Temporal Fusion Transformer (TFT) proposed by Lim et al. [1] is one of the most popular transformer-based model for time-series forecasting. In summary, TFT combines gating layers, an LSTM recurrent encoder, with multi-head attention layers for a multi-step forecasting strategy decoder. For more details on the Nixtla’s TFT implementation visit this link. In this notebook we show how to train the TFT model on the Texas electricity market load data (ERCOT). Accurately forecasting electricity markets is of great interest, as it is useful for planning distribution and consumption. We will show you how to load the data, train the TFT performing automatic hyperparameter tuning, and produce forecasts. Then, we will show you how to perform multiple historical forecasts for cross validation. You can run these experiments using GPU with Google Colab. Open In Colab

​
1. Libraries

!pip install neuralforecast

import pandas as pd

​
2. Load ERCOT Data

The input to NeuralForecast is always a data frame in long format with three columns: unique_id, ds and y:
  • The unique_id (string, int or category) represents an identifier for the series.
  • The ds (datestamp or int) column should be either an integer indexing time or a datestamp ideally like YYYY-MM-DD for a date or YYYY-MM-DD HH:MM:SS for a timestamp.
  • The y (numeric) represents the measurement we wish to forecast. We will rename the
First, read the 2022 historic total demand of the ERCOT market. We processed the original data (available here), by adding the missing hour due to daylight saving time, parsing the date to datetime format, and filtering columns of interest. 
Y_df = pd.read_csv('https://datasets-nixtla.s3.amazonaws.com/ERCOT-clean.csv')
Y_df['ds'] = pd.to_datetime(Y_df['ds'])
Y_df.head()
unique_iddsy
0ERCOT2021-01-01 00:00:0043719.849616
1ERCOT2021-01-01 01:00:0043321.050347
2ERCOT2021-01-01 02:00:0043063.067063
3ERCOT2021-01-01 03:00:0043090.059203
4ERCOT2021-01-01 04:00:0043486.590073

​
3. Model training and forecast

First, instantiate the AutoTFT model. The AutoTFT class will automatically perform hyperparamter tunning using Tune library, exploring a user-defined or default search space. Models are selected based on the error on a validation set and the best model is then stored and used during inference. To instantiate AutoTFT you need to define:
  • h: forecasting horizon
  • loss: training loss
  • config: hyperparameter search space. If None, the AutoTFT class will use a pre-defined suggested hyperparameter space.
  • num_samples: number of configurations explored.
from ray import tune

from neuralforecast.auto import AutoTFT
from neuralforecast.core import NeuralForecast
from neuralforecast.losses.pytorch import MAE

import logging
logging.getLogger("pytorch_lightning").setLevel(logging.WARNING)
Tip Increase the num_samples parameter to explore a wider set of configurations for the selected models. As a rule of thumb choose it to be bigger than 15. With num_samples=3 this example should run in around 20 minutes. 
horizon = 24
models = [AutoTFT(h=horizon,
                  loss=MAE(),
                  config=None,
                  num_samples=3)]
Tip All our models can be used for both point and probabilistic forecasting. For producing probabilistic outputs, simply modify the loss to one of our DistributionLoss. The complete list of losses is available in this link
Important TFT is a very large model and can require a lot of memory! If you are running out of GPU memory, try declaring your config search space and decrease the hidden_size, n_heads, and windows_batch_size parameters. This are all the parameters of the config: 
config = {
      "input_size": tune.choice([horizon]),
      "hidden_size": tune.choice([32]),
      "n_head": tune.choice([2]),
      "learning_rate": tune.loguniform(1e-4, 1e-1),
      "scaler_type": tune.choice(['robust', 'standard']),
      "max_steps": tune.choice([500, 1000]),
      "windows_batch_size": tune.choice([32]),
      "check_val_every_n_epoch": tune.choice([100]),
      "random_seed": tune.randint(1, 20),
}
The NeuralForecast class has built-in methods to simplify the forecasting pipelines, such as fit, predit, and cross_validation. Instantiate a NeuralForecast object with the following required parameters: Then, use the fit method to train the AutoTFT model on the ERCOT data. The total training time will depend on the hardware and the explored configurations, it should take between 10 and 30 minutes. 
nf = NeuralForecast(
    models=models,
    freq='h')

