File size: 14,187 Bytes
938949f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 | """
TimeSeriesPredictor: lag-based direct multi-horizon forecasting for
photosynthesis rate A. Uses daytime-session indexing to handle 12h+
nighttime gaps, with per-horizon models (XGBoost / GradientBoosting).
Each growing season (May-Sep) is handled independently — sessions, lags,
and targets never cross the off-season gap (Oct-Apr).
"""
from __future__ import annotations
from typing import Optional
import numpy as np
import pandas as pd
from sklearn.ensemble import GradientBoostingRegressor
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
try:
from xgboost import XGBRegressor
_HAS_XGB = True
except ImportError:
_HAS_XGB = False
# Horizons: name -> (steps within session | None, calendar days | None)
HORIZONS = {
"15min": {"steps": 1, "days": None},
"1hour": {"steps": 4, "days": None},
"1day": {"steps": None, "days": 1},
"1week": {"steps": None, "days": 7},
"1month": {"steps": None, "days": 30},
}
LAG_COLS = ["A", "ghi_w_m2", "air_temperature_c"]
LAG_STEPS = [1, 2, 3, 4, 8, 12]
ROLLING_WINDOWS = [4, 12]
MAX_GAP_MINUTES = 30
class TimeSeriesPredictor:
"""Train one model per forecast horizon using lag features."""
def __init__(self):
self.models: dict[str, object] = {}
self.feature_cols: Optional[list[str]] = None
self.results: dict[str, dict] = {}
# ------------------------------------------------------------------
# Season splitting
# ------------------------------------------------------------------
@staticmethod
def assign_season(df: pd.DataFrame, ts_col: str = "timestamp_utc") -> pd.DataFrame:
"""Add a 'season' column (year of each row's growing season)."""
out = df.copy()
ts = pd.to_datetime(out[ts_col], utc=True)
out["season"] = ts.dt.year
return out
# ------------------------------------------------------------------
# Session identification
# ------------------------------------------------------------------
@staticmethod
def identify_sessions(df: pd.DataFrame, ts_col: str = "timestamp_utc") -> pd.DataFrame:
"""Assign session_id to contiguous daytime blocks (gap <= MAX_GAP_MINUTES).
Sessions are identified within each season independently."""
out = df.copy()
out = out.sort_values(ts_col).reset_index(drop=True)
ts = pd.to_datetime(out[ts_col], utc=True)
gap = ts.diff().dt.total_seconds() / 60
out["session_id"] = (gap > MAX_GAP_MINUTES).cumsum()
return out
# ------------------------------------------------------------------
# Lag / rolling features
# ------------------------------------------------------------------
@staticmethod
def create_lag_features(df: pd.DataFrame) -> pd.DataFrame:
"""Create within-session lags, rolling stats, and previous-session summary.
Previous-session features only link sessions within the same season."""
out = df.copy()
# Per-session lags and rolling stats
for col in LAG_COLS:
if col not in out.columns:
continue
for lag in LAG_STEPS:
col_name = f"{col}_lag{lag}"
out[col_name] = out.groupby("session_id")[col].shift(lag)
for w in ROLLING_WINDOWS:
out[f"{col}_rmean{w}"] = out.groupby("session_id")[col].transform(
lambda s: s.shift(1).rolling(w, min_periods=1).mean()
)
if col == "A":
out[f"{col}_rstd{w}"] = out.groupby("session_id")[col].transform(
lambda s: s.shift(1).rolling(w, min_periods=1).std()
)
# Previous-session summary for A (within same season)
if "A" in out.columns and "season" in out.columns:
sess_stats = out.groupby("session_id").agg(
season=("season", "first"),
mean_A=("A", "mean"),
max_A=("A", "max"),
)
# Shift within season so first session of each season gets NaN
sess_stats["prev_sess_mean_A"] = sess_stats.groupby("season")["mean_A"].shift(1)
sess_stats["prev_sess_max_A"] = sess_stats.groupby("season")["max_A"].shift(1)
out = out.merge(
sess_stats[["prev_sess_mean_A", "prev_sess_max_A"]],
left_on="session_id", right_index=True, how="left",
)
# Fill NaN lags at session start with prev-session end values (within season)
for col in LAG_COLS:
if col not in df.columns:
continue
sess_end = df.groupby("session_id").agg(
last_val=(col, "last"),
)
if "season" in df.columns:
sess_season = df.groupby("session_id")["season"].first()
sess_end["season"] = sess_season
sess_end["prev_end"] = sess_end.groupby("season")["last_val"].shift(1)
else:
sess_end["prev_end"] = sess_end["last_val"].shift(1)
out = out.merge(
sess_end[["prev_end"]].rename(columns={"prev_end": f"_prev_end_{col}"}),
left_on="session_id", right_index=True, how="left",
)
for lag in LAG_STEPS:
lag_col = f"{col}_lag{lag}"
if lag_col in out.columns:
out[lag_col] = out[lag_col].fillna(out[f"_prev_end_{col}"])
out.drop(columns=[f"_prev_end_{col}"], inplace=True)
return out
# ------------------------------------------------------------------
# Horizon targets
# ------------------------------------------------------------------
@staticmethod
def create_horizon_target(df: pd.DataFrame, horizon_name: str,
ts_col: str = "timestamp_utc") -> pd.Series:
"""Create target column for a given horizon.
