You need to fiddle with the bp-nn training settings to obtain *more* stable results.
You'll never get identical forecasts with a bpnn because of the way the network weights are randomised during the network's training phase.