--- title: N-HiTS keywords: fastai sidebar: home_sidebar nb_path: "nbs/models_nhits__nhits.ipynb" ---
A new model for long-horizon forecasting which incorporates novel hierarchical interpolation and multi-rate data sampling techniques to specialize blocks of its architecture to different frequency band of the time-series signal. It achieves SoTA performance on several benchmark datasets, outperforming current Transformer-based models by more than 25%. Paper available at https://arxiv.org/abs/2201.12886
import pandas as pd
from neuralforecast.data.datasets.epf import EPF
from neuralforecast.data.tsloader import TimeSeriesLoader
import pylab as plt
from pylab import rcParams
plt.style.use('seaborn-whitegrid')
plt.rcParams['font.family'] = 'serif'
FONTSIZE = 19
# Load and plot data
Y_df, X_df, S_df = EPF.load_groups(directory='./data', groups=['NP','FR'])
fig = plt.figure(figsize=(15, 6))
plt.plot(Y_df[Y_df['unique_id']=='NP'].ds, Y_df[Y_df['unique_id']=='NP'].y.values, color='#628793', linewidth=0.4)
plt.ylabel('Price [EUR/MWh]', fontsize=19)
plt.xlabel('Date', fontsize=15)
plt.show()
mc = {}
mc['model'] = 'n-hits'
mc['mode'] = 'simple'
mc['activation'] = 'SELU'
mc['n_time_in'] = 24*3
mc['n_time_out'] = 24
mc['n_x_hidden'] = 8
mc['n_s_hidden'] = 0
mc['stack_types'] = ['identity', 'identity', 'identity']
mc['constant_n_blocks'] = 1
mc['constant_n_layers'] = 2
mc['constant_n_mlp_units'] = 256
mc['n_pool_kernel_size'] = [4, 2, 1]
mc['n_freq_downsample'] = [24, 12, 1]
mc['pooling_mode'] = 'max'
mc['interpolation_mode'] = 'linear'
mc['shared_weights'] = False
# Optimization and regularization parameters
mc['initialization'] = 'lecun_normal'
mc['learning_rate'] = 0.001
mc['batch_size'] = 1
mc['n_windows'] = 32
mc['lr_decay'] = 0.5
mc['lr_decay_step_size'] = 2
mc['max_epochs'] = 1
mc['max_steps'] = None
mc['early_stop_patience'] = 20
mc['eval_freq'] = 500
mc['batch_normalization'] = False
mc['dropout_prob_theta'] = 0.0
mc['dropout_prob_exogenous'] = 0.0
mc['weight_decay'] = 0
mc['loss_train'] = 'MAE'
mc['loss_hypar'] = 0.5
mc['loss_valid'] = mc['loss_train']
mc['random_seed'] = 1
# Data Parameters
mc['idx_to_sample_freq'] = 1
mc['val_idx_to_sample_freq'] = 1
mc['n_val_weeks'] = 52
mc['normalizer_y'] = None
mc['normalizer_x'] = 'median'
mc['complete_windows'] = False
mc['frequency'] = 'H'
print(65*'=')
print(pd.Series(mc))
print(65*'=')
mc['n_mlp_units'] = len(mc['stack_types']) * [ mc['constant_n_layers'] * [int(mc['constant_n_mlp_units'])] ]
mc['n_blocks'] = len(mc['stack_types']) * [ mc['constant_n_blocks'] ]
mc['n_layers'] = len(mc['stack_types']) * [ mc['constant_n_layers'] ]
from neuralforecast.experiments.utils import create_datasets
train_dataset, val_dataset, test_dataset, scaler_y = create_datasets(mc=mc,
S_df=S_df, Y_df=Y_df, X_df=X_df,
f_cols=['Exogenous1', 'Exogenous2'],
ds_in_val=294*24,
ds_in_test=728*24)
train_loader = TimeSeriesLoader(dataset=train_dataset,
batch_size=int(mc['batch_size']),
n_windows=mc['n_windows'],
shuffle=True)
val_loader = TimeSeriesLoader(dataset=val_dataset,
batch_size=int(mc['batch_size']),
shuffle=False)
test_loader = TimeSeriesLoader(dataset=test_dataset,
batch_size=int(mc['batch_size']),
shuffle=False)
mc['n_x'], mc['n_s'] = train_dataset.get_n_variables()
model = NHITS(n_time_in=int(mc['n_time_in']),
n_time_out=int(mc['n_time_out']),
n_x=mc['n_x'],
n_s=mc['n_s'],
n_s_hidden=int(mc['n_s_hidden']),
n_x_hidden=int(mc['n_x_hidden']),
shared_weights=mc['shared_weights'],
initialization=mc['initialization'],
activation=mc['activation'],
stack_types=mc['stack_types'],
n_blocks=mc['n_blocks'],
n_layers=mc['n_layers'],
n_mlp_units=mc['n_mlp_units'],
n_pool_kernel_size=mc['n_pool_kernel_size'],
n_freq_downsample=mc['n_freq_downsample'],
pooling_mode=mc['pooling_mode'],
interpolation_mode=mc['interpolation_mode'],
batch_normalization = mc['batch_normalization'],
dropout_prob_theta=mc['dropout_prob_theta'],
learning_rate=float(mc['learning_rate']),
lr_decay=float(mc['lr_decay']),
lr_decay_step_size=float(mc['lr_decay_step_size']),
weight_decay=mc['weight_decay'],
loss_train=mc['loss_train'],
loss_hypar=float(mc['loss_hypar']),
loss_valid=mc['loss_valid'],
frequency=mc['frequency'],
random_seed=int(mc['random_seed']))
from pytorch_lightning.callbacks import EarlyStopping
early_stopping = EarlyStopping(monitor="val_loss",
min_delta=1e-4,
patience=mc['early_stop_patience'],
verbose=False,
mode="min")
trainer = pl.Trainer(max_epochs=mc['max_epochs'],
max_steps=mc['max_steps'],
gradient_clip_val=1.0,
progress_bar_refresh_rate=10,
log_every_n_steps=500,
check_val_every_n_epoch=1,
callbacks=[early_stopping])
trainer.fit(model, train_loader, val_loader)
model.return_decomposition = False
outputs = trainer.predict(model, val_loader)
print("outputs[0][0].shape", outputs[0][0].shape)
print("outputs[0][1].shape", outputs[0][1].shape)
print("outputs[0][2].shape", outputs[0][2].shape)
Y_forecast_df = Y_df[Y_df['ds']<'2016-12-27']
Y_forecast_df.tail()
X_forecast_df = X_df[X_df['ds']<'2016-12-28']
X_forecast_df.tail()
model.return_decomposition = False
forecast_df = model.forecast(Y_df=Y_forecast_df, X_df=X_forecast_df, S_df=S_df, batch_size=2)
forecast_df