CHANGELOG.md

Changelog

Each ## [x.y.z] section becomes the GitHub release notes for tag vx.y.z, created automatically by .github/workflows/ci.yml after tests pass on the default branch. Bump custom_components/daikin_mpc/manifest.json version and add a section here for every release. 0.2.0 was never tagged (superseded by 0.2.1 the same day).

[0.6.1] - 2026-10-11

Fixed

• HA warning "state class 'measurement' … impossible considering device class ('energy')" for Predicted heat pump electricity 24h: the forecast no longer has the energy device class (that class only allows meter totals); it keeps the measurement state class and its long-term statistics.

[0.6.0] - 2026-10-10

Added

• Measured heat gains. Heat that can be measured is counted separately from the learned "base gains":

  • inverter losses: new input, using sensor.solis_loss_power (PV + battery − AC output). It is counted 100 % as heat (winter balance 0.8 ± 0.4), about 100 W standby plus 5 % of throughput;
  • sun through the windows, from the weather-station radiation (1.5 W per W/m²);
  • hot-water cylinder standing loss (1.5 W/K × tank − room).

  Forecasts use hour-of-day profiles and Solcast. A new Heat gains now sensor shows the breakdown and whether every source is available. See docs/heat_gains.md.

• House electricity input, using sensor.house_power (AC-side house load from docs/ha/packages/energy.yaml). Heat pump, EV charger and loft IT are subtracted. It counts 0 % as heat by default: last winter's daily balance showed no measurable effect (0.03 ± 0.08), and 0.7 made the fit clearly worse. The share is an option and will be learned from clean data later.

Changed

• Priors depend on the configured gain sources:

  • with measured inverter losses: UA 90 W/K, base gains 70 W (winter 2025/26 daily balance);
  • otherwise unchanged at 94 W/K and 440 W.

  Existing installations: after mapping the new inputs, set Heat loss coefficient 90 and Base internal gains 70 under Configure, if you ever saved the options.

• Battery-power loss estimate (4.5 %) is now only a fallback without an inverter-loss sensor.

Fixed (review round 10, before release)

• Learned base gains, profiles and the night regression restart when the gain sources change. The day of the change is not learned from.
• An unavailable gain source makes the hour incomplete instead of adding 0 W. Days with incomplete gains are not learned from.
• An unavailable heat-pump, EV or outside-load meter no longer turns their consumption into household heat.
• Free-cooling capacity estimate corrected for measured gains.
• Measured gains no longer count towards the "no heating that day" check.
• Sun forecast only with a radiation sensor.
• Stale house/inverter readings (> 2 h) rejected.
• Stricter loading of saved profiles.

[0.5.0] - 2026-10-08

Added

• Thermal capacity learned from free-cooling nights. With the heat pump off, the night-time cooling rate measures the building's time constant directly; with UA from the daily learner this gives C. Clean night hours only (22:00–06:00, compressor off, no hot water, indoor–outdoor difference ≥ 6 K), ≈ 50-day memory; used once at least 4 nights identify it. The Thermal capacity sensor shows the source, the time constant and the number of nights. See docs/capacity_learning.md.

Changed

• Default thermal capacity 5.5 kWh/K (was 3.0): with heating off, the model cooled the house about twice as fast as it really cools (time constant 32 h vs ≈ 60 h in last winter's no-heating nights). If you ever saved the options, set "Thermal capacity" to 5.5 there (Daikin MPC → Configure).
• Changed defaults/options now apply while nothing has been learned yet (previously the first saved state kept the old values).
• M3 replay with the new capacity: optimised vs simple timer +28 % on the raw tariff (was +17 %).

[0.4.8] - 2026-10-07

Ninth review round (Codex) — attacks on the two accounting rules of 0.4.7: 3 defects, 4 test gaps; no wrongly rejected day found in normal operation (incl. both clock-change days). See docs/review_2026-10.md.

Fixed

• Saved state is a detached snapshot: HA writes it later in a worker thread while polling continues; the in-progress hour's totals were live references, so a poll in between could be counted twice after a restart.
• A meter reset is a discontinuity: the energy between the last reading and the reset is unknown, so that date is no longer used for learning (it was counted as complete with energy missing).
• Clock going backwards: samples are ignored until time passes the last one, and the current date is not used for learning (repeated hours had inflated a 24-hour energy balance).

Tests

• New tests: learning switched off changes nothing, RT-modulation bounds, exact thermostat thresholds, standby needs three off hours, plus regression tests for each fix. tools/mutation_check.py: 40 mutations, all caught.

[0.4.7] - 2026-10-07

Eighth review round (Codex): 4 code defects, 2 test gaps, and the finding that the restart/day completeness logic was still a set of patches. See docs/review_2026-10.md.

