diff --git a/.env.example b/.env.example index 7e3fae7..7f6effb 100644 --- a/.env.example +++ b/.env.example @@ -4,4 +4,11 @@ INFLUXDB_ORG= INFLUXDB_BUCKET=weather STATION_LATITUDE= STATION_LONGITUDE= -STATION_ELEVATION= \ No newline at end of file +STATION_ELEVATION= + +# Chronos (prévision) +# CHRONOS_MODEL=amazon/chronos-t5-small +# CHRONOS_CONTEXT=336 +# CHRONOS_HORIZON=96 +# CHRONOS_RESAMPLE=1h +# CHRONOS_SAMPLES=20 diff --git a/docs/11 - Modèle Chronos/figures/chronos_errors_amazon__chronos-t5-base.png b/docs/11 - Modèle Chronos/figures/chronos_errors_amazon__chronos-t5-base.png new file mode 100644 index 0000000..1af1042 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_errors_amazon__chronos-t5-base.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_errors_amazon__chronos-t5-mini.png b/docs/11 - Modèle Chronos/figures/chronos_errors_amazon__chronos-t5-mini.png new file mode 100644 index 0000000..03f2edd Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_errors_amazon__chronos-t5-mini.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_errors_amazon__chronos-t5-small.png b/docs/11 - Modèle Chronos/figures/chronos_errors_amazon__chronos-t5-small.png new file mode 100644 index 0000000..287fe7f Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_errors_amazon__chronos-t5-small.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_forecast_amazon__chronos-t5-base.png b/docs/11 - Modèle Chronos/figures/chronos_forecast_amazon__chronos-t5-base.png new file mode 100644 index 0000000..c9c6579 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_forecast_amazon__chronos-t5-base.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_forecast_amazon__chronos-t5-mini.png b/docs/11 - Modèle Chronos/figures/chronos_forecast_amazon__chronos-t5-mini.png new file mode 100644 index 0000000..9768628 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_forecast_amazon__chronos-t5-mini.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_forecast_amazon__chronos-t5-small.png b/docs/11 - Modèle Chronos/figures/chronos_forecast_amazon__chronos-t5-small.png new file mode 100644 index 0000000..74f2be7 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_forecast_amazon__chronos-t5-small.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_holdout6_errors.png b/docs/11 - Modèle Chronos/figures/chronos_holdout6_errors.png new file mode 100644 index 0000000..2431071 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_holdout6_errors.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_models_comparison.png b/docs/11 - Modèle Chronos/figures/chronos_models_comparison.png new file mode 100644 index 0000000..03204a1 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_models_comparison.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_multi_rain_rate.png b/docs/11 - Modèle Chronos/figures/chronos_multi_rain_rate.png new file mode 100644 index 0000000..e1f65b2 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_multi_rain_rate.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_multi_temperature.png b/docs/11 - Modèle Chronos/figures/chronos_multi_temperature.png new file mode 100644 index 0000000..e3ae929 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_multi_temperature.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_multi_wind_speed.png b/docs/11 - Modèle Chronos/figures/chronos_multi_wind_speed.png new file mode 100644 index 0000000..b32ba04 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_multi_wind_speed.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_tuned_rain_rate.png b/docs/11 - Modèle Chronos/figures/chronos_tuned_rain_rate.png new file mode 100644 index 0000000..7c93bac Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_tuned_rain_rate.