A Review of Data-Driven Methods in Building Retrofit and Performance Optimization: From the Perspective of Carbon Emission Reductions

Abstract

In order to reduce the contribution of the building sector to global greenhouse gas emissions and climate change, it is important to improve the building performance through retrofits from the perspective of carbon emission reductions. Data-driven methods are now widely used in building retrofit research. To better apply data-driven techniques in low-carbon building retrofits, a better understanding is needed of the connections and interactions in optimization objectives and parameters, as well as optimization methods and tools. This paper provides a bibliometric analysis of selected 45 studies, summarizes current research hotspots in the field, discusses gaps to be filled, and proposes potential directions for future work. The results show that (1) the building-performance optimization (BPO) process established through physical simulation methods combines the site, retrofit variables, and carbon-related objectives, and the generated datasets are either directly processed using multi-objective optimization (MOO) algorithms or trained as a surrogate model and iteratively optimized using MOO methods. When a sufficient amount of data is available, data-driven methods can be used to develop mathematical models and use MOO methods for performance optimization from the perspective of building carbon emission reductions. (2) The benefits of retrofits are maximized by holistically taking environmental, economic, and social factors into account; from the perspectives of carbon emissions, costs, thermal comfort, and more, widely adopted strategies include improving the thermal performance of building envelopes, regulating HVAC systems, and utilizing renewable energy. (3) The optimization process based on data-driven methods, such as optimization algorithms and machine learning, apply mathematical models and methods for automatic iterative calculations and screen out the optimal solutions with computer assistance with high efficiency while ensuring accuracy. (4) Only 2.2% and 6.7% of the literature focus on the impacts of human behavior and climate change on building retrofits, respectively. In the future, it is necessary to give further consideration to user behaviors and long-term climate change in the retrofit process, in addition to improving the accuracy of optimization models and exploring the generalization and migration capabilities of surrogate models.

1. Introduction

1.1. Background

Climate change and greenhouse gas emissions (GHGEs) have emerged as critical concerns impacting the sustainable development of global human habitations [1]. Global Carbon Emissions (CEs) are increasing at a rate of about 2% annually [2] and have risen by approximately 10 billion tons over the past two decades [3]. The construction industry contributes ~40% to the global CEs [4], making it a primary driver of global warming [5]. Moreover, with current high urbanization rates and buildings’ long lifespans [6], urban construction has entered a phase of stock renewal, necessitating extensive retrofitting of existing buildings now and in the future [7]. Retrofitting buildings represents a practical approach to significantly reducing GHGEs [8] and offering substantial environmental, social, and economic benefits [9], thus plays a pivotal role in advancing long-term sustainable developments [10].

A building retrofit is an intervention to add new materials or elements to existing buildings [11], encompassing methods like upgrading the building envelope, enhancing mechanical systems, and improving operations and management [12]. It serves not only to facelift old buildings but also as a potent means to attain environmental sustainability [13] and building-performance optimization (BPO) [14]. BPO objectives guide the direction of building retrofit strategies [15], and when balancing competing goals such as CEs, energy efficiency, and thermal comfort [16], predictive models are essential for selecting optimal solutions from a range of alternatives [17].

Building-performance prediction models are typically categorized into physical models based on building operational process simulations and mathematical models employing data-driven methods [5]. Physics-based models offer precise and reliable calculations, yet their complexity in input parameters and transient modeling increase computational costs. In contrast, mathematical models provide rapid and efficient performance predictions but require extensive data for model developments and often lack physical interpretations of building-performance parameters [18].

Traditional approaches to retrofitting individual buildings typically rely on architects’ knowledge and experience to enumerate and compare options. However, this often faces challenges due to the vast search space and the difficulties in exploring all potential alternatives exhaustively. Advances in computer technology have introduced data-driven methods to the retrofit process, leveraging computer-assisted sampling and iterative calculations for automated evaluations of retrofitting options to derive optimal solutions [19]. Contrasted with traditional methods, this approach overcomes constraints of time and space [20], thereby enhancing the accuracy and efficiency of decision-making processes [21].

Due to the rising complexity of retrofit standards and objectives, surrogate models (also known as metamodels) have garnered significant attention in recent years for their ability to integrate the strengths of both physics-based and mathematical models. These models are trained using a limited set of simulation data as samples. Following parameter adjustments and validations, they can swiftly generate highly accurate results with appropriate input datasets [22], omitting extensive simulations and substantially cutting down calculation times.

Research on building retrofits focusing on Carbon Emission Reductions (CERs) has extensively utilized BPO based on data-driven methods [23,24,25]. While the optimization procedures in various studies share similarities, differences exist in the methods and tools employed. Thus, it is essential to review existing research, comprehend the development and evolution of the focal points of current research, summarize the characteristics and mechanisms behind the most frequently utilized as well as promising new methods and tools, and provide suggestions and guidance for future research and practice towards the design optimization of CER-based building retrofits.

1.2. Previous Reviews

Previous studies have explored topics related to building retrofits, building CERs, and data-driven methods. Addressing current challenges in building retrofits, scholars have offered theoretical guidance and practical recommendations for implementing retrofit projects, choosing retrofitting tools, and formulating retrofit standards. These efforts encompass perspectives such as sustainable retrofits [14], energy-saving retrofits [26], and decision-making methods [27]. In retrofit projects aimed at CERs, researchers primarily assess the environmental impact across the building’s life cycle [11], emphasizing elements like building envelopes, floor plan layouts, and natural ventilation strategies [28] to minimize both Operational Carbon Emissions (OCEs) and Embodied Carbon Emissions (ECEs). Furthermore, data-driven methods play a crucial role in addressing complex issues such as energy predictions and control within buildings [5], contributing to revealing energy-saving potentials. The surrogate models developed through data-driven methods are applied in Building-Performance Simulations (BPSs) and optimization processes to enhance computational efficiency [22].

1.2.1. A Review of Building Retrofit Research

Jagarajan et al. [13] asserted that green buildings can significantly enhance environmental sustainability across design, construction, and maintenance phases, thereby reducing the environmental influence of existing buildings. Their systematic literature review and analysis of green building retrofit initiatives noted a paucity of research concerning critical factors influencing the implementation of retrofit projects, emphasizing the need to focus on the problems faced by stakeholders, such as the lack of financial incentives from the government and the limited availability of green material and technology. Nielsen et al. [14] evaluated decision support tools applicable to building retrofit projects, discussing their application in establishing sustainable development goals, formulating design strategies, and assessing performance. They advocated the use of sustainable standards to provide clearer criteria for screening existing buildings. In a study focused on China, Liu et al. [26] examined obstacles and challenges in building energy-saving retrofit practices and analyzed the evolving national and local incentive measures. Their work contributed to the theoretical foundation of energy-saving policies and standards, proposing a data-driven decision-making system to guide optimal retrofit measure selections. Pombo et al. [29] conducted a review of residential building retrofit research, highlighting the variability in methods for evaluating energy-efficiency measures. They underscored the importance of adopting a life-cycle approach, including methodologies like life cycle assessments (LCAs) and life-cycle costs (LCCs) to identify optimal solutions and assess the retrofit potential of residential buildings. These studies emphasized the multifaceted challenges and opportunities in green building retrofit initiatives, offered insights into enhancing sustainability practices, and supported decision-making framework developments.