Changed — energy accounting now follows two explicit rules

• Restored all-or-nothing: meter baselines, the in-progress hour and the current date's hours are one accounting state. If any part of the saved state is invalid, none is restored; meters then start new baselines and the affected date is not used for learning (previously a rejected hour could leave a day with missing energy that was still learned from).
• A date counts only if every meter was observed continuously from its start to its end: a meter starting a new baseline (first reading, re-mapping, lost state) invalidates the date it falls in, unless exactly at midnight, and also the previous date when midnight passed since the last sample (re-mapping exactly at midnight previously left the ending day "complete" without its last 5 minutes).

Fixed

• A mapped heating switch that is unavailable no longer counts as "heating on": telemetry is reported incomplete and no heating forecast or recommendation is made until it is readable again.
• Standby power is learned once per day (it was updated twice on days without a COP sample).

Tests

• New tests: recommendation start state follows the thermostat demand signal; standby subtracted for the actual 23/25-hour day length; plus regression tests for each fix. tools/mutation_check.py: 32 mutations, all caught (one more retired as equivalent, with the reason in the script).

[0.4.6] - 2026-10-07

Seventh review round (Codex): 5 code defects, 4 test gaps, see docs/review_2026-10.md.

Fixed

• A day is not used for learning when any meter started counting after midnight (late first reading, meter re-mapped, lost state): previously one counter starting at 00:30, or a re-mapped meter at noon, still produced a "complete" day with missing energy.
• A restart shortly after midnight on a clock-change day no longer reports that day as 24 hours: the installation's time zone is saved and applied to restored times.
• A saved in-progress hour with inconsistent averages is rejected instead of breaking every update.
• Battery charging price: a cheap period cut short by the spring clock change (e.g. 01:30–02:30 keeps 02:00–02:30) is now found.

Tests

• New tests: heat-pump output cap, small meter decreases treated as noise, thermal-capacity learning, exact expiry of Predbat's last slot, plus regression tests for each fix. tools/mutation_check.py now checks 27 mutations; all are caught.

[0.4.5] - 2026-10-07

Sixth review round (Codex): 5 code defects and 12 test-suite gaps, see docs/review_2026-10.md.

Fixed

• After a restart, a first day that was incomplete (integration started mid-day) stayed eligible for learning; the "first day partial" state is now saved. A fresh start during the midnight hour (e.g. 00:30) no longer counts as a complete day: energy must have been observed from the hour's start.
• A restart late on a clock-change day no longer turns its 25/23-hour length into 24 hours (the day's time zone is taken from live data, not from restored timestamps).
• Battery charging price around clock changes: a cheap slot in the non-existent spring hour is no longer counted, and the repeated autumn hour is checked in both its BST and GMT occurrences.
• Restored hour records without temperature means are rejected instead of crashing the next day's aggregation.

Tests

• New tests close the gaps found by the review: stale recommendation cleared and re-planned on recovery, optimiser throttle and retry after failure, startup before entities exist, unit conversion end to end, comfort-period option validation, plan times and diagnostics, WebSocket redirect refusal, full engine state round-trip, clock-change days through the engine (incl. restarts), cost = energy × price, Predbat timestamp offsets, heating switched off through the engine, offline outdoor-temperature coverage.
• tools/mutation_check.py: re-introduces each of 20 fixed defects/gaps and confirms a test fails — all 20 are caught.

[0.4.4] - 2026-10-07

Fifth review round (Codex, 3 findings on 0.4.3), see docs/review_2026-10.md.

Fixed

• HA restarts no longer lose the current day's energy: the in-progress hour and the day's hours are now saved with the learned state, so a restart (e.g. for an update) neither drops energy nor lets a partial day into learning. Without saved state (first start mid-day) the first date is skipped.
• Battery charging price considers every rate change in the preceding 24 h, so short cheap slots (e.g. 15 minutes) are no longer missed.

Tools

• Offline days need valid indoor and outdoor temperatures for all but two hours; otherwise their means and temperature changes are not used.

[0.4.3] - 2026-10-07

Fourth review round (Codex, 4 findings on 0.4.2), see docs/review_2026-10.md.

Fixed

• Energy just before midnight was booked to the next day: each meter increment is now split across hours by the time it covers (e.g. 23:55–00:00 stays on the old day), so daily energy matches the daily temperatures used for learning.
• A day is also withheld from learning when a mapped meter (not only the bridge) is unavailable at midnight.
• Battery charging price: the cheapest rate in the 24 h before each time, using the rate that actually applies then (Predbat where it covers that time, the configured tariff otherwise), so an expired Predbat series no longer sets prices.