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_tuned_temperature.png b/docs/11 - Modèle Chronos/figures/chronos_tuned_temperature.png new file mode 100644 index 0000000..dc90239 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_tuned_temperature.png differ diff --git a/docs/11 - Modèle Chronos/figures/chronos_tuned_wind_speed.png b/docs/11 - Modèle Chronos/figures/chronos_tuned_wind_speed.png new file mode 100644 index 0000000..bce7fc6 Binary files /dev/null and b/docs/11 - Modèle Chronos/figures/chronos_tuned_wind_speed.png differ diff --git a/docs/11 - Modèle Chronos/index.md b/docs/11 - Modèle Chronos/index.md new file mode 100644 index 0000000..78bed7f --- /dev/null +++ b/docs/11 - Modèle Chronos/index.md @@ -0,0 +1,47 @@ +# Modèle Chronos-2 (foundation model HF) + +Objectif : tester un modèle de prévision généraliste récent (**Chronos-2**, Amazon) en zéro-shot sur notre station. On resample la température à l’heure, on coupe les dernières 96 h pour évaluer la prévision, et on compare le forecast à l’observé. + +### Mise en route + +Installez les dépendances (`pip install -r requirements.txt`), puis lancez `run_chronos.py`. Le script lit `data/weather_minutely.csv`, resample la température en pas horaire, prend les 336 dernières heures en contexte et les 96 suivantes en cible. Il charge `amazon/chronos-t5-small` par défaut (modifiable via `CHRONOS_MODEL`), génère 20 échantillons dont on prend la moyenne, calcule MAE/RMSE, sauvegarde les CSV `chronos_forecast_.csv` et `chronos_errors_.csv`, et produit les figures associées dans `docs/11 - Modèle Chronos/figures/`. + +Pour comparer plusieurs tailles, lancez `run_chronos.py` avec différents `CHRONOS_MODEL` (mini/small/base), puis `compare_chronos.py` agrège les CSV et trace la comparaison. + +### Paramètres + +Modifiables via variables d’environnement : `CHRONOS_MODEL` (défaut `amazon/chronos-t5-small`), `CHRONOS_CONTEXT` (336 h), `CHRONOS_HORIZON` (96 h), `CHRONOS_RESAMPLE` (`1h`), `CHRONOS_SAMPLES` (20). + +### Résultats comparés (mêmes données, horizon 96 h) + +![Comparaison des tailles Chronos](./figures/chronos_models_comparison.png) + +Sur la même fenêtre de validation locale, nous avons testé trois tailles : `chronos-t5-mini`, `chronos-t5-small` et `chronos-t5-base`. Le modèle **small** est ressorti devant (MAE ≈ 3,68 °C, RMSE ≈ 4,53 °C), les versions mini et base étant derrière (MAE ≈ 4,18–4,24 °C, RMSE ≈ 5,3–5,6 °C). Autrement dit, monter en taille n’a pas amélioré la prévision à 96 h sur ces données locales ; la version small offre le meilleur compromis précision/poids. + +![Trajectoire prédit vs observé – mini](./figures/chronos_forecast_amazon__chronos-t5-mini.png) + +![Trajectoire prédit vs observé – small](./figures/chronos_forecast_amazon__chronos-t5-small.png) + +![Trajectoire prédit vs observé – base](./figures/chronos_forecast_amazon__chronos-t5-base.png) + +### Lecture et portée + +Pour coller à nos horizons cibles, `run_chronos_multi.py` évalue Chronos-small sur 1 h, 6 h et 24 h pour la température, le vent et la pluie (horaire uniquement ; l’horizon 10 minutes n’est pas couvert par Chronos qui est pré-entraîné en pas horaire). Les figures `chronos_multi_temperature.png`, `chronos_multi_wind_speed.png` et `chronos_multi_rain_rate.png` illustrent où le modèle est le plus fiable : à 1 h, la température reste sous ~1,3 °C de MAE et le vent sous ~0,6 (unités du jeu) ; à 6 h, l’erreur grimpe modérément (≈2 °C temp., ≈3 km/h vent) ; à 24 h, elle dépasse 4–6 (°C/ km/h). Sur la pluie, le F1 reste nul à 1 h/6 h et ne monte qu’à ~0,15 à 24 h, signe que le modèle “foundation” horaire ne capture pas bien les occurrences locales rares. Les figures individuelles (`chronos_forecast_.png`, `chronos_errors_.png`) permettent de voir la trajectoire prédit vs observé et l’erreur par horizon. + +Au total, Chronos-small fournit un signal exploitable sur la température et un peu sur le vent pour des horizons courts à intermédiaires, mais reste faible sur la pluie et se dégrade nettement au-delà de 24 h. Une calibration locale, davantage de contexte ou une cible adaptée (pluie binaire calibrée) seraient nécessaires pour en faire un outil de prévision robuste sur toutes les variables. + +### Réglages prudents (contexte 288 h, horizon limité à 64 h, 100 échantillons) + +![Chronos small réglé – température](./figures/chronos_tuned_temperature.png) + +![Chronos small réglé – vent](./figures/chronos_tuned_wind_speed.png) + +![Chronos small réglé – pluie](./figures/chronos_tuned_rain_rate.png) + +Avec `run_chronos_tuned.py`, on réduit le contexte (288 h) et l’horizon maximum (64 h) tout en augmentant les échantillons (100). Sur la même fenêtre