DARTS多模型集成回归训练报错：无法构建有效训练样本

问题背景 #

• 数据集：40条月度时间序列数据，包含1个目标列和9个协变量列，需预测2022年11月的目标值
• 单个DARTS模型（XGBoost、LightGBM、CatBoost、GRU）单独训练正常，但组合成RegressionEnsembleModel时抛出ValueError

可复现代码 #

import numpy as np
import pandas as pd
from darts import TimeSeries
from darts.models import XGBModel, LightGBMModel, CatBoostModel, BlockRNNModel, RandomForest, RegressionEnsembleModel
from darts.metrics import SymmetricMeanAbsolutePercentageError

col1 = list(np.arange(1,40))
col2 = list(np.arange(41,80))
col3 = list(np.arange(81,120))
col4 = list(np.arange(121,160))

month = pd.date_range('2019-08-01','2022-10-01',freq='1MS').strftime("%Y-%b").tolist()

dt = {'col1': col1,'col2':col2,'col3':col3,'target':col4,'month':month}
df = pd.DataFrame(dt)
df['month'] = pd.to_datetime(df['month'])
df.index = df['month']
df = df.drop(['month'],axis = 1)


X_train_t = TimeSeries.from_dataframe(df,value_cols=list(df.columns.drop(["col3"]) ),freq="1MS")
y_train_t = TimeSeries.from_dataframe(df,freq = '1MS',value_cols = 'target')

#Initializing model and fitting
torch_metrics = SymmetricMeanAbsolutePercentageError()
model1 = XGBModel(lags= [-1,-2,-3], lags_past_covariates=[-1,-2,-3],random_state = 42,output_chunk_length=1)
model2 = LightGBMModel(lags= [-1,-2,-3], lags_past_covariates=[-1,-2,-3],random_state = 42,output_chunk_length=1)
model3 = CatBoostModel(lags= [-1,-2,-3], lags_past_covariates=[-1,-2,-3], lags_future_covariates=None, output_chunk_length=1,random_state = 42)
model4 = BlockRNNModel(input_chunk_length=12,model = 'GRU',dropout=0.2,random_state = 42,output_chunk_length = 1)
forecasting_models = [model1,model2,model3,model4]

regression_model  =RandomForest(lags=None, lags_past_covariates=None, lags_future_covariates=[0], output_chunk_length=1, add_encoders=None, n_estimators=100, max_depth=None)
model_ensemble = RegressionEnsembleModel(forecasting_models = forecasting_models, regression_train_n_points =38,regression_model = regression_model)
history = model_ensemble.fit(y_train_t,past_covariates = X_train_t)

pred = model_ensemble.predict(n=1, series = y_train_t,past_covariates = X_train_t)

报错信息 #

ValueError                                Traceback (most recent call last)
/tmp/ipykernel_27/3395707719.py in <module>
     12 regression_model  =RandomForest(lags=None, lags_past_covariates=None, lags_future_covariates=[0], output_chunk_length=1, add_encoders=None, n_estimators=100, max_depth=None)
     13 model_ensemble = RegressionEnsembleModel(forecasting_models = forecasting_models, regression_train_n_points =38,regression_model = regression_model)
---> 14 history = model_ensemble.fit(y_train_t,past_covariates = X_train_t)
     15 
     16 pred = model_ensemble.predict(n=1, series = y_train_t,past_covariates = X_train_t)

/opt/conda/lib/python3.7/site-packages/darts/models/forecasting/regression_ensemble_model.py in fit(self, series, past_covariates, future_covariates)
    119                 series=forecast_training,
    120                 past_covariates=past_covariates,
---> 121                 future_covariates=future_covariates,
    122             )
    123 

/opt/conda/lib/python3.7/site-packages/darts/models/forecasting/forecasting_model.py in _fit_wrapper(self, series, past_covariates, future_covariates)
   1820             future_covariates=future_covariates
   1821             if self.supports_future_covariates
-> 1822             else None,
   1823         )
   1824 

/opt/conda/lib/python3.7/site-packages/darts/models/forecasting/xgboost.py in fit(self, series, past_covariates, future_covariates, val_series, val_past_covariates, val_future_covariates, max_samples_per_ts, **kwargs)
    219             future_covariates=future_covariates,
    220             max_samples_per_ts=max_samples_per_ts,
---> 221             **kwargs,
    222         )
    223 

/opt/conda/lib/python3.7/site-packages/darts/models/forecasting/regression_model.py in fit(self, series, past_covariates, future_covariates, max_samples_per_ts, n_jobs_multioutput_wrapper, **kwargs)
    481 
    482         self._fit_model(
---> 483             series, past_covariates, future_covariates, max_samples_per_ts, **kwargs
    484         )
    485 

/opt/conda/lib/python3.7/site-packages/darts/models/forecasting/regression_model.py in _fit_model(self, target_series, past_covariates, future_covariates, max_samples_per_ts, **kwargs)
    363 
    364         training_samples, training_labels = self._create_lagged_data(
---> 365             target_series, past_covariates, future_covariates, max_samples_per_ts
    366         )
    367 