Tools

• Offline days must have all calendar hours (day length from local midnights); temperature changes are only taken between two complete consecutive days.

[0.4.2] - 2026-10-07

Third review round (Codex, 5 findings on 0.4.1), see docs/review_2026-10.md.

Fixed

• A bridge outage that spans midnight no longer lets the earlier day into learning before the outage is resolved: a day is closed only if the meter counters were readable at the boundary.
• Battery-aware price: the charging cost now comes from Predbat's rates when available (the fallback tariff's cheap rate was used even when Predbat's cheapest rate was higher).
• Re-mapping a meter entity no longer books the difference between two meters as energy: counter baselines are stored with their entity and dropped when the mapping changes. (After this update the baselines restart once; the current day is not used for learning.)
• Stored COP bins with heat but no electricity are rejected.

Tools

• Offline daily energy requires a valid counter reading at both day boundaries (series starting or ending mid-day no longer yield partial totals).

[0.4.1] - 2026-10-07

Second review round (Codex, 10 findings on 0.4.0), see docs/review_2026-10.md.

Fixed

• Optimiser could report "cannot be held" although a compliant plan existed: state merging kept only the cheapest path per temperature bin; it now also keeps the warmest and coolest. Replay: days below 20 °C 4 → 2 (the remaining two are physically unavoidable in the model). Runtime ≈ 0.7 s per plan.
• "Feasible" and the reported room range now both cover the planned period (not the given start value).
• Meter counters are only read while the bridge heartbeat is fresh, so a frozen bridge cannot hide an outage; a counter gap across midnight now invalidates both days.
• Stored COP and pending-day state fully validated and typed (e.g. negative standby or text values were accepted and could break later updates).
• Malformed weather responses no longer interrupt the update; Predheat comparison values are converted and range-checked like room temperatures.

Tools

• ha_client.py: WebSocket connections that were redirected to another origin are refused before the token is sent (requires websockets>=14).
• Offline daily energy keeps increments across same-day gaps (booked to the next reading) and drops days touched by a gap across midnight, instead of silently undercounting.
• tools/codex_review.md: run Codex with stdin closed (it otherwise waits for input).

[0.4.0] - 2026-10-07

Fixes from an adversarial code review (Codex), see docs/review_2026-10.md.

Fixed

• Daylight saving (clocks go back 25 Oct): all elapsed-time arithmetic now in UTC. Previously the repeated 01:00 hour merged into one, sample intervals went negative, forecasts and plan times shifted by an hour, and 23/25-hour days were treated as 24 h in the building learner and offline fit.
• Meter counters after a restart/outage: increments across a gap are no longer booked into the first hour after it; across midnight they are dropped and that day is not used for learning (counter timestamps are now persisted).
• Flow-temperature offset is learned only during space heating (DHW and defrost had shifted it by several K).
• Building learner only pairs consecutive days.
• Hard comfort limits are now strict: plans are ranked by time outside 20–22 °C first, then cost, so no price can buy a violation; "feasible" no longer has a hidden 0.05 K tolerance.
• Standby no longer counted twice in the electricity forecast: the COP now excludes standby (prior recomputed: 3.06 at 1 °C … 4.82 at 13 °C); standby is added once. Learned COP data is reset (new basis).
• Units: entities in Wh/MWh, kW, °F/K, L/h or m³/h are converted; unsupported units are reported instead of misread. Weather forecasts use the entity's temperature unit; NaN or implausible values are ignored.
• Weather points in the past no longer distort the outdoor-temperature forecast.
• Malformed Predbat rate attributes fall back to the configured tariff instead of failing the update.
• Corrupt stored state never prevents start-up (each section falls back to priors; covariance checked).
• Recommendations are cleared as soon as telemetry becomes invalid, and recomputed on the next poll instead of after 15 minutes (e.g. after HA restarts before the bridge is back); plan times refer to when the plan was made. Schedule text marks the next day ("18:00–18:00 next day").
• Optimiser validates input lengths (no crash on short price/forecast series).

Changed

• The bridge heartbeat entity is now required (P1P2MQTT publishes on change only; without it a frozen bridge cannot be detected). Map it via Reconfigure if not set.

Tools

• ha_client.py: exact read-only endpoint allowlist (the /api/ prefix admitted GET webhooks); redirects refused so the token is never forwarded.
• ha_export.py: naive dates are Europe/London (option --tz), not the machine's time zone.
• Offline fit: hourly windows never span gaps or DHW exclusions; counter increments from statistics sum (reset-adjusted); actual day lengths. Replays handle DST days; Predheat replay notes that fitted/calibrated hold-out errors are optimistic (physical parameters fitted on all days).

