Single-step-ahead recurrent neural networks

Although we took a little detour to check out how DL regression models can be used to train the same global models we learned about in Chapter 10, Global Forecasting Models, now we are back to looking at DL models and architectures specifically built for time series. And as always, we will look at simple one-step-ahead and local models first before moving on to more complex modeling paradigms. In fact, we have another chapter (Chapter 15, Strategies for Global Deep Learning Forecasting Models) entirely devoted to techniques we can use to train global DL models.

Now, let’s bring our attention back to one-step-ahead local models. We saw recurrent neural networks (RNNs) (vanilla RNN, long short-term memory (LSTM), and gated recurrent unit (GRU)) as a few blocks we can use for sequences such as time series. Now, let’s see how we can use them in an end-to-end (E2E) model on the dataset we have been working on (the London smart...