Neural Network NARX - too high error

Hello,

I want a Neural Network which simulate the evolution of a pendulum attached to a cart with external forces. Thus i generate data from ODE (20 samples of 2000 timesteps of 4 elements) and use them to train a NARX neural network (using the nnstart tool). Once the network is trained, i want to use the neural network to predict the evolution of the system. So i use the nets (from the genrated matlab code which allows to evaluate the next step)

nets = removedelay(net);

Then i simulate the evolution with a constant external input

Simtime=2; % Simulation time
Ts=0.01;  % Sampling time
X=cell(2,1);
X{1}=[7]'; % Constant external input
X{2}=[0 0 0.2 0]'; % Initial states
states_nn=zeros(Simtime/Ts,4); 
states_nn(1,:)=initial_cond;
for i=1:Simtime/Ts
[Y,~,~] = sim(nets,X); 
X{1}=[1]; 
X{2}=Y{1};
states_nn(i+1,:)=Y{1}';
end

But when i compare neural network evolution and the ODE evolution results are bad. I tried to midify number of layer, size of layer, number of epoch, performance function, activation function but nothing leads to significant better results.

I read and tried many things but nothing semmes to work. Do you have any advice ? Thank you