With the development of Global Navigation Satellite System (GNSS), satellite-based positioning technology has become an important data source for aviation navigation. However, in scenarios involving unmanned urban air mobility (UAM) applications, satellite positioning is susceptible to multipath (MP) and non-line-of-sight (NLOS) signals leading to deterioration in positioning accuracy, posing a challenge to aircraft safety. To address this problem, a proposed method utilizes the K-LSTM model for satellite positioning error estimation. Firstly, the K-means clustering method is used to detect MP/NLOS signals. Secondly, investigating the relationship between satellite observations and positioning errors in different environments and extending the network model. This extension involves adding a droupout layer, a ReLU layer, a fully-connected layer, and a regression layer on top of the Long Short-Term Memory (LSTM) neural network. Finally, using the extended LSTM model to estimate and correct the localization error caused by MP/NLOS signals. The experimental results reveal that in the static urban canyon environment, the localization errors of the clustered MP/NLOS signals are 0.6m, 0.9m, and 1.0m in the east, north, and up directions, respectively, after the correction by the extended LSTM model. Additionally, the localization errors in the dynamic reflection environment are 1.5m, 1.0m, and 2.5m in the east, north, and up directions, respectively. These results demonstrate significant enhancements in localization accuracies compared to the pre-correction errors.