AI and Big Data-Empowered Low-Carbon Buildings: Challenges and Prospects

Abstract

Reducing carbon emissions from buildings is crucial to achieving global carbon neutrality targets. However, the building sector faces various challenges, such as low accuracy in forecasting, lacking effective methods of measurements and accounting in terms of energy consumption and emission reduction. Fortunately, relevant studies demonstrate that artificial intelligence (AI) and big data technologies could significantly increase the accuracy of building energy consumption prediction. The results can be used for building operation management to achieve emission reduction goals. For this, in this article, we overview the existing state-of-the-art methods on AI and big data for building energy conservation and low carbon. The capacity of machine learning technologies in the fields of energy conservation and environmental protection is also highlighted. In addition, we summarize the existing challenges and prospects for reference, e.g., in the future, accurate prediction of building energy consumption and reasonable planning of human behavior in buildings will become promising research directions.

1. Introduction

Global warming caused by carbon emissions has received increasing attention since the beginning of the 21st century. The fifth assessment report of the Intergovernmental Panel on Climate Change (IPCC) in 2014, under the United Nations, revealed that the global average surface temperature had risen to 14.6 °C, marking a 0.69 °C increase compared to the 20th century average temperature [1]. The impact of climate warming on the natural environment, human-habitable areas, and ecological systems has been profound. In response to this critical issue, the “Paris Agreement” proposes stringent measures to control the global average temperature increase within 2 °C this century, with a focus on limiting the rise to within 1.5 °C above the pre-industrial level [2].

Exploring the reason, many countries have undergone rapid economic development, urbanization, and industrialization since the beginning of the 21st century. However, this progress has led to a significant surge in carbon emissions. By 2019, direct carbon emissions from fossil fuels, including coal and natural gas, had reached an alarming 600 million tons of carbon dioxide. These alarming statistics underscore the urgent need for China to take decisive action in reducing its carbon footprint. To accomplish the targets outlined in the Paris Agreement, countries worldwide must strive for carbon neutrality by 2050. Several nations, such as the United Kingdom, France, and New Zealand, have already enshrined this goal in legislation. Additionally, others, including South Korea, the European Union, and Spain, are currently legislating for carbon neutrality by 2050. Moreover, countries like Japan, South Korea, and Austria have made policy pledges toward achieving this objective. Recognizing the urgency of combating climate change, countries around the world, including China, have increasingly focused on carbon neutrality and emission reduction. Countries have shown tremendous efforts in addressing their carbon footprint, especially in the building industry.

The global building industry plays a significant role in carbon emissions, accounting for a considerable portion of the total greenhouse gas emissions worldwide [3,4]. Recent studies indicate that buildings are responsible for approximately 39% of global carbon dioxide emissions, making it a crucial sector for implementing sustainable practices and achieving carbon neutrality. The Global Status Report on Building and Construction 2022, published by the United Nations Environment Program (UNEP), assessed and analyzed the share of the construction industry in global carbon emissions. According to this report, the carbon emissions and energy consumption of the construction sector are 38% and 35%, respectively. These figures are shown in Figure 1 and Figure 2.

In order to obtain the above calculations, certain equations are commonly utilized in the calculation of carbon emissions. These equations play a crucial role in enhancing our comprehension and quantification of the impact that diverse processes and activities exert on carbon emissions and global warming. Equations (1) to (3) are commonly used formulas for calculating GHG.

Equation (1) is the carbon cycle equation, where C is the carbon content of the atmosphere, E is the carbon emissions released into the atmosphere, and $\tau$ is the carbon lifetime. Equation (1) serves to characterize the dynamics of carbon, enabling the measurement of carbon cycle dynamics and stocks at a large scale.

Equation (2) is employed to compare the global warming potential (GWP) of different gases. Equation (2) is the Intergovernmental Panel on Climate Change (IPCC) equation for calculating the global warming potential (GWP) of a GHG, where $GW{P}_i$ is the GWP of GHG i, $R{F}_i\left(t\right)$ is the radiative forcing of GHG i at time t, and $R{F}_{C{O}_2}\left(t\right)$ is the radiative forcing of CO₂ at time t.

Equation (3) is the equation for calculating the carbon footprint of a product or activity, where $CF$ is the carbon footprint, $E{F}_i$ is the emission factor for GHG i, $Q_i$ is the quantity of GHG i emitted, and n is the number of GHGs considered.

In contrast to the first two equations, which are utilized for calculating carbon emissions on a broad scale, the third equation is specifically employed for calculating the carbon footprint of a particular activity or product:

$$\frac{dC}{dt}=E-\frac{C}{\tau }$$

(1)

$$GW{P}_i={\int }_{t_1}^{t_2}\frac{R{F}_i\left(t\right)}{R{F}_{C{O}_2}\left(t\right)},dt$$

(2)

$$CF=\sum _{i=1}^n(E{F}_i\times Q_i)$$

(3)

With the accelerated urbanization in China, a large number of construction tasks and production activities not only consume substantial energy but also result in significant greenhouse gas emissions during the construction process. Since 2001, the annual increase in urban and public building areas has reached approximately 1 billion square meters, with cooling, heating, and domestic hot water consumption exceeding 60% of the building’s overall energy consumption.

