The problem of poor quality of training of an NARXNET model

In the course of research, I realized that I could not solve the problem of poor quality of training of an artificial neural network model of the NARX type.

Such networks successfully simulate the dynamics of simulation models in the form of transfer functions. But when the network was built on data from production, it does not want to repeat the data in any way or repeats them with a large average error. The time series of product temperature parameters at the entrance to the apparatus and exit from the apparatus are used as data. The selected parameters are functionally related, but with a delay.

Yesterday I noticed that a similar result is obtained on the simulation model if they is connected to a serial circuit.

1. For my time series during training, I disabled the divide of data into training, test and validation samples;

2. The number of layers and input and output delays is selected approximately;

3. The learning algorithm is a cycle until the MSE of the network reached on the test data of the specified eps. In each iteration of the loop, a NARX network object with random initial weights is generated. The open network is trained, the network is closed, the second training of the network takes place in a closed state with a fixed number of training epochs (200) without initial initialization of weights and on the same training data. The MSE of the resulting network is used as an indicator of network quality.

4. All other parameters of the NARX network are used by default.

Could you please help me?