locale, la température s’améliore nettement : MAE ~0,75 °C à 1 h, ~1,27 °C à 6 h, ~3,40 °C à 24 h (vs 1,33/2,02/4,84 auparavant). Le vent progresse surtout à 24 h (≈2,39 contre ~6,38 auparavant), même si le 1 h est moins bon que la première passe. La pluie reste instable : le F1 peut atteindre 0,22–0,28 à 24–48 h mais les scores courts sont peu fiables. Limiter l’horizon à 64 h, raccourcir le contexte et lisser par davantage d’échantillons améliorent donc la température et le vent, mais ne suffisent pas à rendre la pluie prédictible. + +### Dernier essai “comme en vrai” : prévision sur les 6 dernières heures + +![Chronos small – erreurs 6 h holdout](./figures/chronos_holdout6_errors.png) + +On a masqué les 6 dernières heures de la série et demandé à Chronos-small (contexte 288 h, 50 échantillons) de prévoir température, vent (vitesse + direction), humidité et pression, puis comparé aux valeurs réelles. Moyennes des erreurs absolues sur ces 6 pas : température ~0,84 °C, vent ~1,2 km/h, direction ~3,9°, pression ~5,3 hPa, humidité ~24 %. Les erreurs restent très contenues sur température/vent et la direction est raisonnable (quelques degrés) ; la pression dérive (5 hPa) et l’humidité est clairement décalée. Ce test confirme que Chronos peut fournir des tendances crédibles à très court terme (<6 h) sur température et vent, mais pas sur l’humidité (et on a vu plus haut que la pluie reste peu fiable). diff --git a/docs/11 - Modèle Chronos/scripts/compare_chronos.py b/docs/11 - Modèle Chronos/scripts/compare_chronos.py new file mode 100644 index 0000000..a87ba9e --- /dev/null +++ b/docs/11 - Modèle Chronos/scripts/compare_chronos.py @@ -0,0 +1,68 @@ +# scripts/compare_chronos.py +from __future__ import annotations + +from pathlib import Path +import re + +import matplotlib.pyplot as plt +import pandas as pd + +DOC_DIR = Path(__file__).resolve().parent.parent +DATA_DIR = DOC_DIR / "data" +FIG_DIR = DOC_DIR / "figures" + + +def _load_forecasts() -> pd.DataFrame: + records = [] + pattern = re.compile(r"chronos_forecast_(.+)\.csv") + for csv in DATA_DIR.glob("chronos_forecast_*.csv"): + match = pattern.match(csv.name) + if not match: + continue + model = match.group(1).replace("__", "/") + df = pd.read_csv(csv) + if not {"y_true", "y_pred"}.issubset(df.columns): + continue + err = df["y_pred"] - df["y_true"] + records.append( + { + "model": model, + "mae": err.abs().mean(), + "rmse": (err.pow(2).mean()) ** 0.5, + } + ) + return pd.DataFrame(records) + + +def _plot_comparison(df: pd.DataFrame, output_path: Path) -> None: + output_path.parent.mkdir(parents=True, exist_ok=True) + fig, ax = plt.subplots(figsize=(7, 4)) + x = range(len(df)) + ax.bar([i - 0.15 for i in x], df["mae"], width=0.3, label="MAE") + ax.bar([i + 0.15 for i in x], df["rmse"], width=0.3, label="RMSE") + ax.set_xticks(list(x)) + ax.set_xticklabels(df["model"], rotation=15) + ax.set_ylabel("Erreur (°C)") + ax.set_title("Chronos T5 – comparaison des tailles") + ax.grid(True, linestyle=":", alpha=0.4, axis="y") + ax.legend() + fig.tight_layout() + fig.savefig(output_path, dpi=150) + plt.close(fig) + + +def main() -> None: + df = _load_forecasts() + if df.empty: + raise SystemExit("Aucune sortie chronos_forecast_*.csv trouvée dans data/. Lancez run_chronos.py d'abord.") + df_sorted = df.sort_values("mae") + summary_path = DATA_DIR / "chronos_summary.csv" + df_sorted.to_csv(summary_path, index=False) + _plot_comparison(df_sorted, FIG_DIR / "chronos_models_comparison.png") + print(df_sorted.to_string(index=False, float_format=lambda x: f"{x:.3f}")) + print(f"✔ Sauvegardé : {summary_path}") + print(f"✔ Figure : {FIG_DIR / 'chronos_models_comparison.png'}") + + +if __name__ == "__main__": + main() diff --git a/docs/11 - Modèle Chronos/scripts/run_chronos.py b/docs/11 - Modèle Chronos/scripts/run_chronos.py new file mode 100644 index 0000000..4fa68c9 --- /dev/null +++ b/docs/11 - Modèle Chronos/scripts/run_chronos.py @@ -0,0 +1,152 @@ +# scripts/run_chronos.py +from __future__ import annotations + +import os +from pathlib import Path +import sys +from typing import Tuple + +import matplotlib.pyplot as plt +import numpy as np +import pandas as pd +import torch + +PROJECT_ROOT = Path(__file__).resolve().parents[3] +if str(PROJECT_ROOT) not in sys.path: + sys.path.insert(0, str(PROJECT_ROOT)) + +from meteo.dataset import load_raw_csv + +try: + from chronos import ChronosPipeline +except ImportError as exc: # pragma: no cover - guidance if deps missing + raise SystemExit( + "chronos-forecasting est manquant. Installez-le (pip install chronos-forecasting) " + "puis relancez." + ) from exc + + +CSV_PATH = PROJECT_ROOT / "data" / "weather_minutely.csv" +DOC_DIR = Path(__file__).resolve().parent.parent +DATA_DIR = DOC_DIR / "data" +FIG_DIR = DOC_DIR / "figures" + +MODEL_ID = os.getenv("CHRONOS_MODEL", "amazon/chronos-t5-small") +CONTEXT_LEN = int(os.getenv("CHRONOS_CONTEXT", "336")) # 14 jours (1H) +HORIZON_LEN = int(os.getenv("CHRONOS_HORIZON", "96")) # 4 jours (1H) +RESAMPLE_RULE = os.getenv("CHRONOS_RESAMPLE", "1h") +NUM_SAMPLES = int(os.getenv("CHRONOS_SAMPLES", "20")) # échantillons stochastiques + + +def _load_series(csv_path: Path, target_col: str, rule: str) -> pd.Series: + df = load_raw_csv(csv_path) + if target_col not in df.columns: + raise SystemExit(f"Colonne absente dans le CSV : {target_col!r}") + series = ( + df[target_col] + .resample(rule) + .mean() + .interpolate(limit_direction="both") + ) + return series.dropna() + + +def _split_context_target(series: pd.Series, context_len: int, horizon_len: int) -> Tuple[pd.Series, pd.Series]: + needed = context_len + horizon_len + if len(series) < needed: + raise SystemExit(f"Pas assez de données après resampling : {len(series)} < {needed} (context+horizon).") + window = series.iloc[-needed:] + context = window.iloc[:context_len] + target = window.iloc[context_len:] + return context, target + + +def _plot_forecast(target: pd.Series, forecast: pd.Series, output_path: Path) -> None: + output_path.parent.mkdir(parents=True, exist_ok=True) + fig, ax = plt.subplots(figsize=(10, 4)) + ax.plot(target.index, target.values, label="Observé", linewidth=2) + ax.plot(forecast.index, forecast.values, label="Chronos-2 (préd.)", linewidth=2, linestyle="--") + ax.set_title("Prévision Chronos-2 vs observation") + ax.set_xlabel("Date") + ax.set_ylabel("Température (°C)") + ax.grid(True, linestyle=":", alpha=0.4) + ax.legend() + fig.autofmt_xdate() + fig.tight_layout() + fig.savefig(output_path, dpi=150) + plt.close(fig) + + +def _plot_errors(errors: pd.Series, output_path: Path) -> None: + output_path.parent.mkdir(parents=True, exist_ok=True) + fig, ax = plt.subplots(figsize=(8, 4)) + ax.plot(errors.index, errors.values, marker="o") + ax.set_title("Erreur absolue par horizon (heures)") + ax.set_xlabel("Horizon (h)") + ax.set_ylabel("Erreur absolue (°C)") + ax.grid(True, linestyle=":", alpha=0.4) + fig.tight_layout() + fig.savefig(output_path, dpi=150) + plt.close(fig) + + +def main() -> None: + if not CSV_PATH.exists(): + raise SystemExit(f"Fichier introuvable : {CSV_PATH}") + + series = _load_series(CSV_PATH, target_col="temperature", rule=RESAMPLE_RULE) + context, target = _split_context_target(series, CONTEXT_LEN, HORIZON_LEN) + + model_slug = MODEL_ID.replace("/", "__") + print(f"Contexte : {len(context)} points ({context.index[0]} -> {context.index[-1]})") + print(f"Cible : {len(target)} points ({target.index[0]} -> {target.index[-1]})") + print(f"Modèle HF : {MODEL_ID}") + + pipeline = ChronosPipeline.from_pretrained( + MODEL_ID, + device_map="auto", # CUDA/MPS/CPU automatique + dtype="auto", + ) + + context_array = context.to_numpy(dtype=float) + context_tensor = torch.tensor(context_array, dtype=torch.float32) + forecasts = pipeline.predict( + [context_tensor], # batch de 1 série (Tensor attendu) + prediction_length=HORIZON_LEN, + num_samples=NUM_SAMPLES, + ) + forecast_mean = forecasts.mean(0) + # Si batch dimension présente, on prend la première série + if forecast_mean.ndim == 2: + forecast_mean = forecast_mean[0] + forecast_mean = np.asarray(forecast_mean).ravel() + + forecast_index = pd.date_range(target.index[0], periods=HORIZON_LEN, freq=RESAMPLE_RULE) + forecast_series = pd.Series(forecast_mean, index=forecast_index, name="y_pred") + target_series = target.rename("y_true") + + # Métriques + abs_errors = (forecast_series - target_series).abs() + mae = abs_errors.mean() + rmse = np.sqrt(((forecast_series - target_series) ** 2).mean()) + print(f"MAE : {mae:.3f} °C | RMSE : {rmse:.3f} °C") + + # Sauvegardes + DATA_DIR.mkdir(parents=True, exist_ok=True) + out_csv = DATA_DIR / f"chronos_forecast_{model_slug}.csv" + per_horizon_csv = DATA_DIR / f"chronos_errors_{model_slug}.csv" + + pd.concat([target_series, forecast_series], axis=1).to_csv(out_csv, index=True) + pd.DataFrame({"horizon_h": np.arange(1, HORIZON_LEN + 1), "abs_error": abs_errors.values}).to_csv(per_horizon_csv, index=False) + + # Figures + _plot_forecast(target_series, forecast_series, FIG_DIR / f"chronos_forecast_{model_slug}.png") + _plot_errors(abs_errors.rename("abs_error").rename_axis("h"), FIG_DIR / f"chronos_errors_{model_slug}.png") + + print(f"✔ Sauvegardé : {out_csv}") + print(f"✔ Sauvegardé : {per_horizon_csv}") + print(f"✔ Figures : {FIG_DIR / f'chronos_forecast_{model_slug}.png'} ; {FIG_DIR / f'chronos_errors_{model_slug}.png'}") + + +if __name__ == "__main__": + main() diff --git a/docs/11 - Modèle Chronos/scripts/run_chronos_holdout6.py b/docs/11 - Modèle Chronos/scripts/run_chronos_holdout6.py new file mode 100644 index 0000000..cac62c0 --- /dev/null +++ b/docs/11 - Modèle Chronos/scripts/run_chronos_holdout6.py @@ -0,0 +1,147 @@ +# scripts/run_chronos_holdout6.py +from __future__ import annotations + +from pathlib import Path +import sys +from typing import Dict + +import matplotlib.pyplot as plt +import numpy as np +import pandas as pd +import torch + +PROJECT_ROOT = Path(__file__).resolve().parents[3] +if str(PROJECT_ROOT) not in sys.path: + sys.path.insert(0, str(PROJECT_ROOT)) + +from meteo.dataset import load_raw_csv + +try: + from chronos import ChronosPipeline +except ImportError as exc: + raise SystemExit("chronos-forecasting manquant : pip install -r requirements.txt") from exc + + +CSV_PATH = PROJECT_ROOT / "data" / "weather_minutely.csv" +DOC_DIR = Path(__file__).resolve().parent.parent +DATA_DIR = DOC_DIR / "data" +FIG_DIR = DOC_DIR / "figures" + +MODEL_ID = "amazon/chronos-t5-small" +RESAMPLE_RULE = "1h" +CONTEXT_H = 288 # 12 jours +HOLDOUT_H = 6 +NUM_SAMPLES = 50 + +TARGETS = { + "temperature": {"kind": "reg", "unit": "°C"}, + "wind_speed": {"kind": "reg", "unit": "km/h"}, + "wind_direction": {"kind": "angle", "unit": "deg"}, + "humidity": {"kind": "reg", "unit": "%"}, + "pressure": {"kind": "reg", "unit": "hPa"}, +} + + +def _load_series(target: str) -> pd.Series: + df = load_raw_csv(CSV_PATH) + if target not in df.columns: + raise SystemExit(f"Colonne absente : {target}") + return ( + df[target] + .resample(RESAMPLE_RULE) + .mean() + .interpolate(limit_direction="both") + .dropna() + ) + + +def _angular_error(y_true: pd.Series, y_pred: pd.Series) -> pd.Series: + diff = (y_pred - y_true).abs() % 360 + return diff.apply(lambda x: min(x, 360 - x)) + + +def _plot_errors(df: pd.DataFrame, output_path: Path) -> None: + output_path.parent.mkdir(parents=True, exist_ok=True) + fig, ax = plt.subplots(figsize=(8, 4)) + for target in df["target"].unique(): + sub = df[df["target"] == target] + ax.plot(sub["h"], sub["abs_error"], marker="o", label=target) + ax.set_xlabel("Heure du horizon (1-6)") + ax.set_ylabel("Erreur absolue (unité cible)") + ax.set_title("Chronos small – erreurs sur les 6 dernières heures (holdout)") + ax.grid(True, linestyle=":", alpha=0.4) + ax.legend() + fig.tight_layout() + fig.savefig(output_path, dpi=150) + plt.close(fig) + + +def main() -> None: + if not CSV_PATH.exists(): + raise SystemExit(f"Fichier introuvable : {CSV_PATH}") + + pipeline = ChronosPipeline.from_pretrained( + MODEL_ID, + device_map="auto", + dtype="auto", + ) + + rows: list[Dict[str, object]] = [] + + for target, meta in TARGETS.items(): + series = _load_series(target) + if len(series) < CONTEXT_H + HOLDOUT_H: + raise SystemExit("Pas assez de données pour contexte+holdout.") + + context = series.iloc[-(CONTEXT_H + HOLDOUT_H) : -HOLDOUT_H] + holdout = series.iloc[-HOLDOUT_H:] + + context_tensor = torch.tensor(context.to_numpy(dtype=float), dtype=torch.float32) + forecasts = pipeline.predict( + [context_tensor], + prediction_length=HOLDOUT_H, + num_samples=NUM_SAMPLES, + ) + forecast_mean = forecasts.mean(0) + if forecast_mean.ndim == 2: + forecast_mean = forecast_mean[0] + forecast_series = pd.Series( + np.asarray(forecast_mean).ravel(), + index=holdout.index, + ) + + if meta["kind"] == "angle": + err = _angular_error(holdout, forecast_series) + else: + err = (forecast_series - holdout).abs() + + for i, (ts, e) in enumerate(err.items(), start=1): + rows.append( + { + "target": target, + "timestamp": ts, + "h": i, + "abs_error": float(e), + "unit": meta["unit"], + } + ) + + out_prefix = f"{target}_{MODEL_ID.replace('/', '__')}_holdout6" + DATA_DIR.mkdir(parents=True, exist_ok=True) + pd.concat([forecast_series.rename("y_pred"), holdout.rename("y_true")], axis=1).to_csv( + DATA_DIR / f"chronos_holdout6_forecast_{out_prefix}.csv", + index=True, + ) + + df_errors = pd.DataFrame(rows) + err_path = DATA_DIR / "chronos_holdout6_errors.csv" + df_errors.to_csv(err_path, index=False) + _plot_errors(df_errors, FIG_DIR / "chronos_holdout6_errors.png") + + print(df_errors.groupby("target")["abs_error"].mean().to_string(float_format=lambda x: f"{x:.3f}")) + print(f"✔ Sauvegardé : {err_path}") + print("✔ Figure : figures/chronos_holdout6_errors.png") + + +if __name__ == "__main__": + main() diff --git a/docs/11 - Modèle Chronos/scripts/run_chronos_multi.py b/docs/11 - Modèle Chronos/scripts/run_chronos_multi.py new file mode 100644 index 0000000..3ed8286 --- /dev/null +++ b/docs/11 - Modèle Chronos/scripts/run_chronos_multi.py @@ -0,0 +1,163 @@ +# scripts/run_chronos_multi.py +from __future__ import annotations + +from pathlib import Path +import sys +from typing import Iterable + +import matplotlib.pyplot as plt +import numpy as np +import pandas as pd +import torch + +PROJECT_ROOT = Path(__file__).resolve().parents[3] +if str(PROJECT_ROOT) not in sys.path: + sys.path.insert(0, str(PROJECT_ROOT)) + +from meteo.dataset import load_raw_csv + +try: + from chronos import ChronosPipeline +except ImportError as exc: + raise SystemExit("chronos-forecasting manquant : pip install -r requirements.txt") from exc + + +CSV_PATH = PROJECT_ROOT / "data" / "weather_minutely.csv" +DOC_DIR = Path(__file__).resolve().parent.parent +DATA_DIR = DOC_DIR / "data" +FIG_DIR = DOC_DIR / "figures" + +MODEL_ID = "amazon/chronos-t5-small" +RESAMPLE_RULE = "1h" # on reste sur l'heure pour rester aligné avec le pré-entraînement Chronos +CONTEXT_H = 336 # 14 jours +HORIZONS_H = (1, 6, 24) # horizons demandés (10 min exclu car modèle horaire) +NUM_SAMPLES = 50 + +TARGETS = { + "temperature": {"kind": "reg"}, + "wind_speed": {"kind": "reg"}, + "rain_rate": {"kind": "rain"}, +} + + +def _load_series(target: str) -> pd.Series: + df = load_raw_csv(CSV_PATH) + if target not in df.columns: + raise SystemExit(f"Colonne absente : {target}") + s = df[target].resample(RESAMPLE_RULE).mean().interpolate(limit_direction="both") + return s.dropna() + + +def _prepare_window(series: pd.Series, context_h: int, horizon_h: int) -> tuple[np.ndarray, pd.Series]: + needed = context_h + horizon_h + if len(series) < needed: + raise SystemExit(f"Pas assez de données pour {needed} heures.") + window = series.iloc[-needed:] + context = window.iloc[:context_h] + target = window.iloc[context_h:] + return context.to_numpy(dtype=float), target + + +def _plot_metrics(df: pd.DataFrame, target: str, output_path: Path) -> None: + output_path.parent.mkdir(parents=True, exist_ok=True) + fig, ax = plt.subplots(figsize=(6, 4)) + if df["kind"].iloc[0] == "reg": + ax.plot(df["horizon_h"], df["mae"], marker="o", label="MAE") + ax.plot(df["horizon_h"], df["rmse"], marker="o", label="RMSE") + ax.set_ylabel("Erreur (°C)" if target == "temperature" else "Erreur (unité)") + else: + ax.plot(df["horizon_h"], df["f1"], marker="o", label="F1") + ax.plot(df["horizon_h"], df["brier"], marker="o", label="Brier") + ax.set_ylabel("Score") + ax.set_xlabel("Horizon (heures)") + ax.set_title(f"Chronos (small) – {target}") + ax.grid(True, linestyle=":", alpha=0.4) + ax.legend() + fig.tight_layout() + fig.savefig(output_path, dpi=150) + plt.close(fig) + + +def main() -> None: + if not CSV_PATH.exists(): + raise SystemExit(f"Fichier introuvable : {CSV_PATH}") + + pipeline = ChronosPipeline.from_pretrained( + MODEL_ID, + device_map="auto", + dtype="auto", + ) + + rows: