mirror of
https://github.com/lynchaos/ashvale-station.git
synced 2026-09-12 12:47:49 +00:00
Reported from a real station after 1.5 days: a six hour temperature forecast of 53 C in a 24 C room, and 9 C at one day, both carrying a plus or minus of 0.43. Confidently wrong is the one failure this project is supposed to refuse. Root cause. fit() replayed history into the live RLS on every retrain tick and never reset, so 453 grid rows had produced 64,676 updates in a day and a half. RLS with forgetting reads every update as fresh evidence, so the model believed it had a hundred times the data it had: P collapsed, in-sample error looked excellent, and the weights drifted without bound in directions the data never excited. Measured: cond(P) 3.1e9 and ||theta|| 1680 against a median |theta| of 1.67. A refit now starts from the prior, which makes retraining idempotent. Across 25 refits on the real data ||theta|| holds at 11.35, drifting 0.03, where before it grew without limit. The two largest weights were sin_doy and cos_doy at +1174 and +1191. Annual harmonics were in the design matrix from the first sample, where they are near-constant, near-collinear with each other and with the bias, and a rank-deficient regressor is what RLS answers with enormous cancelling weights. They are now held at zero until the record spans the same 120 days the climatology fit already requires, because a day and a half of data says nothing whatsoever about the season. Also raised the standardiser's variance floor from 1e-8, which only caught a bit-exactly constant column, to 1e-3. A feature that merely barely moves was being divided by its own noise. The conformal calibrators and Hedge weights are deliberately not reset by a refit: those are earned from scored forecasts, not from this regression. Backtest unchanged within noise, coverage still 89 to 91 across all 18 heads. Four regression tests added, including that refitting the same history twice must give the same model.
175 lines
6.9 KiB
Python
175 lines
6.9 KiB
Python
#!/usr/bin/env python3
|
|
# Copyright 2026 Kemal Yaylali
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
"""Rolling-origin backtest. The only number that decides whether to ship.
|
|
|
|
Protocol, strictly walk-forward:
|
|
|
|
1. Build the 5-minute feature grid from stored telemetry.
|
|
2. Split at `--train-frac`. Fit the ensemble and the climatology on the
|
|
first part only.
|
|
3. Walk the second part one step at a time. At each step, forecast,
|
|
record the error, and only then let the model learn from the target
|
|
that has just matured. No target is ever visible before its time.
|
|
4. Report MAE against three baselines:
|
|
persistence the value now
|
|
climatology the harmonic fit
|
|
the ensemble
|
|
|
|
Skill = 1 - MAE_model / MAE_persistence. A positive number means the
|
|
model earns its electricity. A negative number at a given horizon is not
|
|
a failure of the exercise, it is the exercise working: ship persistence
|
|
at that horizon and stop pretending.
|
|
|
|
python scripts/evaluate.py --train-frac 0.6
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import argparse
|
|
import sys
|
|
import time
|
|
from pathlib import Path
|
|
|
|
import numpy as np
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
|
|
|
|
from ashvale.config import load_config # noqa: E402
|
|
from ashvale.features import build_features # noqa: E402
|
|
from ashvale.models.climatology import HarmonicClimatology # noqa: E402
|
|
from ashvale.models.nowcast import NowcastEnsemble # noqa: E402
|
|
from ashvale.storage import Store, resample # noqa: E402
|
|
|
|
|
|
def horizon_label(seconds: int) -> str:
|
|
if seconds < 3600:
|
|
return f"{seconds // 60}m"
|
|
if seconds < 86400:
|
|
return f"{seconds // 3600}h"
|
|
return f"{seconds // 86400}d"
|
|
|
|
|
|
def main() -> None:
|
|
ap = argparse.ArgumentParser(description=__doc__,
|
|
formatter_class=argparse.RawDescriptionHelpFormatter)
|
|
ap.add_argument("--train-frac", type=float, default=0.6)
|
|