[0.3.1] - 2026-10-07

Added

• Preferred temperature periods (option, default 07:00-09:00=21, 18:00-24:00=21): the optimiser aims for these temperatures (soft target); outside them only the lowest allowed temperature (default 20 °C) applies. Hard limits 20–22 °C unchanged.

Changed

• Default lowest room temperature 20.0 °C (was 20.5 °C).
• Replay compares against a simple timer schedule and reports the shortfall against the preferred periods.

[0.3.0] - 2026-10-07

Added — M3 shadow optimiser (recommendations only, nothing is sent to the heat pump)

• Every 15 minutes a 24 h plan of hourly room setpoints (0.5 °C steps) that minimises electricity cost while keeping the room between 20.0 and 22.0 °C (configurable). No fixed night setback: lower night setpoints only when they save money.
• Prices from Predbat (predbat.rates, predbat.rates_export, including VPP events) with the configured tariff as fallback; round-trip efficiency from Predbat's loss settings.
• Battery- and export-aware pricing (default): energy outside the cheap window is valued at min(import, max(cheap rate / efficiency, export rate)), i.e. the opportunity cost of not exporting. Raw import tariff selectable.
• New sensors: Recommended room setpoint (with the hourly plan), Expected saving 24h (model estimate vs the current schedule), Recommendation (plain-language reason).
• Options: lowest/highest room temperature, price basis, fallback import/export tariffs.
• tools/m3_replay.py and docs/m3_replay_report.md: replay on winter 2025/26.

[0.2.7] - 2026-10-07

Fixed

• Predheat comparison ignores Predheat values older than 30 minutes, so a disabled or stopped Predheat (whose last states remain in HA) is no longer scored as if it were a current forecast.

[0.2.6] - 2026-10-07

Changed

• Plain-language names and help texts for every field in the setup and Reconfigure forms (e.g. Room thermostat asking for heat instead of Thermostat heating demand). No functional change.

[0.2.5] - 2026-10-07

Added

• Optional Thermostat heating demand role (suggested binary_sensor.bridge0_mode_valve_zone_main, verified in a live test): the forecast starts from the actual thermostat state instead of inferring it from the compressor (which lags demand by several minutes and cycles on water temperature). Existing installs: Reconfigure; the field is pre-filled.

[0.2.4] - 2026-10-07

Fixed

• Suggested "DHW active" entity is now binary_sensor.bridge0_dhw_dhw_demand (on only while the cylinder is being heated). binary_sensor.bridge0_dhw_dhw means "DHW enabled", is always on, and made every hour count as DHW. Existing installs: Reconfigure and change the DHW active field.

[0.2.3] - 2026-10-07

Fixed

• Setup/reconfigure no longer rejects the Predheat comparison entities when Predbat is not running (its predheat.* states only exist while Predbat runs); they are always pre-filled.

[0.2.2] - 2026-10-07

Added

• Reconfigure (Settings → Devices & services → Daikin MPC → ⋮ → Reconfigure): edit the entity mapping in place without losing learned state. Unmapped optional roles are pre-filled with verified defaults.

Fixed

• Predheat comparison entities are entered as text, because predheat.* states are not registered entities and the entity picker could drop them.

[0.2.1] - 2026-10-07

Added

• Optional external heat-pump meter roles (ext_w, ext_kwh) and the Daikin DHW electricity counter (dhw_elec_kwh), pre-filled in the setup form.
• Live COP-by-outdoor-temperature learner on the external-meter basis (includes standby and pump), monotone, with forgetting and fallback to the prior curve.
• Standby power learned from compressor-off hours; new sensor.daikin_mpc_standby_power.

Changed

• sensor.daikin_mpc_predicted_heat_pump_electricity_24h (renamed) now includes standby; attributes space_heating_kwh_24h, standby_kwh_24h, cop_source.
• COP prior recomputed from winter 2025/26 with the external meter (2.93 at 1 °C … 3.53 at 13 °C).

Upgrade note

Remove and re-add the integration so the new optional entities appear in the setup form.

[0.2.0] - 2026-10-07

Added

• First Home Assistant release (shadow mode, observation only, no writes to the heat pump).
• Config/options flow with verified, pre-filled entity mapping; 5-minute coordinator; bridge heartbeat.
• Hour/day aggregation, bounded daily learner for heat loss, gains and thermal capacity.
• 24 h room-temperature and heating-electricity forecast (thermostat gating, RT modulation, 25 °C LWT floor).
• Self-scoring of 1/3/6 h forecasts, including Predheat comparison; persistence; diagnostics.