variance, as the extended and ventilation layers already did: it stays at 0
for `heat` while the relay never cycles and below 0.8 for `loss` and `bias`
while the indoor-outdoor difference is steady. Readiness is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Confirmed: learned models were matched only by zone key, so replacing a
zone's sensors, changing sensor weights, moving or replacing a heat
source or turning a window kept coefficients learned for a different
configuration, silently.
compatibility.py fingerprints exactly what each learned component
depends on: the zone measurement (sensors and relative weights; equal
scaling changes nothing), each zone's heat sources, the neighbour's
measurement, the zone's windows, window orientation (direction term
only), the recovery time (recovery only) and other zones' openings
(cross term only). Comfort weight, target offset and names are not
included. Heat-source baselines have their own fingerprint (sensor and
included sources).
On restore, a changed measurement recalibrates the zone: base and
extended loss/heat/bias keep their values as a warm start, but their
uncertainty, counts and readiness start over, other terms return to the
prior, and the zone's forecast statistics are dropped. Otherwise only
the changed terms are reset; everything else, including other zones,
is kept. Each recalibration is logged (Status model_changes, info log)
and saved at once. Data from before fingerprints (0.4.0) is assumed to
match its current configuration and is fingerprinted at the first
start, so the next edit is detected even before a learned sample.
Tests cover comfort changes, sensor replacement and weights, window
orientation and size, moved and replaced heat sources, re-added zones,
0.4.0 data and a second restart, plus an HA Core test through real
subentry updates.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Electrical equipment heats its room, and a gaming session can add
hundreds of watts, which the 3-parameter model attributed to the boiler
or the bias term.
heat_sources.py adds heat sources as config subentries ("Add heat
source"): an active-power sensor (W/kW; VA and % rejected), the zone,
the share that becomes heat, an automatic baseline (20th percentile of
24 h, so the persistent UPS load is not double counted with the bias
term) or a manual one, an optional "included in" parent (the UPS total
minus a separately measured PC, one level deep) and a first-order
thermal response time. The heat during each interval is the power held
from its start, like the radiator lag. After startup or an outage a
source's response must run two time constants before it counts as
settled.
zone_model.py layers each zone: the base model is kept unchanged as the
fallback and comparison, and an extended FeatureLearner adds "internal"
(kW above baseline after the response) and neighbour coupling for
adjacent zones. It learns only intervals where all its inputs are
known, and is used for forecasts once as trained as the base. The
optimizer accepts a multi-zone step model; forecasts let extra power
fade to baseline with a configurable half-life (default 30 min).
On a synthetic office with thermostat cycling and gaming sessions the
extended model recovers the electrical, boiler and loss coefficients
within 10 %. Status lists each source's power, effective power,
baseline, variable and responding heat; models show the extended
coefficients with confidence. Saved models gain optional zone_models
and heat_sources keys.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each temperature sensor used to be its own "room", so a room with two
sensors counted twice and comfort importance was tied to sensors.
zones.py adds Zone: a name, one or more sensors with sensor weights, a
comfort weight, an optional target offset and adjacent zones (made
symmetric). Sensor weights only combine readings; comfort weights only
steer the MPC; neither touches the learned thermal coefficients.
Zones are config subentries (HA's native "Add zone" / edit / delete on
the integration page), with validation that a sensor belongs to one
zone and is not an outdoor source, and a sensor-weight step for
multi-sensor zones. Initial setup still asks for room sensors and
weights and creates one zone per sensor. Configure no longer asks for
rooms or weights.
Config entry version 3. Migrating from 2 turns each room sensor into a
one-sensor zone with its comfort weight and the sensor's entity ID as
the zone key, so learned models, forecast entities (now linked to their
zone, so deleting a zone removes them) and accuracy statistics carry
over. Older versions cannot load a version 3 entry; rolling back needs
the HA backup.
A zone learns only across polls that used the same sensors (a sensor
dropping out shifts the mean by its offset); the new poll outcome
"sensors_changed" reports this. Status renames: input_quality.zones
(per-sensor detail nested), excluded_zones, retired_zone_models; adds
zone_names. Invalid stored comfort weights now fall back to 1.0, so the
invalid_weights status is gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Electrical heat, ventilation, wind and inter-zone exchange each need
their own coefficient next to the existing loss/heat/bias, and each must
stay physically bounded and not claim precision it does not have.
learning.py adds FeatureLearner: exponentially weighted RLS over named
features with per-feature bounds, priors at 0 for new effects, the same
covariance windup bound as the base model, and a linked bound that keeps
the directional wind term from cancelling more than the isotropic one.
