Kaggle Solo第一！

簡介

第一名方案主要由4個XGBoost模型（2個不同的標簽）以及2個GRU模型（2個不同模型）組成。

驗證策略

使用前500天數據進行訓練並且將剩余的時間用做驗證。

特征

# client_df
client_df = client_df.drop(["date"]).sort(["data_block_id"], descending=False)
df = df.join(client_df.select(["county", "is_business", "product_type", "data_block_id","eic_count","installed_capacity"]), how="left", on=["county", "is_business", "product_type", "data_block_id"])
# electricity
electricity_df = electricity_df.drop(["origin_date"]).rename({"forecast_date":"datetime"}).with_columns([(pl.col("datetime").str.to_datetime()+pl.duration(days=1)).alias("datetime"), (pl.col("euros_per_mwh").abs()+0.1).alias("euros_per_mwh"),])
df = df.join(electricity_df[["data_block_id","datetime","euros_per_mwh"]], how="left", on=["data_block_id","datetime"])
# fw_df 
fw_df = fw_df.with_columns(
(pl.col("origin_datetime").str.to_datetime()+pl.duration(hours="hours_ahead")).alias("datetime"),
pl.col("latitude").cast(pl.datatypes.Float32).round(1),
pl.col("longitude").cast(pl.datatypes.Float32).round(1),
pl.col("data_block_id").cast(pl.datatypes.Int64),
).join(weather_station_to_county_mapping.drop(["county_name"]).with_columns(
pl.col("latitude").cast(pl.datatypes.Float32).round(1),
pl.col("longitude").cast(pl.datatypes.Float32).round(1),
), how="left",on=["longitude","latitude"]).drop(["longitude","latitude","origin_datetime"])
fw_df = fw_df.group_by(["county","datetime","data_block_id"]).agg([pl.mean(col).alias("fw_{}".format(col)) for col in forecast_weather_cols]).with_columns([pl.col("county").cast(pl.datatypes.Int64),
pl.col("data_block_id").cast(pl.datatypes.Int64),])
df = df.join(fw_df, how="left", on=["county","datetime","data_block_id"]).with_columns([(pl.col("installed_capacity")*pl.col("fw_surface_solar_radiation_downwards") / (pl.col("fw_temperature") + 273.15)).alias("fw_new_feature"),])
# hw_df 
hw_df = hw_df.with_columns(
(pl.col("datetime").str.to_datetime()+pl.duration(days=1)).alias("datetime"),
pl.col("latitude").cast(pl.datatypes.Float32).round(1),
pl.col("longitude").cast(pl.datatypes.Float32).round(1),
pl.col("data_block_id").cast(pl.datatypes.Int64),
).join(weather_station_to_county_mapping.drop(["county_name"]).with_columns(
pl.col("latitude").cast(pl.datatypes.Float32).round(1),
pl.col("longitude").cast(pl.datatypes.Float32).round(1),
), how="left",on=["longitude","latitude"]).drop(["longitude","latitude"])
hw_df = hw_df.group_by(["county","datetime","data_block_id"]).agg([pl.mean(col).alias("hw_{}".format(col)) for col in historical_weather_cols]).with_columns([ pl.when(pl.col("datetime").dt.hour()>10).then(pl.col("datetime")+pl.duration(days=1)).otherwise(pl.col("datetime")).alias("datetime"),
pl.col("county").cast(pl.datatypes.Int64),
pl.col("data_block_id").cast(pl.datatypes.Int64),])
df = df.join(hw_df, how="left", on=["county","datetime","data_block_id"])


Target

df = df.with_columns([
pl.when(pl.col('target').shift(2).over(['prediction_unit_id','hour']).is_not_null()).then(pl.col('target').shift(2).over(['prediction_unit_id','hour'])).when(pl.col('target').shift(3).over(['prediction_unit_id','hour']).is_not_null()).then(pl.col('target').shift(3).over(['prediction_unit_id','hour'])).when(pl.col('target').shift(4).over(['prediction_unit_id','hour']).is_not_null()).then(pl.col('target').shift(4).over(['prediction_unit_id','hour'])).when(pl.col('target').shift(2).mean().over(['datetime']).is_null()).then(pl.col('target').mean()).otherwise(pl.col('target').shift(2).mean().over(['datetime'])).fill_null(0.0).cast(pl.Float32).alias('target_shift_tmp'),
pl.when(pl.col('data_block_id')<=1).then(pl.col('installed_capacity').shift(-2).over(['prediction_unit_id','hour'])).when(pl.col('installed_capacity').shift(1).over(['prediction_unit_id','hour']).is_not_null()).then(pl.col('installed_capacity').shift(1).over(['prediction_unit_id','hour'])).when(pl.col('installed_capacity').shift(2).over(['prediction_unit_id','hour']).is_not_null()).then(pl.col('installed_capacity').shift(2).over(['prediction_unit_id','hour'])).when(pl.col('installed_capacity').shift(3).over(['prediction_unit_id','hour']).is_not_null()).then(pl.col('installed_capacity').shift(3).over(['prediction_unit_id','hour'])).when(pl.col('installed_capacity').mean().over(['datetime']).is_null()).then(pl.col('installed_capacity').mean()).otherwise(pl.col('installed_capacity').mean().over(['datetime'])).cast(pl.Float32).alias('installed_capacity_tmp'),
])
df = df.with_columns([ pl.when(pl.col('installed_capacity').is_null()).then(pl.col('installed_capacity_tmp')).otherwise(pl.col('installed_capacity')).clip(10,None).cast(pl.Float32).alias('installed_capacity_tmp'), #clip(10,None) 
]).with_columns([
pl.when(pl.col('installed_capacity_tmp')<2*pl.col('target_shift_tmp').abs()).then(2*pl.col('target_shift_tmp').abs()).otherwise(pl.col('installed_capacity_tmp')).cast(pl.Float32).alias('installed_capacity_tmp'),
])
df = df.with_columns([
((pl.col('target')-pl.col('target_shift_tmp')) / pl.col('installed_capacity_tmp')).alias('target2'),
(pl.col('target') / pl.col('installed_capacity_tmp')).alias('target3'),
])

模型

線上學習策略

線上學習策略可以使用更多數據，這對public learderboard幫助較小，但private leaderboard可能會帶來更多的幫助。

參考文獻

1. https://www.kaggle.com/competitions/predict-energy-behavior-of-prosumers/discussion/472793

2. https://www.kaggle.com/competitions/predict-energy-behavior-of-prosumers/overview

作者：hyd