Enhancing Cost Prediction and Estimation Techniques for Sustainable Building Maintenance and Future Development

Abstract

Building maintenance is crucial, yet predicting financial resources for it remains challenging, particularly during the design and construction phases. This research aims to analyze and synthesize existing studies on maintenance cost estimation, with a focus on identifying key trends, methodologies, and sustainability considerations. The review finds that most studies emphasize educational and office buildings, while limited attention has been given to infrastructure such as bridges and roads. Moreover, growing attention is being directed toward early-stage maintenance cost estimation and integrating sustainability principles into cost prediction models. The findings underscore that incorporating sustainability factors in maintenance planning enhances long-term performance, reduces lifecycle costs, and supports future-ready building management. The study concludes by highlighting the need for more comprehensive, sustainability-oriented frameworks to improve the accuracy and applicability of maintenance cost estimation in the built environment.

1. Introduction

Not so long ago, the construction cost was expressed only through the capital costs of acquiring land ownership, design costs, obtaining a building permit, and building construction costs. Today, there is more and more talk about the construction costs that will appear after the building is used, i.e., after obtaining a use permit [1]. All buildings are built to last as many years as possible, and the economic life of a building is an average period of 30 to 50 years [1,2]. Of course, the economic life is different for a family house, a building, or a bridge. Service life is the time after construction/installation during which the building or its parts meet or exceed the required behaviour or property [3]. The economic life is reached when the building no longer meets requirements from an economic point of view, i.e., when there is a variant that meets the intended function at lower costs [1]. However, proper and timely maintenance is essential for the structure (building, bridge, road) to be usable during its lifetime and to be used safely [4]. Adequate and regular maintenance refers to implementing a suitable, economically efficient strategy that anticipates potential defects or failures, thereby highlighting the importance of preventive maintenance [5].

Maintenance is carried out so that the building or some part of it retains its original purpose and so that all its users can be guaranteed safe use [6], that is, in order to meet the basic requirements for the building, such as mechanical resistance and stability, safety in the event of fire, preservation of hygiene, health and the environment, ensuring safety and accessibility during use, noise protection, energy management and heat conservation, and sustainable use natural sources [7]. In the life cycle of a building, the period of use of the building lasts the longest, and for this reason, proper maintenance of the building is essential because this is where costs appear that may not be economical. Nowadays, it is increasingly visible that engineers, architects, and investors pay more and more attention to reducing the costs of maintenance and use of buildings or, more importantly, to reducing the total costs of projects, while this was not the case earlier because attention was paid exclusively to reducing construction costs [8]. Maintenance activities comprise various systematic measures designed to uphold the operational performance of a facility, safeguard user safety, and maintain the asset in satisfactory condition to prolong its lifespan [4,9,10]. If maintenance were not carried out on time and to a sufficient extent, the building would collapse relatively quickly. Building maintenance itself, especially when managing several buildings (e.g., city housing fund, university building group), should be carefully programmed in order to optimize investments and achieve optimal allocation of funds to priority operations [9].

Building maintenance costs include the costs of necessary work, materials, and other related costs that occur when maintaining a defined level of easement of a building, and include the costs of corrective, preventive, and reactive maintenance of the entire building or its parts [11].

Given the current research trends, it is important to acknowledge the significance of sustainability within the context of building maintenance [12,13]. Integrating sustainability into the maintenance of buildings goes beyond just saving money; it encompasses the wider effects on the environment and society. Sustainable maintenance approaches strive to minimize resource utilization and decrease waste generation, hence reducing the carbon emissions associated with buildings during their entire lifespan [14,15]. This includes the utilization of sustainable materials for repairs, the incorporation of energy-conserving equipment, and the adherence to environmental rules during maintenance operations. Furthermore, the practice of sustainable maintenance encourages the reuse and reutilization of building materials, thereby improving the durability and flexibility of constructions. Integrating sustainability into maintenance techniques not only improves the economic lifespan of buildings but also promotes environmental conservation and the well-being of future generations [16,17]. This comprehensive strategy guarantees that buildings maintain their functionality, safety, and environmental responsibility by connecting maintenance methods with global sustainability objectives [18].

In order to be able to plan financial resources for maintenance for the next year or the next few years, it is essential to estimate and determine the maintenance costs because proper maintenance of buildings is only possible with sufficient funds. By using a model for estimating building maintenance costs in the early design phase, it is possible to rationalize maintenance costs, plan the exact future costs, and prevent or reduce the impact of business interruptions due to maintenance needs.

Despite the growing awareness of the importance of maintenance, research on cost prediction and estimation has mostly focused on construction rather than post-construction phases. Construction cost estimation models are generally based on well-defined quantities, materials, and labor inputs within a limited project timeframe, whereas maintenance cost prediction must account for long-term uncertainty, degradation rates, user behavior, and the impact of environmental conditions. This fundamental difference makes maintenance cost forecasting inherently more complex and less predictable. Furthermore, few studies have comprehensively reviewed and compared the existing approaches to maintenance cost estimation, particularly from a sustainability perspective. Given that maintenance activities contribute significantly to a building’s total life-cycle cost and environmental footprint, there is a pressing need to consolidate current knowledge, identify methodological gaps, and propose directions for more sustainable and data-driven prediction models. Therefore, this review aims to address this gap by critically analyzing the existing literature on maintenance cost estimation techniques, highlighting their limitations, and emphasizing how integrating sustainability principles can enhance accuracy, efficiency, and long-term value in building management.

