共変量に欠損があるとChronos-2の予測はどう変わるか

import pandas as pd
from sklearn.datasets import fetch_openml

bike_sharing = fetch_openml(
    "Bike_Sharing_Demand",
    version=2,
    as_frame=True,
)

df = bike_sharing.frame.copy()
df["year_num"] = df["year"].astype(int)
df["month_num"] = df["month"].astype(int)
df["hour_num"] = df["hour"].astype(int)
df["day_num"] = (
    df.groupby(["year_num", "month_num"], sort=False)["hour_num"]
    .transform(lambda s: s.diff().lt(0).cumsum() + 1)
)
df["日時"] = pd.to_datetime(
    {
        "year": df["year_num"] + 2011,
        "month": df["month_num"],
        "day": df["day_num"],
    }
) + pd.to_timedelta(df["hour_num"], unit="h")
df = (
    df.rename(columns={"temp": "気温", "count": "利用台数"})
    .assign(時系列ID="自転車シェアリング")
    .sort_values(["時系列ID", "日時"])
    .reset_index(drop=True)
)

prediction_start = pd.Timestamp("2012-12-31 00:00:00")
context_df = df[df["日時"] < prediction_start][["時系列ID", "日時", "気温", "利用台数"]].copy()
future_df = df[df["日時"] >= prediction_start][["時系列ID", "日時", "気温"]].copy()

from chronos import Chronos2Pipeline

pipeline = Chronos2Pipeline.from_pretrained(
    "amazon/chronos-2",
    device_map="cpu",
)

pred_chronos = pipeline.predict_df(
    df=context_df,
    future_df=future_df,
    prediction_length=24,
    quantile_levels=[0.1, 0.5, 0.9],
    id_column="時系列ID",
    timestamp_column="日時",
    target="利用台数",
    freq="h",
)

from autogluon.timeseries import TimeSeriesPredictor

train = context_df.rename(
    columns={"時系列ID": "item_id", "日時": "timestamp"}
)
known = future_df.rename(
    columns={"時系列ID": "item_id", "日時": "timestamp"}
)

predictor = TimeSeriesPredictor(
    prediction_length=24,
    target="利用台数",
    freq="h",
    known_covariates_names=["気温"],
    quantile_levels=[0.1, 0.5, 0.9],
).fit(
    train_data=train,
    hyperparameters={"Chronos2": {"model_path": "amazon/chronos-2"}},
)

pred_autogluon = (
    predictor.predict(data=train, known_covariates=known)
    .reset_index()
    .rename(columns={"item_id": "時系列ID", "timestamp": "日時"})
)

from darts import TimeSeries
from darts.models import Chronos2Model
from darts.utils.likelihood_models import QuantileRegression

series = TimeSeries.from_dataframe(
    df=context_df.assign(利用台数=context_df["利用台数"].astype("float32")),
    time_col="日時",
    value_cols="利用台数",
    freq="h",
)
cov_df = pd.concat(
    [context_df[["日時", "気温"]], future_df[["日時", "気温"]]],
    ignore_index=True,
).sort_values("日時")
future_cov = TimeSeries.from_dataframe(
    df=cov_df.assign(気温=cov_df["気温"].astype("float32")),
    time_col="日時",
    value_cols="気温",
    freq="h",
)
model = Chronos2Model(
    input_chunk_length=8192,
    output_chunk_length=24,
    likelihood=QuantileRegression(quantiles=[0.1, 0.5, 0.9]),
    hub_model_name="amazon/chronos-2",
    pl_trainer_kwargs={"accelerator": "cpu"},
)
model.fit(series=series, future_covariates=future_cov)
forecast = model.predict(
    n=24,
    future_covariates=future_cov,
    predict_likelihood_parameters=True,
)
pred_darts = forecast.to_dataframe().reset_index().rename(
    columns={
        "利用台数_q0.100": "0.1",
        "利用台数_q0.500": "0.5",
        "利用台数_q0.900": "0.9",
    }
)
pred_darts["時系列ID"] = "自転車シェアリング"

import japanize_matplotlib
import matplotlib.dates as mdates
import matplotlib.pyplot as plt

history = context_df.tail(48)
actual_df = df[df["日時"] >= prediction_start][["時系列ID", "日時", "利用台数"]].copy()
observed = pd.concat(
    [history[["日時", "利用台数"]], actual_df[["日時", "利用台数"]]],
    ignore_index=True,
).sort_values("日時")

forecast_colors = {
    "chronos-forecasting": "#d62728",
    "AutoGluon-TimeSeries": "#1f77b4",
    "Darts": "#ff7f0e",
}
preds = [
    (pred_chronos, "chronos-forecasting"),
    (pred_autogluon, "AutoGluon-TimeSeries"),
    (pred_darts, "Darts"),
]

