Data-driven insights for improved heating and cooling predictions: Impact of input parameters on multivariate deep learning algorithms using geothermal borehole field data

Abstract

The current research examines the effect of input parameters on the accuracy of deep learning models for forecasting borehole heat exchanger (BHE) temperatures. The study investigates the impact of different input parameters on the selection and performance of various multivariate deep AI models, such as LSTM, GRU, CNN, and BD-LSTM, within the context of geothermal borehole coupled heat pump systems. Real borehole-field data is used to train models with variable number of input parameters, aiming to predict 24-hour-ahead performance for meeting heating and cooling requirements. Key input parameters are identified using significance level test of Explainable AI techniques like Shapley Additive Explanations (SHAP) and sensitivity analysis was performed to assess their impact on model predictions. The results indicate that BD-LSTM outperforms other models in forecasting accuracy, followed by LSTM, GRU, and 1D-CNN, establishing BD-LSTM as the most accurate. The number of input parameters significantly affects forecasting accuracy, but more parameters do not necessarily improve performance. For instance, adding solar radiation as an input did not significantly alter BD-LSTM and BLSTM's forecasting accuracy, with MAPEs of 2.69% and 3.26%, respectively. Conversely, GRU's accuracy decreased (MAPE: 3.36%), while 1D-CNN's improved (MAPE: 2.92%). The SHAP analysis shows that the significance ranking of input parameters for the top-performing BD-LSTM model remains same for Case-1 and Case-3 but with different impact of magnitude. The study concludes that the selection of practically measurable, relevant input parameters is vital for accurate temperature forecasting in geothermal applications, a finding that extends beyond existing research which often emphasizes algorithm selection over input parameter optimization.