Energy & Buildings publishes our systematic review on advancements in supervised machine learning for outdoor thermal comfort, highlighting scales, applications, and data types to refine predictive methodologies and address urban climatic challenges.

Nature AI Lab
2025-01-06

We are glad to share our new paper:
Luo, Tianze., Chen, Mingze*. Advancements insupervised machine learning foroutdoor thermal comfort: Acomprehensive systematic review ofscales, applications, and data types. Energy & Buildings 2025 | https://doi.org/10.1016/j.enbuild.2024.115255.[download]

💡 ABSTRACT:
Supervised Machine Learning (SML) has a proven track record in addressing the complexities of urban studies. To enhance the accuracy and efficiency of current urban outdoor thermal environment research, many scholars had employed SML methods. However, the current status and limitations of SML applications in this field remained unclear. This paper offered a systematic review of SML use in urban studies based on 58 publications. It recorded the topic, use case, application, data type, and method of each paper, providing statistical insights into trends and evolution. The studies were categorized into three scales, six application types, and six data types. We found that mesoscale studies are the most popular, accuracy comparison is the most common application type, and multiple data sources are the most common input. The Random Forest (RF) algorithm was the most frequently used across categories. A detailed review showed how SML is applied in various outdoor thermal comfort (OTC) studies, suggesting that future research should cover wider study areas and longer time spans. Qualitative methods should also be incorporated into SML research as complementary tools. We discussed SML methods, highlighting algorithm synergy and portability as key areas for future OTC research. The limitations of current applications, such as high computing costs, parameter adjustment complexity, and tedious pretreatment, were identified, along with development directions to address these issues. For input data, we recommended using more dimensional datasets and advanced preprocessing models to enhance prediction accuracy and depth. In conclusion, this review demonstrated SML’s potential to effectively solve outdoor thermal comfort problems. It refines and summarizes existing workflows and applications, offering three propositions for future development in this field.

💡 Background and Significance:
With the intensification of global climate change, optimizing urban thermal environments has become crucial for enhancing livability and addressing extreme weather challenges. Outdoor Thermal Comfort (OTC) is a key indicator of urban environmental quality, yet research in this area faces challenges such as data complexity, regional variability, and fragmented methodologies. Supervised Machine Learning (SML), with its robust data processing and predictive capabilities, offers innovative solutions for OTC research, enabling multi-scale analysis from individual to regional levels.

🔍 Research Scope and Objectives:
This paper systematically reviews 58 publications, analyzing their topics, use cases, applications, data types, and methodologies, while providing statistical insights into trends and evolution. The studies are categorized into three scales, six application types, and six data types. The review also highlights current limitations in computational costs, parameter adjustment complexity, and data preprocessing.

🎯 Research Framework:
Scales of Study: Divided into micro, meso, and macro levels, with mesoscale studies being the most common, focusing on typical urban units such as streets and parks.
Application Types: Six core applications, including feature extraction, precise prediction, model training, and clustering analysis.
Data Types: Includes numerical data (e.g., skin temperature), spatiotemporal data (e.g., surface temperature), satellite data (e.g., Landsat imagery), and multi-source data.

🔍 Findings:
Dominant Algorithms: Random Forest (RF) is the most widely used, accounting for 37.5% of studies, and excels in classification and regression