Building and training the multivariate neural architecture

Let’s start with the simplest problem: only generating the predictions for one time series, cluster_12. To do this, we train a feedforward neural network to predict the next value of cluster_12 given the past 30 values of all 30 time series in the dataset. The solution workflow, named Multivariate_Training_one_output and reported in Figure 13.2, covers the classic steps to import and prepare the data and then build, train, and evaluate the neural network:

Figure 13.2 – The workflow for multivariate time series forecasting

After importing the data with the Table Reader node and sorting it to ensure its sequential order, all the time series are partitioned, using the first 70% for the training set and the remaining 30% for the test set. Next, all the values are normalized to the [0, 1] interval. So far, this is very similar to what we have done in the other chapters.

Now we want to...