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)
101 lines
4.4 KiB
Python
101 lines
4.4 KiB
Python
"""
|
|
Shadow 예측기 — 히스토리 리플레이 백테스트.
|
|
|
|
선행: c6111_data.pkl. 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.ensemble import GradientBoostingRegressor
|
|
from sklearn.metrics import r2_score, mean_absolute_error
|
|
|
|
BASE = "/home/windpacer/projects/hc900_ax/scripts/analysis/"
|
|
FEATURES = ["feed", "product", "T_C"]
|
|
SMOOTH = 40
|
|
TRAIN_FRAC = 0.70
|
|
|
|
|
|
class SteamPredictor:
|
|
"""운전점 학습 + 밸브 역특성(flow→OP)."""
|
|
def fit(self, df_train):
|
|
ops = (df_train.set_index("dtat").resample("6h").median(numeric_only=True)
|
|
.dropna(subset=["steam_flow", "feed"]))
|
|
ops = ops[ops["feed"] > 50]
|
|
self.model = GradientBoostingRegressor(n_estimators=200, max_depth=2,
|
|
learning_rate=0.05, random_state=0)
|
|
self.model.fit(ops[FEATURES].values, ops["steam_flow"].values)
|
|
# 밸브 역특성: OP = poly(flow) (단조, 3차)
|
|
self.inv = np.polyfit(df_train["steam_flow"], df_train["steam_op"], 3)
|
|
return self
|
|
|
|
def predict_flow(self, X):
|
|
return self.model.predict(X)
|
|
|
|
def flow_to_op(self, flow):
|
|
return np.clip(np.polyval(self.inv, flow), 0, 100)
|
|
|
|
|
|
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")
|
|
# 입력 평활 (실제 shadow도 노이즈 평활 사용)
|
|
for c in FEATURES:
|
|
df[c + "_s"] = df[c].rolling(SMOOTH, min_periods=1, center=True).median()
|
|
|
|
cut = df["dtat"].quantile(TRAIN_FRAC)
|
|
tr, te = df[df["dtat"] <= cut], df[df["dtat"] > cut]
|
|
print(f"학습 {tr.dtat.min()}~{tr.dtat.max()} ({len(tr)}) "
|
|
f"shadow(held-out) {te.dtat.min()}~{te.dtat.max()} ({len(te)})")
|
|
|
|
pred = SteamPredictor().fit(tr)
|
|
# OOD(학습 운전envelope 밖) 게이트: 입력이 학습 1~99% 범위 밖이면 '저신뢰→오퍼레이터 폴백'
|
|
lo, hi = tr[FEATURES].quantile(0.01), tr[FEATURES].quantile(0.99)
|
|
print(f"학습 envelope: " + ", ".join(f"{c}[{lo[c]:.0f},{hi[c]:.1f}]" for c in FEATURES))
|
|
|
|
def in_env(d):
|
|
return ((d[FEATURES] >= lo) & (d[FEATURES] <= hi)).all(axis=1)
|
|
|
|
for name, d in [("학습기간", tr), ("★held-out shadow", te)]:
|
|
Xs = d[[c + "_s" for c in FEATURES]].values
|
|
pf = pred.predict_flow(Xs)
|
|
po = pred.flow_to_op(pf)
|
|
ao = d["steam_op"].values
|
|
env = in_env(d).values
|
|
within = np.mean(np.abs(po - ao) <= 2.0) * 100
|
|
print(f"\n[{name}] OOD(범위밖)={100*(~env).mean():.1f}%")
|
|
print(f" 전체 OP MAE={mean_absolute_error(ao,po):.2f} |Δ|≤2%={within:.1f}%")
|
|
if env.sum() > 50:
|
|
print(f" in-envelope OP MAE={mean_absolute_error(ao[env],po[env]):.2f} "
|
|
f"|Δ|≤2%={np.mean(np.abs(po[env]-ao[env])<=2)*100:.1f}% ← shadow가 신뢰구간에서 조언")
|
|
d = d.assign(pred_flow=pf, pred_op=po, ood=~env)
|
|
if name.startswith("★"):
|
|
te = d
|
|
|
|
# 플롯: held-out 시계열 오버레이 + OP 비교 + 오차분포
|
|
fig, ax = plt.subplots(3, 1, figsize=(16, 11))
|
|
s = te.iloc[::20]
|
|
ax[0].plot(s.dtat, s["steam_flow"], lw=.6, label="actual steam flow")
|
|
ax[0].plot(s.dtat, s["pred_flow"], lw=.6, c="r", label="predicted")
|
|
ax[0].set_title("held-out shadow: steam flow actual vs predicted"); ax[0].legend(fontsize=8)
|
|
ax[1].plot(s.dtat, s["steam_op"], lw=.6, label="actual operator OP")
|
|
ax[1].plot(s.dtat, s["pred_op"], lw=.6, c="r", label="predicted OP")
|
|
ax[1].set_ylabel("OP %"); ax[1].set_title("operator OP vs shadow-predicted OP"); ax[1].legend(fontsize=8)
|
|
err = te["pred_op"] - te["steam_op"]
|
|
ax[2].hist(err, bins=80); ax[2].axvline(0, c="k", lw=.5)
|
|
ax[2].set_title(f"OP error (pred-actual): median {err.median():+.2f}%, std {err.std():.2f}%")
|
|
fig.tight_layout(); fig.savefig(BASE + f"{args.prefix}_shadow.png", dpi=95)
|
|
print(f"\n플롯 저장: {BASE}{args.prefix}_shadow.png")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|