"""
C-6111 Shadow 예측기 — 히스토리 리플레이 백테스트 (플랜 §7 shadow 진입).

학습기간 운전점으로 `스팀유량=f(피드,제품,목표T_C)` 학습 → held-out 미래기간을
매 시점 리플레이하여 예측 스팀→(밸브 역특성)→예측 OP 를 산출, **실제 오퍼레이터 OP와 비교**.
"이 예측기를 shadow로 돌렸다면 오퍼레이터 손과 얼마나 일치했나" 를 정직 검증.

선행: c6111_data.pkl. 포팅대상(추후 C# live shadow)은 동일 로직.
"""
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"]   # 깨끗한 인과/목표 입력 (§16.6)
SMOOTH = 40                              # 입력 평활 20분(운전점 성격 유지)
TRAIN_FRAC = 0.70                        # 앞 70% 기간 학습, 뒤 30% held-out shadow


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():
    df = pd.read_pickle(BASE + "c6111_data.pkl")
    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 + "c6111_shadow.png", dpi=95)
    print(f"\n플롯 저장: {BASE}c6111_shadow.png")


if __name__ == "__main__":
    main()