Comparative study of machine learning models based on different data preprocessing methods in geothermal heat flow prediction in China

Abstract

Geothermal heat flow (GHF) estimates could be derived from a variety of geological and geophysical features using machine learning techniques. The dataset is crucial for developing GHF prediction models. Previous studies have used different methods to process the global GHF dataset. However, few studies have explored the influence of various data preprocessing methods on GHF predictions.
In this paper, we investigate the influence of four datasets processed in different ways on GHF prediction in China utilizing the GBRT algorithm. The results showed that more features are needed for the model performance to converge after the dataset is processed on average. Additionally, averaging the GHF data will significantly affect the model's judgment on the importance of features. Models constructed with the averaged dataset exhibit lower GHF estimations and a narrower prediction range. The prediction accuracy of GBRT models on the test set might be greatly increased by low-pass filtering the GHF data. A further comparison revealed that the model created using the dataset that was just subjected to low-pass filtering had the best ability to predict GHF in China.
In the end, we provide eight fresh GHF maps for China. All of them have high GHF in Tibet and eastern China and low GHF in northwest China. Additionally, our models predict a few hot regions where there are no GHF observations, which might be advantageous for geothermal exploration.