ap.add_argument("--hours", type=float, default=24 * 60)
|
|
ap.add_argument("--config", default=None)
|
|
args = ap.parse_args()
|
|
|
|
cfg = load_config(args.config)
|
|
store = Store(cfg.storage.db_path)
|
|
|
|
raw = store.window(args.hours, ["ts", "temp_smooth", "hum_smooth", "press_slp", "lux"])
|
|
if raw["ts"].size < 200:
|
|
print("Not enough history. Run: python scripts/simulate.py --days 14")
|
|
return
|
|
|
|
grid_ts, cols = resample(
|
|
raw["ts"],
|
|
{"temperature": raw["temp_smooth"], "humidity": raw["hum_smooth"],
|
|
"pressure": raw["press_slp"], "lux": raw["lux"]},
|
|
cfg.model.grid_s,
|
|
)
|
|
X, valid = build_features(grid_ts, cols["temperature"], cols["humidity"],
|
|
cols["pressure"], cols["lux"], cfg.model.grid_s,
|
|
cfg.site.latitude, cfg.site.longitude,
|
|
cfg.model.climatology_min_days_annual)
|
|
|
|
n = grid_ts.size
|
|
split = int(n * args.train_frac)
|
|
span_days = (grid_ts[-1] - grid_ts[0]) / 86400.0
|
|
print(f"grid rows : {n} ({span_days:.2f} days at {cfg.model.grid_s}s)")
|
|
print(f"train / test : {split} / {n - split}")
|
|
|
|
clim = HarmonicClimatology(cfg.model.targets,
|
|
min_days_annual=cfg.model.climatology_min_days_annual)
|
|
clim.fit(grid_ts[:split], {k: v[:split] for k, v in cols.items() if k in cfg.model.targets},
|
|
valid[:split])
|
|
|
|
ens = NowcastEnsemble(cfg.model.targets, cfg.model.horizons_s, cfg.model)
|
|
t0 = time.time()
|
|
ens.fit(X[:split], valid[:split],
|
|
{k: v[:split] for k, v in cols.items() if k in cfg.model.targets},
|
|
clim, grid_ts[:split])
|
|
print(f"fit : {time.time() - t0:.1f}s\n")
|
|
|
|
per_step = cfg.model.grid_s
|
|
results = {}
|
|
|
|
for target in cfg.model.targets:
|
|
y = cols[target]
|
|
for h in cfg.model.horizons_s:
|
|
steps = max(int(round(h / per_step)), 1)
|
|
errs, pers, clims, covered = [], [], [], []
|
|
head = ens.heads[(target, h)]
|
|
|
|
for i in range(split, n - steps):
|
|
if not valid[i] or not np.isfinite(y[i]) or not np.isfinite(y[i + steps]):
|
|
continue
|
|
x = ens.scaler.transform(X[i:i + 1])[0]
|
|
anchor = float(y[i])
|
|
truth = float(y[i + steps])
|
|
cd = 0.0
|
|
if clim.ready:
|
|
cd = float(clim.predict(target, np.array([grid_ts[i] + h]))[0]
|
|
- clim.predict(target, np.array([grid_ts[i]]))[0])
|
|
pred = head.predict(x, anchor, cd)
|
|
errs.append(truth - pred["mu"])
|
|
pers.append(truth - anchor)
|
|
clims.append(truth - (anchor + cd))
|
|
covered.append(1.0 if pred["lo"] <= truth <= pred["hi"] else 0.0)
|
|
head.learn(x, anchor, truth, cd) # learn only after scoring
|
|
|
|
if len(errs) < 5:
|
|
continue
|
|
e = np.abs(errs)
|
|
p = np.abs(pers)
|
|
c = np.abs(clims)
|
|
results[(target, h)] = {
|
|
"mae": e.mean(), "persistence": p.mean(), "climatology": c.mean(),
|
|
"skill": 1.0 - e.mean() / max(p.mean(), 1e-9),
|
|
"bias": float(np.mean(errs)),
|
|
"coverage": float(np.mean(covered)),
|
|
"n": len(errs),
|
|
"weights": {k: round(float(v), 2) for k, v in
|
|
zip(("pers", "clim", "rls"), head.weights)},
|
|
}
|
|
|
|
units = {"temperature": "C", "humidity": "%", "pressure": "hPa"}
|
|
header = f"{'target':<12}{'lead':>6}{'MAE':>9}{'persist':>9}{'clim':>9}{'skill':>8}{'cover':>7}{'bias':>8} weights"
|
|
print(header)
|
|
print("-" * len(header))
|
|
for target in cfg.model.targets:
|
|
for h in cfg.model.horizons_s:
|
|
r = results.get((target, h))
|
|
if not r:
|
|
continue
|
|
flag = " <-- persistence wins" if r["skill"] < 0 else ""
|
|
print(f"{target:<12}{horizon_label(h):>6}{r['mae']:>9.3f}{r['persistence']:>9.3f}"
|
|
f"{r['climatology']:>9.3f}{r['skill'] * 100:>7.1f}%{r['coverage'] * 100:>6.0f}%"
|
|
f"{r['bias']:>+8.3f} {r['weights']}{flag}")
|
|
print()
|
|
|
|
# Driven off the dict rather than hardcoded, so adding a target cannot leave
|
|
# the units line silently describing the wrong columns.
|
|
print("units: " + ", ".join(f"{t} {units[t]}" for t in cfg.model.targets if t in units))
|
|
print("coverage should sit near 90% if the conformal calibration is honest.")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|