Serving as one of the important error sources in radio geodesy technology, the tropospheric delay correction aims to enhance the measurement accuracy of the earth observation technology in light of its accurate estimation. Given the difference of zenith tropospheric delay (ZTD) parameters estimated by the Global Pressure and Temperature 3 (GPT3) model in different regions, this paper used the ZTD calculated by the Global Navigation Satellite System (GNSS) to correct the residual of GPT3-derived ZTD, introduced Back Propagation (BP) neural network to establish the functional relationship between the ZTD residual and the geographic location of the station, and constructed the regional ZTD (RZTD) model that corporated the compensation of the ZTD residual. The RZTD model can estimate the ZTD parameters in real time by merely inputting the station position (longitude, latitude, and geodetic height) and time. The GNSS data of 19 continuously operating reference (CORS) stations in Hong Kong from 2018 to 2020 were selected for the experiment, and the accuracy of GNSS ZTD was verified using radiosonde (RS) data. It was found that the root mean square error (RMSE) and bias (Bias) of ZTD estimated by GNSS were 0.013 and 0.001 m, respectively. Taking the ZTD calculated by GNSS and RS as the true value, the accuracy of the RZTD model was verified. It was observed through the experiment that the average RMSE and Bias of the model were 0.013/0.001 m and 0.022/-0.003 m, respectively, exhibiting good stability in different seasons. The statistical results revealed that the RZTD model outperformed the GPT3 model with a 72% improvement in accuracy, which could meet the demand for real-time correction of regional tropospheric delay.