1.2.2. A Review of Building CER Retrofit Research

Vilches et al. [11] conducted a comprehensive review on building retrofits and environmental assessments using LCA methods. Their findings highlight that the construction and operational phases are most frequently analyzed, with a primary focus on impact categories such as the global warming potential (GWP) and embodied energy. Mostafavi et al. [28] reviewed 48 studies from 2005 to 2020 investigating energy and carbon performance in high-rise buildings across diverse climates. They identified the significant potential for energy-saving through improvements in envelope designs, floor plan optimizations, and the use of natural ventilation. This study also emphasized strategies to reduce OCEs and ECEs by enhancing thermal performance and integrating recycled materials into building constructions. Li et al. [30] examined sustainable retrofit practices in subtropical high-density cities like Hong Kong, emphasizing the variability of CER outcomes across different building types and environmental conditions. For high-rise and multi-story residential buildings, this study underscored the importance of targeting energy-efficiency improvements in public areas and optimizing building envelopes to meet CER goals. Aghamolaei et al. [31] conducted a research study on decarbonization strategies for university campuses under various climatic conditions. Their research emphasized the critical role of implementing low-carbon and energy-saving measures across campus infrastructures. This review identified key parameters including spatial planning, landscape integration, renewable energy adoption, building envelope enhancement, and sustainable transportation as pivotal for achieving CER goals. Their study also anticipated the Internet of Things (IoTs) and the integration of data-driven technologies to facilitate campus decarbonization efforts through enhanced planning and design processes. These studies contribute to understanding effective strategies and technologies aiming at reducing CEs in buildings and urban environments, providing insights into tailored approaches for different building types and climatic contexts.

1.2.3. A Review of the Application of Data-Driven Methods in Building-Performance Analyses

Wei et al. [5] conducted a thorough review of data-driven methods applied in building energy analyses, encompassing predictive and classification techniques. Their findings underscore the efficacy of these methods in energy load forecasts and modeling, facilitating comprehensive assessments of macro-level energy-saving potentials aligned with consumer demands and the formulation of sustainable urban development strategies. Sun et al. [32] reviewed multi-step energy forecasts using data-driven models. Their comprehensive analysis encompassed feature engineering, model types, and anticipated outcomes, proposing future research directions in predictive capabilities and energy model controls. Roman et al. [16] systematically reviewed the literature on metamodels based on Artificial Neural Networks (ANNs) and BPS. They detailed the methodology for generating ANN-based metamodels, covering stages from sample construction to model training and validation. The authors suggest that for the reproducibility of research results, developers of the metamodel should try to publicize sample databases, which will facilitate more in-depth analyses and wide applications in the future. Westermann et al. [22] explored the application of metamodels in sustainable building design research. They highlighted statistical models as effective surrogates for detailed simulations, offering computational efficiency and mitigating barriers in BPS. Their study synthesized successful applications of metamodels in architectural designs to guide their practical implementation. These studies contribute to advancing methodologies in energy predictions and sustainable analyses through the application of data-driven techniques and metamodeling approaches.

Existing reviews primarily focus on enhancing building energy efficiency or providing decision-making guidance, with limited analyses from the perspective of CERs in building retrofit approaches. Those that discuss CERs tend to compare retrofit strategies from the perspective of different building types or building LCAs, with less discussion on the application of data-driven methods. Reviews of data-driven methods and metamodel applications emphasize energy predictions and sustainable building designs yet rarely integrate these approaches into a comprehensive framework for BPO.

Therefore, based on a systematic review, this article explores the relevance of CERs, building retrofits, and data-driven methods. It introduces research hotspots, compares various data-driven methods applied to enhancing building performance, discusses potentials and challenges in their application in building CER retrofits, and provides suggestions on the selection of appropriate optimization methods from the perspective of CERs for different retrofit conditions and scenarios. It aims to offer an overview of the BPO process using data-driven methods in CER-based retrofits and provide a timely and valuable reference for future research.

1.3. Outline and Structure of This Review

The rest of this article is structured as follows: Section 2 outlines the literature search and selection process (Section 2.1) and methods deployed for bibliometric analyses (Section 2.2) and summarizes the general process of BPO based on data-driven methods (Section 2.3). Section 3 details the methodology for constructing BPO models, covering the simulation method (Section 3.1) and the machine learning method (Section 3.2) and performance indicators and optimization parameters selected during the optimization process. Section 4 summarizes the optimization methods and tools (Section 4.1), as well as the decision-making methods for post-optimization processing (Section 4.2). Building upon these sections, Section 5 discusses the current status and challenges of building retrofit methods (Section 5.1, Section 5.2 and Section 5.3) and outlines future research directions (Section 5.4). Section 6 provides a summary of the entire article.

2. Literature Screening and Bibliometric Analysis

2.1. Literature Search and Screening

The literature search and screening process followed the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [33], ensuring a transparent, consistent, and comprehensive systematic literature review (SLR) [34]. This process comprised four phases: identification, screening, eligibility, and inclusion. This study conducted an extensive literature search across prominent databases in natural sciences and engineering research, including Scopus, Web of Science (WoS), and Science Direct (SD) [35]. To maximize the number of retrieved articles, the initial search avoided terms like “data-driven”, which were later applied during literature screening. Keywords encompassed three categories related to “building retrofit”, “optimization”, and “carbon emission”, alongside their synonyms, as outlined in

Table 1. Literature search keywords.

Term	Keywords
Term 1	“building renovation” OR “building reconstruction” OR “building retrofit *” OR “building refurbishment” OR “building repairment” OR “building restoration” OR “building upgrade” OR “building renewal” OR “building improvement” OR “building reformation”
Term 2	multi-objective OR multi-criteria OR optimization
Term 3	“carbon emission” OR “carbon mitigation” OR “CO2 emission” OR “CO2 mitigation” OR “greenhouse gas” OR “global warming” OR “environmental impact” OR “sustainable development”

Note: The wildcard “*” indicates a fuzzy search and is used to replace words with multiple spelling variants, such as “retrofit” and “retrofitting”.

. Each keyword group was combined using the Boolean operator “AND” during the search process.

Firstly, filtering criteria were applied based on the publication year, document type, and language during the initial search. Secondly, duplicate articles were removed, and the remaining results underwent screening based on title, abstract, and keywords. Articles that aligned with the scope of this paper were then subjected to a full-text review. Finally, all references were meticulously checked, and relevant ones meeting the criteria were included. The screening process adhered to the following criteria:

• To ensure timeliness, articles published within the last 20 years (2003–2023) were selected. Journal articles, known for their rigorous peer-review process [36], were prioritized for their representation and impact in this field [37]. Conference papers, dissertations, and non-English language publications were excluded, retaining only English-language journal articles. Additionally, to facilitate the focus on research methodologies and processes, review papers were omitted from consideration.
• The literature reviewed addresses CEs during the operational phase of buildings, covering life-cycle carbon emissions (LCCEs), the GWP, and building environmental effects. Studies focusing solely on energy consumption (EC) without converting it to CE objectives were excluded due to differences in concepts and calculation methodologies. Similarly, the literature exclusively discussing CEs during building construction or other phases was also omitted from consideration.
• The research must incorporate one or more data-driven methods, such as statistical analyses, optimization algorithms, or machine learning (ML) to optimize building performance. Studies that merely listed and compared retrofit plans without employing these methods were excluded.
• The literature must address the impacts of envelope retrofit measures on the overall building performance. Studies exclusively focusing on the operations of building mechanical systems, energy structure predictions, and similar topics were excluded. Additionally, articles concentrating solely on specific local building components like curtain wall retrofits and structural seismic performance optimizations were also excluded.

