Prediction Research on the Ground Temperature Variation Caused by Ground Source Heat Pump Based on Different Intelligent Algorithms

Ground temperature prediction is critical for ensuring the long-term operational stability and energy efficiency of ground-source heat pump (GSHP) systems, which are essential for the sustainable utilization of shallow geothermal energy and the stability of the underground ecological environment. However, existing models are predominantly developed for shallow depths (typically within 5 m) for agricultural or permafrost applications, leaving the 50–100 m depth range relevant to GSHP systems underexplored. This study addresses this gap by constructing differentiated prediction models for two typical data types: multi-parameter short-period data and single-parameter long-period data. For Sample I, random forest and support vector regression (SVR) models were developed with hyperparameters optimized by the Sparrow Search Algorithm. For Sample II, ARIMA and SARIMA models were established. A standardized preprocessing workflow integrating box plot-based outlier detection, Newton interpolation, and Min-Max normalization was applied to both datasets. The results show that random forest significantly outperforms SVR, while SARIMA with seasonal components substantially improves upon ARIMA by capturing annual ground temperature periodicity. These findings provide quantitative guidance for model selection in GSHP engineering, particularly for the 50–100 m depth range, enabling accurate prediction of deep ground temperature for thermal balance assessment and early warning of thermal imbalance risks, thereby supporting the sustainable operation of GSHP systems and the efficient utilization of shallow geothermal energy.