Assessing AImodel performance in time-series GNSS data analysis with different neural network structures

Abstract

Artificial intelligence is widely used in time series data analysis in general, and specifically for GNSS time series data. The performance of each AI model used for analyzing GNSS time series data depends on the selection of the optimization function, loss function, the number of nodes in the hidden layers, and the number of epochs. The GRU (Gated Recurrent Unit) deep learning model has been proven to perform well in time series prediction. This paper presents the results of evaluating the performance of the GRU model with different parameter selections mentioned above. The input data for the model is the vertical coordinate component from the HYEN CORS station from 10/8/2019 to 18/3/2022, which is the result of analyzing GNSS data collected at this station using the Gamit/Globk software. The processing results show that when using the Adam optimizer and MSE loss function, the model’s performance decreases rapidly as the number of nodes in the hidden layer reduces from 200÷100. In this case, the model's performance metrics include an R2decrease from 85÷20%, and the MAE value increases from 3.77÷8.37 mm. When replacing the MSE loss function with the Huber loss function, the model's performance significantly improves, with the R2 increasing by 7%, and the MAE value decreasing from 3.77÷3.21mm. This is a relatively high performance for predicting data using an AI model with a training-to-testing ratio of 60÷40%