Figure 1 illustrates the literature search and screening process, which was applied to retrieve 611 articles from the database. Following screening and supplementation based on the outlined criteria, 45 articles were selected for a detailed analysis. Based on their content, the relevant studies are summarized in terms of building location and type, optimization objectives and variables, and data-driven methods such as machine learning, sensitivity analyses, optimizations, and decision-making methods, as shown in

Table 2. Summary of previous research studies (optimization objectives and variables).

Refs.	Location	Building Type	Optimization Objective	Optimization Variable
[23]	UK	Office building	LCCs, LCEs, and LCCEs	Insulation material area of roof and exterior wall, equipment and energy system, PV panel area, and solar thermal device.
[24]	Switzerland	Residential building	CEs and ACs	U-/R-value of roof, exterior wall, and ground; window type; equipment and energy system; PV system; and solar thermal device.
[25]	Italy	Office building	ECs, TDHs, GCs, and GHGEs	Surface material characteristics of roof and exterior wall, insulation material thickness of roof and exterior wall, window type, equipment and energy system, sunshade component, and PV panel angle and area.
[38]	UK	Office building	LCCEs and OCE	Insulation material type of roof and exterior wall, equipment and energy system, and solar thermal device.
[39]	UK	Office building	LCCs, LCEs, and LCCEs	Insulation material type and area of roof and exterior wall, window-to-wall ratio, equipment and energy system, PV panel area, and solar thermal device.
[40]	China	Shopping mall	OCEs	U/R-value of exterior wall, Glass material, Sunshade component and equipment and energy system.
[41]	Iran	Residential building	CEs and TDHs	Insulation material thickness of roof and exterior wall, insulation material thickness and type of ground, window type, airtightness, and equipment and energy system.
[42]	Canada	Office building	ECs and CEs	Insulation material type of roof, exterior wall, and floor; window type; airtightness; and equipment and energy system.
[43]	Finland	Office building	LCCs, RCs, CEs, and TDHs	Insulation material thickness of roof and exterior wall, window type, sunshade component, equipment and energy system, and PV system.
[44]	Iran	Residential building	ECs and the GWP	Insulation material type and thickness of exterior wall and exterior wall type (combination of different materials).
[45]	Canada	Educational building	ECs, LCCs, and LCAs	Type of roof and exterior wall, glass material, airtightness, window opening percentage, and equipment and energy system.
[46]	Korea	Residential building	RCs, LCCs, LCCEs, and CERs	Insulation material type and thickness of exterior wall, window type, sunshade component, and equipment and energy system.
[47]	France	Educational building	ECs, TDHs, RCs, and CEs	Type of roof, floor, ground, and interior and exterior wall; window type, and sunshade component.
[48]	Europe	Residential building	ECs, RCs, OCs, and CEs	Surface material characteristics of roof and exterior wall, window type, sunshade component, sunspace, building form, PV panel angle and area, and solar thermal device.
[49]	Finland	Residential building	ECs, LCCs, and CEs	Insulation material thickness of roof and exterior wall, window type, door material, PV panel area, solar thermal device, and equipment and energy system.
[50]	Iran	Residential building	CEs, WCs, LCCs, and TDHs	Insulation material type and thickness of roof and exterior wall, glass material, filling gas, PV panel area, and equipment and energy system.
[51]	China	Residential building	CEs, TDHs, and GCs	Surface material characteristics, insulation material type and thickness of roof and exterior wall, window type, sunshade component, sunspace, and PV panel angle and area.
[52]	China	Residential building	ECs, RCs, and CERs	Insulation material type of roof, exterior wall and floor, glass material, window-to-wall ratio, and sunspace.
[53]	Estonia	Residential building	GCs, ECs, and LCCEs	Insulation material thickness of exterior wall, surface material characteristics of roof, window type, door material, and building form.
[54]	Korea	Educational building	ECs, CEs, RCs, and TDHs	Type of roof, floor, ground, ceiling, and interior and exterior wall; window type; and equipment and energy system.
[55]	China	Residential building	The GWP, LCCs, and TDHs	Insulation material type and thickness of roof and exterior wall, window type, window-to-wall ratio, and sunshade component.
[56]	China	Residential building	ECs, LCCEs, and LCCs	Insulation material type and thickness of floor and exterior wall, glass material, window-to-wall ratio, sunshade component, and Airtightness.
[57]	UK	Residential building	LCCEs and LCCs	Insulation material thickness, exterior wall type, and window-to-wall ratio.
[58]	Sweden	Residential building	LCEs, LCCEs, and LCCs	Insulation material type and thickness of exterior wall, roof, and ground and window type.
[59]	Switzerland	Residential building	ACs CEs	U-/R-value of roof, floor, and exterior wall; window type; PV panel area; and solar thermal device, and equipment and energy system.
[60]	Canada	Residential building	LCCEs and LCCs	Insulation material type of ceiling and exterior wall, window frame material, door material, airtightness, and equipment and energy system.
[61]	UK	Non-domestic building	BER	Type of roof and exterior wall, window type, and equipment and energy system.
[62]	Italy	Residential building	ECs, OCs, RCs, and CEs	Insulation material thickness of roof, floor, and exterior wall; surface material characteristics of roof and exterior wall; PV panel angle and area; glass material; sunshade component; building form; sunspace; and solar thermal device.
[63]	Iran	Residential building	CEs and TDHs	Insulation material thickness of roof, ground, and exterior wall; window type, airtightness, and equipment and energy system.
[64]	Denmark	Residential building	ECs, the GWP, OCs, and RCs	Insulation material type and thickness of interior wall, insulation material type and thickness of roof and exterior wall, surface material characteristics of roof and exterior wall, window frame material, glass material, PV panel area, solar thermal device, and equipment and energy system.
[65]	Italy	Residential building	RCs, OCs, ECs, and CEs	Insulation material thickness of roof, floor, and exterior wall; surface material characteristics of exterior wall; PV panel angle and area; sunshade component, building form, sunspace, and solar thermal device.
[66]	Bosnia and Herzegovina	Residential building	ECs, CEs, and RCs	Insulation material thickness of ceiling and exterior wall, window type, and equipment and energy system.
[67]	China	Office building	ECs, CEs, and TDHs	PV panel angle and area and equipment and energy system.
[68]	Switzerland	Residential building	LCCs and GHGEs	Type of roof and exterior wall, window type, airtightness, PV system, solar thermal device, and equipment and energy system.
[69]	Germany	Residential building	ACs and CEs	Type of roof and exterior wall, window type, PV system, solar thermal device, and equipment and energy system.
[70]	China	Office building	ECs, CEs, and OCs	U-/R-value of roof and exterior wall, window type, and equipment and energy system.
[71]	Greece	Residential building	GHGEs and LCCs	Insulation material thickness of roof, ground, and exterior wall; window type, PV system, solar thermal device, and equipment and energy system.
[72]	China	Office building	LCCEs	Insulation material thickness of roof and exterior wall, surface material characteristics of exterior wall, window type, PV panel area, and equipment and energy system.
[73]	China	Educational building	ECs and LCCEs	Type of roof, floor, and exterior wall; filling gas; building form; insulation material thickness of floor and exterior wall; window frame material; glass material; building form; insulation material thickness of roof; window-to-wall ratio; sunshade component; PV panel area; and equipment and energy system.
[74]	USA	Residential building	GHGEs, WCs, the SQOL, and LCCs	U-/R-value of roof, ceiling and exterior wall, glass material, window-to-wall ratio, and equipment and energy system.
[75]	UK	Residential building	LCCEs and LCCs	Type of roof, floor, ceiling, and interior and exterior wall and window type.
[76]	Canada	Educational building	ECs, LCCs, and LCAs	Type of roof and exterior wall, glass material, window frame material, window-to-wall ratio, airtightness, window opening percentage, and equipment and energy system.
[77]	Canada	Office building	ECs, ECEs, and LCCs	Type of roof and exterior wall, glass material, window frame material, window-to-wall ratio, airtightness, sunshade component, and equipment and energy system.
[78]	China	Office building	ECs, CEs, and LCCs	Insulation material type and thickness of roof and exterior wall and window type.
[79]	Switzerland	Residential building	LCCs and LCAs	Insulation material type of ceiling and exterior wall; insulation material thickness of ceiling, floor, and exterior wall; glass material; and window frame material.

