predheat_replay_report.md

Predheat replay and calibration (M1)

Generated 2026-10-07 by tools/predheat_replay.py. Window 2025-11-01 … 2026-03-31, 150 complete days (74 train / 76 test, split by ISO-week parity). Outdoor sensor to_ecowitt, measured (perfect forecast). Mean measured heating electricity on test days: 5.14 kWh/day.

Hold-out error

Set	Elec MAE kWh/day	Elec bias kWh/day	Room RMSE K (hourly)	Room MAE @6 h K	Room MAE @24 h K
current	0.98	+0.40	1.90	0.65	2.56
fitted	0.84	-0.11	1.16	0.82	1.16
calibrated	0.92	+0.49	0.75	0.32	0.79

Parameters

Parameter	current	fitted	calibrated
heat_loss_watts	156	94	94
heat_loss_degrees	0.0194	0.0315	0.02
heat_gain_static	400	440	440
heat_min_power	500	680	1500
flow_difference_target	5	5	1
heat_cop	3.68	3.81	3.81
hysteresis	1	1	1.5
hysteresis_off	1	1	0.1091
implied C (kWh/K)	8.0	3.0	4.7

Calibrated parameters at a search bound (structural mismatch, treat with caution): heat_loss_degrees, heat_min_power, flow_difference_target, hysteresis.

Recommended predheat: values (fitted set: lowest hold-out electricity error)

    heat_loss_watts: 94
    heat_loss_degrees: 0.0315
    heat_gain_static: 440
    heat_min_power: 680
    flow_difference_target: 5.0
    heat_cop: 3.81
    hysteresis: 1.00
    hysteresis_off: 1.00
    heat_pump_efficiency:
      -20: 1.93
      -18: 1.97
      -16: 2.02
      -14: 2.07
      -12: 2.11
      -10: 2.2
      -8: 2.3
      -6: 2.39
      -4: 2.48
      -2: 2.57
      0: 2.66
      2: 2.83
      4: 2.97
      6: 3.14
      8: 3.28
      10: 3.28
      12: 3.63
      14: 3.81
      16: 3.81
      18: 3.81
      20: 3.81

Limitations: Predheat does not model the Daikin's RT modulation, the overnight deviation schedule or DHW interruptions; the daily counters have 1 kWh resolution; hourly room temperatures are means.