Studies show that China’s construction industry accounts for 30% of the country’s total greenhouse gas emissions, with building energy consumption comprising 40% of all energy consumption [6]. In the UK, buildings account for 46% of all CO₂ emissions, while in the US, the figure is 40%, and in Australia, it is 27%. Globally, building energy consumption accounts for about 36% of the total global final energy consumption, and this energy consumption also contributes about 40% of the total global carbon emissions. Therefore, reducing carbon emissions in the construction sector is paramount for China and the world to achieve their established carbon neutrality goals. The construction industry is responsible for a significant proportion of greenhouse gas emissions. However, this also means that the sector has immense potential and can make a substantial contribution to reducing carbon emissions.

This study explores the use of artificial intelligence in the construction industry with a specific focus on methods for AI-assisted emission reduction and low-carbon operations. Additionally, the article summarizes the main approaches for carbon emission accounting, which are essential for quantifying the final results. Leveraging relevant technologies and tools, this paper proposes machine learning-based solutions to support the green and low-carbon development of the construction industry as depicted in Figure 3. The outline structure of the article is shown in Figure 4.

The article is timely since AI and big data are currently hot research topics. However, there is still an urgent need for basic research to improve building energy conservation and promote low-carbon practices. To address this need, we conducted a thorough analysis of the existing literature to identify potentially important trends and hotspots in this research field. Moving forward, we aim to promote the development of this field further.

2. AI and Big Data Technology in Building Energy Saving and Carbon Reduction

This section mainly describes the data preprocessing methods commonly used in the building sector, the use of neural networks in building energy prediction, and deep learning for realizing low-carbon buildings, respectively.

Table 1. The commonly used methods for data preprocessing.

Main Method	Advantages	Disadvantages
Normalization/standardization	It can avoid the mismatch of different feature weights and improve the convergence speed and effect of the model.	Outliers can affect normalization, and normalization can make some eigenvalues too small.
Common algorithms for data cleaning: regression imputation, hash-based deduplication, cluster-based outlier detection	Can improve data quality and avoid negative effects on the model.	Data cleaning may delete useful information, while filling missing values may introduce noise.
Common algorithms for data augmentation: random scaling, random speed perturbation, sentence rearrangement	Can increase the diversity of training data and improve model generalization ability.	Data augmentation may introduce noise, and increasing the number of training data also increases computational cost.
Feature selection	Can reduce the dimensionality of the feature space and improve model efficiency and generalization ability.	Feature selection may delete useful information and reduce model accuracy.
Feature scaling	It can avoid the mismatch of different feature weights and improve the convergence speed and effect of the model.	It may lead to errors and reduced model accuracy.

presents the current data preprocessing methods along with their respective advantages and disadvantages.

2.1. Data Preprocessing and Analysis

AI and big data technologies provide powerful methods to preprocess and analyze building-relevant data. It is important to assist building administrators and energy managers in making decisions. By optimizing energy usage strategies, this approach allows them to minimize carbon emissions to the greatest extent possible.

Faced with the complex and changeable nature of building energy consumption data, appropriate data preprocessing methods can address these challenges. However, the lack of interdisciplinary research has raised concerns about building carbon emissions. Therefore, identifying and preparing meaningful and reliable modeling data has always been a significant challenge, as only reasonable and high-quality data can enable neural networks to produce reliable decision-making results. Relevant research suggests that data preprocessing can be categorized into four main areas: data integration, data cleaning, data reduction, and data transformation [7].

Data integration mainly deals with issues like data redundancy and format contradictions. Such problems can be efficiently addressed using tools like Excel without the need for machine learning algorithms [8,9]. Data cleaning and data reduction are the main focus of the discussion. Data cleaning involves problems such as adding default values and identifying and removing outliers. The missing data can be filled with the K-nearest neighbors (KNN) algorithm. KNN is a supervised learning classification algorithm that selects the K-nearest neighbors for classification and judgment by calculating the distance between the sample data and the training data. Figure 5 demonstrates the basic principles of the KNN algorithm.

Zhou et al. [10] successfully utilized this method to identify and correct outliers in the actual operational data of ground-source heat pump units. Data reduction is another technique used to handle multi-dimensional building data. Too many data dimensions can make it challenging to express the final result and negatively impact model training intuitively. Using dimensionality reduction algorithms like principal component analysis (PCA) [11] or wavelet analysis can help eliminate unnecessary dimensions while retaining key data content.

Śmieja et al. [12] proposed a general, theoretically justified mechanism for processing missing data using neural networks. The K-means clustering algorithm can be employed to identify and correct outliers. Abnormal data may include repeated values and data that could be more logical, such as values that are too large or too small. The K-means algorithm is an unsupervised machine learning method that automatically iterates to obtain the best cluster class when the number of clusters m is set. Utilizing the clustering algorithm allows for the successful separation of valid data from abnormal data, achieving effective data cleaning.

The engineering data-filling problem can also be addressed using machine learning methods, such as support vector machine (SVM) regression and Bayesian techniques [13]. Proximity algorithms and machine learning-related approaches are the mainstream methods for data filling. With the continuous advancement of related research and the development of artificial intelligence, methods of filling missing data using neural networks have emerged. Zhang et al. utilized the backpropagation neural network (BPNN) to predict building energy consumption values and fill in missing data, resulting in significant improvements.

