Atomic Clock Ensemble Using ARIMA-LSTM Time Series Prediction Model

Jahnvi Verma, Thejesh N. Bandi

Abstract:	Accurate and stable clock ensemble generation is critical for Positioning, Navigation, and Timing (PNT) systems, where timevarying noise processes and environmental effects influence clock performance. Conventional ensemble methods typically rely on fixed or slowly varying weighting schemes, limiting their ability to adapt to dynamic changes in clock instability. This paper presents a hybrid framework that integrates auto-regressive integrated moving average (ARIMA) modeling, long short-term memory (LSTM) networks, and dynamic weighting based on rolling overlapping Allan variance (oAVAR) for multiclock ensemble formation. In the proposed approach, each clock’s fractional frequency series is first modeled using ARIMA to capture linear stochastic behavior. The residuals are then modeled using an LSTM network to learn nonlinear temporal dependencies, improving prediction accuracy. To form the ensemble, a time-dependent weighting scheme is introduced using rolling oAVAR evaluated across multiple averaging times. Results demonstrate improved stability of the hybrid ensemble compared to individual clocks and non-adaptive methods, particularly under time-varying noise conditions. The computational efficiency of the proposed approach supports scalability to large datasets and near real-time applications.
Published in:	Proceedings of the ION 2026 Pacific PNT Meeting April 13 - 16, 2026 Hilton Waikiki Beach Honolulu, Hawaii
Pages:	70 - 77
Cite this article:	Verma, Jahnvi, Bandi, Thejesh N., "Atomic Clock Ensemble Using ARIMA-LSTM Time Series Prediction Model," Proceedings of the ION 2026 Pacific PNT Meeting, Honolulu, Hawaii, April 2026, pp. 70-77. https://doi.org/10.33012/2026.20579
Full Paper:	ION Members/Non-Members: 1 Download Credit Sign In