Calendar-day targets only match within the same season."""
h = HORIZONS[horizon_name]
ts = pd.to_datetime(df[ts_col], utc=True)
if h["steps"] is not None:
# Within-session shift
target = df.groupby("session_id")["A"].shift(-h["steps"])
else:
# Calendar-day match within same season
days = h["days"]
target_ts = ts + pd.Timedelta(days=days)
if "season" in df.columns:
# Build per-season lookup so targets don't cross seasons
target = pd.Series(np.nan, index=df.index)
for season, grp in df.groupby("season"):
grp_ts = pd.to_datetime(grp[ts_col], utc=True)
lookup = pd.Series(grp["A"].values, index=grp_ts)
lookup = lookup[~lookup.index.duplicated(keep="first")]
grp_target_ts = (grp_ts + pd.Timedelta(days=days)).dt.floor("15min")
matched = grp_target_ts.map(lookup)
target.loc[grp.index] = matched.values
else:
target_ts_rounded = target_ts.dt.floor("15min")
lookup = pd.Series(df["A"].values, index=ts)
lookup = lookup[~lookup.index.duplicated(keep="first")]
target = target_ts_rounded.map(lookup)
target = target.reset_index(drop=True)
return target
# ------------------------------------------------------------------
# Feature columns
# ------------------------------------------------------------------
def _get_feature_cols(self, df: pd.DataFrame) -> list[str]:
"""Return numeric feature columns, excluding targets / metadata."""
exclude = {"A", "timestamp_utc", "time", "source", "session_id",
"target", "season"}
cols = [c for c in df.select_dtypes(include=[np.number]).columns if c not in exclude]
return cols
# ------------------------------------------------------------------
# Train / evaluate
# ------------------------------------------------------------------
def train_all_horizons(self, df: pd.DataFrame, train_ratio: float = 0.75) -> pd.DataFrame:
"""Train one model per horizon, treating each season independently.
Within each season the first ``train_ratio`` rows are used for
training and the remainder for testing. Training data from all
seasons is pooled to fit a single model per horizon, and test
data from all seasons is pooled for evaluation. Per-season
metrics are also reported.