After identifying outliers, cleaning data, and replacing missing values on the data, decomposing the data will make the results more intuitive. Data decomposition is a commonly used method for time series data processing.

Most building energy consumption data are in the form of time series data. Figure 6 shows that time series data consist of three main components: trend, seasonality, and residual or white noise. The trend represents the long-term and gradual “direction” of the time series data. It reflects the overall increasing or decreasing pattern over time and indicates the sustained trend of the data. Seasonality represents the periodic patterns in the time series over a certain period. Seasonality is significant for building data, as summer cooling and winter heating exhibit highly seasonal energy consumption behavior. Therefore, the energy consumption of buildings fluctuates periodically with different seasons. Residual or white noise refers to the random fluctuations in the time series data that the trend and seasonality cannot explain. It represents the random and irregular fluctuations in the time series data beyond the trend and seasonality. Residuals may indicate energy consumption fluctuations caused by weather, human behavior, or other uncertainties. If random factors do not dominate the entire dataset, extensive processing is usually unnecessary. However, if there are excessive random fluctuations in the data, it is essential to consider separating the residuals and applying smoothing techniques.

Xie et al. [14] used Kalman filtering to filter their battery-timing data to predict the battery’s future capacity. Deng et al. [15] used GPR (Gaussian process regression) to smooth the residual part of the time series, and their model performance with GPR was significantly better than the model performance without smoothing under the same experimental conditions.

In addition to analyzing residuals, seasons, and trends, further processing is required for construction data. Building operation data are commonly collected using sensors or publicly available statistics, like the National Statistical Yearbook. Regardless of the data collection method, building energy consumption data typically exhibit large volumes, multi-dimensionality, and complexity. Simultaneously, the statistical process involves various types of buildings, unexpected exceptional circumstances, instrument failures, extreme environmental disturbances, etc., making building energy consumption data susceptible to abnormal values and default entries.

2.2. Building Energy Consumption Detection and Prediction Based on Machine Learning Method

Developing a reliable model to predict the growth of energy consumption could provide a valuable reference for decision makers at various levels. It enables decision makers, e.g., government entities and enterprises, to design appropriate environmental policies and strategies, effectively controlling environmental issues and promoting sustainable practices.

Researchers have previously utilized classical statistical and econometric methods to model and predict carbon emissions growth. However, the effectiveness of these methods relies on the availability and reliability of independent variables. Since carbon emissions modeling involves chaotic, non-stationary, and nonlinear variables, classical statistical and econometric methods are inadequate for capturing such complex behavior. Artificial neural networks (ANNs), on the other hand, offer a solution to nonlinear modeling through multiple network layers, providing robust prediction results by accurately approximating the nonlinear input–output relationship. ANNs excel at handling noisy data, accommodating multiple variables with nonlinear, linear, and unknown interactions, and making reliable generalizations. Building energy consumption is often represented as time series data, typically focusing on energy consumption of heating, ventilation and air conditioning (HVAC) due to its significance, time series, and nonlinear characteristics.

For this, Sendra-Arranz et al. [16] developed a prediction model for a heat pump system in the north, employing the long short-term memory neural network (LSTM) method. The LSTM model outperformed ordinary models, such as support vector regression (SVR) and regression trees (RT), demonstrating excellent predictive stability. However, relying solely on a single algorithm may only sometimes meet the desired prediction accuracy requirements. To address this challenge, Wang et al. [17] proposed three algorithms: mathematical morphological clustering (MMC), improved tree species algorithm (ITSA), and generalized regression neural network (GRNN). These algorithms focus on selecting heat load factors, data clustering, and neural network construction. Combining these algorithms provided a comprehensive set of solution ideas, improving the model’s convergence speed and stability in complex situations. Alex et al. [18] employed a backpropagation (BP) neural network to estimate country-wide building carbon emissions for regional building energy predictions. After extensive research, they selected nine characteristics: economic growth, energy consumption, population size, research and development, urbanization, finance, foreign direct investment, industrialization, and globalization. Based on these characteristics, a three-layer neural network model was constructed, successfully predicting quarterly carbon emissions for Australia, Brazil, the United States, China, and India, with an accuracy rate exceeding 80%. Sensitivity analysis revealed that population had the most significant impact on China’s carbon emissions, research influenced Australia’s emissions the most, energy consumption played a crucial role in India’s emissions, and urbanization significantly affected Brazil’s emissions.

To address the challenge of limited generalization in existing building energy consumption prediction models, Yu et al. [19] proposed a transfer learning-based approach. By leveraging transfer learning, different processing operations can be selected based on task similarity and the number of new samples, effectively reducing data requirements across related fields and achieving improved results simultaneously. The proposed method utilizes the structural transfer of hidden layers to share building energy consumption information and extract deep features. Finally, a reinforcement learning method was employed to predict building energy consumption in Fremont, California, resulting in approximately 20% higher prediction accuracy. Figure 7 illustrates the basic concept of transfer learning.

