Getting Started with Modeltime GluonTS
In modeltime.gluonts: 'GluonTS' Deep Learning

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  warning = FALSE, 
  message = FALSE,
  fig.align = "center",
  out.width = "90%",
  fig.width = 7,
  fig.height = 5
)

Making an N-Beats Ensemble

Let's get started by making an N-BEATS ENSEMBLE. N-BEATS is a powerful algorithm that has shown exceptional results^1^. In a matter of minutes, you'll generate the 4 forecasts shown below. If you'd like to improve your time series forecasting abilities, then please take my High-Performance Time Series Course.

library(tidyverse)
library(modeltime)
modeltime_forecast_tbl <- read_rds("modeltime_forecast_tbl.rds")

modeltime_forecast_tbl %>%
  plot_modeltime_forecast(
    .facet_ncol   = 2, 
    .facet_scales = "free",
    .interactive  = FALSE
  )

Libraries

Load the following libraries.

library(modeltime.gluonts)
library(tidymodels)
library(tidyverse)
library(timetk)

Installation

Next, set up the Python Environment with install_gluonts(). You only need to run this one time, and then you are good to go.

install_gluonts()

Troubleshooting Installation

Python Environment setup is always fun. Here are a few recommendations if you run into an issue.

Check to make sure Conda is available using reticulate::conda_version(). If no conda version is returned, then use reticulate::install_miniconda() to install. Then (re-)run install_gluonts().
Check if GluonTS (Python) is available using reticulate::py_module_available("gluonts"). If this returns TRUE, then your installation has succeeded in building the environment, but you may have other issues like missing C++ build tools (next).
Windows 10 error: Microsoft Visual C++ is required. Here are the instructions for installing the C++ tools needed.
Other installation issues. Please file a GitHub issue here.

Time Series Data

We'll use the m4_hourly dataset, which contains 4 hourly time series.

data <- m4_hourly %>%
  select(id, date, value) %>%
  group_by(id) %>%
  mutate(value = standardize_vec(value)) %>%
  ungroup()

data

We'll create the forecast region using future_frame(). We are forecasting 1 week (24x7 timestamps) into the future.

HORIZON <- 24*7

new_data <- data %>%
  group_by(id) %>%
  future_frame(.length_out = HORIZON) %>%
  ungroup()

new_data

Making an N-Beats Ensemble Model

We'll create a model using the N-BEATS ENSEMBLE algorithm using the nbeats() function.

This is a univariate modeling algorithm that ensembles (combines) multiple N-BEATS Deep Learning models.
We select the Ensemble Version of N-BEATS by setting the engine to gluonts_nbeats_ensemble.
The meta-learner uses 3 sub-models by adjusting the lookback_length. We can include more sub-models adjusting the bagging_size.

model_fit_nbeats_ensemble <- nbeats(
  id                    = "id",
  freq                  = "H",
  prediction_length     = HORIZON,
  lookback_length       = c(HORIZON, 4*HORIZON),
  epochs                = 5,
  num_batches_per_epoch = 15,
  batch_size            = 1 
) %>%
  set_engine("gluonts_nbeats_ensemble") %>%
  fit(value ~ date + id, data)

An NBEATS ENSEMBLE is produced.

model_fit_nbeats_ensemble

knitr::include_graphics("nbeats_model.jpg")

Forecasting

With a model in hand, we can simply follow the Modeltime Workflow to generate a forecast for the multiple time series groups.

modeltime_forecast_tbl <- modeltime_table(
  model_fit_nbeats_ensemble
) %>%
  modeltime_forecast(
    new_data    = new_data,
    actual_data = data,
    keep_data   = TRUE
  ) %>%
  group_by(id)

We can visualize the forecast with plot_modeltime_forecast().

modeltime_forecast_tbl %>%
  plot_modeltime_forecast(
    .conf_interval_show = FALSE, 
    .facet_ncol         = 2, 
    .facet_scales       = "free",
    .interactive        = FALSE
  )

Saving and Loading Models

GluonTS models will need to "serialized" (a fancy word for saved to a directory that contains the recipe for recreating the models). To save the models, use save_gluonts_model().

Provide a directory where you want to save the model.
This saves all of the model files in the directory.
Note that N-BEATS models can be VERY LARGE.

model_fit_nbeats_ensemble %>%
  save_gluonts_model(path = "nbeats_ensemble_model", overwrite = TRUE)

You can reload the model into R using load_gluonts_model().

model_fit_nbeats_ensemble <- load_gluonts_model("nbeats_ensemble_model")

Learning More

Time series is changing. Businesses now need 10,000+ time series forecasts every day. This is what I call a High-Performance Time Series Forecasting System (HPTSF) - Accurate, Robust, and Scalable Forecasting.

I teach how to build a HPTFS System in my High-Performance Time Series Forecasting Course. By taking my course, you will learn:

Time Series Machine Learning (cutting-edge) with Modeltime - 30+ Models (Prophet, ARIMA, XGBoost, Random Forest, & many more)
NEW - Deep Learning with GluonTS (Competition Winners)
Time Series Preprocessing, Noise Reduction, & Anomaly Detection
Feature engineering using lagged variables & external regressors
Hyperparameter Tuning
Time series cross-validation
Ensembling Multiple Machine Learning & Univariate Modeling Techniques (Competition Winner)
Scalable Forecasting - Forecast 1000+ time series in parallel
and more.