MAINT-28 fix round 5: while the pause holds, the batteries give only what the guard needs
checks / unit (pull_request) Failing after 38m55s
infra-pin / fresh (pull_request) Successful in 1s

H-1: one site limit while the pause holds (the period's largest floor, never under this cycle's), so the other pack no longer gives for the house. New live-try test (A 11 %, B 25 %, 90 min at 6.5 kW) red on 07d23d2. M-1: a 45-min rig-shape test that goes red when the settled check is forced True. M-2/L-2: DECISIONS. CHANGELOG line. test_safety84 n2_charge_banked re-pinned (quarters 3349 and 3950 W, aim 3950 W).

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
glenn schrooyen
2026-11-14 16:40:54 +01:00
co-authored by Claude
parent 9a76f1a968
commit de06a2c6f5
6 changed files with 93 additions and 14 deletions
+4
View File
@@ -1,5 +1,9 @@
# Changelog
### Not in a release yet
- During a real peak, a battery in its maintenance charge now stays paused instead of switching between charging and giving, and the batteries give only what keeps the peak under the limit ([[MAINT-28]]).
## 0.22.0
**A minor bump: a battery's monthly charge now gets the cheap night, the plan knows about the 7-day wait, a slow meter is followed more gently, and one peak-guard stop no longer costs a night charge its whole period. No option changes.**
@@ -14,12 +14,12 @@ Why 300 W: the house wanders +-50 W; once the charge resumes it settles at (aim
The flag lives only while the charge-only branch is taken: every cycle that reaches the maintenance loop starts it at False, so a maintenance that ended, a mode change or a missing aim cannot carry a stale True into the next charge (test_maint28.py part 1b). `cycle()` has two returns before that loop, and the flag is kept through them on purpose: "inputs missing" and "readings late" (SAFETY-128) are holds of under `stale_after` seconds that write nothing and decide nothing, so the next full cycle sees the same charge and the same peak. A missed poll must not end a peak pause; a reset there would restart the charge on it (`_maint_shape_suspended` is kept through them the same way).
It only ever keeps the maintenance charge off. No cut-back is delayed or weakened: the guard's claim is filed as before. While the pause is held the loop works as at 02eac3b, so A also gives for the house (quarters then sit under the aim, A's energy goes to the load: 1.67 kWh in the 45 min below, 0.45 kWh on 1b5f7c9). Measured alternative, A capped at 0 W while held and the guard idle (review scratch, same 45 min run): A gives 0.44 kWh in the load instead of 1.66, quarters 5919/5915 W, but A switches give/idle 32 times in 40 min. The uncapped give costs ~1.2 kWh round trip per 45-min peak (a few cents) and ~50 min more maintenance charge, far inside the 36 h timeout, so the steady give is kept.
It only ever keeps the maintenance charge off. No cut-back is delayed or weakened: the guard's claim is filed as before. While the pause is held, the pack in CHARGE gives only its share of the guard's floor (0 W if the guard holds it at its reserve), and the batteries together give only what the guard needs (see Fix-forward, F-2 and H-1). The first version let A give for the house; it was dropped.
## Numbers (sim, real `cycle()`, A rated 2500 W)
House 3000 -> 6500 W for 45 min -> 3000 W, +-50 W every 5 s, banked month peak 6000 W (aim 5950 W), forecast = quarter average so far plus the rest at the present grid (`scratch/MAINT-28/dev/m28_repro.py`; `test_maint28.py` runs the same shape for 15 min to keep the suite short).
| 45 min load | base 02eac3b | 1b5f7c9 | MAINT-28 |
| 45 min load | base 02eac3b | 1b5f7c9 | MAINT-28 (f3f9483, without the lag) |
|---|---|---|---|
| A charge/give reversals | 0 | 52 | 1 (the first trip) |
| A charging seconds in the load (of 2700) | 0 | 367 | 301 (before the first trip) |
@@ -40,14 +40,16 @@ The sim had no battery-reading lag, and the first version said the 300 W line is
**F-1: the pause ends only on a settled estimate.** `Controller._batts_settled()` is False from a step UP of a unit's command (`MAINT_STEP_W` = 100 W more give or less charge, the SAFETY-128 pending size) until its reading changed `PLANT_SETTLE_S` (6 s) after the step, or `BATT_OBLIGED_S` (17 s) have passed. `peakguard.shaping_held(..., settled)` then keeps the pause. A step down makes the estimate err high, which only holds the pause, so it waits for nothing. Measured in the 45 min sim, pause on: settled, the estimate was at most 146 W under the real house (0 cycles over 300 W); not settled, up to 1933 W under (725 of 1611 cycles over 300 W). Sim, 45 min, house 6500 W: 83 of 83 releases on a stale estimate (f3f9483) -> 0; flips 148 -> 1 (1b5f7c9 210). Charge is back 4 s after the load (12 s in test_maint28b), not 0: the first settled estimate. No cut-back is touched.