Li et al. [20] applied the support vector machine (SVM) method to predict cooling energy consumption in office buildings in China, specifically focusing on the summer season and utilizing local climate data. They used one month of data for training and an additional four months for testing, comparing the results with neural network training. This study demonstrated the significant practical application potential of SVM in modeling building cooling energy consumption. However, it also highlighted the impact of feature selection on model performance, requiring further investigation. Feature selection poses more challenges in building energy consumption than model selection, as different research goals necessitate distinct feature selections. Geothermal and coal consumption considerations include outdoor temperature, solar radiation, hot water outlet temperature, building area, number of occupants, and electrical equipment. Energy performance analysis focuses on ratios of north, south, east, and west window glass, glass type, roof shape, roof glass, external weather conditions, and building envelope design. Power demand prediction emphasizes factors like air temperature, solar radiation, average heat transfer coefficient of building walls, average thermal inertia index of building walls, roof heat transfer coefficient, and building size coefficient [21].

Recent research has explored various approaches, including autoregressive models and attentional mechanisms, to address the growing interest in accurate building energy prediction. One example of an autoregressive model is the TimeGrad method as proposed by [22]. This method utilizes a multivariate probabilistic time series forecasting approach that estimates the data distribution at each time step by approximating its gradient. By combining diffusion models with time series forecasting, TimeGrad has proven to be a state-of-the-art adversarial competition approach that outperforms existing models in typical time series tasks.

Attention mechanisms have also gained significant interest in building energy prediction, particularly since the success of the Transformer model. In a related article by [23], the effects of three different attention mechanisms, including the time domain, frequency domain, and wavelet transform, were compared. The article provides a detailed analysis of the applicable effects of different attention mechanisms, offering valuable insights.

In addition, Khan et al. proposed a two-stage power forecasting model that utilizes an end-to-end hybrid residual CNN with stacked LSTM models for power prediction of buildings [24]. This model combines convolutional and recurrent neural networks to capture both spatial and temporal dependencies in the data, leading to improved accuracy in power forecasting. Also, based on LSTM research, a related study developed a hybrid convolutional neural network (CNN) with a long short-term memory autoencoder (LSTM-AE) model for future energy prediction of residential and commercial buildings [25]. This research focuses on energy forecasting using data from smart meters for proper energy management in buildings. After experimenting with Korean commercial building data, the model achieved the best root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) results.

More novel research proposes a hybrid electricity load forecasting (ELF) model that incorporates COVID-19 data to improve prediction accuracy [26]. The model utilizes convolutional layers, gated recurrent nets, and a self-attention module for feature selection, resulting in better generalization for predicting electricity consumption (EC) patterns. The model outperforms existing models, with average reductions of 0.56% and 3.46% in mean square error (MSE), 1.5% and 5.07% in RMSE, and 11.81% and 13.19% in MAPE over pre- and post-pandemic data. These findings significantly affect improving ELF algorithms during pandemics and other events that disrupt historical data patterns.

By exploring different approaches to building energy prediction, including autoregressive models, attentional mechanisms, and hybrid models, researchers are advancing the field toward more accurate and effective methods for energy management in buildings.

In summary, smaller datasets of construction time series data often employ statistical models, including autoregressive (AR) models, moving average (MA) models, autoregressive moving average (ARMA) models, autoregressive integrated moving average (ARIMA) models, etc. Deep learning models are commonly used for larger datasets, including multilayer perceptron, recurrent neural networks, long and short-term memory neural networks, and AR models with attention mechanisms. The support vector machine (SVM) model is known to require longer training times for more extensive training sets and is typically more suitable for smaller sample sizes. On the other hand, the long short-term memory (LSTM) model exhibits superior performance on long-term sequential data; however, its parameter tuning process is more complex. Transfer learning, or migration learning, has been shown to reduce data dependency effectively, but its success is contingent upon the similarity between the target and source domains. Meanwhile, residual convolutional neural networks (CNNs) are better equipped to address the problem of vanishing gradients but require more extensive modifications when applied to non-image data. Finally, the current state-of-the-art sequence modeling model is the Transformer, but its suitability for tasks related to the construction domain may only sometimes be optimal. Ultimately, selecting an appropriate modeling method requires careful consideration of each model’s unique characteristics and the specific requirements of the task at hand. The corresponding comparisons in terms of models, literature and fields are summarized in

Table 2. Common-use models for forecasting building energy consumption.

Model	Referring Studies	Field
Autoregressive (AR) model	Developing a hybrid time series artificial intelligence model to forecast energy use in buildings [27]	Traditional statistical model
Moving average (MA) model	Machine Learning, Deep Learning and Statistical Analysis for forecasting building energy consumption—A systematic review [28]	Traditional statistical model
Autoregressive moving average (ARMA) model	Application of Combined Model Based on Empirical Mode Decomposition, Deep Learning, and Autoregressive Integrated Moving Average Model for Short-Term Heating Load Predictions [29]	Traditional statistical model
Vector autoregressive (VAR) model	Forecast of electricity consumption in the Cameroonian residential sector by Grey and vector autoregressive model [30]	Traditional statistical model
Autoregressive integrated moving average (ARIMA) model	Application of Combined Model Based on Empirical Mode Decomposition, Deep Learning, and Autoregressive Integrated Moving Average Model for Short-Term Heating Load Predictions [29]	Traditional statistical model
Multi-layer perceptron (MLP)	Modelling carbon emission intensity: Application of artificial neural Network [18]	Machine learning and deep learning
Recurrent neural network (RNN)	Assessment of deep recurrent neural network-based strategies for short-term building energy predictions [31]	Machine learning and deep learning
Long and short-term memory network (LSTM)	Forecasting building energy consumption: Adaptive long-short term memory neural networks driven by genetic algorithm [32]	Machine learning and deep learning
Autoregressive neural networks	Autoregressive Denoising Diffusion Models for Multivariate Probabilistic Time Series Forecasting [22]	Machine learning and deep learning
Attention model	First De-Trend then Attend: Rethinking Attention for Time Series Forecasting [23]	Machine learning and deep learning

.