2. Methodology

The methodology for reviewing and analyzing cost prediction and estimation techniques for sustainable building maintenance followed a systematic and structured approach. Relevant studies were identified, filtered, and reviewed from reputable scientific databases, including Web of Science and Scopus, as well as selected institutional repositories to ensure comprehensive coverage of both international and regional research. The review focused on peer-reviewed journal articles and conference papers published in English and Croatian, addressing the estimation or prediction of building maintenance costs with an emphasis on sustainability, life-cycle cost management, and predictive modeling. Studies published between 1961 and 2025 were included to capture both foundational works and the most recent advancements in data-driven and artificial intelligence-based approaches. A keyword-based search strategy was developed using combinations of terms such as “building maintenance cost estimation,” “maintenance prediction,” “life-cycle cost,” “sustainable maintenance,” “cost modeling,” “regression analysis,” and “artificial neural networks.” The search results were screened in multiple stages: first, by title and abstract to remove duplicates and unrelated studies, and second, by full-text review to assess methodological quality and relevance to the research objectives. Only studies that presented empirical data, analytical frameworks, or predictive models applicable to building maintenance were included, while non-peer-reviewed reports and non-scientific sources were excluded. Each selected publication was carefully analyzed to identify its research focus, building type, data source, estimation or prediction technique. The reviewed studies included a range of building categories—such as educational, residential, offices, and infrastructure facilities—and employed diverse methodologies, including statistical regression models, artificial intelligence and machine learning techniques, and life-cycle cost analyses. Comparative evaluation of these studies allowed for the identification of key trends, limitations, and emerging opportunities for integrating sustainability principles into maintenance cost prediction. The synthesis of findings focused on highlighting methodological patterns, gaps in data availability and standardization, and the growing role of intelligent modeling techniques in achieving sustainable maintenance management. The results of this review provide a foundation for proposing future directions toward more comprehensive, sustainability-oriented frameworks for predicting and estimating building maintenance costs.

3. Methods for Development of Cost Prediction Models

Regression and artificial neural networks are the two most common methods used to develop building maintenance cost prediction models. Both methods will be briefly explained in the following two sections.

3.1. Regression

Regression analysis is generally employed to determine how independent variables influence or predict a dependent variable [19]. Given known inputs for the independent variables, regression modeling enables the prediction or estimation of the dependent variable’s value [20]. Model construction aims to produce a tool that simplifies complex realities while maintaining adequate accuracy. It should be comprehensible and practical, yet detailed enough to account for essential influencing variables.

Regression models are used to analyse data from unplanned experiments derived from observations and uncontrolled phenomena or historical records [21,22].

Regression models are used for several reasons, namely:

• Descriptiveness—describes the strength of the connection between input and output variables.
• Alignment—according to independent variables related to dependent (output) variables.
• Predicting—determination of risk factors that affect output (dependent) variables.
• Cost—sometimes, the collection of dependent variables can be costly, but the collection of independent variables is not.
• Time—since the output (dependent) variables occur in the future, it is difficult to know the values of these variables in the present, but using regression analysis, they become available immediately [23,24].
• Explanation—the regression model estimates the influence of the independent variables on the dependent variable, i.e., the result [25].
• Adjustment and process control [26].

The purpose of the model is to observe the relationships between two or more observed variables, which can be established by simply comparing the series members. A relationship between phenomena exists if an increase in one variable is followed by an increase in another and vice versa. In the first case, the direction of the relationship between phenomena or variables is positive, and in the second case, the direction of the relationship is negative [27].

Regression analysis fits a function in a partial data set. Linear regression is fitting data with linear functions. The above is done using the method of least squares [28]. The direction of the linear trend, which is laid between the set of (original) data points by the method of least squares, should be placed so that the sum of the deviations of the original values of the trend is equal to zero and that the sum of the squares of these deviations is minimal [29]. The least squares line is also called the regression line, so that’s where the name regression comes from [28].

Regression analysis involves constructing a model that relates a dependent variable to one or more independent variables for the purpose of estimation or prediction. The approach is defined as simple regression when only one predictor is included and as multiple regression when several predictors are incorporated [30].

3.2. Artificial Neural Networks

Artificial Neural Networks (ANNs) constitute a non-traditional computational framework that models the way the human brain and nervous system process information, offering enhanced capability for pattern recognition and complex data analysis [31]. A simple and accurate definition of ANNs was formulated in 1990 by Alexander and Morton. According to this definition, a neural network is a “massively parallel distributed processor made up of simple processing units having a natural propensity for storing experiential knowledge and making it available for use, and is similar to the brain in the way it acquires and stores knowledge” [5,32,33]. ANNs are structured in layered architectures made up of computational nodes. Each node functions similarly to a neuron, processing input data through weighted connections that either enhance or suppress signals, thereby determining the contribution of each input variable to the network’s predictive function [34]. In a standard neural network architecture, numerous interconnected neurons serve as computational processors that output continuous activation values. Activation of input neurons occurs through external stimuli, whereas activation of deeper neurons results from weighted connections with earlier network layers [35]. ANN is a method commonly used for learning [36]. Compared with traditional analytical techniques, Artificial Neural Networks demonstrate superior performance when the underlying input–output mapping of a system is nonlinear or not explicitly characterized [37].

