LSTM keras very poor post-workout prognosis, what am I doing wrong?

Question

Neural networks

There is a neuron, built on tf, keras libraries with one LSTM layer, 1 fully connected and 1 output dimension
of input data (N, 100, 2)
dimension of output data (1, 2)

(N, 2) the grid seems to cope with the graphs, but what about the forecast outside the dataset, I can’t form N amount of data because I don’t know about them.

Did so took the last one hundred values ​​from each row received the dimension (1, 100, 2) predicted (1, 2) by a new value for each row. Further, from (1, 100, 2) of the first and second rows, I throw out one first value each and supplement the predicted ones from the end and again get (1, 100, 2) but new data from the end and so 100 iterations. Of course, I expected something similar to the previous 100 values, at least some kind of curvature of the graph, but I get almost two lines; of course, there are small deviations of these lines, but this is still not what I expected.
Does anyone know how to do it right? Everywhere they show on the test sample how wonderful they are, because the data already exists, as it were, and they form these rows in one package, shifting each row by N + 1, but how do I prepare the package, because I am limited to 1 step forward since the end of the dataset, even if I don’t predict by one value, but several will not change the essence of the problem