3. Machine Learning-Based Approaches for Building Low Carbon

This section focuses on several ways in which machine learning can be employed to decrease energy consumption in buildings. It is divided into the applications of machine learning in power grid control, reinforcement learning in system optimization, machine learning in the power market, and the use of machine learning in electric vehicles. In all of these domains, machine learning has the potential to curtail carbon emissions in buildings through intelligent control, peak shaving and valley filling, and other approaches to optimize the allocation of clean energy utilization.

As mentioned above, after predicting building carbon emissions through machine learning, sensitivity analysis can be performed based on the prediction results. Measures can then be taken to target the most sensitive features, aiming to achieve the goal of building carbon reduction and low carbon emissions. Su et al. [33] analyzed the carbon emissions of residential buildings in Fujian province using the STIRPAT (stochastic impacts by regression on population, affluence and technology) model and proposed four measures to reduce carbon emissions: installing energy-saving exterior window glass, implementing external wall insulation, promoting LED lights, and popularizing solar water heaters.

The penetration rate was estimated to reach 0.45 in 2020 with huge emission reduction potential that can reach 101,600 tons of carbon dioxide. In the case of 0.7, 0.3 and 0.15, the emission reduction can reach 54,100, 99,000, and 42,200 tons of carbon dioxide, respectively. In addition, AI technology can also contribute to reducing building emissions, such as using computer vision technology to identify the distribution of people and control building electrical appliances to avoid waste of electricity in unmanned areas. Alternatively, a camera is used to perform motion recognition, predict people’s behavior in the building, and autonomously control building electrical appliances such as air conditioners to avoid the waste of cooling or heat caused by constant temperature settings.

However, this kind of method still has the following shortcomings:

• The distribution of people in the building has a significant influence on the energy consumption of the building, but currently, only relying on the camera to capture has limitations: the recognition effect of the dark area is not ideal; when the line of sight is blocked, it is inaccurate; the camera is affected by the installation position; and there are visual blind spots. Sometimes, camera installation is not allowed due to privacy protection [34].
• In addition, due to the large randomness of human behavior trajectories, the current work is not solid enough on the analysis and research based on human behavior.
• To achieve significant building energy conservation and carbon reduction goals, it is necessary to precisely control a large number of building electrical appliances in the building. This requires the collection of the electricity consumption information of local units in the building (such as user rooms), but the detailed electricity consumption information may hide users’ privacy data [35], for example, the user’s activity pattern can be inferred by analyzing the user’s daily electricity consumption.

Faced with these problems, various important solutions in the form of energy systems have been proposed, such as smart grids, integrated energy systems (IES), etc. The aim is to unify decentralized energy and power systems, optimize the resource allocation structure, and meet the low-carbon needs of buildings [36]. The advantages of reinforcement learning, based on model-free automatic decision making in the context of energy saving and efficient energy rationing requirements, make it an important strategic support for building sustainable energy and power systems. The following are some of the approaches that may be applied.

3.1. Grid Generation Control Approach

In energy system generation control, modern power systems are experiencing greater volatility and uncertainty due to the spread of renewable energy sources and power electronics. Some studies have used deep reinforcement learning (DRL) algorithms to derive optimal power flow (OPF) solutions for real-time AC (alternating current) by adjusting several controllable variables to meet the operating constraints [37].

One study proposes an innovative algorithm called the three-network double-delay actor–critic (TDAC) algorithm for automatic generation control, which is based on deep reinforcement learning [38]. This algorithm addresses the challenges of stochastic perturbations to the grid caused by renewable energy sources. The TDAC algorithm enhances the conventional actor–critic approach by incorporating an incentive heuristic mechanism, resulting in improved efficiency and quality of exploration for automatic generation control.

3.2. System Operation Optimization Approach

Li et al. [39] present an optimal dispatch model for isolated microgrids for grid dispatch optimization problems. The model utilizes an automatic reinforcement learning-based multi-period forecasting technique for renewable generation and load. The authors introduce the multi-cycle forecasting method, incorporating prioritized experience replay automated reinforcement learning (PER-AutoRL). By using corrected forecast values for grid dispatch decisions, this approach effectively reduces the grid’s operating costs. Furthermore, their study highlights the positive impact of involving demand-side users in optimizing the system operation to a greater extent.

The paper by Lu et al. [40] proposes an innovative approach to address the demand response challenges in smart grid systems using a real-time incentive-based algorithm. The method combines two powerful techniques: RL and deep neural networks (DNNs). The combination of RL and DNN allows the algorithm to autonomously determine the appropriate incentive rates for users without explicit knowledge of the underlying system dynamics. By predicting the future electricity load patterns accurately and dynamically adjusting the incentive rates, the algorithm aims to stabilize grid generation and consumption. Doing so can reduce peak demand, improve grid efficiency, and lower electricity costs.

