Short-Timescale Gravitational Microlensing Events Prediction with ARIMA-LSTM and ARIMA-GRU Hybrid Model

Abstract

Astronomers hope to give early warnings based on light-detection data when some celestial bodies may behave abnormal in the near future, which provides a new method to detect low-mass, free-floating planets. In particular, to search short-timescale microlensing (ML) events from high-cadence and wide-field survey in real time, we combined ARIMA with LSTM and GRU recurrent neural networks (RNN) to monitor all the observed light curves and to alert before abnormal deviation. Using the good linear fitting ability of ARIMA and the strong nonlinear mapping ability of LSTM and GRU, we can form an efficient method better than single RNN network on accuracy, time consuming and computing complexity. ARIMA can reach smaller alerting time and operating time, yet costing high false prediction rate. By sacrificing 15% operating time, hybrid models of ARIMA and LSTM or GRU can achieve improved 14.5% and 13.2% accuracy, respectively. Our work also provide contrast on LSTM and GRU, while the first type is commonly used for time series predicting systems, the latter is more novel. We proved that in the case of abnormal detection of light curves, GRU can be more suitable to apply to as it is less time consuming by 8% while yielding similar results as LSTM. We can draw a conclusion that in the case for short-timescale gravitational microlensing events prediction, hybrid models of ARIMA-LSTM and ARIMA-GRU perform better than separate models. If we concentrate more on accuracy, ARIMA-LSTM is the best option; on the other hand, if we concentrate more on time consuming, ARIMA-GRU can save more time.