,

Table 3. Summary of previous research studies (machine learning and sensitivity analysis method).

Refs.	Location	Building Type	Machine Learning Method (Accuracy)	Sensitivity Analysis Method
[23]	UK	Office building	-	-
[24]	Switzerland	Residential building	ANN (R2 = 0.94)	-
[25]	Italy	Office building	-	-
[38]	UK	Office building	-	-
[39]	UK	Office building	-	LSA
[40]	China	Shopping mall	-	LSA
[41]	Iran	Residential building	-	GSA (DOE)
[42]	Canada	Office building	-	LSA
[43]	Finland	Office building	-	-
[44]	Iran	Residential building	-	-
[45]	Canada	Educational building	ANN (MSE1 = 0.016 and MSE2 = 0.056)	-
[46]	Korea	Residential building	-	-
[47]	France	Educational building	-	-
[48]	Europe	Residential building	-	-
[49]	Finland	Residential building	-	-
[50]	Iran	Residential building	-	-
[51]	China	Residential building	-	GSA (PCC and SRRC)
[52]	China	Residential building	-	-
[53]	Estonia	Residential building	-	-
[54]	Korea	Educational building	-	-
[55]	China	Residential building	DNN (R2 > 0.99, CV (RMSE) ≤ 1%, and NMBE ≤ 0.2%)	GSA
[56]	China	Residential building	-	-
[57]	UK	Residential building	-	-
[58]	Sweden	Residential building	-	-
[59]	Switzerland	Residential building	-	-
[60]	Canada	Residential building	-	-
[61]	UK	Non-domestic building	GBRT (RMSE = 1.7%)	LSA
[62]	Italy	Residential building	-	GSA (SRRC)
[63]	Iran	Residential building	-	GSA (DOE)
[64]	Denmark	Residential building	-	-
[65]	Italy	Residential building	-	GSA (SRRC)
[66]	Bosnia and Herzegovina	Residential building	-	-
[67]	China	Office building	-	GSA (SRC)
[68]	Switzerland	Residential building	-	-
[69]	Germany	Residential building	-	-
[70]	China	Office building	-	GSA (Morris)
[71]	Greece	Residential building	-	-
[72]	China	Office building	-	-
[73]	China	Educational building	ANN (MRE = 1.57% R2 = 0.94)	-
[74]	USA	Residential building	-	-
[75]	UK	Residential building	-	-
[76]	Canada	Educational building	-	-
[77]	Canada	Office building	MVLR and MARSs (MAPE = 0.2–1.8%)	-
[78]	China	Office building	-	-
[79]	Switzerland	Residential building	Gaussian process modelling (Kriging)	GSA (Sobol)

and

Table 4. Summary of previous research studies (optimization and decision-making method).

Refs.	Location	Building Type	Optimization Method	Decision-Making Method
[23]	UK	Office building	PSO	-
[24]	Switzerland	Residential building	MILP	-
[25]	Italy	Office building	NSGA-II	-
[38]	UK	Office building	PSO	-
[39]	UK	Office building	PSO	-
[40]	China	Shopping mall	Regression	-
[41]	Iran	Residential building	NSGA-II	-
[42]	Canada	Office building	-	-
[43]	Finland	Office building	Pareto-Archive and NSGA-II	-
[44]	Iran	Residential building	Fitness Comparison	-
[45]	Canada	Educational building	NSGA-II	-
[46]	Korea	Residential building	iMOO score	-
[47]	France	Educational building	NSGA-II	-
[48]	Europe	Residential building	aNSGA-II and pNSGA-II	Utopia point
[49]	Finland	Residential building	Pareto-Archive and NSGA-II	-
[50]	Iran	Residential building	prNSGA-III	TOPSIS
[51]	China	Residential building	SPEA2	Utopia point
[52]	China	Residential building	-	Entropy method (Weight of CERs is 30.95%)
[53]	Estonia	Residential building	Regression	-
[54]	Korea	Educational building	NSGA-II/III and MOEA/D	-
[55]	China	Residential building	NSGA-II	TOPSIS (Weight of the GWP is 37.29%)
[56]	China	Residential building	NSGA-II	-
[57]	UK	Residential building	NSGA-II	-
[58]	Sweden	Residential building	NSGA-II	-
[59]	Switzerland	Residential building	GA and MILP	-
[60]	Canada	Residential building	NSGA	-
[61]	UK	Non-domestic building	GA	-
[62]	Italy	Residential building	aNSGA-II	Utopia point
[63]	Iran	Residential building	EWSOA	-
[64]	Denmark	Residential building	Omni-Optimizer	Utopia point
[65]	Italy	Residential building	aNSGA-II	Utopia point
[66]	Bosnia and Herzegovina	Residential building	NSGA-III	Desirability function (Weight of CEs is 30%)
[67]	China	Office building	NSGA-II	-
[68]	Switzerland	Residential building	ϵ-constraint	-
[69]	Germany	Residential building	ϵ-constraint	-
[70]	China	Office building	-	-
[71]	Greece	Residential building	MOGA	-
[72]	China	Office building	NOP and MILP	-
[73]	China	Educational building	SEGA	-
[74]	USA	Residential building	GA	-
[75]	UK	Residential building	NSGA-II	-
[76]	Canada	Educational building	NSGA-II	-
[77]	Canada	Office building	-	-
[78]	China	Office building	AHP	-
[79]	Switzerland	Residential building	NSGA-II	-

.

2.2. Literature Statistics and Bibliometric Analysis

We present a statistical analysis on the quantitative trends, research locations (Figure 2), types of buildings studied, and data-driven methods used (Figure 3) across all reviewed articles. Figure 2 illustrates that due to rapid advancements in computing applications and the growing maturity of building retrofit research [27], significant research contributions have emerged over the past decade. Over time, the volume of publications has steadily increased, underscoring the field’s substantial research value and potential for development [80]. Geographically, the majority of studies originate from Europe (n = 21) and Asia (n = 17), with a limited representation of regions such as Africa, South America, and Oceania. Notably, there has been a surge in research output from China since 2022. The location of research and local climate conditions play pivotal roles in building retrofit strategies [81], and the diverse environmental and climatic contexts in Europe and Asia contribute to the breadth of study cases reported in these regions [82].

Figure 3 reveals that residential buildings dominate the literature, reflecting their large scale, high EC, and pollution levels and pressing the need for CER retrofits. Among public buildings, office buildings exhibit more pronounced retrofit effects compared to other commercial buildings due to their stable occupancy intensity and schedule, contrasting with the high population mobility typical of commercial spaces. In addition, each article adopts one or more types of data-driven methods. More than 90% of the reviewed studies use optimization methods for analysis, while the application of machine learning methods is less than 20%, which needs further exploration.