Buildings play a significant role in global energy consumption, attracting academic attention [41]. However, in the absence of a thermodynamic model for the building, the number of occupants and the outdoor temperature, efforts to maintain comfort often result in substantial energy wastage. To address this, the optimization problem can be transformed into a Markovian decision process. Utilizing an attention multi-intelligent deep reinforcement learning model offers a potential solution to achieve efficient modulation without relying on explicit modeling [42].

From an electricity cost perspective, several methods [43] combine electricity price forecasting models with multi-intelligent reinforcement learning. These methods generate agents for decentralized decision optimization concerning various electrical products. The approach efficiently manages energy consumption for multiple appliances, minimizing building electricity costs.

Another study [44,45,46] additionally considers the charging and discharging of the building’s energy storage devices and appliance energy consumption. Deep reinforcement learning methods minimize energy costs while meeting energy requirements.

3.3. Application for Electricity Markets

With the increasing energy demand, there has been a notable interest focused on studying the electricity energy market and how energy prices affect system operation [47]. Electricity providers can mitigate profit losses caused by fluctuating renewable energy generation by investing in partial energy storage. For instance, certain wind power producers have implemented the A3C algorithm for optimizing their bidding strategy [48].

In district energy systems, service providers purchase energy and sell it to customers, making a profit. Implementing dynamic pricing can be challenging due to customer information and system uncertainty. Du et al. [46] trained DNNs in a supervised manner to learn electricity trading information, combined with RL algorithms to develop an electricity pricing strategy. This approach minimizes peak-to-average ratios while improving returns. Other studies have considered service provider profitability and customer costs [49]. Customers aim to minimize costs while meeting electricity demand, and service providers strive to maximize profits.

3.4. Applications for Electric Vehicles

Electric vehicles, as representatives of new energy transportation, are widely admired for their high efficiency and energy-saving advantages [50]. Given the fluctuating electricity prices and the unpredictable nature of commuting behavior, optimizing electric vehicles’ charging and energy use efficiency has been the subject of extensive research. Li et al. [51] proposed a DRL-based method to minimize the charging cost of electric vehicles. They utilized an RNN network to extract effective time information from the electricity price sequence, aiding the deep deterministic policy gradient (DDPG) algorithm in making decisions on the charging and discharging system.

In other research, the focus is on optimizing specific aspects of the vehicle. DRL algorithms are used to learn shift strategies to enhance engine performance [52]. However, it is crucial to note that the DRL algorithm may not be suitable for learning intermittent control strategies as indicated in the article.

4. Big Data Driven Carbon Emission Accounting

There are three mainstream methods for carbon emission accounting, namely the emission factor method, mass balance method and actual measurement method. Among them, the emission factor method is suitable for three levels of macro, micro and meso, and is currently the most widely used carbon emission accounting method. The basic principle of the mass balance method is based on the law of conservation of energy. The material input into the system must be the output quality, which is mainly used in the process of industrial production. Both the mass balance method and the actual measurement method are suitable for carbon emission accounting at the micro level [53]. The emission factor method is mostly used for carbon emissions in the building sector. There are two primary methods for carbon emission accounting in buildings: top–down and bottom–up approaches. The top–down method involves estimating the overall building energy consumption and carbon emissions. On the other hand, the bottom–up method starts by calculating the hourly energy consumption of individual buildings and then scales up to regional levels for carbon emission calculations [54]. The two methods correspond to regional building carbon emission accounting and single building carbon emissions accounting, respectively. Each approach has its advantages and disadvantages, described in the following sections.

4.1. Regional Building Accounting Method

Regional buildings mainly refer to the overall buildings in large-scale regions, such as the world, countries, provinces, and cities. The purpose of such building accounting is to provide a theoretical basis for government policies. For regional buildings, a top–down approach is mainly used. The main input variables of the top–down model include the economy, energy prices, and GDP, and the output results are the relationship between energy consumption, carbon emissions, and the economy. Such methods focus more on macroeconomic factors and need more research on the technical or physical details of the impact of building energy consumption [55]. Wang et al. [56] used the input–output model to calculate the carbon emissions of buildings in China from 2004 to 2012. The data source China Statistical Yearbook mainly selected the construction industry, electricity, gas and water production and supply industry, agriculture, forestry, and animal husbandry. Fishery, transportation, warehousing, and postal industry are compared with the five industries of wholesale, retail and accommodation, and catering industry. The whole research process does not involve knowledge in the field of construction but completes the carbon emission of buildings from an economic point of view. After accounting, it is concluded that every one percentage point change in the per capita industrial added value of the construction industry will comprehensively cause a change in carbon emissions by 21.29, 25.80, and 25.69 percentage points.

The bottom–up method of regional buildings draws on the accounting method of the life cycle of a single building. It mainly selects two stages: the construction stage and the operation stage. The construction stage includes the carbon emissions in the entire range of building materials production, transportation, and construction. The carbon emissions in the building operation stage mainly refer to the greenhouse gas emissions caused by energy consumption in heating, air conditioning, and lighting. The method considers details such as temperature and humidity, building performance, terminal equipment, and operating characteristics, and based on the energy consumption of typical buildings, predicts and simulates regional, regional, and even national-scale building energy demand and then calculates carbon emissions, taking the framework distance of the China building carbon emission model (CBCEM) as shown in Figure 8.