nf.fit(df=Y_df)
Finally, use the predict method to forecast the next 24 hours after the training data and plot the forecasts. 
Y_hat_df = nf.predict()
Y_hat_df.head()

c:\Users\ospra\miniconda3\envs\neuralforecast\lib\site-packages\utilsforecast\processing.py:384: FutureWarning: 'H' is deprecated and will be removed in a future version, please use 'h' instead.
  freq = pd.tseries.frequencies.to_offset(freq)
c:\Users\ospra\miniconda3\envs\neuralforecast\lib\site-packages\utilsforecast\processing.py:440: FutureWarning: 'H' is deprecated and will be removed in a future version, please use 'h' instead.
  freq = pd.tseries.frequencies.to_offset(freq)

Predicting: |          | 0/? [00:00<?, ?it/s]
unique_iddsAutoTFT
0ERCOT2022-10-01 00:00:0038600.757812
1ERCOT2022-10-01 01:00:0036871.199219
2ERCOT2022-10-01 02:00:0035505.500000
3ERCOT2022-10-01 03:00:0034781.691406
4ERCOT2022-10-01 04:00:0034647.484375
Plot the results with matplot lib 
import matplotlib.pyplot as plt

fig, ax = plt.subplots(1, 1, figsize = (10, 3))
plot_df = pd.concat([Y_df.tail(24*5).reset_index(drop=True), Y_hat_df]).set_index('ds') # Concatenate the train and forecast dataframes
plot_df[['y', 'AutoTFT']].plot(ax=ax, linewidth=2)

ax.set_title('Load [MW]', fontsize=12)
ax.set_ylabel('Monthly Passengers', fontsize=12)
ax.set_xlabel('Date', fontsize=12)
ax.legend(prop={'size': 10})
ax.grid()

​
4. Cross validation for multiple historic forecasts

The cross_validation method allows you to simulate multiple historic forecasts, greatly simplifying pipelines by replacing for loops with fit and predict methods. See this tutorial for an animation of how the windows are defined. With time series data, cross validation is done by defining a sliding window across the historical data and predicting the period following it. This form of cross validation allows us to arrive at a better estimation of our model’s predictive abilities across a wider range of temporal instances while also keeping the data in the training set contiguous as is required by our models. The cross_validation method will use the validation set for hyperparameter selection, and will then produce the forecasts for the test set. Use the cross_validation method to produce all the daily forecasts for September. Set the validation and test sizes. To produce daily forecasts set the forecasting set the step size between windows as 24, to only produce one forecast per day. 
val_size  = 90*24 # 90 days x 24 hours
test_size = 30*24 # 30 days x 24 hours
fcst_df = nf.cross_validation(df=Y_df, val_size=val_size, test_size=test_size,
                                n_windows=None, step_size=horizon)
Finally, we merge the forecasts with the Y_df dataset and plot the forecasts. 
Y_hat_df = fcst_df.reset_index(drop=True)
Y_hat_df = Y_hat_df.drop(columns=['y','cutoff'])

plot_df = Y_df.merge(Y_hat_df, on=['unique_id','ds'], how='outer').tail(test_size+24*7)

plt.figure(figsize=(20,5))
plt.plot(plot_df['ds'], plot_df['y'], c='black', label='True')
plt.plot(plot_df['ds'], plot_df['AutoTFT'], c='blue', label='Forecast')
plt.axvline(pd.to_datetime('2022-09-01'), color='red', linestyle='-.')
plt.legend()
plt.grid()
plt.plot()

​
Next Steps

In Challu et al [2] we demonstrate that the N-HiTS model outperforms the latest transformers by more than 20% with 50 times less computation. Learn how to use the N-HiTS and the NeuralForecast library in this tutorial.

​
References

[1] Lim, B., Arık, S. Ö., Loeff, N., & Pfister, T. (2021). Temporal fusion transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 37(4), 1748-1764.. [2] Cristian Challu, Kin G. Olivares, Boris N. Oreshkin, Federico Garza, Max Mergenthaler-Canseco, Artur Dubrawski (2021). N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting. Accepted at AAAI 2023.