A bibliometric analysis provides insights into the current status and developmental trends within a specific research field, capturing its dynamics and directions [83]. Utilizing scientific meteorological mapping tools allows for a detailed analysis of the literature [84]. Tools such as CiteSpace [85] enable citation burst analyses, visually depicting significant changes and trends in the evolution of a research field [86].

CiteSpace 6.3.R1 was employed to conduct a citation burst analysis (Figure 4). A citation burst analysis highlights trends in disciplinary developments based on the evolution of subject terms [87]. Figure 4 displays the top nine keywords sorted by their initiation year, with the red-line segments indicating periods of significant citation [88]. The early literature extensively discussed “genetic algorithm” (2011–2015), whereas recent attention has been paid to keywords like “embodied carbon” and “machine learning” (2021–2023), indicating a gradual shift towards addressing CEs in building retrofits because of global warming. Moreover, data-driven methods such as “machine learning” began gaining attention in building CER retrofit research around 2021.

2.3. General Process of BPO Based on Data-Driven Methods

Through a literature review and analysis, it was determined that in the CER retrofit, BPO based on data-driven methods typically adheres to a consistent process. While minor variations exist in the specifics of methods across the literature, the fundamental steps remain the same. These steps are the acquisition of building information, the construction of BPO models, and the screening and decision-making of optimization solutions, as illustrated in Figure 5.

Figure 5 outlines the general process as follows: (1) A suitable retrofit building type is selected, appropriate samples are classified and selected, and pertinent data, including the building form and indoor environmental parameters, etc., are collected through investigation. (2) A physical model is constructed using the above data through simulation methods, and the model accuracy is validated through a comparison with actual data, such as EC records. Optimization parameters and objectives are identified, functional relationships are established between parameters and objectives through simulations and calculations, and an optimization model is developed. Based on the computational efficiency of the model, the choice of whether to use data-driven methods, such as machine learning, is made to further transform it into a surrogate model. When the amount of data is sufficient, a mathematical model can be directly generated based on the collected building information data, combined with optimization objectives and parameters. (3) Optimization algorithms and decision-making methods are integrated to select the retrofit strategy and achieve the optimal solution aligned with retrofit requirements.

3. The Construction BPO Models

The BPO process, shown in Figure 5, can be divided into two categories: (1) the establishment of physics-based BPS protocols (or workflows) based on multiple combinations of site and building energy-carbon-related variables (parameters), and the generated datasets can be either directly processed in MOO algorithms or trained and validated as a surrogate model and then processed with MOO and (2), the development of a mathematical model using regression or a machine learning method, with a sufficient amount of available empirical data, and this is processed with MOO for the optimal solution(s) from the perspective of CERs.

3.1. Optimization Model Based on Physical Simulation Methods

3.1.1. BPS Tools

To construct an optimization model using physical simulation methods, simulation tools are needed to compute building-performance metrics, such as EC calculations based on heat transfer or thermodynamic theories [89]. BPS tools are being made to be more user-friendly by equipping the simulation engines with certain kinds of graphical user interfaces (GUIs) [90]. An analysis of the BPS tools used in the reviewed papers (

Table 5. The use of BPS tools in the literature.

Simulation Tool	References
Designbuilder	[25,40,41,44,45,46,50,63,66,68,73,76,77,78]
TRNSYS	[39,47,67,71]
SIMEB	[42]
IDA ICE	[43,49,53]
SketchUp—OpenStudio	[48,52,57,62,65,70]
Grasshopper—Honeybee	[51,55,56,58,72]
EnergyPlus	[54,59,74,75]
HOT2000	[60]

) indicated that DesignBuilder [91], OpenStudio [92] (a plugin for SketchUp [93]), and Honeybee [94] (a plugin for Grasshopper [95]) are the primary tools used in simulation research. A comparative study evaluated the strengths and weaknesses of widely used simulation tools and recommended employing a combination of tools and comprehensively assessing data needs and model suitability to enhance modeling accuracies [96].

3.1.2. Model Calibration

After the physical model is constructed, it is usually necessary to calibrate the model based on monitored weather data and field-measured energy data [97]. The primary calibration process involves adjusting input parameters like heating and cooling set temperatures, indoor occupancy schedules, etc., guided by user experiences and energy consumption records [98].

The model’s accuracy is typically assessed by comparing errors between simulated and actual ECs, commonly evaluated against mean bias errors (MBEs) and coefficient of variation of root mean square error CV (RMSE) criteria as defined by ASHRAE Guideline 14 [99], IPMVP [100], and FEMP [101] (

Table 6. Commonly used error-evaluation indicators.

Evaluation Indicators	Guideline	Monthly Criteria	Hourly Criteria	References
MBE	ASHRAE	±5%	±10%	[46,51,55,56,67]
	IPMVP	-	±5%
	FEMP	±5%	±10%
CV (RMSE)	ASHRAE	15%	30%	[46,50,51,55,56,67,77]
	IPMVP	-	20%
	FEMP	15%	30%

). A smaller error indicates a greater similarity between the physical model and the operational conditions of the reference building [102]. It is worth noting that the above calibration criteria do not take into account the uncertainty of input parameters, which can be further explored through methods such as energy auditing or Bayesian optimization [103].

3.2. Surrogate and Mathematical Models

Unlike physical simulation methods, machine learning methods use mathematical analyses to establish a functional relationship between input and output data in a black-box model. In the BPO process, the surrogate model converted from the optimization model is applied to the optimization process for pre-processing steps [104], which can expedite the generation of samples [105] and decrease computational costs while preserving accuracy [106]. By utilizing historical weather data, energy profiles, and existing building information, the mathematical model generates output values for each retrofit objective [23], which can be effectively applied to large-scale bottom-up building retrofit analyses, such as community and urban scales [24]. However, it demands extensive and accurate historical and statistical data as a foundation [107]. Moreover, predictive variables are typically limited to certain thermo-physical or equipment parameters, constraining the ability to assess the potential for energy savings from retrofits or technological upgrades [108].

Machine learning has emerged as a crucial tool in recent years, leveraging artificial intelligence (AI) to tackle complex tasks and processes in the building sector [109]. It learns from vast datasets and a relatively small number of input features [110] and utilizes learned information for predictions [111]. The process of converting optimization models into surrogate models using machine learning methods typically involves four main steps: data collection and sampling, data pre-processing and model selection, model training and hyperparameter optimization, as well as model validation [112]:

• The samples of input data are selected using sampling methods, and output data are computed through physical model simulations. Typically, the sample size should range from 10 to 100 times the number of input parameters [113], although this can vary depending on the model’s complexity [114].
• Continuous and discrete variables are distinguished, and both input and output variables are standardized and normalized to ensure a comparability of the data. Typically, the correctly formatted data are divided into training and testing sets in an 80%-to-20% ratio [115], after which an appropriate mathematical model is selected [116].
• The model is trained, and to prevent overfitting (where the model performs well on the training datasets but struggles with out-of-sample data), hyperparameter optimization is needed to balance variances and biases. When choosing hyperparameters, strategies like a grid search can be employed [117], with cross-validations serving as the scoring method [118].
• The model is validated, and various metrics are chosen to assess its accuracy. Common evaluation indicators include the mean absolute percent error (MAPE), mean absolute error (MAE), and CV (RMSE) [119], with CV (RMSE) being particularly favored, due to its ability to provide a unitless measurement, which facilitates straightforward comparisons of indicators [120].