For the top–down approach, Langevin [58] proposed a SCOT (social cognitive optimization for customizable training) model to calculate the carbon emissions of buildings in the United States, selected the annual energy demand of each climate zone in the United States as the calculation scale, simulated the energy consumption of typical buildings, and evaluated the impact of energy-saving measures on the impact on carbon emissions. Villamar et al. designed the ELENA) model [59] for the national characteristics of Ecuador, which comprehensively considered the renewable energy and electricity demand of six energy-consuming sectors (transportation, buildings, commerce, industry, agriculture, and others).

For bottom–up methods, models such as Invert/EE-Lab, ECCABS, RE-BUILDS, and CoreBee are commonly used in Europe and the United States [60,61,62,63,64]. However, due to the complexity of bottom–up models that consider architectural details, such as electrical, air conditioning, lighting, etc., as well as variations in living habits, latitudes, geographical environments, and populations in different countries, applying a single model universally becomes challenging. Yang et al. proposed the CBCEM to address this limitation for building carbon emission accounting in China [57]. CBCEM emphasizes that research in China has predominantly focused on energy consumption rather than carbon emissions. To overcome this limitation, CBCEM subdivides the model framework into final energy consumption types for each building type in each climate region. It uses this refined framework to calculate the carbon emissions of buildings. Notably, CBCEM concentrates solely on the building operation stage, recognizing that emissions during this phase constitute the majority of total building emissions.

The CBCEM model divides China into three regional types: the north, hot summer and warm winter, and the south. It also categorizes buildings into three types: public buildings, urban residential buildings, and rural residential buildings. Energy consumption terminals include heating, cooling, lighting, equipment, hot water, and six types of kitchens. The three-story structure contains 42 computational paths sufficient to generalize all the basic scenarios for different climate levels, building types, and end uses.

4.2. Single Building Accounting Method

The carbon emission accounting of a single building mainly adopts the building life cycle method. In relevant domestic research, the building life cycle can be divided into four stages: the building design stage, the construction stage, the building operation stage, and the building demolition stage [65]. The carbon emission accounting of individual buildings often adopts the method of carbon emission intensity factor and converts all greenhouse gases into a carbon dioxide equivalent for calculation. According to the IPCC global warming potential (GWP), six gases of carbon dioxide, methane, nitrous oxide, hydrofluorocarbons, perfluorocarbons, and sulfur hexafluoride can be converted into carbon dioxide according to the corresponding GWP value emissions.

In addition, considering that the whole life cycle of a building is divided into four stages, the division of the carbon emission accounting boundary is a factor that has an important impact on the results, which needs to be clarified and standardized. From the perspective of the whole life cycle of the building, the design stage needs to consider the emissions of office equipment. The materialization stage includes the carbon emissions of building materials production and transportation, the carbon emissions of construction machinery, and the operation stage includes heating, cooling, electricity for household equipment, decoration, hot water, lighting, and other carbon emissions. The main consideration in the dismantling phase is the carbon emission caused by the engineering equipment. After determining the emission boundary, the accounting can be carried out in stages according to the whole life cycle.

Design stage: The design stage mainly includes the energy required in this stage, the consumption in the office. Compared with the overall carbon emissions of the building, the emissions in this stage are small and can be ignored [66].

Materialization stage: The materialization stage includes the whole process of the building, from the extraction of raw materials, production, transportation, and construction to the completion and delivery of the building. The factors that need to be considered in the production of building materials are the type of building materials, the consumption of building materials, the emission factor, and the recovery factor. During the transportation process of building materials, it is essential to take into account factors such as fuel consumption per 100 km, the total transportation distance, and the carbon emission factor associated with the specific means of transportation used. The factors that need to be considered in the construction process include the type of construction machinery, the number of construction machinery, and the carbon emission factor. Relevant data can be consulted in the “National Unified Construction Machinery Shift Expenses”.

Operation phase: The building carbon emissions in the operation phase include two aspects—building use carbon emissions and building maintenance and renovation. During the usage stage, it is important to take into account factors such as the average energy consumption, service life, and carbon emission factor. The carbon emissions of refurbishment and maintenance can be calculated by multiplying the total carbon emissions in the operation phase by an appropriate factor based on experience or data.

Demolition treatment stage: Considering the relatively long time from construction to demolition, it is difficult for relevant researchers and research projects to continue to pay attention, and there are few domestic and foreign related studies. If the conditions are met, construction machinery in the demolition stage can be considered: shift consumption, number of construction machinery, and recyclable carbon emissions. But a more common practice is to directly convert the carbon emissions of the demolition stage into a percentage calculation of the physicalization stage. According to relevant research, the carbon emission in the demolition stage can be calculated as 8–10% in the physical and chemical stages [67,68].

However, in the carbon emission accounting of individual buildings, there are often problems, such as unclear emission boundaries, differences in the division of carbon emission inventories, and different life cycle stages. Therefore, there is a lack of recognized calculation standards for single-building accounting, and the comparability between similar studies is weak and needs to continue to be studied and solved. The current carbon accounting method is more concerned with the accounting of historical carbon emissions but has not deeply explored the prediction mechanism for the upcoming carbon emissions of buildings. Knowing the upcoming carbon emissions can provide guidance for the control and distribution of carbon emissions. In addition, at present, the overall carbon reduction is mainly accounted for, but the carbon emission information of each component is ignored. Therefore, it is difficult to accurately predict the overall carbon emission effect. Related studies also suggest that the traditional linear life cycle accounting method is not suitable for sustainability and circularity requirements, each stage does not exist independently, nor can it be analyzed independently [69].

