how to reduce error in ANN ? MAPE error

I am using 3 year load data of electricity and using NARX model for load prediciton for day ahead, one month , three month data using trainning function(trainbr) with different combination of 7 inputs( day of month, weekend, month of year, temp, RH......etc), number of nuerons(10,20,30,40) and one output load for training. but the still MAPE error is high in prdiciton but low in training with good regression value of 0.956. how to reduce MAPE and can anyone suggest the sensivity of model with different parameters like inputs, neurons, layers etc. currently MAPE is in the range of 20-50% with different scenarios.

Thanks