Machine learning for building energy system analysis

Sammanfattning: Buildings account for approximately 40% of the global energy, and Heating, Ventilation, and Air Conditioning (HVAC) contributes to a large proportion of building energy consumption. Two main negative characteristics that contribute to performance degradation and energy waste in an HVAC system are inappropriate control strategies and faults. These two negative characteristics are the focus areas of this thesis. As a type of HVAC system, district heating (DH) plays an important role in sustainable thermal energy production and dominates the heat market in Sweden. However, analysis of district heating has not been well studied in the literature, therefore, district heating substations are selected as the target systems for analysis in this thesis, where night setback control setting, and abrupt fault/anomaly detection are the two problem domains analyzed. One of the main reasons for these problems is the lack of effective methods for detecting such negative characteristics. The knowledge gaps found in these two problem domains are that studies focusing on night setback identification are not found in the literature, and while deep learning (DL) methods show great potential in terms of fault detection and diagnosis (FDD), they are not systematically evaluated, and the application of deep learning methods for HVAC fault detection and diagnosis is scarce. Therefore, this thesis aims at addressing these research gaps. Different machine learning (ML) methods are investigated for night setback identification in the first study using a relatively small dataset. The main contributions of this thesis to microdata analysis are that it proposes a new perspective of sequential data analysis by converting the original time series data into corresponding images. Thus the original research problem is shifted from time series analysis to image analysis, which enables the use of transfer learning to improve model performance. In addition, the proposed approach can be customized and generalized to analyze other building energy systems as well as solving other research problems in the energy domain. The second part of the thesis contributes to the understanding of how data is analysed by performing a meta-analysis. It not only analyses the effectiveness of deep learning methods for performing HVAC FDD quantitively, but also investigates outlier studies and publication bias. These analysis results give future researchers in this field an unbiased perspective and help them avoid overestimating effectiveness, either due to outlier studies or publication bias. The result shows that bidirectional long short-term memory (BDLSTM) with attention mechanism outperforms other benchmarking methods included in this study. However, the labeling process used in this study is difficult, since by inspecting the original time series data, the dataset is balanced, whereas in real life, there are much fewer substations with night setback settings enabled. Most importantly, this approach is difficult to generalize to a large number of substations. To address these drawbacks, the second study reformulated the research problem from time series to image analysis by converting the original time series data into corresponding heat map images. Then, a transfer learning approach is proposed to classify these images. Results show that MobilenetV2 outperforms other benchmark models in terms of f1 score, while Squeezenets are relatively lighted weighted in terms of training time and model sizes and works well This approach provides a feasible solution that can be used for a relatively large, real-world, and imbalanced dataset with overall high accuracy. It is also concluded that night setback patterns are more easily observed from heatmap images, compared to the original time series. Therefore, the proposed approach reduces the complexity of data labeling, and, even if the model fails, it is flexible and easy to switch to manual judgment. In terms of fault detection and diagnosis, the third study reviews and analyzes deep learning methods for HVAC fault detection and diagnosis in a systematic way. It is concluded that long short-term memory (LSTM) and 1D convolutional neural network (CNN) are effective methods for HVAC FDD by design without extra data transformation, while 2D CNNs gain their popularity in recent years due to the increased diversity of collected data. Meta-analysis results show that deep learning methods are effective for HVAC fault detection and diagnosis. However, publication bias exists, which suggests small studies without significant results are not published, hence, the effectiveness of the included studies can be overestimated. Based on the conclusion from the systematic review, an LSTM-based model is proposed for abrupt faults/anomaly detection. Results show that performance varies based on the threshold values. A lower threshold value results in a better recall, while a higher threshold value results in better precision, and there is a trade-off between recall and precision. Therefore, the choice of the threshold depends on the cost of investigating false alarms versus that of missed out faults.