In constructing surrogate models, various machine learning methods have been utilized in relevant studies of Table 3, including ANNs [24], gradient boosting regression trees (GBRTs) [61], and multiple adaptive regression splines (MARSs) [77], with ANNs being the predominant approach. An ANN is composed of multiple nodes interconnected in a network [121], including input, hidden, and output layers [122]. This method excels in revealing intricate relationships between input and output data [123], outperforming comparable ML algorithms in processing high-dimensional or nonlinear data, as well as exhibiting a superior prediction accuracy and generalization capabilities [124]. Sharif et al. employed an ANN to develop a surrogate model trained and validated using EC data from DesignBuilder, which was applied thereafter to assess ECs, LCCs, and LCAs in retrofit planning. Their findings demonstrate a strong correlation between surrogate and simulation model outputs, and the computation time can be significantly reduced from 170 h to 150 s [45].

In addition, two articles in the literature chose to explore building retrofitting strategies by establishing mathematical models. Studies [24,61] explored retrofit strategies using 12,806 residential buildings in Zurich and 4900 non-residential buildings in the UK, respectively. The researchers selected a GBRT and an ANN as machine learning methods to directly establish mathematical models, respectively, and the calculation accuracy of the models reached over 90%.

3.3. Optimization Objectives and Parameters

In the process of constructing optimization models, it is necessary to determine optimization objectives and parameters according to the research problem and purpose. The optimization objectives can reflect the building performance before and after the retrofit, while the optimization parameters represent the retrofit means and methods to be taken.

3.3.1. Building-Performance Indicators

The articles reviewed in this paper focus on optimization objectives related to building CEs, encompassing LCCEs and OCEs, among others. Given that OCEs constitute more than 70% of the life cycle [125], some studies calculate CEs specifically during the operational phase rather than across the entire life cycle [30]. However, retrofit measures involve additional materials and components, leading to increased ECEs [126], which must be carefully considered when assessing environmental impacts [29]. Furthermore, certain indicators can gauge the built environment performance and indirectly reflect CEs, such as the GWP, which is used to measure the heat-trapping potential of greenhouse gases in the atmosphere and is expressed as CO₂-eq in calculations [127].

The built environment profoundly impacts personnel comfort, productivity, and well-being [128]. Therefore, the process of building retrofits must integrate environmental, economic, and social criteria to maximize sustainable benefits [129]. Research indicates that ECs, costs, and thermal comfort are the most frequently selected objective functions in optimization studies [130]. Social criteria, such as thermal comfort and indoor environmental quality, enhance the quality of life and satisfaction [131]. Figure 6 and Figure 7 illustrate a heatmap distribution and frequency ranking, showing the co-occurrence of building CE-related objectives with other research goals in the relevant literature of Table 2. It is evident that optimizing building CEs typically considers enhancing ECs, reducing retrofit costs (RCs) and operational costs (OCs), and improving indoor thermal comfort at the same time. There is also significant research focusing on optimizing CEs and costs throughout the building life cycle.

3.3.2. Optimization Parameters and Variables

The CER retrofit of buildings necessitates a comprehensive consideration of multiple variables to enhance the overall building performance. The retrofit variables of the literature listed in Table 2 typically fall into two categories: envelope-related variables and other variables (Figure 8). Envelope variables encompass vertical opaque envelopes (interior and exterior walls and doors); horizontal opaque envelopes (ground, floor, ceiling, and roof); as well as transparent envelopes (window-related variables). Others primarily include building forms (such as atria and corridors); sunspaces; sunshade components; building equipment and energy systems (HVAC, lighting, and ventilation systems); and renewable energy systems (like photovoltaic (PV) modules and solar thermal devices). Unlike new building designs, considerations regarding building orientation and form parameters are generally unnecessary in retrofits. Figure 8 illustrates that discussions in the literature often focus on building equipment and energy systems, the window thermal performance, the thickness and type of exterior wall insulations, the thickness of roof insulations, and the area covered by PV panels. Transparent enclosures, although typically smaller in area, exhibit greater heat gains and losses compared to opaque enclosures [132], contributing approximately 20% to 40% of the total [127]. Moreover, rural areas often present greater potential for implementing renewable energy technologies (RETs), facilitated by factors such as accessibility and environmental conditions [133]. Therefore, active and passive measures should be comprehensively utilized during the retrofit to make full use of renewable energy while improving the building performance.

A sensitivity analysis is used to evaluate the correlation between input and output parameters within a model, assessing how changes in inputs influence outputs [134]. It can also be regarded as an optimization search method [135]. For complex scenarios with multiple parameters, a sensitivity analysis can rank the importance of variables [136] and narrow parameter ranges before optimization to simplify model complexity [137]. Before a sensitivity analysis, data sampling is essential to ensure suitable samples, which is often accomplished using Latin hypercube sampling (LHS) [138]. This is a method of stratified random sampling [139] that maintains the representativeness and balance of data while significantly reducing the sample size [140].

Table 7. Classification and application of sensitivity analysis methods.

Sensitivity Analysis Method		References
LSA		[39,40,42,61]
GSA	Metamodel-based method	[41,63]
	Regression-based method	[51,62,65,67]
	Variance-based method	[55,79]
	Density-based method	[55]
	Screening-based method	[70]

shows the classification of sensitivity analysis methods and their applications in the literature, with global sensitivity analyses, such as the regression-based standardized regression coefficient (SRC) method [67], the variance-based Sobol method [79], and the screening-based Morris method [70], dominating. The selection of sensitivity analysis methods should consider both model applicability and computational costs comprehensively.

3.3.3. Constraints on Optimization Objectives and Parameters

Retrofitting existing buildings has certain constraints such as geometric shapes, site layouts, functional zones, and structural limitations [141]. These constraints confine the range of feasible variables [73], compelling designers to select retrofit measures within these boundaries. Additionally, constraints often stem from regulatory norms, industry standards, or empirical data. For instance, the total area of solar heater and rooftop PV panels cannot exceed the overall roof area [39]. Design specifications also govern aspects like window-to-wall ratios [55], while budgetary limits constrain retrofit project costs [60]. For CE targets, the ECEs of retrofit measures need to be less than the OCEs that are reduced as a result of an improved building performance. The constraints restrict the optimization range of complex retrofit parameters, prompting the efficient search for solutions that meet requirements.

4. Optimization and Decision-Making Process

4.1. Optimization Process Based on Data-Driven Methods

By reviewing the optimization objectives and parameters, it becomes evident that the process of decarbonizing building retrofits involves numerous variables impacting the building performance. To enhance efficiency, a systematic and effective optimization process is essential [142]. Mathematical methods such as optimization algorithms can be used to sample, screen, and iteratively refine retrofit strategies, so as to efficiently identify optimal solutions. Figure 9 illustrates the uses of various optimization methods in the reviewed articles listed in Table 4, with the non-dominated sorting genetic algorithm II (NSGA-II) being the most frequently employed [143].

The exhaustive search method ensures finding the optimal solution but faces an exponentially growing search space with increasing variables, which is often impractical for many real-world optimization problems. A case-by-case analysis through sample screening still relies on enumerated comparisons after sampling and is less utilized in building optimizations [47]. A study employed statistical methods like a regression analysis to model relationships between objective functions and retrofit variables [144]. However, due to the complexity and nonlinearity of BPO problems, the applicability of statistical models may be limited to certain scenarios [145]. To mitigate computational costs and achieve automated iterations and the selection of the optimal solution, researchers have increasingly turned to Evolutionary Algorithms (EAs) for solving optimization problems.