list[dict[str, object]] = [] + + for target, meta in TARGETS.items(): + series = _load_series(target) + # On prend la plus grande fenêtre (max horizon) + context_arr, target_series = _prepare_window(series, CONTEXT_H, max(HORIZONS_H)) + + context_tensor = torch.tensor(context_arr, dtype=torch.float32) + forecasts = pipeline.predict( + [context_tensor], + prediction_length=max(HORIZONS_H), + num_samples=NUM_SAMPLES, + ) + forecast_mean = forecasts.mean(0) + if forecast_mean.ndim == 2: + forecast_mean = forecast_mean[0] + forecast_series = pd.Series( + np.asarray(forecast_mean).ravel(), + index=pd.date_range(target_series.index[0], periods=max(HORIZONS_H), freq=RESAMPLE_RULE), + ) + + for h in HORIZONS_H: + y_true = target_series.iloc[:h] + y_pred = forecast_series.iloc[:h] + if meta["kind"] == "reg": + mae = float((y_pred - y_true).abs().mean()) + rmse = float(np.sqrt(((y_pred - y_true) ** 2).mean())) + rows.append( + {"target": target, "kind": "reg", "horizon_h": h, "mae": mae, "rmse": rmse} + ) + else: + y_true_bin = (y_true > 0).astype(int) + y_pred_bin = (y_pred > 0).astype(int) + if y_true_bin.sum() == 0: + f1 = 0.0 + else: + tp = ((y_true_bin == 1) & (y_pred_bin == 1)).sum() + fp = ((y_true_bin == 0) & (y_pred_bin == 1)).sum() + fn = ((y_true_bin == 1) & (y_pred_bin == 0)).sum() + prec = tp / (tp + fp) if tp + fp > 0 else 0.0 + rec = tp / (tp + fn) if tp + fn > 0 else 0.0 + f1 = 2 * prec * rec / (prec + rec) if prec + rec > 0 else 0.0 + # Brier approximé en traitant y_pred comme proba clampée [0,1] + proba = y_pred.clip(0, 1) + brier = float(((proba - y_true_bin) ** 2).mean()) + rows.append( + {"target": target, "kind": "cls", "horizon_h": h, "f1": float(f1), "brier": brier} + ) + + # Sauvegarde forecast/target complets pour inspection + out_prefix = f"{target}_{MODEL_ID.replace('/', '__')}" + DATA_DIR.mkdir(parents=True, exist_ok=True) + pd.concat([forecast_series.rename("y_pred"), target_series.rename("y_true")], axis=1).to_csv( + DATA_DIR / f"chronos_multi_forecast_{out_prefix}.csv", + index=True, + ) + + df_metrics = pd.DataFrame(rows) + DATA_DIR.mkdir(parents=True, exist_ok=True) + metrics_path = DATA_DIR / "chronos_multi_metrics.csv" + df_metrics.to_csv(metrics_path, index=False) + + for target in df_metrics["target"].unique(): + sub = df_metrics[df_metrics["target"] == target].sort_values("horizon_h") + _plot_metrics(sub, target, FIG_DIR / f"chronos_multi_{target}.png") + + print(df_metrics.to_string(index=False, float_format=lambda x: f"{x:.3f}")) + print(f"✔ Sauvegardé : {metrics_path}") + print("✔ Figures par cible dans figures/chronos_multi_.png") + + +if __name__ == "__main__": + main() diff --git a/docs/11 - Modèle Chronos/scripts/run_chronos_tuned.py b/docs/11 - Modèle Chronos/scripts/run_chronos_tuned.py new file mode 100644 index 0000000..0fca602 --- /dev/null +++ b/docs/11 - Modèle Chronos/scripts/run_chronos_tuned.py @@ -0,0 +1,157 @@ +# scripts/run_chronos_tuned.py +from __future__ import annotations + +from pathlib import Path +import sys + +import matplotlib.pyplot as plt +import numpy as np +import pandas as pd +import torch + +PROJECT_ROOT = Path(__file__).resolve().parents[3] +if str(PROJECT_ROOT) not in sys.path: + sys.path.insert(0, str(PROJECT_ROOT)) + +from meteo.dataset import load_raw_csv + +try: + from chronos import ChronosPipeline +except ImportError as exc: + raise SystemExit("chronos-forecasting manquant : pip install -r requirements.txt") from exc + +CSV_PATH = PROJECT_ROOT / "data" / "weather_minutely.csv" +DOC_DIR = Path(__file__).resolve().parent.parent +DATA_DIR = DOC_DIR / "data" +FIG_DIR = DOC_DIR / "figures" + +# Réglages plus prudents : contexte raccourci, horizon 64 max, plus d'échantillons +MODEL_ID = "amazon/chronos-t5-small" +RESAMPLE_RULE = "1h" +CONTEXT_H = 288 # 12 jours +MAX_HORIZON_H = 64 # <=64 conseillé +HORIZONS_H = (1, 6, 24, 48) # on reste cohérents avec nos jalons +NUM_SAMPLES = 100 + +TARGETS = { + "temperature": {"kind": "reg"}, + "wind_speed": {"kind": "reg"}, + "rain_rate": {"kind": "rain"}, +} + + +def _load_series(target: str) -> pd.Series: + df = load_raw_csv(CSV_PATH) + if target not in df.columns: + raise SystemExit(f"Colonne absente : {target}") + s = df[target].resample(RESAMPLE_RULE).mean().interpolate(limit_direction="both") + return s.dropna() + + +def _prepare_window(series: pd.Series, context_h: int, horizon_h: int) -> tuple[np.ndarray, pd.Series]: + needed = context_h + horizon_h + if len(series) < needed: + raise SystemExit(f"Pas assez de données pour {needed} heures.") + window = series.iloc[-needed:] + context = window.iloc[:context_h] + target = window.iloc[context_h:] + return context.to_numpy(dtype=float), target + + +def _plot_metrics(df: pd.DataFrame, target: str, output_path: Path) -> None: + output_path.parent.mkdir(parents=True, exist_ok=True) + fig, ax = plt.subplots(figsize=(6, 4)) + if df["kind"].iloc[0] == "reg": + ax.plot(df["horizon_h"], df["mae"], marker="o", label="MAE") + ax.plot(df["horizon_h"], df["rmse"], marker="o", label="RMSE") + ax.set_ylabel("Erreur (°C)" if target == "temperature" else "Erreur (unité)") + else: + ax.plot(df["horizon_h"], df["f1"], marker="o", label="F1") + ax.plot(df["horizon_h"], df["brier"], marker="o", label="Brier") + ax.set_ylabel("Score") + ax.set_xlabel("Horizon (heures)") + ax.set_title(f"Chronos (small, réglages prudents) – {target}") + ax.grid(True, linestyle=":", alpha=0.4) + ax.legend() + fig.tight_layout() + fig.savefig(output_path, dpi=150) + plt.close(fig) + + +def main() -> None: + pipeline = ChronosPipeline.from_pretrained( + MODEL_ID, + device_map="auto", + dtype="auto", + ) + + rows: list[dict[str, object]] = [] + + for target, meta in TARGETS.items(): + series = _load_series(target) + context_arr, target_series = _prepare_window(series, CONTEXT_H, MAX_HORIZON_H) + + context_tensor = torch.tensor(context_arr, dtype=torch.float32) + forecasts = pipeline.predict( + [context_tensor], + prediction_length=MAX_HORIZON_H, + num_samples=NUM_SAMPLES, + ) + forecast_mean = forecasts.mean(0) + if forecast_mean.ndim == 2: + forecast_mean = forecast_mean[0] + forecast_series = pd.Series( + np.asarray(forecast_mean).ravel(), + index=pd.date_range(target_series.index[0], periods=MAX_HORIZON_H, freq=RESAMPLE_RULE), + ) + + for h in HORIZONS_H: + y_true = target_series.iloc[:h] + y_pred = forecast_series.iloc[:h] + if meta["kind"] == "reg": + mae = float((y_pred - y_true).abs().mean()) + rmse = float(np.sqrt(((y_pred - y_true) ** 2).mean())) + rows.append( + {"target": target, "kind": "reg", "horizon_h": h, "mae": mae, "rmse": rmse} + ) + else: + y_true_bin = (y_true > 0).astype(int) + y_pred_bin = (y_pred > 0).astype(int) + if y_true_bin.sum() == 0: + f1 = 0.0 + else: + tp = ((y_true_bin == 1) & (y_pred_bin == 1)).sum() + fp = ((y_true_bin == 0) & (y_pred_bin == 1)).sum() + fn = ((y_true_bin == 1) & (y_pred_bin == 0)).sum() + prec = tp / (tp + fp) if tp + fp > 0 else 0.0 + rec = tp / (tp + fn) if tp + fn > 0 else 0.0 + f1 = 2 * prec * rec / (prec + rec) if prec + rec > 0 else 0.0 + proba = y_pred.clip(0, 1) + brier = float(((proba - y_true_bin) ** 2).mean()) + rows.append( + {"target": target, "kind": "cls", "horizon_h": h, "f1": float(f1), "brier": brier} + ) + + out_prefix = f"{target}_{MODEL_ID.replace('/', '__')}_tuned" + DATA_DIR.mkdir(parents=True, exist_ok=True) + pd.concat([forecast_series.rename("y_pred"), target_series.rename("y_true")], axis=1).to_csv( + DATA_DIR / f"chronos_tuned_forecast_{out_prefix}.csv", + index=True, + ) + + df_metrics = pd.DataFrame(rows) + DATA_DIR.mkdir(parents=True, exist_ok=True) + metrics_path = DATA_DIR / "chronos_tuned_metrics.csv" + df_metrics.to_csv(metrics_path, index=False) + + for target in df_metrics["target"].unique(): + sub = df_metrics[df_metrics["target"] == target].sort_values("horizon_h") + _plot_metrics(sub, target, FIG_DIR / f"chronos_tuned_{target}.png") + + print(df_metrics.to_string(index=False, float_format=lambda x: f"{x:.3f}")) + print(f"✔ Sauvegardé : {metrics_path}") + print("✔ Figures par cible (réglages prudents) dans figures/chronos_tuned_.png") + + +if __name__ == "__main__": + main() diff --git a/requirements.txt b/requirements.txt index bebe89d..a1577c9 100644 --- a/requirements.txt +++ b/requirements.txt @@ -15,6 +15,7 @@ astral # Modèles statistiques / ML scikit-learn statsmodels +chronos-forecasting # Gestion propre des secrets/config (.env) python-dotenv