Wpływ typu sieci neuronowej i sposobu przygotowania danych na wynik prognozowania poprawek UTC - UTC(PL)

W pracy omówiono wyniki badań wpływu typu sieci (MLP, RBF), sposobu przygotowania i rozmiaru wektora danych wejściowych na wynik prognozowania poprawek UTC-UTC(PL) dla atomowego wzorca czasu i częstotliwości. Najkorzystniejsze wyniki prognozowania poprawek osiągnięto dla sieci neuronowych typu RBF, w których dane wejściowe stanowiły wektory zawierające wartości odchyleń od trendu. Otrzymane błędy prognoz nie przekraczają wartości š4ns, co pozwoliło osiągnąć w porównaniu z dotychczasowym sposobem prognozowania opartym na metodzie regresji (prowadzonym w GUM) ponad dwukrotnie lepsze wyniki prognoz.

In the paper there are presented the results of investigations on the influence of type of a neural network (MLP, RBF), the way of preparation and size of the input vector on prediction of the UTC-UTC(PL) corrections for the atomic clock realising the national atomic time scale UTC(PL) at the Central Office of Measures (COM). UTC(PL) is the basis for reproducing the time and frequency units and determining the official time in Poland. At the first stage of research, the time series characterising the time instability of the atomic clock in relation to UTC was a basis for proper preparation of the groups of input data for the assumed types of neural networks (Fig. 1). For the process of learning the neural networks as well as further prediction, the input data was assumed to be formed into 30 or 60 element vectors (Fig. 2). At the second stage, the input data were formed into the vectors containing 30 consecutive values of the deviation of the time series from the trend and, additionally, the linear regression coefficients (Fig. 1). The best results of predicting the corrections were achieved for the RBF neural networks in which the input data were vectors of 30 consecutive values of the deviation from the trend and the directional coefficient of linear regression. The obtained errors of the prediction did not exceed the values of š4ns (Tab. 2), which enabled achieving more than two times better results of the prediction compared with the present way of prediction based on the regression method (used in COM).