**F-2 and the cap.** Owner 2026-11-13 ("1" = cap): while the pause holds, the pack in maintenance CHARGE gives only what the guard claims for the peak, at any SoC; other packs unchanged. Owner 2026-11-13 ("ok max", relayed by the manager): accepted "while the hold is on, the pack in maintenance CHARGE gives at most the largest floor the peak guard claimed in the current quarter" (about 0.37 kWh per 45-min peak in the sim; quarters at or under the aim).
**F-2 and the cap.** Owner 2026-11-13 ("1" = cap): while the pause holds, the pack in maintenance CHARGE gives only what the guard claims for the peak, at any SoC; other packs unchanged (replaced by H-1 below). Owner 2026-11-13 ("ok max", relayed by the manager): accepted "while the hold is on, the pack in maintenance CHARGE gives at most the largest floor the peak guard claimed in the current quarter" (about 0.37 kWh per 45-min peak in the sim; quarters at or under the aim).
**H-1 (review 2026-11-14, owner decision 2026-11-14): the batteries together give only what the guard needs.** With the cap on A alone, the loop gave for the whole house with B (Cheap hours): in the live-try shape (A 11 % in CHARGE, B 25 %, house 6.5 kW for 90 min) B gave 4.6 kW, went from 25 to 10 % in 43 min, and two quarters were 6502 and 6499 W against the month peak 6000 W. So while the pause holds, `_m28_share()` also files one site limit (source `charge`, `P_MAINT_SHAPE`, per-site `_m28_site`): `hi` = the largest floor of the period, never under this cycle's floor. The guard's claim is not cut: `hi` is at least the floor, `P_PEAK` outranks `P_MAINT_SHAPE`, and `_peak_on_able` still places the whole floor. Outside the hold nothing is filed. Live try, same shape (test_maint28b): quarters 5357 to 5911 W (aim 5950), B 0.86 kWh and ends at 20.6 % (1b5f7c9: 1.0 kWh, 19.8 %), B switches give/idle once, A never. The re-pinned `n2_charge_banked` quarters are 3349 and 3950 W (aim 3950).
The rule as built (`Controller._m28_share()`): while the pause holds (N>=2, month aim above the cap) the pack gets a unit limit `hi` = the largest part of the guard's floor placed on it in this 15-minute period. Its part in a cycle is the floor less what the other units are placed at in this cycle's split without that limit, so a part the guard's split puts on the other pack is not this pack's. A pack the guard holds at its reserve (`peak_guard.held`) gets 0 W and its part starts again from 0 W. The largest part is kept per unit (`_m28_hi`, keyed by the unit) and dropped in a new period or when the pause ends; only one pack can be in maintenance CHARGE anyway (`maintenance.arbitrate()`, one slot per site).
Why the largest of the period and not the present part: the guard alone lands the quarter 9.5-10 W over its aim in the sim (SAFETY-126: its own projection ends on 5950.0 W but the meter reading it counts lags the true grid), and with the present part only, test_maint28 part 1 ends at 5960.1 W; the largest of the period lets the loop keep A up and it ends at 5925 W. Why not `hi` = 0 W: with the other pack under its reserve the guard needs this one (test_maint28b, house 12 kW, B at 18 %: A gives 1.07 kWh for the guard).
Tests (test_maint28b, 13-min load, A 5 kW): A at 49 %: SoC drop 0.64 point (the owner's 0.37 kWh per 45 min is 0.70 point here; the check is <= 0.70). Both packs able, B on Cheap hours: B is placed nothing, A's limit is the period's largest floor (3087 W) and A gives 3081 W. B on Solar first (gives for the house): A's limit 3294 W against a period's largest floor of 6284 W, under it by 500 W or more in 564 of 569 paused cycles; A still gives 3294 W, so the quarter ends 3843 W, far under the aim: that is the cost of the largest-of-period rule when the other pack gives for the house.