4. Research on Predicting and Estimating Building Maintenance Costs

4.1. Comprehensive Cost Estimation and Predicting in Construction Maintenance

Estimating and predicting costs is a daily process. One of the most common ways of predicting construction costs is the creation of cost estimates. In the cost sheet, a detailed description is made, and the quantity is entered for each item from the proof of measures; in the price analysis, the price of each item is calculated, and, finally, by multiplying with the total quantity and price, the price for a specific position of the works is obtained. After that, all positions by type of work are added up. In the end, all types of work are added up, and the total building price is obtained. The better the costs are calculated, estimated, or predicted, the lower the risk for the investor or any other person that they will not have enough funds to carry out the intended work (construction, maintenance).

Cost predicting is an essential process in any business because it precedes price calculations and resource allocation in the project life cycle. In fact, obtaining input data for estimating costs is difficult, especially when the scope of work is poorly known, which can lead to poor and rough estimates. Of course, the more the scope of the project is known, the greater the possibilities for generating more accurate estimates because more project specifications are available [38].

When estimating construction production costs, costs should be known and monitored at the point of origination. In addition to the above, costs can be observed over time so that they can be fixed or variable. Investments are affected by various influences, such as manufacturers and suppliers of construction materials and equipment, banks, financial organizations, inspections, design companies, various companies related to the provision of construction services, utility companies, the investor, and his consultant [39].

Cost estimation is developing an approximation of the funds needed to complete project activities. The key advantage of this process is that it determines the amount of costs needed to complete the project. Cost estimates are predictions based on information known at a particular point in time, and they include the identification and consideration of alternative costs for project initiation and completion [40].

Cost estimates should be continuously reviewed and updated throughout the project to incorporate new information and validate prior assumptions. As the project advances through its life cycle, the precision of cost projections typically improves. For instance, during the early conceptual phase, preliminary estimates may exhibit an accuracy range of approximately −25% to +75% [41]. Later in the project, as more data becomes known, the final estimates could be more accurate and the accuracy range could be −5% to +10% [40].

Furthermore, during the project’s planning, creation, and development, it is necessary to take care of the possibilities, that is, maintenance costs, right at the beginning of the project. In the early stages of the project, such as during the feasibility study and preliminary design preparation, the project’s costs are low compared to the whole project. In this initial period, it is possible to influence, with low costs, the project’s total costs. As the project enters the construction phase, the changes are more and more expensive than the total cost. At the end of construction, when the building is already built, the costs of changes are high [42].

The ability to influence the final parameters of the project without significantly affecting the costs is most significant at the beginning of the project. It decreases as the project progresses toward completion. The costs of making changes and correcting errors tend to increase significantly as the project nears completion. Although the stated determinants remain present to a certain extent in almost all project life cycles, they are not always present to the same extent [40].

The curve showing the cost of changes over time is called the Boehm Curve after Dr. Barry Boehm, a computer science researcher, who discovered that the average cost of fixing defects increases exponentially the longer it takes to find a defect [43].

In the initial periods of project implementation, methods for quick cost estimation are mainly used. These methods rely on simpler cost estimation models:

• rough estimate based on capacity (number of beds in a hospital, number of rooms in a hotel, etc.);
• assessment by elements or matrix of works-functional groups of works (e.g., for a residential building, preparatory works, foundations, structure, roof, etc.);
• models of cost-effective positions of works (It is based on the Pareto rule, which was named after the Italian economist Vilfredo Pareto, who, based on his research, noted that 80% of the world’s wealth is held by 20% of people [44]) which says that about 20% of the positions define a large part of the total costs, about 80%;
• parametric (regression) models that use formulas that connect costs and one or more characteristics of the object such as height or length [45].

An important prerequisite for predicting and estimating maintenance costs is to have data on building specifications and maintenance costs from the past [46]. Historical maintenance cost data is the most important tool available when preparing and planning a maintenance budget [47]. However, anyone who has ever been involved in estimating construction costs is aware of the difficulty of the job because such estimates are usually based on historical cost data, updated based on experience, and adjusted based on expected economic trends [48]. Therefore, estimates of costs in the future are just that—estimates and it is not certain when these costs will occur nor if they will occur at all [48].

4.2. Previously Developed Models for Predicting and Estimating Building Maintenance Costs

In their work, Bouabaz and Horner proposed a model that links the bridge pavement surface and the repair costs by analysing historical costs. They investigated cost-significant items and analysed 37 cost estimates, where they analysed an average of 40 items for each cost estimate. They concluded that without quality data from the past, it is not possible to determine and adjust the maintenance budget [49]. One of the previously mentioned authors—Bouabaz, in collaboration with Hamami, created a bridge maintenance cost estimation model based on an ANN. Data from 40 projects were used, and the network provided high accuracy. The authors concluded that the developed model is more precise and easier to use than elemental and parametric models, significantly saving time [50].

Asadi et al. dealt with the assessment of the costs of the life cycle of bridges in their paper. They analysed 14 bridges in Chicago. The results indicate that the neural network model offers strong potential for accurate prediction of bridge life-cycle costs. Its main strength is the ability to model complex, nonlinear behaviours commonly observed in infrastructure assets [51]. In their work, Shi et al. used historical data on reinforced concrete bridges on 11 expressways operated by the Shaanxi Province Transportation Group in China. Two models for calculating regular maintenance costs based on linear regression and time series analysis were proposed [52].

Rowan et al. examined the maintenance costs of wastewater treatment plants (WWTPs). They sent 750 questionnaires, and 321 were validly filled in and used in further analysis. The annual maintenance and use costs were analysed concerning the average annual daily flow, the number of inhabitants served, and the equivalent of inhabitants [53].