An EA is a population-based random search algorithm [146] known for its faster computational speed, greater accuracy, and enhanced adaptability compared to direct search methods. Including genetic algorithms (GAs), the ant colony optimization (ACO), and the particle swarm optimization (PSO), an EA is effective in navigating the complex search spaces related to building optimization problems [147].

A GA is effective for solving nonlinear and discontinuous problems in building retrofits [148]. It enables a simultaneous search across multiple points in space, thereby reducing the possibility of converging to the local minimum [149]. NSGA-II stands out as one of the most efficient EAs [150], though its performance diminishes with problems involving four or more objectives [151]. NSGA-III [152] offers notable advantages over NSGA-II by enhancing its capability to handle multiple objectives and achieve superior solution distributions [153]. Son et al. examined the optimization of ECs, CEs, RCs, and thermal discomfort hours (TDHs) in public building retrofits, comparing NSGA-II, PSO, multi-objective evolutionary algorithms based on decomposition (MOEA/D), and NSGA-III algorithms. Their findings indicate that NSGA-III exhibits superior diversity and convergence [54]. Furthermore, research has shown that the strength Pareto evolutionary algorithm2 (SPEA2) and NSGA-II perform similarly in solving MOO problems [154], but SPEA2 tends to outperform NSGA-II in high-dimensional objective spaces involving three to four objectives [155].

Similar to GAs, PSO belongs to population-based evolutionary algorithms with meta-heuristic characteristics. However, it does not employ computational processes such as crossovers and mutations [156]. The principle of PSO involves particles searching for optimal solutions by continuously moving through multidimensional search spaces [157].

An optimization analysis necessitates numerous evaluations to achieve near-optimal solutions, leading to relatively extended processing times. Hence, the performance of optimization algorithms plays a crucial role in the efficiency of the optimization process, prompting the selection of suitable methods tailored to specific research requirements [90].

Comprehensive optimization tools integrate multiple algorithms, harmonizing the optimization process.

Table 8. The use of optimization tools in the literature.

Optimization Tool	References	Optimization Tool	References
MATLAB	[25,50,72,74]	Python	[48,56,58,60,62,65,71,73,77]
SPSS	[40]	Octopus	[51]
jEPlus + EA	[41,57,67]	JESS + JEA	[63]
MOBO	[43,49,64]	CPLEX	[68]
Excel-VBA	[46]	Gurobi	[69]
MultiOpt	[47]	PLOOTO	[75]

illustrates the application of optimization tools in the literature and reveals that Python and MATLAB are the predominant general programming tools utilized in research.

Python is a high-level interpreted programming language. Its ease of learning and vast library of tools have popularized its use in data science and programming [158]. In the context of BPO, Python serves as an application platform for data-driven methods, making the implementation of optimization algorithms more convenient and feasible. For instance, Shadram et al. [58] utilized the pymoo package in Python to implement the NSGA-II algorithm and focused on optimizing the life-cycle energy consumption (LCE), LCCE, and LCC of a conceptualized apartment in various locations in Sweden through retrofit strategies. This study integrated performance simulation and optimization processes using Python and EnergyPlus extension plugins within Grasshopper, establishing a complete and fast optimization process.

MATLAB [159] is a versatile tool for algorithm development, data visualization, and program interaction. The MATLAB Optimization Toolbox facilitates collaborations with other programs, making it an ideal numerical environment for optimization tasks that rely on third-party simulation programs. However, the programming language demands of MATLAB and the complexity of integrating simulation programs can pose challenges for architects. Specialized optimization platforms, like jEPlus + EA [160], can integrate energy simulation programs effectively. Despite offering fewer optimization algorithms, they excel in combining simulation tools and optimization methods with less programming demands, thus enhancing accessibility for architects.

Figure 10 illustrates the relationship among the simulation tools, optimization algorithms, and optimization tools as utilized in the literature. This figure highlights the broad applicability and interaction of tools and methods such as DesignBuilder, NSGA-II, and Python, facilitating diverse applications in research.

4.2. Solution Set Evaluation and Decision-Making Methods

For MOO problems, due to conflicting objectives constrained by each other, the resultant solutions often form a set rather than a single solution, known as the non-dominated or Pareto optimal solution set. The spatial distribution of this set is referred to as the Pareto front. Figure 11a,b illustrate typical examples of the Pareto front for two and three objectives, respectively [161]. Throughout the optimization process, each generation of solution sets requires an evaluation based on metrics such as convergence, uniformity, and diversity [162], with assessment indicators including the inverted generational distance (IGD) [163] and the hypervolume index (HI) [164], among others. The quality of the solution set improves as it approaches the true Pareto frontier, exhibiting better uniformity and a wider distribution.

Each solution within the Pareto front represents a non-dominated outcome, necessitating post-processing decision-making methods to identify the optimal solution. The technique for order preference by similarity to an ideal solution (TOPSIS) [165] and the Utopia-point method are two commonly employed approaches. TOPSIS ranks solutions by evaluating their distances to both positive and negative ideal solutions, accommodating different weights assigned to objectives based on decision-makers’ preferences to derive optimal solutions. Song et al. used the entropy-based method to assign weights to indicators and calculated the information utility value of each indicator based on information entropy, which avoids the influence of human factors compared with subjective empowerment methods. According to the entropy-based method, the GWP, LCCs, and TDHs were assigned weights of 37.29%, 34.17%, and 28.54%, respectively, and were applied to the TOPSIS calculation. The GWP of the final optimized solution can be reduced by 2720 kg/m² (about 61%) [55].

The Utopia point serves as an ideal reference within the search space, representing the optimal solutions across all objective functions. Solutions on the Pareto front are assessed by their proximity to this point, with the closest solution indicating the optimal compromise when all objectives are equally weighted. Figure 12a,b depict examples of the TOPSIS [166] and Utopia point [167] methods in 2D coordinates.

5. Discussions

5.1. The Research Status Quo

The increasing prevalence of relevant research in recent years may be attributed to a deeper understanding and increased interest in the significance of CERs within the building construction industry. Advances in science and technology and the widespread adoption of computing have facilitated the continuous evolution of BPS technologies and mathematical algorithms, thereby rendering related research more achievable.

Most research has predominantly been conducted in Europe. Achieving the CER target by 2050 necessitates the extensive decarbonization retrofitting of large-scale existing buildings [168]. Moreover, with the fixed form and orientation of buildings slated for retrofitting, emphasis can be placed on optimizing the envelope and systems, thereby facilitating the application of optimization techniques [21].

Building CER projects typically involve multiple objectives and variables. Integrating BPS with data-driven methods such as optimization algorithms and ML can streamline the process, minimizing redundant tasks and facilitating a quicker and more accurate identification of optimal solutions.

5.2. Optimization and Surrogate Models

Using physical simulation methods to establish an optimization model can accurately reflect the relationship between building parameters and performance and evaluate the performance improvement potential, but the simulation time is relatively long. The use of surrogate models significantly enhances computational speed, and there exist variations in the selection and training of mathematical models across studies. For instance, the number of nodes in the hidden layer of an ANN varies depending on task objectives [169]. Likewise, the division ratio of the datasets into training and testing sets lacks uniformity, and the result reliability may hinge on the training and testing data used in surrogate model developments [135]. Therefore, the choice of model and computational approach should be determined by each task’s unique characteristics, such as the data volume and prediction goals [170].