"""
self.feature_cols = self._get_feature_cols(df)
seasons = sorted(df["season"].unique())
rows = []
for horizon_name in HORIZONS:
# Collect train/test splits per season
all_X_train, all_y_train = [], []
all_X_test, all_y_test = [], []
season_train_n: dict[int, int] = {}
season_test: dict[int, tuple] = {}
for season in seasons:
sdf = df[df["season"] == season].copy()
target = self.create_horizon_target(sdf, horizon_name)
sdf = sdf.copy()
sdf["target"] = target.values
sdf = sdf.dropna(subset=self.feature_cols + ["target"])
if len(sdf) < 30:
continue
n = int(len(sdf) * train_ratio)
if n < 10 or len(sdf) - n < 5:
continue
X_tr = sdf[self.feature_cols].iloc[:n]
y_tr = sdf["target"].iloc[:n]
X_te = sdf[self.feature_cols].iloc[n:]
y_te = sdf["target"].iloc[n:]
all_X_train.append(X_tr)
all_y_train.append(y_tr)
all_X_test.append(X_te)
all_y_test.append(y_te)
season_train_n[season] = len(X_tr)
season_test[season] = (X_te, y_te)
if not all_X_train or not all_X_test:
print(f" {horizon_name}: insufficient data across seasons, skipping")
continue
X_train = pd.concat(all_X_train)
y_train = pd.concat(all_y_train)
X_test = pd.concat(all_X_test)
y_test = pd.concat(all_y_test)
model = self._make_model()
model.fit(X_train, y_train)
self.models[horizon_name] = model
# Overall metrics
pred = model.predict(X_test)
rmse = float(np.sqrt(mean_squared_error(y_test, pred)))
mae = float(mean_absolute_error(y_test, pred))
r2 = float(r2_score(y_test, pred))
self.results[horizon_name] = {
"predictions": pred, "y_test": y_test.values,
"rmse": rmse, "mae": mae, "r2": r2,
"n_train": len(X_train), "n_test": len(X_test),
}
rows.append({
"horizon": horizon_name, "season": "all",
"approach": "time_series",
"RMSE": round(rmse, 4), "MAE": round(mae, 4), "R2": round(r2, 4),
"n_train": len(X_train), "n_test": len(X_test),
})
print(f" {horizon_name} [all]: RMSE={rmse:.4f} MAE={mae:.4f} "
f"R²={r2:.4f} (train={len(X_train)}, test={len(X_test)})")
# Per-season metrics
for season, (X_te_s, y_te_s) in season_test.items():
pred_s = model.predict(X_te_s)
rmse_s = float(np.sqrt(mean_squared_error(y_te_s, pred_s)))
mae_s = float(mean_absolute_error(y_te_s, pred_s))
r2_s = float(r2_score(y_te_s, pred_s))
rows.append({
"horizon": horizon_name, "season": str(season),
"approach": "time_series",
"RMSE": round(rmse_s, 4), "MAE": round(mae_s, 4),
"R2": round(r2_s, 4),
"n_train": season_train_n[season],
"n_test": len(X_te_s),
})
print(f" {horizon_name} [{season}]: RMSE={rmse_s:.4f} "
f"MAE={mae_s:.4f} R²={r2_s:.4f} (test={len(X_te_s)})")
return pd.DataFrame(rows)
def get_comparison_with_baseline(
self, baseline_metrics: pd.DataFrame
) -> pd.DataFrame:
"""Combine TS horizon results with cross-sectional baseline into one table."""
ts_rows = []
for horizon_name, res in self.results.items():
ts_rows.append({
"horizon": horizon_name, "season": "all",
"approach": "time_series",
"RMSE": round(res["rmse"], 4),
"MAE": round(res["mae"], 4),
"R2": round(res["r2"], 4),
})
ts_df = pd.DataFrame(ts_rows)
# Best cross-sectional model
if not baseline_metrics.empty:
best_idx = baseline_metrics["RMSE"].idxmin()
best = baseline_metrics.loc[best_idx]
bl_rows = []
for h in HORIZONS:
bl_rows.append({
"horizon": h, "season": "all",
"approach": f"cross_sectional ({best['model']})",
"RMSE": round(float(best["RMSE"]), 4),
"MAE": round(float(best["MAE"]), 4),
"R2": round(float(best["R2"]), 4),
})
bl_df = pd.DataFrame(bl_rows)
return pd.concat([ts_df, bl_df], ignore_index=True)
return ts_df
# ------------------------------------------------------------------
# Internals
# ------------------------------------------------------------------
@staticmethod
def _make_model():
if _HAS_XGB:
return XGBRegressor(
n_estimators=300, max_depth=4, learning_rate=0.05,
min_child_weight=10, reg_alpha=0.1, reg_lambda=1.0,
n_jobs=-1, random_state=42,
)
return GradientBoostingRegressor(
n_estimators=300, max_depth=4, learning_rate=0.05,
min_samples_leaf=10, random_state=42,
)
|