鲁棒预测

!pip install neuralforecast

import logging
import numpy as np
import pandas as pd

import matplotlib.pyplot as plt
from random import random
from random import randint
from random import seed

from neuralforecast import NeuralForecast
from neuralforecast.utils import AirPassengersDF

from neuralforecast.models import NHITS
from neuralforecast.losses.pytorch import MQLoss, DistributionLoss, HuberMQLoss

from utilsforecast.losses import mape, mqloss
from utilsforecast.evaluation import evaluate

logging.getLogger("pytorch_lightning").setLevel(logging.ERROR)

# Original Box-Cox AirPassengers 
# as defined in neuralforecast.utils
Y_df = AirPassengersDF.copy() 
plt.plot(Y_df.y)
plt.ylabel('Monthly Passengers')
plt.xlabel('Timestamp [t]')
plt.grid()

# Here we add some artificial outliers to AirPassengers
seed(1)
for i in range(len(Y_df)):
    factor = randint(2, 4)
    if random() > 0.97:
        Y_df.loc[i, "y"] += factor * Y_df["y"].std()

plt.plot(Y_df.y)
plt.ylabel('Monthly Passengers + Noise')
plt.xlabel('Timestamp [t]')
plt.grid()

# Split datasets into train/test 
# Last 12 months for test
Y_train_df = Y_df.groupby('unique_id').head(-12)
Y_test_df = Y_df.groupby('unique_id').tail(12)
Y_test_df

horizon = 12
level = [50, 80]

# Try different hyperparmeters to improve accuracy.
models = [NHITS(h=horizon,                           # Forecast horizon
                input_size=2 * horizon,              # Length of input sequence
                loss=HuberMQLoss(level=level),    # Robust Huber Loss
                valid_loss=MQLoss(level=level),   # Validation signal
                max_steps=500,                       # Number of steps to train
                dropout_prob_theta=0.6,              # Dropout to robustify vs outlier lag inputs
                #early_stop_patience_steps=2,        # Early stopping regularization patience
                val_check_steps=10,                  # Frequency of validation signal (affects early stopping)
                alias='Huber',
              ),
          NHITS(h=horizon,
                input_size=2 * horizon,
                loss=DistributionLoss(distribution='Normal', 
                                      level=level), # Classic Normal distribution
                valid_loss=MQLoss(level=level),
                max_steps=500,
                #early_stop_patience_steps=2,
                dropout_prob_theta=0.6,
                val_check_steps=10,
                alias='Normal',
              )
          ]
nf = NeuralForecast(models=models, freq='M')
nf.fit(df=Y_train_df)
Y_hat_df = nf.predict()

# By default NeuralForecast produces forecast intervals
# In this case the lo-x and high-x levels represent the 
# low and high bounds of the prediction accumulating x% probability
Y_hat_df

fig, ax = plt.subplots(1, 1, figsize = (20, 7))
plot_df = pd.concat([Y_train_df, Y_hat_df]).set_index('ds') # Concatenate the train and forecast dataframes
plot_df[['y', 'Huber-median', 'Normal-median']].plot(ax=ax, linewidth=2)

ax.set_title('Noisy AirPassengers Forecast', fontsize=22)
ax.set_ylabel('Monthly Passengers', fontsize=20)
ax.set_xlabel('Timestamp [t]', fontsize=20)
ax.legend(prop={'size': 15})
ax.grid()

df_metrics = Y_hat_df.merge(Y_test_df, on=['ds', 'unique_id'])
df_metrics.rename(columns={'Huber-median': 'Huber'}, inplace=True)

metrics = evaluate(df_metrics,
                   metrics=[mape, mqloss],
                   models=['Huber', 'Normal'],
                   level = [50, 80],
                   agg_fn="mean")

metrics