Session: 01-05: AI for Energy Sustainability V

Paper Number: 129687

129687 - Distribution Shift Problem in Artificial Intelligence Model of Smart Building: Concept, Impact, and Solutions 

Abstract: 

Building sector consumes nearly 40% of global energy consumptions as well as 30% of carbon emissions, which plays an essential role in the ongoing energy transition. In recent years, artificial intelligence (AI) has become one of the key technologies in building energy conservation. The advanced machine learning and deep learning algorithms provide feasible solutions for several tasks in smart building, such as on-site renewable energy generation forecast, building energy consumption forecast, intelligent control, and predictive maintenance of building energy systems. However, the existing artificial intelligence models were mostly studied in a static manner, where the test data or online data were assumed to follow a similar data distribution with training data. In the real-world case, the online data might follow a significant different data distribution due to the dynamic of building, e.g., the change of outdoor environment, working condition, control logic, and building user behaviors. Under this case, the well-trained deep learning model will result in poor performance and be useless. This phenomenon is named as the distribution shift problem.

In this study, we will systematically introduce the concept, impact, and solutions of distribution shift problem based on our recent publications. Three promising solutions, including domain adaptation, physics-informed machine learning, and uncertainty quantification, are highlighted and illustrated with several representative examples. For the domain adaptation method, it is to minimize the feature distribution discrepancy between training data and online data. We apply the domain adaptation algorithm in the predictive maintainence problem of building energy system. And the result demonstrate that the domain adaptation algorithm can significanltly improve the performance of fault diagnosis model under distribtuion shift scenarios. Another important technology to solve distribtuion shift problem is the physics-informed machine learning, which is to encode prior physical knowledge into machine learning model to improve their extrapolation performance. We will introduce the physics-constrained cooperative learning for the reference model development of building energy system. Two physical laws, including energy conservation and mass conservation, will be transformed as a new loss function of deep learning model. And our result show that the proposed method can improve the model extrapolation performance for a large margin. And finally, we will introduce the uncertainty quantification method of deep learning model and illustrate the importance of uncertainty quantification in handling distribution shift problem. The investigtaed method can capture both data uncertainty and model uncertainty. The experimental result demonstrate that the quantification of model uncertainty is essentital for sovling distribution shift problem. This study might provide a practice guideline for the development of robust artificial intelligence model in smart building energy management.

Presenting Author: Xinbin Liang Shanghai Jiao Tong University

Presenting Author Biography: Xinbin Liang is currently a PhD student in Shanghai Jiao Tong University, and his research interest is the application of artificial intelligence in smart building energy management. He has published several paper in Applied energy, Energy conversion and management, and Energy.

Authors:

Xinbin Liang Shanghai Jiao Tong University
Zhimin Du Shanghai Jiao Tong University