Tests (test_maint28b, 13-min load, A 5 kW): A at 49 %: SoC drop 0.64 point (the owner's 0.37 kWh per 45 min is 0.70 point here; the check is <= 0.70). Both packs able, B on Cheap hours: B is placed nothing, A's limit is the period's largest floor (3087 W) and A gives 3081 W. B on Solar first (it gives for the house): A's limit stays under the period's largest floor by 500 W or more in 740 of 744 paused cycles, and the quarters stay under the aim.
Sim, rig shape, 45 min, house 6500 W: A in CHARGE at 49 %: A gives 0.17 kWh (f3f9483 0.44), SoC 50.5 -> 49.0 %, charge/give flips 1, give/idle switches 3, quarters 5358 1663 1899 2572 4476 W, charge back 4 s after the load; B (Cheap hours, not in maintenance) gives 3.04 kWh for the house (f3f9483 0.32). B at 18.2 % in CHARGE: f3f9483 gave 1.89 kWh to 10.0 %; now 0.00 kWh, 18.2 %, quarters 5948 and 5952 W (the guard alone, on A). test_safety84 n2_charge_banked: banked quarters 3349 and 3217 W against the aim 3950 W.
Sim, rig shape, 45 min, house 6500 W (load from minute 10), A in CHARGE at 49 %, B 30 % on Cheap hours: A gives 0.17 kWh (f3f9483 0.44), SoC 50.5 -> 49.0 %, charge/give flips 1, give/idle switches 1, quarters 5358 5897 5911 5176 4477 W, charge back 4 s after the load; B gives 0.23 kWh (before H-1 3.04, f3f9483 0.32). test_safety84 n2_charge_banked: banked quarters 3349 and 3950 W against the aim 3950 W.
Mutants, each red: the freshness check dropped (always settled); the cap dropped; the other units' placement dropped (part = the whole floor); one cap for the site instead of per unit; the 0 W for a held pack dropped (all in `test_maint28b.py`); the period's largest dropped (present part only: `test_maint28.py` part 1 quarter 5960 W).
Mutants, each red: the freshness check dropped (always settled); the cap dropped; the other units' placement dropped (part = the whole floor); one cap for the site instead of per unit; the 0 W for a held pack dropped (all in `test_maint28b.py`); the period's largest dropped (present part only: `test_maint28.py` part 1 quarter 5960 W); `_batts_settled()` replaced by True at the hold in ctl_cycle.py (the 45-min rig shape in `test_maint28b.py`, one stale release); the site limit dropped (the live try in `test_maint28b.py`: quarters 6502 and 6499 W, B 2.91 kWh to 10.0 %).
+16 -1
View File
@@ -747,7 +747,9 @@ class TradeMixin:
reserve) set to the largest part of the guard's floor left to it in
this 15-minute period: the floor less what the other units are placed
at in this cycle's split without the cap. Kept per unit in `_m28_hi`,
dropped while the unit is not paused or is held.
dropped while the unit is not paused or is held. Also one site limit,
the period's largest floor, so the other pack does not give for the
house (`_m28_site`).
DECISIONS/MAINT-28-hold-the-pause-in-a-peak.md"""
floor = max((c.lo for c in claims if c.source == peakguard.SOURCE),
default=0.0)
@@ -755,6 +757,8 @@ class TradeMixin:
held = {self.actuator_of("maintenance", inv) for inv in self.inverters
if g is not None and inv.name in g.held}
q, seen = int(now.timestamp() // 900), set()
site = any(c.source == "maintenance" and c.reason == M28_REASON
and c.actuator != ACTUATOR_SITE for c in claims)
for i, c in enumerate(claims):
if c.source != "maintenance" or c.reason != M28_REASON \
or c.actuator == ACTUATOR_SITE or c.actuator in held:
@@ -771,6 +775,17 @@ class TradeMixin:
# A unit not paused (or held at its reserve) starts again from 0 W.
for u in set(self._m28_hi) - seen:
del self._m28_hi[u]
