# How to decide the shape of the input of an LSTM layer with temporally combined data

I'm working on a time series with 8 features, where at the beginning of each day,

• features 1 to 5 are unknown,
• features 6 to 8 are known, and
• features 1, 2 are the targets to be predicted.

Currently, I'm using 3-lag time steps, so the columns of the data matrix looks like this:

feature_1(t-3) feature_2(t-3) ... feature_7(t-1) feature_8(t-1) feature_6(t) feature_7(t) feature_8(t)

As far as I understand, the shape of the input of an LSTM layer should be (num_rows, num_lags, num_features). However, this shape does not fit my data matrix. My current idea is to use two branches where the first one takes historical data with shape (num_rows, 3, 8) as input and the second one takes data of time t with shape (num_rows, 3) as input. Then, I'll feed the first branch into LSTM and the second branch into Dense and concatenate them later on.

I'm wondering whether my idea is feasible. If not, what are the common methods to deal with this situtation?