Strange values when dropping column from a supervised learning dataset

Question

I am trying to predict bike availability. Initially I've defined this problem as a time series. To predict the bikes I'm transforming my set to a supervised learning problem and then feeding it to the NN.

On a first approach the supervised problem I defined it as the following

day_of_year(t-1) | time(t-1) | weekday(t-1) | weekday(t-1) | free_bikes(t-1) | free_bikes(t)

But after thinking for a while. In the future when I want to predict bikes for the future, not near future, I'd like to know the availability for example the next day or next week so I don't want to feed all the time-steps to the nn. Instead I've come up with this:

day_of_year(t-1) | time(t-1) | weekday(t-1) | weekday(t-1) | free_bikes(t)

I have removed free_bikes(t-1) as an input to the problem.

After this the predicted values are messed up, the predicted values are absurdly incorrect and I don't fully understand why is so.

Links:

• Predicted values
• Output of the script
• Model

Answer 1

Firstly, if you change the input dimensions, you need to retrain the model.

Your model has a input dimension of (batch_size,1,3). So it is one timestep and three features per time step. This doesnt match the input dimensions you described in your question.

Please specify the problem setup.