# Site: while the pause holds, the packs together give at most the
# period's largest floor, never under this cycle's (the guard's claim
# and P_PEAK outrank this one anyway).
if not site:
self._m28_site = (None, 0.0)
return
hi = max(floor, self._m28_site[1] if self._m28_site[0] == q else 0.0)
self._m28_site = (q, hi)
claims.append(Claim.limit("charge", P_MAINT_SHAPE, hi=hi,
reason=M28_REASON,
actuator=ACTUATOR_SITE))
def _m28_was(self, u, now) -> float:
"""MAINT-28: unit `u`'s cap so far in this 15-minute period, 0 W in
+1
View File
@@ -508,6 +508,7 @@ class Controller(InputsMixin, WritesMixin, TradeMixin, CycleMixin, AlarmsMixin,
self._maint_peak_hold = False # MAINT-28
self._m28_step = {} # MAINT-28: unit -> (command, step at)
self._m28_hi = {} # MAINT-28: unit -> (period, cap)
self._m28_site = (None, 0.0) # MAINT-28: (period, site cap)
# OPT-11: battery modes, off by default (SAFETY-10 section 1). Each
# battery's mode and on/off switch live in the store, not in options,
# so a change applies at once (`pack_use`, `pack_mode`).
+58 -3
View File
@@ -28,6 +28,9 @@ peak forecast like the live sensor, house +-50 W every 5 s, 3000 W then
4. Unchanged: the MAINT-24 live shape (A 1475 W, B no grid trade), with the
same lag on.
5. `peakguard.shaping_held()` with `settled`, and `Controller._batts_settled()`.
6. H-1 (owner 2026-11-14): while the pause holds the batteries together give
only what the guard needs. The live try (A 11 %, B 25 %, 90 min at 6.5 kW)
and, for the F-1 wiring, the 45-min rig shape.
Mutations: DECISIONS/MAINT-28-hold-the-pause-in-a-peak.md.
Read MUTATION.md before you report a survivor.
"""
@@ -69,11 +72,11 @@ PRE, LOAD, POST = 120.0, 780.0, 120.0 # s (CI time: short runs)
RESERVE = 20.0 # peak_soc_floor default
def stepped_house(base_w, hi_w):
def stepped_house(base_w, hi_w, load=LOAD, pre=PRE):
rnd = random.Random(28)
out = [(sc.UTC0.timestamp() - 10, base_w)]
for t in range(0, int(PRE + LOAD + POST), 5):
w = hi_w if PRE <= t < PRE + LOAD else base_w
for t in range(0, int(pre + load + POST), 5):
w = hi_w if pre <= t < pre + load else base_w
out.append((T0.timestamp() + t, w + rnd.uniform(-50.0, 50.0)))
return out
@@ -206,6 +209,27 @@ check("after the load A's charge is back within 20 cycles (1 s each: a 10 s "
"register period plus its settle time, 16 s)",
back is not None and back <= 20)
# M-1 (review 2026-11-14): the F-1 wiring, `_batts_settled()` read at the hold
# in ctl_cycle.py. With `settled=True` there the 13-min load above still passes;
# the rig shape (the load starts 10 min into the quarter, 45 min long) does not.
RIG_PRE, RIG_LOAD = 600.0, 2700.0
rows, st, c = scen(house=stepped_house(3000.0, 6500.0, RIG_LOAD, RIG_PRE),
dur=RIG_PRE + RIG_LOAD + POST)
load45 = [r for r in rows if RIG_PRE <= r["t"] < RIG_PRE + RIG_LOAD]
rel45 = [(r, s_) for p, r in zip(rows, rows[1:])
if RIG_PRE <= r["t"] < RIG_PRE + RIG_LOAD and p["hold"]
and not r["hold"]
for s_ in [r["est"] is not None and r["est"] < LINE
and r["house"] >= LINE]]
print(f" A 49 %, 45 min: flips {flips(load45)}, releases {len(rel45)} "
f"(stale {sum(1 for _r, s_ in rel45 if s_)}), quarters "
f"{[round(a[0] / a[1]) for a in st['quarters'].values()]}")
check("45 min: A flips charge/give 3 times or fewer (f3f9483: 148)",
flips(load45) <= 3)
check("45 min: no release of the pause on a stale estimate (f3f9483: 83 of 83)",
not any(s_ for _r, s_ in rel45))
check("45 min: every quarter at or under the month aim", quarters_ok(st))
# ============================================================================ #
print("\n2. the cap: while the pause holds, A gives at most the largest part "
"of the guard's floor placed on it this period")
@@ -292,6 +316,37 @@ check("the guard places its claim on the other pack: A under its reserve "
sum(r["u"] for r in rows if r["u"] > 0) / 3.6e6 < 0.05
and min(r["soc"] for r in rows) >= 18.0)
