docs: add notebook

examples/energy_forecasting.py

@@ -0,0 +1,204 @@
+# ---
+# jupyter:
+#   jupytext:
+#     cell_metadata_filter: -all
+#     formats: ipynb,py
+#     text_representation:
+#       extension: .py
+#       format_name: light
+#       format_version: '1.5'
+#       jupytext_version: 1.16.1
+#   kernelspec:
+#     display_name: .venv
+#     language: python
+#     name: python3
+# ---
+
+# +
+import altair as alt
+import polars as pl
+
+# Needed for larger datasets
+alt.data_transformers.enable("vegafusion")
+
+# alt.renderers.enable("browser")
+# -
+
+X = pl.read_csv("X_train.csv")
+y = pl.read_csv("Y_train.csv")
+X = X.with_columns(pl.col("Time").str.to_datetime("%d/%m/%Y %H:%M"))
+X_y = X.join(y, on="ID")
+X_WF1 = X.filter(pl.col("WF") == "WF1")
+X_WF1.filter(pl.col("Time").dt.hour() == 0)
+X_WF1["Time"].max()
+X_y_WF1 = X_y.filter(pl.col("WF") == "WF1")
+
+alt.Chart(X_WF1).mark_point().encode(
+    x="Time", y="NWP1_00h_D-2_U", tooltip=alt.Tooltip("Time", format="%H:%M")
+).interactive()
+(
+    list(
+        X_WF1.filter(pl.col("NWP1_00h_D-2_U").is_not_null())["Time"].dt.hour().unique()
+    ),
+    list(
+        X_WF1.filter(pl.col("NWP1_06h_D-2_U").is_not_null())["Time"].dt.hour().unique()
+    ),
+    list(
+        X_WF1.filter(pl.col("NWP1_12h_D-2_U").is_not_null())["Time"].dt.hour().unique()
+    ),
+    list(
+        X_WF1.filter(pl.col("NWP1_18h_D-2_U").is_not_null())["Time"].dt.hour().unique()
+    ),
+)
+with pl.Config(tbl_rows=-1):
+    print(X_WF1)
+
+# +
+# Complementary data
+X_comp = pl.read_csv("WindFarms_complementary_data.csv", separator=";")
+X_comp = X_comp.filter(pl.col("Time (UTC)").is_not_null())
+X_comp = X_comp.with_columns(pl.col("Time (UTC)").str.to_datetime("%d/%m/%Y %H:%M"))
+
+(
+    alt.Chart(
+        X_comp.filter(
+            (pl.col("Wind Farm") == "WF1") & (pl.col("Wind Turbine") == "TE1")
+        ).with_columns(
+            (pl.col("Wind direction (�)") - pl.col("Nacelle direction (�)")).alias(
+                "Nacelle misalignment (deg)"
+            )
+        )
+    )
+    .mark_point()
+    .encode(x="Time (UTC)", y="Nacelle misalignment (deg)")
+)
+
+
+# +
+# Statistics of power production
+X_y_WF1["Production"].describe()
+
+# Histogram
+(
+    alt.Chart(X_y_WF1)
+    .mark_bar()
+    .encode(alt.X("Production", bin=alt.Bin(step=0.5)), y="count()")
+    .properties(width=800)
+)
+
+# The distribution of production is heavily right skewed. The median is 0.82 MW.
+# According to Our World in Data 2017 (https://ourworldindata.org/scale-for-electricity), a French person consumes 0.019 MWh/day
+# -
+
+# Total production for the month
+(
+    alt.Chart(X_y_WF1)
+    .mark_line()
+    .encode(x="yearmonth(Time):T", y="sum(Production)")
+    .properties(width=800)
+)
+
+
+# +
+
+# There's a big drop in December 2018, compared to November and January. Is it because demand dropped, or because the data was corrupted, or because the wind farms were in maintenance?
