Exploring the training results of machine learning models using different batch sizes and epochs: A case study with GNSS time series data

Abstract

This study applies the GRU (Gated Recurrent Unit) model when selecting different values of batch-size, namely 16, 32, and 64, with varying epochs of 20, 50, 100, 150, and 200. The input data comprises observations collected by two GNSS CORS stations from the VNGEONET network, namely HYEN and CTHO, spanning from August 10, 2019, to March 18, 2022. Initially, GNSS CORS data is processed using Gamit/Globk software to obtain the Up-component, which serves as the input data for the GRU model. The research results indicate that the statistical performance metrics of the model, such as RMSE and MAE, decrease while the F-Score increases when the batch-size decreases and the epoch value increases. In cases where the Up-component exhibits irregular variations (seasonal fluctuations), the performance of the GRU model is subpar, with an F-Score of 0 observed when batch-size values are 32 and 64 and epoch value is 20. For data following the pattern of CTHO CORS station, the GRU model performs exceptionally well when batch-size is 16 and epoch is 200. However, the forecasting performance is low for data from HYEN CORS station, indicating the need for further investigation in the future.