Boussabaine and Kirkham proposed an innovative simulation-based approach to modelling maintenance costs in local authority sports facilities in Great Britain. The results show that the gross area, the pool size, and the number of users is the key factors affecting maintenance costs in sports centre buildings. The testing and validation results showed that multi-regression linear models can model maintenance costs. The results of that work can be used to manage and determine the costs of maintaining sports facilities similar to those analysed in that work [54].

The paper by Zarembski and Patel presents the approach used in developing the methodology for estimating the costs of maintaining railway corridors for mixed passenger and freight traffic at higher speeds. All elements are included: track, bridges, buildings, communications, and signals. The final costs are presented as a set of cost matrices in terms of total cost per mile of track and cover a range of traffic and track configuration combinations, with minimum and maximum costs developed for each element in the cost matrices [55]. Also, a doctoral dissertation by Ling deals with railway renovation and maintenance costs. This dissertation aims to present a structured methodology that estimates the costs of renovating and maintaining railway infrastructure when quantitative data on costs in the project’s life cycle are missing. The model is implemented within the prototype of the software tool [56].

In the paper [57], authors Bello and Loftness analyzed the state of public school infrastructure in the United States and emphasized the urgent need for increased investment in maintenance. Existing methodologies for estimating maintenance needs were critically evaluated, and a new model known as the Bello–Loftness plant value model, was developed to more accurately reflect the true requirements of school facilities. This model incorporates a standardized adjustment to account for deferred maintenance and introduces a prioritization framework for addressing maintenance backlogs. Additionally, the study explored appropriate custodial and maintenance staffing levels, noting that current staffing was below recommended standards despite growing building complexity and aging infrastructure. By providing practical guidelines for both investment and staffing, the research offers a comprehensive approach to improving the physical condition of school facilities, with the ultimate goal of supporting better health, safety, and educational outcomes for students and staff [57].

In his doctoral dissertation, author Krstić created models for estimating the costs of maintenance and used them in the example of the buildings of the Josip Juraj Strossmayer University in Osijek. The possibility of applying multiple linear regression (MLR) was determined, and the determinants of the building that most affect these costs were defined. The adopted model for estimating usage costs has three variables: the area of sanitary facilities, office premises, and the average number of employees. The determinant by which it is possible to estimate the costs of maintaining the use of faculty buildings is the area of communication (corridors, entrance vestibules, and halls) [46].

The authors Krstić and Marenjak investigated the possibility of collecting data from the past about the costs of maintenance and use of the buildings of Josip Juraj Strossmayer University in Osijek [11]. Questionnaires were created on maintenance and use costs, general data, and building operation, all in order to determine a database of independent variables (general construction and use determinants) and dependent variables (maintenance and use costs). Regression analysis was used. Also, the same authors described the development and validation of a model of the average annual maintenance costs and use of university buildings that have similar determinants in Osijek. They showed that it is possible to predict the annual costs of maintenance and use of buildings of similar purpose [58].

Author Liu created a maintenance and usage cost estimation model for office buildings. The Analytic Hierarchy Process (AHP) method was used, and the program prototype was coded in Visual Basic in the Microsoft Windows environment. The system he created provides numerical and graphical reports [59].

Also, for office buildings in Malaysia, Shah et al. created models in the SPSS program (short for Statistical Package for the Social Sciences) using MLR and concluded that the regression model can be used in practice [60].

Kwon et al. created a model for estimating the costs of maintaining residential buildings based on the Genetic Algorithm (GA) and Case Based Reasoning (CBR). Data from 1978 to 2009 were used, and 90 cases were analysed [61].

Authors Li and Guo investigated the maintenance costs of four university buildings in their two papers. The models were made using simple linear regression (SLR), MLR, and ANN, and the one made using an ANN turned out to be the best model [62,63].

Mahmoud et al. made a model for predicting the costs of historic buildings. The paper combined a literature review, cost modelling, survey, and case studies and used MLR. Model validation results show that the maintenance cost prediction model has an accuracy of about 93% in predicting annual maintenance costs for historic buildings based on building age, gross floor area, and building performance index [64].

Based on data from 30 years of school building management budgets, Lee and Jeon created a model to estimate the cost of maintaining elementary schools. The research results can be used as an objective criterion for long-term and realistic school maintenance budget planning [65].

The paper [66] proposes a model that can determine maintenance and repair costs using statistical analysis of actual cost records. The costs of maintenance and repairs in educational institutions are investigated to determine key performance indicators (KPIs) and develop a cost estimation model for integrated facilities management.

In his doctoral dissertation, Nipp researched 34 buildings at Tennessee Martin University in Martin, Tennessee, USA. He analysed buildings on a university campus and created a regression model based on historical cost data for ten years. It emphasizes the importance of having historical maintenance data available and points out that no cost estimation expression provides 100 percent certainty in calculations and budgeting, but it can be beneficial [67].

In the work of the authors, Gausmann et al., the application of machine learning algorithms is presented: multi-layer perceptron-neural network and k-means algorithm to estimate the level of services required for the preservation of highways in Brazil. A database was created that contains data for the history of regular maintenance, a catalogue of road solutions, and price lists. Machine learning algorithms were applied and evaluated, and it was concluded that the k-means algorithm provides the best estimate of maintenance costs for Brazilian highways [68].

In her doctoral dissertation, Gudac Hodanić made a model for estimating the life cycle costs of pontoons and anchoring systems of pontoons and vessels. Machine learning methods were applied: random forest (RF), ANN, support vectors, and vector lifting. In the paper, the basis for estimating the life cycle costs is set, which is based on the climatological and geographical determinants of the location where the pontoons are placed, as well as on the determinants that are determined during the design of the marina [69].

