I am creating a neural network using tensorflow that predicts the energy consumption of a vehicle. Originally, I planned on normalizing all of the features from 0 to 1 using the scikit-learn object MinMaxScaler. However, one of the features, altitude change, contains both positive and negative values. In addition, the label, energy consumption, can be negative due to the energy returned to the battery from regenerative braking. I need my model to respond differently to these negative values as opposed to the meaningless negative values of other features like temperature.
Is it best practice to instead normalize some/all features from -1 to 1? Or should I be standardizing all the features instead (even if they may not be normally distributed)?