fig, ax = plt.subplots(figsize=(14, 6), dpi=100)
ax.plot(
    observed["日時"],
    observed["利用台数"],
    label="実測値",
    color="black",
    marker="o",
    markersize=3,
)
for pred_df, method_name in preds:
    ax.plot(
        pd.to_datetime(pred_df["日時"]),
        pred_df["0.5"],
        label=method_name,
        color=forecast_colors[method_name],
        marker="o",
        markersize=4,
    )
ax.axvline(prediction_start, color="gray", linestyle="--", label="予測開始")
ax.set_xlim(observed["日時"].min(), observed["日時"].max())
ax.set(title="過去と未来の気温が欠損した24時間予測", xlabel="日時", ylabel="利用台数（台）")
ax.xaxis.set_major_formatter(mdates.DateFormatter("%m/%d %H:%M"))
ax.grid(alpha=0.3)
ax.legend()
plt.xticks(rotation=45)
plt.tight_layout()
fig.savefig("過去と未来の気温が欠損した24時間予測.png")

from chronos import Chronos2Pipeline

pipeline = Chronos2Pipeline.from_pretrained(
    "amazon/chronos-2",
    device_map="cpu",
)

pred_chronos = pipeline.predict_df(
    df=context_df,
    future_df=future_df,
    prediction_length=24,
    quantile_levels=[0.1, 0.5, 0.9],
    id_column="時系列ID",
    timestamp_column="日時",
    target="利用台数",
    freq="h",
)

from autogluon.timeseries import TimeSeriesPredictor

train = context_df.rename(
    columns={"時系列ID": "item_id", "日時": "timestamp"}
)
known = future_df.rename(
    columns={"時系列ID": "item_id", "日時": "timestamp"}
)

predictor = TimeSeriesPredictor(
    prediction_length=24,
    target="利用台数",
    freq="h",
    known_covariates_names=["気温"],
    quantile_levels=[0.1, 0.5, 0.9],
).fit(
    train_data=train,
    hyperparameters={"Chronos2": {"model_path": "amazon/chronos-2"}},
)

pred_autogluon = (
    predictor.predict(data=train, known_covariates=known)
    .reset_index()
    .rename(columns={"item_id": "時系列ID", "timestamp": "日時"})
)

from darts import TimeSeries
from darts.models import Chronos2Model
from darts.utils.likelihood_models import QuantileRegression

series = TimeSeries.from_dataframe(
    df=context_df.assign(利用台数=context_df["利用台数"].astype("float32")),
    time_col="日時",
    value_cols="利用台数",
    freq="h",
)
cov_df = pd.concat(
    [context_df[["日時", "気温"]], future_df[["日時", "気温"]]],
    ignore_index=True,
).sort_values("日時")
future_cov = TimeSeries.from_dataframe(
    df=cov_df.assign(気温=cov_df["気温"].astype("float32")),
    time_col="日時",
    value_cols="気温",
    freq="h",
)
model = Chronos2Model(
    input_chunk_length=8192,
    output_chunk_length=24,
    likelihood=QuantileRegression(quantiles=[0.1, 0.5, 0.9]),
    hub_model_name="amazon/chronos-2",
    pl_trainer_kwargs={"accelerator": "cpu"},
)
model.fit(series=series, future_covariates=future_cov)
forecast = model.predict(
    n=24,
    future_covariates=future_cov,
    predict_likelihood_parameters=True,
)
pred_darts = forecast.to_dataframe().reset_index().rename(
    columns={
        "利用台数_q0.100": "0.1",
        "利用台数_q0.500": "0.5",
        "利用台数_q0.900": "0.9",
    }
)
pred_darts["時系列ID"] = "自転車シェアリング"

import japanize_matplotlib
import matplotlib.dates as mdates
import matplotlib.pyplot as plt

history = context_df.tail(48)
actual_df = df[df["日時"] >= prediction_start][["時系列ID", "日時", "利用台数"]].copy()
observed = pd.concat(
    [history[["日時", "利用台数"]], actual_df[["日時", "利用台数"]]],
    ignore_index=True,
).sort_values("日時")

forecast_colors = {
    "chronos-forecasting": "#d62728",
    "AutoGluon-TimeSeries": "#1f77b4",
    "Darts": "#ff7f0e",
}
preds = [
    (pred_chronos, "chronos-forecasting"),
    (pred_autogluon, "AutoGluon-TimeSeries"),
    (pred_darts, "Darts"),
]

fig, ax = plt.subplots(figsize=(14, 6), dpi=100)
ax.plot(
    observed["日時"],
    observed["利用台数"],
    label="実測値",
    color="black",
    marker="o",
    markersize=3,
)
for pred_df, method_name in preds:
    ax.plot(
        pd.to_datetime(pred_df["日時"]),
        pred_df["0.5"],
        label=method_name,
        color=forecast_colors[method_name],
        marker="o",
        markersize=4,
    )
ax.axvline(prediction_start, color="gray", linestyle="--", label="予測開始")
ax.set_xlim(observed["日時"].min(), observed["日時"].max())
ax.set(title="過去と未来の気温を埋めた24時間予測", xlabel="日時", ylabel="利用台数（台）")
ax.xaxis.set_major_formatter(mdates.DateFormatter("%m/%d %H:%M"))
ax.grid(alpha=0.3)
ax.legend()
plt.xticks(rotation=45)
plt.tight_layout()
fig.savefig("過去と未来の気温を埋めた24時間予測.png")