Abstract: Elevated duct is a natural phenomenon occurring in the troposphere, where signals undergo superrefraction within the duct and can achieve long-distance, low-loss propagation, making it a promising breakthrough point in marine communication research. To support marine communication applications and improve the prediction accuracy of duct parameters, this paper utilizes high-resolution radiosonde data from the SPARC (Stratosphere-troposphere Processes and their Role in Climate) project, specifically focusing on the elevated ducting strength time series from the SAN JUAN, MAJURO, and TRUK INTL stations recorded between 2016 and 2022. By introducing the Neural Hierarchical Interpolation for Time Series (N-HiTS) method, a high-precision long-term prediction model for elevated ducting strength was constructed, with single-site modeling and multi-site validation performed. The results show that for the observational data at the TRUK INTL station, the model captures both long-term trends and local fluctuation features in the strength series, achieving a root mean square error (RMSE) of 6.35 M-units and a mean absolute error (MAE) of 4.31 M-units. For the validation stations, SAN JUAN and MAJURO, the RMSE values are 7.35 M-units and 6.54 M-units, respectively. The proposed model exhibits strong generalization capability in the long-term prediction of elevated duct strength, providing robust support for the analysis of elevated duct characteristic parameters and the study of wireless propagation effects under this mechanism.