Speaker
Description
Achieving accurate performance forecasting of borehole heat exchanger (BHE) is essential for optimizing ground source heat pump systems, enabling optimal control, and facilitating energy-efficient operations with enhanced sustainability of the built environment. It is challenging to eliminate redundant input sensors without compromising the accuracy of the deep machine learning models. The performance of deep neural networks has been extensively studied, but there is still a lack of clear understanding regarding the influence of model architecture, the number of input data sensors, and their accurate identification on the performance of multistep forecasting for ground source heat pump applications. This study aims to fill this gap by investigating and quantifying the impact of these factors on multivariate hybrid deep learning algorithms used for forecasting BHE outlet temperature. Moreover, the significance of incorporating a recent development in deep learning to pay selective attention to the input data i.e., attention-based mechanisms in LSTM-CNN and CNN-LSTM architectures is investigated. The most essential input parameters for the data-driven AI models are determined via an importance interpretability analysis by using Explainable-AI local-method i.e., Shapley Additive Explanations (SHAP) and global-explanation methods i.e., permutation feature importance method and Friedman statistical test. The findings highlight the efficacy of attention mechanisms in capturing temporal dependencies in LSTM-CNN-At and spatial patterns in CNN-LSTM-At, may not necessarily enhance their multistep forecasting capabilities for the borehole field data in comparison to LSTM-CNN architecture. The 24 hours ahead forecasting results show that the order of accuracy is LSTM-CNN> LSTM-CNN-At> CNN-LSTM> CNN-LSTM-At. The analysis of Ex-AI for feature importance at global and local level show that increasing the number of input parameters has the potential to enhance the multi-step forecasting accuracy of the proposed hybrid models. Moreover, it is highlighted that depending on the construction of specific model layers, it is possible to eliminate redundant borehole field sensors for data measurement without compromising the forecasting accuracy of the hybrid deep learning models. Friedman statistical test shows that the performance order based on average rank from best to worst as follows: LSTM-CNN model with an average rank of 1.33 (best performing), LSTM-CNN-At with a rank of 2.0, CNN-LSTM with a rank of 3.00, and CNN-LSTM-At with a rank of 3.66 (worst performing). For the study implications, the proposed multistep prediction hybrid model (LSTM-CNN) can assist the industry in predicting 24 hours ahead of BHE output, the demand-side management of heating and cooling, and building operations.
| Conference Topic Areas | Track9: Smart Energy Storage, Integration and Utilization |
|---|