Tijanić Štrok researched maintenance management in public primary and secondary schools in Primorje-Gorski Kotar County. A maintenance management and mathematical models were developed to estimate maintenance costs based on regression analysis. The model mentioned above for effective maintenance of public educational buildings determines the actions taken during the planning, organizing, leading, and controlling maintenance activities. The doctoral dissertation concludes that the implementation of such a model could improve current maintenance management processes [70].

In his doctoral dissertation, Obradović researched the costs of maintaining sewer systems in Croatia. A database was created based on the collected data, and models were expressed using linear regression. It has been shown that the determinants that have the most significant influence on the maintenance costs of sewerage systems are:

• the total length of the sewerage system,
• the total number of connections to the sewerage system,
• the number of pumping stations, and
• the average annual volume of wastewater removal.

It was concluded that using the model could improve the efficiency of the maintenance of sewer systems, given that it is possible to plan the budget for maintenance costs annually [71,72].

The study by Gurmu and Pourdadash Miri developed a machine learning–based approach to estimate building project costs using large datasets from the Victorian Building Authority in Australia. Focusing on multiple building classes defined under the National Construction Code, the researchers tested five regression models—decision tree, random forest, gradient boosting, linear regression, and k-nearest neighbour—to identify the most accurate cost prediction method. Among these, the decision tree model achieved the highest reliability with an R² value of 0.66. Key cost-influencing variables included building class, gross floor area, and wall materials, with building class emerging as the most significant predictor of cost. The study provides valuable insights for improving cost estimation accuracy in the early stages of building projects [73].

Mahpour created a method using machine learning algorithms to improve the precision of estimating building maintenance costs, minimize the risk of overestimating these costs, and reduce resource waste in maintenance. The author emphasizes that this methodology is adaptable to various applications, including anomaly detection, feature engineering, cost estimation, and validation. The primary advantage of this methodology is its applicability to any building type, regardless of geographic location [74].

Plebankiewicz and Grącki proposed an MLR model to predict annual renovation costs in educational buildings. Initially, they analysed 17 University of Krakow buildings built between 1830 and 2017 using data from 2017 to 2020, focusing on building age, number of floors, and usable area. In the second phase, the study expanded to 55 buildings in Krakow and Lodz, built between 1830 and 2014, using the period since the last renovation instead of building age for standardization. The study concluded that renovation costs could be predicted using the usable area and the time since the last major renovation, leading to a trend equation for precise short-term cost predictions [75].

Hauashdh et al. developed a predicting model to estimate the quantity of building defects and the corresponding costs using historical data on defects and the age of the buildings. The study included a total of 40 buildings and utilized MLR analysis. The Python programming language was used to develop a predictive model. The purpose of the model was to verify and confirm that the outputs were logical and aligned with the predicted results. This was achieved by applying the Pearson product-moment correlation coefficient to measure the correctness of the model’s output in relation to the variables [76].

Abuhussain and Baghdadi proposed a novel framework for estimating maintenance and operation costs in construction projects using an Emotional Artificial Neural Network (EANN). Based on data from 313 experts in Ha’il, Saudi Arabia, the EANN outperformed a conventional ANN in both accuracy and reliability. The results demonstrate the model’s potential to improve cost predictions, reduce budgeting uncertainty, and support sustainable building management [77].

Obradović et al. estimated the maintenance and operation costs of a school building using the factor method that is used to calculate the estimated service life (ESL) of an element or assembly under certain conditions. The paper presented the development of a 15-year maintenance plan and program for the building, covering the period from 2024 to 2038 [78].

The study by Wu et al. proposed a hybrid forecasting model to enhance the accuracy of cost prediction in green building projects. Using a dataset of 150 certified green buildings in China, the researchers integrated multiple machine learning algorithms—including Random Forest, XGBoost, Support Vector Machine, and Artificial Neural Network—with feature selection and optimization techniques. This hybrid approach significantly improved predictive performance, achieving an R² value of 0.93, indicating high reliability. Key cost drivers included project size, structure type, certification level, and energy system type. While the study focused on construction cost estimation rather than maintenance, its findings demonstrate the potential of hybrid AI models in achieving more precise cost forecasting for sustainable construction projects [79].

4.3. Summary of Previous Research

A summary of the research on the development of models for predicting and estimating building maintenance costs is provided in

Table 1. Summarized chronological presentation of research on predicting and estimation models of maintenance costs of different types of buildings (Data from [71]).