# H-1 (review 2026-11-14, owner decision): while the pause holds, the packs
# together give only what the guard needs. The live try: A 11 % in CHARGE (under
# its reserve), B 25 % on Cheap hours, 90 min at 6.5 kW.
LIVE = 5400.0
rp.CAP_KWH = 20.48 # the rig's B; A gives nothing here
rows, st, c = scen(soc=11.0, other=25.0,
house=stepped_house(3000.0, 6500.0, LIVE, RIG_PRE),
dur=RIG_PRE + LIVE + POST)
rp.CAP_KWH = 15.36
qs = [a[0] / a[1] for a in st["quarters"].values()]
load = [r for r in rows if RIG_PRE <= r["t"] < RIG_PRE + LIVE]
gb = sum(r["o"] for r in load if r["o"] > 0) / 3.6e6
def switches(k):
g = [("give" if r[k] > 50 else "idle") for r in load]
return sum(1 for x, y in zip(g, g[1:]) if x != y)
print(f" live try: quarters {[round(q) for q in qs]}, B gives {gb:.2f} kWh, "
f"B ends {rows[-1]['soc_o']:.1f} %, give/idle switches A "
f"{switches('u')}, B {switches('o')}")
check("live try: every quarter at or under the aim + 10 W (07d23d2: 6502 and "
"6499 W against the month peak 6000 W)", max(qs) <= AIM + 10.0)
check("live try: B gives about its guard part, under 1.5 kWh (1b5f7c9 1.0; "
"07d23d2 4.6 kW, 25 -> 10 % in 43 min)", gb < 1.5)
check("live try: B ends at 19.8 % or more (1b5f7c9 19.8)",
rows[-1]["soc_o"] >= 19.8)
check("live try: give/idle switches of B at most 40 in the 90 min",
switches("o") <= 40)
# ============================================================================ #
print("\n3. mirror: B in CHARGE at 18 % (under its reserve), A on Cheap hours")
rows, st, c = scen(unit="b", soc=18.2, other=30.0,
+6 -4
View File
@@ -487,9 +487,11 @@ print("\n8. the N=1 and N=2 digests")
# (DECISIONS/MAINT-28-hold-the-pause-in-a-peak.md). MAINT-28 fix-forward
# re-pinned it again: while the pause holds A gives at most the largest part
# of the guard's floor placed on it this period, so rows 288-479 (t = 1440 s
# on, 192 of 480) differ: a "peak hold" limit claim on A and A's share. Its
# banked quarters stay under the aim 3950 W (3349 and 3217 W; f3f9483 3349 and
# 2527 W), graded below. The other five are equal.
# on, 192 of 480) differ: a "peak hold" limit claim on A and A's share. Round 5
# (H-1) re-pinned it again: one site limit, the period's largest floor, so the
# batteries give only what the guard needs. Its banked quarters stay at or
# under the aim 3950 W (3349 and 3950 W; f3f9483 3349 and 2527 W), graded
# below. The other five are equal.
PINNED = {
"n1_charge":
"00f0a16b3ac1bd2830e9e0aaac8923fd9d1736c2bc6c543669bdd35b2fb542e9",
@@ -502,7 +504,7 @@ PINNED = {
"n2_charge_dear":
"aa2bbc677a8e2274b25ba1170de1b1635e08a3bd93105e2a38a9bb1575b5b6db",
"n2_charge_banked":
"2da0c3bc9f0ff3857402146efcd6119180a7d4d1cc1297f1c15ec0b8c0104f42",
"358ebd2a23fe7c7f52f441b60391e40c6605e32bcfb109fa0ff4502d4070745e",
}
_nb_rows, _nb = s84.run(**s84.SCENARIOS["n2_charge_banked"])
print(f" n2_charge_banked banked quarters {[round(v, 1) for v in banked(_nb)]}"