Abstract: High-precision priori tropospheric delay can reduce the convergence time of precise point positioning. Based on a highprecision and high-resolution numerical meteorological database, the deep learning method N-BEATS algorithm was used to predict the site-specific tropospheric zenith total delay. Nine IGS tracking stations were selected. It covered 18.5 years from January 2002 to June 2019. Based on the N-BEATS algorithm, three forecast strategies were designed with different input arcs. The first 17.5 years of the entire period were used for model training. The last year’s data was used for validation. The results show that the average forecast residuals of different forecast arcs shorter than 12hours are mostly in the sub-millimeter range. As the forecast arc increases, the average forecast residual increases. The strategy with a longer input arc shows better performance than the other two strategies with shorter arcs. The standard deviations of the forecast residuals of 2hours, 4hours, and 6hours are approximately 5mm, 9mm, and 13mm, respectively.

Key words: N-BEATS, deep learning, time series forecast, tropospheric zenith total delay, tropospheric wet delay, tropospheric hydrostatic delay