Building Type	The Method of Creating a Model/Method of Cost Estimation	Authors and Year	Sample Size	Variables	Reliability	Dependent Variable	Maintenance Frequencies
waste water treatment devices	logarithmic transformations	Rowan, Jenkins and Howells, 1961 [53]	321 sewage treatment plants (from 750 surveyed) across the U.S.	treatment type (primary, standard-rate filter, high-rate filter, activated sludge), flow (MGD), population served, annual O&M cost ($/MGD, $/capita)	based on 5+ years of satisfactory operation; ±1 standard error (≈64% confidence) for estimates	annual operation and maintenance cost (per MGD or per capita)	not directly stated—focuses on annual costs
bridges	MLR	Bouabaz and Horner, 1990 [49]	51 bridge repair projects (14 masonry, 13 masonry-concrete, 24 reinforced concrete)	bridge type, deck area (m2), repair cost (£/m2), age, cost-significant work packages (CSWPs)	accuracy ±10%; r = 0.97 (masonry), r = 0.98 (reinforced concrete); CMF = 0.76–0.82	repair cost (total £ or £/m2)	masonry bridges: ~63 years; reinforced concrete: ~18 years between major repairs
sports facilities	MLR	Boussabaine and Kirkham, 2004 [54]	16 sports centres	gross floor area, swimming pool size, number of users, internal finishes, cladding type and condition, roof covering, structure system, hall size, age, glazing ratio, facility rating, total ground floor area, and annual maintenance cost.	validation through Kolmogorov–Smirnov (KS) goodness-of-fit test; the Weibull distribution found to best represent maintenance cost data.	maintenance cost per year	included response (normal/emergency) and programmed (cyclical and preventive) maintenance categories
railroad	expert system	Ling, 2005 [56]	3 industrial case studies used for validation; focus group with 11 experts	project type, asset type, cost elements, expert pairwise comparisons, alternative weights, known cost criteria	12 of 15 model estimates within expected accuracy; validated through prototype software	estimated renewal and maintenance cost (£)	evaluated within life-cycle context; focuses on early-stage estimation
office buildings	AHP method regression	Liu, 2006 [59]	63 office buildings (data from U.S. General Services Administration database)	building age, floor area, occupancy rate, construction type, location, maintenance history, repair cost records	adjusted R2 ≈ 0.78; prediction error within ±10–15%	annual maintenance and repair cost (USD/m2)	evaluated annually; model supports long-term forecasting of maintenance and repair cycles (1–10 years)
bridges	ANN	Bouabaz and Hamami, 2008 [50]	40 bridge repair projects	cost-significant work packages (CSWPs), cost model factor (CMF), and type of bridge work	ANN model accuracy 96%; mean error −0.25%; correlation coefficient R = 0.998; ±4.1% standard deviation	total bridge repair cost	not directly analysed; focuses on cost estimation for repair works rather than maintenance intervals
railroad infrastructure	cost matrix	Zarembski and Patel, 2010 [55]	multiple U.S. Class I railroads and two detailed segments	annual tonnage, traffic mix, curvature, tie type, class of track, operating speed, track geometry	model accuracy within ±10% of observed costs	annual right-of-way maintenance cost	ties 16–60 years, surfacing 2–4 years
school	plant value methodology (PVM)	Bello and Loftness, 2010 [57]	not mentioned	plant replacement value, Current plant value	not mentioned	annual required maintenance budget	long-term maintenance
university buildings-faculties	MLR	Krstić, 2011 [46]	8 university buildings	number of stories, building age, circulation area,	PRESS; RMSECV; R2	average annual nominal costs of maintenance and operation	12 years
bridges	ANN	Asadi et al., 2011 [51]	14 Chicago bascule bridges (≈800 data sets over ~60 years)	bridge length, width, age, initial cost	optimal ANN; MSE = 0.00111; 60% training/40% testing split	bridge LCC including maintenance, repair, and rehabilitation	65 years
university buildings	MLR	Krstić and Marenjak, 2012 [11]	8 university buildings	number of stories, building age, circulation area,	PRESS; RMSECV; R2	average annual nominal costs of maintenance and operation	12 years
university buildings	SLR, MLR and ANN	Li and Guo, 2012 [62,63]	4 university buildings at National Taiwan University; 8430 maintenance records over a 42-year period	building age, number of floors, number of classrooms, presence of elevators	R2 = 0.90; RMSE ≈ 3–7 USD/m2	annual or accumulated maintenance and renovation cost (USD/m2)	25–35 years
office buildings	MLR	Shah Ali et al., 2013 [60]	133 valid responses (≈33% of surveyed maintenance professionals)	skill & knowledge of labour, spare parts stock, spare parts quality, maintenance interval length, maintenance downtime	Cronbach’s α = 0.741; regression model R2 = 0.532 (53.2%)	maintenance expenditure variance (cost performance)	based on scheduled maintenance intervals—performance optimized by balancing frequency
office buildings	MLR	Mahmoud et al., 2015 [64]	3 buildings	building age, gross floor area, building performance index	Cronbach’s α = 0.72–0.83; R2 = 0.933; model accuracy ≈ 93%	annual maintenance cost	annual
university buildings	MLR	Krstić and Marenjak, 2017 [58]	13 university buildings (Osijek, Croatia), 12 years of data (1998–2010)	building age, number of storeys, hallway area, office area, total area, number of students	best regression model: R2 = 0.895, Adj. R2 = 0.816; validation error ≈ 8–10%	annual maintenance and operation cost	4–12 years
primary schools	MLR	Lee and Jeon, 2017 [65]	60 schools (1011 maintenance works)	years since construction, total floor area, school grade, maintenance budget history (30 years), and 45 itemized maintenance categories	R2 = 0.834 (≈ 83% explanatory power); 90% prediction reliability	annual and cumulative maintenance cost	15–30 years