/opt/conda/lib/python3.7/site-packages/darts/models/forecasting/regression_model.py in _create_lagged_data(self, target_series, past_covariates, future_covariates, max_samples_per_ts)
    334             lags_future_covariates=lags_future_covariates,
    335             max_samples_per_ts=max_samples_per_ts,
---> 336             multi_models=self.multi_models,
    337         )
    338 

/opt/conda/lib/python3.7/site-packages/darts/utils/data/tabularization.py in _create_lagged_data(target_series, output_chunk_length, past_covariates, future_covariates, lags, lags_past_covariates, lags_future_covariates, max_samples_per_ts, is_training, multi_models)
    150             "Unable to build any training samples of the target series "
    151             + (f"at index {idx} " if len(target_series) > 1 else "")
---> 152             + "and the corresponding covariate series; "
    153             "There is no time step for which all required lags are available and are not NaN values.",
    154         )

/opt/conda/lib/python3.7/site-packages/darts/logging.py in raise_if(condition, message, logger)
    102         if `condition` is satisfied
    103     """
--> 104     raise_if_not(not condition, message, logger)
    105 
    106 

/opt/conda/lib/python3.7/site-packages/darts/logging.py in raise_if_not(condition, message, logger)
     76     if not condition:
     77         logger.error("ValueError: " + message)
--> 78         raise ValueError(message)
     79 
     80 

ValueError: Unable to build any training samples of the target series and the corresponding covariate series; There is no time step for which all required lags are available and are not NaN values. 

错误原因分析 #

1. 集成模型训练样本数设置不合理：regression_train_n_points=38超出了部分基础模型的有效样本生成能力。例如BlockRNNModel设置了input_chunk_length=12，需要至少12条历史数据才能生成预测，有效训练样本数应为总样本数 - input_chunk_length = 40-12=28，设置38会导致模型无法生成足够的有效预测样本。
2. 回归模型参数错误：设置lags_future_covariates=[0]，但当前没有提供未来协变量数据，模型尝试获取不存在的未来协变量滞后项，导致样本构建失败。

解决方案 #

1. 调整集成模型训练样本数 #

将regression_train_n_points设置为所有基础模型中最小的有效样本数，这里匹配GRU模型的28。

2. 修正回归模型参数 #

移除未来协变量滞后设置，改为lags_future_covariates=None，因为没有可用的未来协变量数据。

修正后的完整代码 #

import numpy as np
import pandas as pd
from darts import TimeSeries
from darts.models import (XGBModel, LightGBMModel, CatBoostModel, 
                          BlockRNNModel, RandomForest, RegressionEnsembleModel)
from darts.metrics import SymmetricMeanAbsolutePercentageError

# 构建数据集
col1 = list(np.arange(1,40))
col2 = list(np.arange(41,80))
col3 = list(np.arange(81,120))
col4 = list(np.arange(121,160))

month = pd.date_range('2019-08-01','2022-10-01',freq='1MS').strftime("%Y-%b").tolist()

dt = {'col1': col1,'col2':col2,'col3':col3,'target':col4,'month':month}
df = pd.DataFrame(dt)
df['month'] = pd.to_datetime(df['month'])
df.index = df['month']
df = df.drop(['month'],axis = 1)

# 转换为Darts TimeSeries对象
X_train_t = TimeSeries.from_dataframe(df, value_cols=list(df.columns.drop(["col3"])), freq="1MS")
y_train_t = TimeSeries.from_dataframe(df, freq='1MS', value_cols='target')

# 初始化基础预测模型
model1 = XGBModel(lags= [-1,-2,-3], lags_past_covariates=[-1,-2,-3], random_state=42, output_chunk_length=1)
model2 = LightGBMModel(lags= [-1,-2,-3], lags_past_covariates=[-1,-2,-3], random_state=42, output_chunk_length=1)
model3 = CatBoostModel(lags= [-1,-2,-3], lags_past_covariates=[-1,-2,-3], lags_future_covariates=None, output_chunk_length=1, random_state=42)
model4 = BlockRNNModel(input_chunk_length=12, model='GRU', dropout=0.2, random_state=42, output_chunk_length=1)
forecasting_models = [model1, model2, model3, model4]

# 修正回归模型参数：移除未来协变量滞后设置
regression_model = RandomForest(lags=None, lags_past_covariates=None, lags_future_covariates=None, 
                                output_chunk_length=1, add_encoders=None, n_estimators=100, max_depth=None)

# 调整集成模型的训练样本数：匹配GRU模型的有效样本数
model_ensemble = RegressionEnsembleModel(forecasting_models=forecasting_models, 
                                         regression_train_n_points=28, 
                                         regression_model=regression_model)

# 训练集成模型
history = model_ensemble.fit(y_train_t, past_covariates=X_train_t)

# 预测2022年11月目标值
pred = model_ensemble.predict(n=1, series=y_train_t, past_covariates=X_train_t)
print(pred)

内容的提问来源于stack exchange，提问作者Crossfit_Jesus