4306f76ddb
- c6111_extract: roles_for() 동적 생성, COLUMN_EXCEPTIONS per-prefix - c6111_prodmap/shadow/startup/rolling: --data/--prefix CLI 인자 지원 - run_column.py: 5개 컬럼 전 파이프라인 실행 래퍼 - c6111_shutdown.py: detect_cutoffs + shutdown_milestones (lookback 1200) - c6111_operator_assist.py: OOD 게이트 + shadow 리플레이 - c6111_export_model.py: 선형근사 JSON export - SteamAdvisor.cs: Predict+ClassifyMode+InEnvelope (NaN guard, Ood fix) - SteamAdvisorController: GET/POST /api/steam/predict - appsettings.json/Program.cs: DI 등록 - docs: 작업지시서 현황 갱신, 진단보고서 작성 (3 MED/8 LOW, 100% 정확도)
86 lines
3.8 KiB
Python
86 lines
3.8 KiB
Python
"""
|
|
롤링(walk-forward) 재학습 — OOD/외삽 바이어스 해소 데모.
|
|
|
|
형제 컬럼 호환: --data, --prefix CLI 인자.
|
|
"""
|
|
import argparse
|
|
import numpy as np
|
|
import pandas as pd
|
|
import matplotlib
|
|
matplotlib.use("Agg")
|
|
import matplotlib.pyplot as plt
|
|
from sklearn.metrics import mean_absolute_error
|
|
from c6111_shadow import SteamPredictor, FEATURES, BASE, SMOOTH
|
|
|
|
HELDOUT_START = "2026-05-01"
|
|
RETRAIN_EVERY = "1D"
|
|
|
|
|
|
def main():
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--data", default=BASE + "c6111_data.pkl")
|
|
parser.add_argument("--prefix", default="c6111")
|
|
args = parser.parse_args()
|
|
df = pd.read_pickle(args.data)
|
|
df = df[df["mode"] == "PROD"].copy()
|
|
df = df[(df["feed"] > 50) & (df["steam_flow"] > 10) & (df["steam_op"] > 1)
|
|
& df[FEATURES + ["steam_op"]].notna().all(axis=1)].sort_values("dtat")
|
|
# 인과(trailing) 평활 — 미래누설 없음
|
|
for c in FEATURES:
|
|
df[c + "_s"] = df[c].rolling(SMOOTH, min_periods=1).median()
|
|
|
|
ho = pd.Timestamp(HELDOUT_START)
|
|
if df["dtat"].max() < ho:
|
|
print(f"데이터 종료 {df.dtat.max()} < HELDOUT_START({ho}) — 롤링 재학습 불가. (컬럼 가동기간이 5월 이전)")
|
|
return
|
|
|
|
days = pd.date_range(ho, df["dtat"].max(), freq=RETRAIN_EVERY)
|
|
|
|
# 정적 모델: 5월 이전 전체로 1회 학습
|
|
static = SteamPredictor().fit(df[df["dtat"] < ho])
|
|
slo, shi = (df[df["dtat"] < ho][FEATURES].quantile(0.01),
|
|
df[df["dtat"] < ho][FEATURES].quantile(0.99))
|
|
|
|
rows = []
|
|
for d0, d1 in zip(days[:-1], days[1:]):
|
|
day = df[(df["dtat"] >= d0) & (df["dtat"] < d1)]
|
|
if len(day) < 30:
|
|
continue
|
|
train = df[df["dtat"] < d0] # expanding: 그 날 이전 전체
|
|
roll = SteamPredictor().fit(train)
|
|
lo, hi = train[FEATURES].quantile(0.01), train[FEATURES].quantile(0.99)
|
|
Xs = day[[c + "_s" for c in FEATURES]].values
|
|
ao = day["steam_op"].values
|
|
po_r = roll.flow_to_op(roll.predict_flow(Xs))
|
|
po_s = static.flow_to_op(static.predict_flow(Xs))
|
|
ood_r = (~((day[FEATURES] >= lo) & (day[FEATURES] <= hi)).all(axis=1)).mean()
|
|
rows.append(dict(day=d0,
|
|
mae_roll=mean_absolute_error(ao, po_r),
|
|
mae_static=mean_absolute_error(ao, po_s),
|
|
w2_roll=np.mean(np.abs(po_r - ao) <= 2) * 100,
|
|
w2_static=np.mean(np.abs(po_s - ao) <= 2) * 100,
|
|
ood_roll=ood_r * 100))
|
|
r = pd.DataFrame(rows)
|
|
|
|
print(f"=== 5월 held-out, 일별 walk-forward 재학습 ({len(r)}일) ===")
|
|
print(f"정적 모델 : OP MAE {r.mae_static.mean():.2f}% |Δ|≤2% {r.w2_static.mean():.1f}%")
|
|
print(f"롤링 모델 : OP MAE {r.mae_roll.mean():.2f}% |Δ|≤2% {r.w2_roll.mean():.1f}%")
|
|
print(f"롤링 OOD 비율: 첫주 {r.head(7).ood_roll.mean():.0f}% → 마지막주 {r.tail(7).ood_roll.mean():.0f}%")
|
|
print("\n일별(요약):")
|
|
print(r[["day", "mae_static", "mae_roll", "w2_roll", "ood_roll"]]
|
|
.assign(day=r.day.dt.strftime("%m-%d")).round(1).to_string(index=False))
|
|
|
|
fig, ax = plt.subplots(2, 1, figsize=(14, 8), sharex=True)
|
|
ax[0].plot(r.day, r.mae_static, "r.-", label="static (Feb-Apr model)")
|
|
ax[0].plot(r.day, r.mae_roll, "g.-", label="rolling retrain")
|
|
ax[0].axhline(2, color="gray", ls=":", label="2% 허용")
|
|
ax[0].set_ylabel("OP MAE %"); ax[0].legend(); ax[0].set_title("Rolling vs static — adaptation over May")
|
|
ax[1].plot(r.day, r.ood_roll, "b.-"); ax[1].set_ylabel("rolling OOD %")
|
|
ax[1].set_title("OOD fraction (학습 envelope 밖) — 5월 데이터 흡수하며 감소")
|
|
fig.tight_layout(); fig.savefig(BASE + f"{args.prefix}_rolling.png", dpi=95)
|
|
print(f"\n플롯 저장: {BASE}{args.prefix}_rolling.png")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|