Each prior is sized as "worth N typical samples" (20 for new effects,
200 for zone coupling, which is weakly identifiable). Confidence is the
share of a coefficient's information that comes from data; the standard
error uses a bias-corrected running noise estimate. Saved models are
matched by feature name, so adding a heat source, window or neighbour
retrains only that coefficient and keeps the covariance of the rest.
The existing 3-parameter learner now uses the same RLS step, covariance
bound and positive-definiteness check (behavior unchanged; all existing
tests pass).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was no evidence of how accurate the thermal model is: the only
number was an in-sample 5-minute error.
At every poll, record each room's reading, the outdoor temperature, the
radiator-lag estimate and the model parameters. 15, 30, 60 and 120
minutes later, replay the room from that poll with the parameters it
had then (out of sample), the outdoor temperature held constant and the
radiator lag that actually followed from the relay's real history, and
compare with the reading now. A no-change forecast is scored alongside.
Only spans with regular polls and a well-known lag are scored. 7 days
of per-room statistics are saved with the model (optional "evaluation"
payload key; about 18 KB at most).
Four new sensors, Heating MPC Forecast Error 15/30/60/120 min, show the
mean absolute error with bias, persistence error, skill and per-room
detail. On the synthetic lagged plant the 60-minute error is 0.03 °C
against 0.39 °C for persistence.
Room forecast entities gain forecast_15min and plan_15min attributes
and are labelled counterfactual. Volatile attributes are unrecorded.
The model is now saved after scoring, so saved statistics are current.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Status showed accepted sample counts but not why learning was slow.
Each room's model now counts samples it rejected (outliers, implausible
inputs), persisted as an optional rejected_count key and shown as
models.<room>.rejected_samples. Status also counts every poll outcome
since start (sample_results_since_start).
Add tests that a model in the exact shape 0.2.1 wrote restores with its
coefficients, and that today's saved model still has the keys and types
0.2.1's restore code reads, so rolling back keeps the learned model.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Polling only sees the relay's current state and last change, so two or
more changes between polls (e.g. an on/off round trip) made the lag
estimate unknown and paused learning for about half an hour.
Subscribe to the relay's state changes (read-only) and integrate every
recorded change between polls. Because HA delivers these events one
loop iteration late, the current state stays authoritative and fills in
a change not yet recorded. Unknown periods (non on/off states, missed
transitions, event overflow) still restart the estimate with an
explicit error bound. Without the listener, behavior is as before.
HA skips unload callbacks after unexpected setup errors, so setup now
removes the listener itself on any failure and registers it for unload
only once setup succeeds.
Tested on HA Core 2025.12.0 and 2026.2.3.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The lag proxy was reset to the new relay command at every relay change
and after any skipped interval, so it read 1.0 straight after ON and 0.0
straight after OFF. This disabled the lag model around every boiler
cycle: on a synthetic lagged plant cycled by a hysteresis thermostat the
learner's heat gain came out 34% low (0.079 vs 0.12) and the loss term
43% low, and the optimizer started post-transition plans from the wrong
lag.
Track the lag at every poll with a valid relay state, independently of
room sensors. Integrate a single change between polls at its
last_changed time. For ambiguous cases (round trip between polls, relay
outage, restart, clock jump) restart from time-in-state with an explicit
error half-width, and learn only once it is at most 0.1 (about 32 min
after an HA restart). The same synthetic test now learns 0.1195/0.00389
(true 0.12/0.004).
Expose heat_lag, heat_lag_uncertainty and sample_result on Status so
users can see why a poll was or was not learned. The HA test fixture
relay now has an hour-old timestamp; a new test covers a freshly
written relay state.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With forgetting 0.999 and no heat excitation (summer, long OFF periods),
the heat-coefficient variance grew by 1/0.999 per sample without limit:
about 8.7e12 after 104 simulated OFF days. The first ON samples could
then swing coefficients to their bounds, and accumulated asymmetry
approached the absolute 1e-8 restore tolerance, past which a restart
silently reset the whole model.
Symmetrize P and bound its diagonal by the prior after every update
(row/column scaling keeps it positive definite). Restore uses a relative
round-off tolerance and applies the same bound, so wound-up v0.2.1
models load instead of being discarded. Payload schema is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
buzeraandClaude Opus 5.5