+
+(
+    alt.Chart(X_y_WF1.filter(pl.col("Time").dt.month() == 12))
+    .mark_point()
+    .encode(x="Time", y="Production")
+    .properties(width=3000)
+)
+
+# The power production was very near zero for 9 consecutive days, from 12 December to 21 December.
+
+# +
+# Trying out a classic sktime forecasting workflow
+
+import numpy as np
+from sktime.forecasting.compose import make_reduction
+from sklearn.ensemble import RandomForestRegressor
+from sktime.performance_metrics.forecasting import MeanAbsolutePercentageError
+from sktime.split import temporal_train_test_split
+
+# Format the data to be sktime-friendly
+y_train, y_test, X_train, X_test = temporal_train_test_split(
+    y=X_y_WF1["Production"].to_pandas(), X=X_WF1.drop(["ID", "WF", "Time"]).to_pandas()
+)
+
+fh = np.arange(1, len(y_test) + 1)  # forecasting horizon
+regressor = RandomForestRegressor()
+forecaster = make_reduction(
+    regressor,
+    strategy="recursive",
+    window_length=12,
+)
+# -
+
+# Takes a while
+forecaster.fit(y=y_train, X=X_train, fh=fh)
+
+y_pred = forecaster.predict(fh=fh, X=X_test)
+smape = MeanAbsolutePercentageError()
+smape(y_test, y_pred)
+
+# Show predictions with test
+df = pl.DataFrame({"y_pred": y_pred, "y_test": y_test}).with_row_index().melt("index")
+alt.Chart(df).mark_line().encode(x="index", y="value", color="variable")
+# It's not great
+
+alt.Chart(X_y_WF1).mark_line().encode(x="Time", y="Production").properties(
+    width=2000, height=400
+)
+# +
+# Average production depending on the day of the week
+(
+    alt.Chart(X_y_WF1.with_columns(pl.col("Time").dt.weekday().alias("Day of week")))
+    .mark_bar()
+    .encode(x="Day of week", y="mean(Production)")
+    + alt.Chart(X_y_WF1.with_columns(pl.col("Time").dt.weekday().alias("Day of week")))
+    .mark_errorbar(extent="iqr")
+    .encode(x="Day of week", y="Production")
+)
+# 1 is Monday, 7 is Sunday
+# Top production is on Mondays and Sundays, bottom is Thursdays
+
+# Error bars are the IQR
+# +
+# Average production depending on month of the year
+base = alt.Chart(X_y_WF1.with_columns(pl.col("Time").dt.month().alias("Month")))
+(
+    base.mark_bar().encode(x="Month", y="mean(Production)")
+    + base.mark_errorbar(extent="iqr").encode(x="Month", y="Production")
+)
+
+# 1 is January, 12 is December
+# Top production is January by far, bottom is August/September
+# December is low, as mentioned earlier
+
+# The error bars show the inter-quartile range (bottom is 25% quantile, top is 75% quantile)
+# This way we can clearly see that a lot of the data is very close to 0
+# -
+import polars.selectors as cs
+
+nwp1 = X_y_WF1.select(cs.matches("Time") | (cs.matches("NWP1") & cs.matches("_U")))
+alt.Chart(nwp1.melt(id_vars="Time")).mark_point().encode(
+    x="Time", y="value", color="variable"
+).properties(width=5000, height=500)
+X_y_WF1.with_columns(
+    mean_U=pl.mean_horizontal((cs.matches("NWP1") & cs.matches("_U"))),
+    min_U=pl.min_horizontal((cs.matches("NWP1") & cs.matches("_U"))),
+    max_U=pl.max_horizontal((cs.matches("NWP1") & cs.matches("_U"))),
+)
+
+(
+    alt.Chart(nwp1.melt(id_vars="Time")).mark_line().encode(x="Time", y="mean(value)")
+    + alt.Chart(nwp1.melt(id_vars="Time"))
+    .mark_errorband(extent="ci")
+    .encode(x="Time", y="value")
+).properties(width=5000, height=500)
+
+# +
+# Correlation between the different variables
+
+X_y_WF1