university buildings-university campus	MLR	Nipp, 2017 [67]	34 university buildings (11 years data, 2004–2014)	age, size, current plant value (CPV), initial cost, use, capital improvements	R2 = 0.89; within ±10% of other models	annual maintenance costs	Annual
schools	MLR	Kim et al., 2018 [66]	331 educational buildings	building area, age, campus location, tropical cyclone risk, lightning risk, FEMA flood zone	R2 = 0.366, Adj. R2 = 0.355, p < 0.05; cross-validation (PRESS ≈ SSE = 120.5 ≈ 115.3)	maintenance and repair cost ratio (log-transformed)	seasonal pattern identified-summer (34%) most frequent; leakage and weather events top recurring causes
residential buildings	GA and CBR	Kwon et al., 2019 [61]	90 residential building maintenance cases (70 for training, 20 for validation)	building coverage ratio, floor area ratio, number of buildings, floors, households, parking ratio, maintenance area, completion year	CBR + GA model validated with 20 test cases; average case similarity ≈ 90%; MAER = 18.7% (Monte Carlo–adjusted); strong model robustness	maintenance cost for MEP	derived from retrieved cases: typically 2–4 repairs per component over 30 years; maintenance intervals estimated using similarity-based case history
bridges	MLR	Shi et al., 2019 [52]	11 expressways	age, bridge length, number of lanes	R2, adjusted R2	maintenance cost	3–15 years
pontoons and marina anchor systems	RF, ANN, support vectors, gradient lifting	Gudac Hodanić, 2020 [69]	16 marines	sea temperature, wind influence, tidal influence, concession sea area, number of pontoon piers, pier length, wooden walking surface, total number of berths, number of users, number of inspections over ten years, and concession costs	RF highest reliability	annual maintenance and repair cost	10 years
highways	ANN, k-means algorithm	Gaussmann et al., 2020 [68]	50	stretch length, traffic factor, track grade, number of tracks and shoulders, vegetation area, and the extent and number of safety and drainage devices	error rate decreased	service costs related to the runway, vegetation, safety elements (linear and punctual), and drainage devices	annual
primary and secondary schools	MLR	Tijanić Štrok, 2021 [70]	8	average planned annual maintenance costs, area of sanitary facilities	MAPE, R2	average total actual annual maintenance costs	5 years
sewer systems	MLR	Obradović, 2022 [71,72]	13	length of sewer network, annual amount of wastewater discharge, number of pumping plants	R2, Adjusted R2, MAE, AIC	average annual nominal maintenance costs	10 years
residential, commercial, industrial, public, and uninhabitable structures	DTs, RF, GB, MLR, and KNN	Gurmu and Pourdadash Miri, 2023 [73]	initially 113,000 projects from the Victorian Building Authority (2020 data), reduced to 60,000 after cleaning; additional year (2019) added for validation	building area, type (class), materials used in roof, floor, walls, and frame; categorical encoding applied	best model = Decision Tree with R2 = 0.66 and error = 0.42; individual feature accuracy: Building Class 0.34, Walls 0.26, Area 0.26, Frame 0.14, Roof 0.03, Floor 0.01	cost of building projects (categorized by quartiles into four cost ranges)	not covered/not applicable in this paper
Office and residential buildings	machine learning	Mahpour, 2023 [74]	office and residential buildings in Tehran	Building Condition Index (deterioration measure), age, features from building dataset	validated via cross-validation, comparison of models, anomaly treatment scenarios; the machine learning methodology purported to increase accuracy over deterministic/probabilistic models	building maintenance cost (annual)	the paper distinguishes between planned (active/proactive) maintenance and unplanned (reactive/passive) maintenance
university buildings	MLR	Plebankiewicz and Grącki, 2023 [75]	55	last modernization, usable area	R2	renovation costs	5 years
Construction projects	emotional artificial neural network (EANN)	Abuhussain and Baghdadi, 2024 [77]	313 experts (survey-based data)	building age, building area, structure type, heating system, cooling system	Cronbach’s α = 0.87 (survey); EANN R2 = 0.88 (train), 0.84 (test); RMSE = 1.60 ± 0.05 (train), 1.65 ± 0.06 (test); outperforming ANN	annual maintenance and operation cost	grouped by building age (0–10, 10–20, 20–30, 30–40 years); EANN provides lifecycle cost predictions for each stage
various	MLR	Hauashdh et al., 2024 [76].	40	building ages	Pearson product moment correlation coefficient	number of defects (civil, electrical, and mechanical)	3 years
school	factor method, cost calculation in spreadsheets	Obradović et al., 2024 [78]	1	discount rate, inflation, analysis period, types of costs, service life of building elements	estimated costs	present value of maintenance and operation costs	15 years
transport infrastructure (public construction projects, e.g., highways, roads)	FAHP	Canesi and D’Alpaos, 2024 [80]	141 526 Italian public procurement contracts (2008–2021) analysed	probability of risk occurrence, cost impact, 20 combined risk categories, expert judgment weights, project type, and cost escalation	model validated via expert focus group; consistent results; highest-risk category confirmed by case study results	risk-induced cost escalation	not directly analysed; focuses on construction-phase cost and time overruns due to risk events rather than long-term maintenance cycles
Green building projects	hybrid forecasting model combining machine learning algorithms (Random Forest, XGBoost, SVM, ANN)	Wu et al., 2025 [79]	150 green building projects	project size, structure type, green certification level, materials, energy systems, location, and construction duration, among others.	evaluated using R2, RMSE, and MAE—hybrid model achieved R2 = 0.93, outperforming standalone machine learning methods.	total construction cost of green building projects.	not specifically covered in this study (focus was on construction cost prediction, not life-cycle or maintenance analysis)

