How to deal with multivariate time series forecasting with neural networks, considering known values in the future for some of them?

[Marcelino Borges]

I would like to deal with a time series forecasting problem where we have a multivariate time series, involving 5 different time series (4 inputs and 1 output). That is, at each time point we have 5 features.

One of these time series represents an arbitrary human input that changes the process that produces the values of the other 4 time series (that are sensor measurements). The time series are represented below.

In this context, let us suppose that we want to predict the value ahead of a given window of 5 time points in the past. In this context, I would like to consider that I know the value of the arbitrary feature on a subsequent day.

Due to this, I think that it is not a good practice to consider as training inputs tensors of 5X5, because one of the time series would have 6 values (considering that we know the value of the subsequent time point). This is depicted in the following image. The orange cells represent that values that I want to consider for predicting the value in blue.

I think that using encoder-decoder architectures could help in this case, but I'm not sure how to apply it to this problem.

Could someone explain how can I do that? Is there some good reference that I can use for supporting the development of this neural network architecture?