I am training an RBF neural network to forecast the future user requests to a Web server. That is to say, the input variables are weekend (0 or 1), Day of week, hour of the day, the minute of hour and user requests at this time. Moreover, the dataset of 2 weeks of user requests to a web server is being used for training. Weekend variable experience has a rather unusual behavior. The desired value to be predicted in weekends is rather lower than other days, and I think this situation causes the prediction to be less than general. Because the produced predicted values in test phase are usually less than actual.

Firstly, I have better forecasting when I leave out the Weekend variable as an input. Why does it happen?

Secondly, What do you suggest to improve the quality of this neural network prediction?

Lastly, what do you prefer? having more inputs or integrating some relative inputs to one input?

Thank you in advance.