Wpływ dodatkowych danych wejściowych w sieci neuronowej na wynik prognozowania poprawki dla krajowej skali czasu UTC(PL)

W artykule przedstawiono wyniki badań, których celem było sprawdzenie, czy wprowadzenie na wejście sieci neuronowej (SN) typu MLP i GMDH dodatkowych danych wejściowych, charakteryzujących pracę zegarów atomowych, wpłynie korzystnie na wynik prognozowania poprawki dla UTC(PL). Otrzymane wyniki wskazują, że kluczowym aspektem jest dobór SN. Dla SN typu GMDH wpływ dodatkowych danych jest niewielki. SN tego typu osiąga lepsze wyniki w prognozowaniu poprawki dla UTC(PL), niż sieć typu MLP.

The paper presents the results of investigations whose aim was to examine whether the insertion of additional input data for MLP and GMDH neural networks would increase compliance of the UTC(PL) with UTC. The primary input vector in the training process of the network included historical data of measurements of the phase time between the UTC and atomic clock (Cs2). Additional input vectors were made in two ways. The first way results from the rules of determining the UTC time scale, which is computed as a weighted average based on continuous comparisons of more than 300 atomic clocks located in laboratories of many countries around the world. Hence, the input data containing the historical data, characterizing the behavior of two additional atomic clocks whose results are available in the Central Office of Measures (GUM), was added to the study. These clocks are to "imitate" the gait of atomic clocks used to calculate the UTC time scale. The second way is related to the results of previous studies, which indicate that the quality of predicting the corrections for the UTC(PL) depends on the atomic clock gait instability. The measure of this instability is the Allan deviation. That is why, an additional input data containing the Allan deviation gait characterizing the atomic clock Cs2 gait instability was added to the study. Selection of a neural network has very large impact on the results. In the case of the MLP neural network the insertion of additional input data resulted in a smaller value of the prediction error, and in the case of the GMDH networks insertion of an additional input data affected the final prediction result to a small extent.