Problem in training feedforward network for system control

Hello, I am trying to create my first more complicated neural network used to control Magnetic Levitation System. The whole system is described by nonlinear differential equation, which is modeled in MLS block. At first I was trying to force my network to mimic the behaviour of PID, by taking about 100000 samples from 10 seconds of simulation (random numbers were chenging every 0.3s) from this diagram:

and using PDIn and PDout as input and desired output to train my network with this script:

P = PDin';
T = PDout';
net = feedforwardnet(20);
net = configure(net, P, T);
net.trainParam.goal = 1e-8;
net = train(net, P, T);
gensim(net);

And used generated block to replace PD controller. Unfortunately, the network keeps failing and simulated levitated ball is just falling into the infinite distance from the magnet. Can anyone help me in finding out why is this happening?