Additionally, addressing issues such as model generalizations and transferability is needed to foster the widespread adoption of surrogate models. Transfer learning methods offer a viable approach to applying data acquired from previous tasks to similar or related tasks [171], which is particularly effective when the training dataset for the targeted task is not large [172]. The successful application of this approach depends on the similarity of data sources and the feasibility for predicting the building performance [173]. Similarly, in energy prediction research for large-scale buildings, transfer learning can solve problems like data scarcity and time-consuming data collection, enhancing the accuracy of data-driven models [174]. A study used transfer learning methods to apply evaluation models pre-trained using Swedish building datasets to predict the performance of buildings in China, which can effectively support building retrofit decisions in data-scarce regions [175].

The clarification and classification of optimization objectives and parameters can provide researchers with experience and paradigms for retrofits and serve as a reference for sensitivity analyses and variable screening outcomes. While focusing on CERs, the relevant studies also take into account optimization goals in terms of building ECs, thermal comfort, and costs [25]. A study also found that cost parameters are particularly common in MOO problems of private housing retrofits, and householders are more willing to carry out CER retrofits on the premise of economic benefits. Therefore, cost savings are as important as CERs, or even more critical [176]. Thus, achieving a balance among multiple performance indicators is essential in retrofitting efforts.

Retrofit strategies typically consider enhancing the performance of building envelopes and adjusting parameters of energy systems. Currently, there is a heightened focus on utilizing renewable energy sources, particularly PV and solar thermal systems. Additionally, emerging technologies and materials, such as switchable blinds [177], PV walls [178], electrochromic windows [179], phase change materials (PCMs), and cool paints [180], among others, are increasingly being employed to enhance the building performance. The vertical greening system (VGS) on roofs and walls are also noteworthy for their potential to mitigate GHGEs, urban heat islands, and noise pollution [181].

5.3. Optimization Methods and Tools

In building CER problems, the choice of optimization methods varies depending on specific objectives and parameters, necessitating tailored selections based on the actual problem [182]. Furthermore, it is crucial to assess the computational efficiency and stability of these methods. While prioritizing computational speed, it is essential to strike a balance between model simplicity and accuracy [97].

The utilization of general optimization platforms like MATLAB necessitates proficiency in a complex programming language, and the constant switching between modeling and optimization environments also poses an inconvenience. Hence, there is a need to integrate programming languages into toolkits that allow for direct applications. The Grasshopper-based Octopus tool [183] integrates optimization algorithms with building parametric modeling and is widely adopted for current BPS and optimization [184,185,186]. Nevertheless, further exploration is required, particularly in its application to building retrofit processes aimed at achieving CER goals. In addition, regarding the assessment of the solution set quality, a study compared the convergence and diversity using metrics such as IGD and HI [187].

In recent years, scholars have started to focus on integrating optimization algorithms with Building Information Modeling (BIM) to enhance decision-making abilities in building retrofitting [188]. As data-driven methods advance, there is a corresponding need to further develop tools that integrate these methods with simulation software, enabling timely and efficient applications in BPO. Additionally, establishing a database using existing building performance datasets [189] and implementing a decision support system based on data-driven methods can assist decision-makers in selecting optimal retrofit strategies based on real-world conditions [26].

5.4. Future Work

Uncertainties stemming from user behaviors and climate change present formidable challenges in building retrofitting [8]. Factors such as user preferences [190], comfort requirements [191], and energy-saving awareness exert a substantial influence on the building performance, making user behaviors one of the primary contributors to the disparity between BPS and actual operation data [192] and a pivotal factor shaping retrofit strategies [193]. For instance, Li et al. [194] integrated three user-behavior patterns—adjusting sunshade components, managing lighting, and ventilating by opening windows—into the optimization of sunshade devices. Their findings revealed how adaptive user behaviors significantly impact room illumination and the simulation of energy loads, thereby influencing the layout and sizing of sunshade panels. Only one of the reviewed studies discussed the impacts of user behaviors on building CEs [70]. This study focused on public buildings in the hot summer and cold winter region of China and implemented a time-of-use pricing strategy to regulate occupants’ energy-use behaviors, including the temperature set point of air conditioners and the usage time of electrical appliances. This strategy can reduce users’ ECs and CEs by 26.4% and energy expenditure by 26.1%. Current simulation studies often rely on fixed schedules, overlooking active user adjustments and the variability and spontaneity of behaviors. Future studies should incorporate more precise and comprehensive behavioral models into BPS frameworks [142].

A study developed a climate model based on atmospheric CO₂ concentration changes, projecting that the Earth’s surface annual average temperature will rise by 1.5 °C by 2050 and by 2–4 °C by 2080 [195]. Given the long lifespan of buildings, climate change should be factored into energy-saving analyses [196] to ensure the applicability and efficacy of retrofit strategies [197]. Data-driven approaches can be used to generate future weather data and establish dynamic climate models.

The building performance is closely correlated with the outdoor environment [198]. Factors such as community greening, urban heat islands, and global warming alter microclimates, influencing the regional building performance [199]. Therefore, using uniform meteorological parameters for predicting building ECs can lead to inaccuracies [200], highlighting the need to integrate microclimate models into energy-efficiency predictions. Currently, optimization research for building CER retrofits has expanded beyond individual buildings to include local outdoor environments, exploring the effects of tree shading and the PV coverage of building façades on the life-cycle GWP of buildings and outdoor thermal comfort [201]. As such, there is still great potential for research on decarbonization retrofits at the scale of building groups or blocks.

6. Conclusions

This paper provides a review of the application of data-driven methods in building retrofits and performance optimization from the perspective of CERs and also discusses recent trends and advancements in various studies. A comparison of modeling methods, optimization variables, objectives, and decision-making methods across 45 relevant papers revealed the following findings:

• There are usually two workflows to optimize the building performance. One is the workflow of the optimization of the physical simulation (model surrogate) performance: Using the combined input of a building site and energy-carbon-related retrofit variables, a BPO process based on a physical simulation is established. The generated datasets can be either iteratively processed with optimization algorithms directly or trained as a surrogate model, validated, and then processed using the MOO method. The other is the workflow of mathematical modeling-optimization analyses: with sufficient actual field-measured empirical data available, data-driven methods, such as regression or machine learning, are used to develop mathematical models, and multiple objectives are comprehensively optimized from the perspective of building CERs.
• A building retrofit aims to maximize its benefits by integrating environmental, economic, and social considerations. Therefore, alongside CE objectives, factors like costs and thermal comfort should also be taken into account. There are 27 relevant studies in Table 2 related to the comprehensive optimization of three or more objectives, accounting for 60% of the total. Discussions on retrofit parameters should extensively cover aspects such as the thermal performance of the building envelope, building equipment and energy systems, and the utilization of renewable energy sources.
• Data-driven methods applied in optimization enable the sampling, screening, and iterative refinement of retrofit plans using computational tools, facilitating the determination of optimal solutions. The advancement and deployment of surrogate models make simplified mathematical calculations replace complex physical simulations, which further enhance optimization efficiency while ensuring accuracy.
• In the reviewed studies, only 2.2% (1 article) and 6.7% (3 articles) of the total focus on the impacts of human behaviors and climate change on building retrofits, respectively. Future research should delve deeper into the application of data-driven methods in building CER retrofits and BPO, considering user behaviors and variations in retrofit conditions amid long-term climate change scenarios. In addition, more work is needed to improve the accuracy of surrogate models and enhance generalizations and transfer capabilities.

Within the broader context of global warming and building stock renewals, building decarbonizing retrofitting and performance enhancements have emerged as focal points of research. This review highlights the capability of integrating data-driven methods with BPS to screen and determine retrofit plans, showing its significant applicational value in building CERs. It will also make a greater contribution to the CER targets of the building construction industry in the future.