. The building type, the model’s construction, the authors, and the year of making the model are listed.

An overview of the number of published research papers related to building maintenance cost prediction and estimation over time is presented in Figure 1. Although early contributions were sparse and sporadic, e.g., only one paper published in 1961 and another in 1990, interest in the topic has notably increased in recent years. From 2010 onwards, the frequency of publications shows a growing trend, with peaks observed in 2017, 2023 (with three papers published), and 2024 (with four papers published). This suggests a growing recognition of the importance of accurate cost estimation in building maintenance, particularly in the context of aging infrastructure, sustainability demands, and the need for data-driven facility management strategies.

Figure 2 presents the distribution of research studies on maintenance cost estimation according to the type of building or infrastructure analysed.

An overview of the literature (Figure 2) shows that studies on building maintenance cost estimation cover a wide variety of building and infrastructure types. The most frequently analysed are educational facilities, including schools and university buildings (a total of 12 cases), highlighting the growing importance of maintaining public-use buildings with high occupancy rates. This is followed by bridges (4), office buildings (3), residential buildings (2), and rail infrastructure (2). Other building types, such as sports facilities, wastewater treatment devices, sewer systems, highways, green buildings, marina systems and construction projects, are represented with one study each. These findings suggest that although there is interest across different sectors, research is predominantly focused on publicly owned buildings, particularly educational facilities. This may be attributed to their continuous use, high operational costs, and the direct impact of facility conditions on users such as students, teachers, and administrative staff.

5. Discussion

When analysing the previous research on the topic, MLR emerges as the most frequently used method, appearing in 17 studies. This prevalence suggests that MLR is highly regarded for its robustness and effectiveness in handling various types of cost estimation tasks. The use of ANN is also notable, being applied in four different studies, reflecting the growing interest in leveraging machine learning for more complex and potentially more accurate estimations.

Cost-estimating approaches have progressively evolved over time, demonstrating a noticeable transition towards more advanced and computationally demanding procedures. During the period from the 1960s to the 2000s, simpler techniques such as logarithmic transformations and expert systems were frequently utilized. However, beginning in the 2000s, there was a clear rise in the use of MLR and ANN, emphasizing the shift towards data-driven methodologies. In the 2010s, the utilization of MLR became increasingly prevalent, while more sophisticated techniques such as GA, CBR, and hybrid regression models began to attract interest. In the 2020s, there is evident diversification through the implementation of various machine learning approaches and hybrid models, indicating a trend towards integrating multiple analytical techniques to enhance the accuracy and adaptability of maintenance cost estimation. Importantly, while many of these models—such as MLR, ANN, and GA—originate from construction cost estimation research, their application in maintenance cost prediction involves distinct methodological adaptations.

Unlike construction cost estimation, which typically relies on static input data (e.g., quantities, materials, labor, and equipment) associated with one-time project phases, maintenance cost prediction must account for time-dependent variables, degradation processes, usage intensity, and environmental impacts over the building’s lifecycle. As a result, maintenance-focused models often incorporate factors such as building age, condition assessment scores, frequency of interventions, and energy performance indicators. These dynamic and uncertain variables require more flexible modeling techniques—such as ANN, GA, and hybrid approaches—that can capture nonlinear patterns and long-term cost fluctuations. Furthermore, the utilization of various cost assessment techniques differs considerably among different building types. Researchers commonly analyze bridges and university structures using regression and ANN methods, while MLR and the AHP are frequently applied to office buildings. Specialized models have also been developed for other facility types, including sports facilities, railways, schools, residential buildings, highways, sewer systems, pontoons, and marina anchor systems. These models are tailored to the unique maintenance demands and service conditions of each type. The diversity of building applications reinforces the importance of selecting estimation methods that align with the specific operational characteristics, data availability, and sustainability objectives of each facility.

Overall, the reviewed studies demonstrate that although traditional estimation tools such as MLR remain foundational, the adaptation of intelligent and hybrid models has become essential in maintenance cost prediction due to the greater complexity, variability, and sustainability considerations associated with long-term asset management. In this context, sustainability can be operationalized through measurable indicators that link maintenance performance to environmental efficiency—such as energy use intensity (kWh/m²·year), embodied carbon (kgCO₂e/m²), and material recyclability or reuse potential (%). Several recent studies have begun to integrate such indicators into cost prediction frameworks, demonstrating how maintenance expenditure can be correlated with energy performance and resource optimization over the building life cycle. In addition to methodological development, future progress in maintenance cost modeling will depend on strong data governance practices, including standardized data formats, transparent variable definitions, and consistent treatment of missing or uncertain records.

6. Conclusions

The analysis shows a clear progression in the sophistication of cost estimation methods, with a notable shift towards more advanced, data-driven approaches over time. The consistent use of MLR underscores its value in the field, while the growing interest in machine learning techniques points to future directions for further improving estimation accuracy and efficiency. This evolution mirrors broader trends in engineering and construction, where the integration of advanced computational methods is becoming increasingly central to practice.

However, despite the advancements in cost estimation methodologies, there is a critical need for future research to emphasize sustainability. As the construction industry increasingly recognizes the importance of reducing environmental impact, incorporating sustainable practices into cost estimation becomes essential. Future studies should explore methods that not only enhance accuracy and efficiency but also integrate sustainability considerations. This includes evaluating the environmental impact of construction materials, energy consumption, and long-term maintenance practices.

Ultimately, although the field of cost estimation has made significant progress in adopting sophisticated approaches, the next challenge is to incorporate sustainability into these procedures. By focusing on sustainable practices, researchers can contribute to the development of more environmentally responsible construction projects, ensuring that future buildings are not only economically viable but also environmentally sound. This holistic approach will help the construction industry meet global sustainability goals and create a more sustainable future.

Future studies should focus on developing advanced predictive models that incorporate real-time data and machine learning algorithms to more accurately forecast the costs and benefits of sustainable building maintenance practices. These models should consider a comprehensive range of factors, including energy consumption patterns, material durability, and lifecycle costs, as well as the impact of emerging technologies and regulatory changes.