5. Challenges and Prospects

The field of building carbon emissions faces two main challenges. Firstly, there is a need to enhance the accuracy of machine learning methods in predicting building carbon emissions and effectively address various real-world random scenarios, while also ensuring the protection of privacy-related concerns. Secondly, there are issues related to inconsistent standards in the accounting of building carbon emissions.

Addressing the first problem requires a multi-faceted approach. First is the selection of more appropriate building features. For example, temperature, wind speed, rainfall, and UV intensity for training can improve prediction accuracy. Secondly, enhancing data quality by selecting better datasets and increasing their size can further enhance prediction performance. To tackle emission-related challenges, prior probability sensors or a combination of wireless perception technology and intelligent environmental perception technology can help analyze personnel distribution and predict trajectories, thus aiding emission improvement. However, deploying too many indoor sensors may raise privacy concerns, which can be addressed using technologies like federated learning [70], differential privacy, blockchain [35,71] and blinding factors to ensure privacy protection.

Federated learning is a technology that implements machine learning by sharing training model parameters among multiple devices or servers. Unlike traditional centralized machine learning, federated learning keeps data on the local device and only transmits model parameters back to the central server for aggregation. This approach helps protect user privacy because data do not need to be transferred between devices. Differential privacy is a method of quantitative privacy preservation that protects data by adding noise. The basic principle of differential privacy is to add specific noise to the data while keeping the statistical characteristics of the data unchanged so that personal data cannot be identified. This approach can be used for privacy-sensitive data collection and analysis, such as medical records or user behavior data. Blinding is a data-processing technique that preserves privacy by hiding sensitive information. For example, blinding factors can be used to anonymize digital signatures to protect the signer’s identity while maintaining the signature’s validity. Another application is to replace research subjects’ names and other identifying information with codes to protect their privacy when conducting experiments.

In the field of building carbon emission accounting, there are several problems with the accounting method. Firstly, the accounting boundary needs to be clarified. Secondly, the accounting data are numerous and miscellaneous due to the various types of buildings. Lastly, there is a need for uniform accounting standards in this domain. Furthermore, macro guidance standards often lack specific accounting methods, while micro standards may only be partially applicable. To address the issue of non-uniform building accounting standards, the integration of Internet of Things and big data technology can be employed to create a monitoring and management platform. This platform can facilitate real-time evaluation, monitoring, and regulation of power consumption, carbon emissions, and the potential for carbon reduction. The effect of carbon emission reduction can be visualized to avoid much calculation work. Possible methods include comparing different carbon emission reduction measures and their emission reduction effects through a histogram. Such a graph makes it easy to see which measures are most effective and which need further improvement. Additionally, the graph can be broken down into different sectors, such as transport, energy, and buildings, to better understand the impact of emissions reductions. Another example is an interactive map showing the effects of emission reductions in different regions. This visualization allows one to compare regions and discover the potential for reductions in each region. Such a map could compare emissions data for different regions with other data, such as population, economic activity levels, and energy use. Through these visualization tools, people can better understand the effect of carbon emission reduction measures to formulate specific emission reduction plans.

Establishing a big data platform for monitoring building operation data is significant to solve the current problems in carbon emission calculations. The platform can monitor the data of various equipment and systems in the building in real time. It can also analyze these data with the design and use of the building to identify potential energy waste or opportunities for emission reduction. Benefits of the program include the following:

• Collect more comprehensive data: Traditional methods of calculating carbon emissions may need more key data, such as actual energy use. Big data platforms can collect more comprehensive data to provide more comprehensive emission reduction calculations.
• Automated calculation process: The big data platform can use modern computing technology to quickly select different accounting standards to automate the calculation process and provide real-time data analysis and reporting, improving calculation efficiency and accuracy.
• Promote sustainable development: By establishing a big data platform to supervise building operation data, real-time monitoring and the accurate calculation of building carbon emissions can be realized, and energy waste and emission reduction opportunities can be identified. The platform will help the construction industry and relevant government agencies better understand the energy use of buildings, develop more effective emission reduction plans, and promote the sustainable development of the construction industry.

6. Conclusions

This article discusses standard methods of construction data preprocessing, such as data decomposition, clarity, augmentation, and other techniques. The paper provides a summary of common machine learning and deep learning methods used for predicting building energy consumption. Additionally, potential applications and risks of artificial intelligence technology in the field of deep learning are explored to aid in reducing carbon emissions. The article also explains the fundamental concepts of building life cycle theory and outlines standard methods for accounting for building carbon emissions at both the single building and regional levels.

In summary, the main results of this paper are as follows: (1) The paper introduces the preprocessing method for building carbon emission data, which involves utilizing KNN to fill in missing data and employing an algorithm to detect anomalies in building carbon emission data. Additionally, the paper discusses the utilization of the PCA dimensionality reduction algorithm or wavelet analysis algorithm for reducing data dimensionality. Finally, a comparison of the advantages and disadvantages of common data preprocessing methods is presented. (2) Various artificial intelligence methods commonly used to predict carbon emissions are summarized, such as the LSTM memory neural network, GRNN, migration learning, ANN, and multi-model collaborative training using federated learning, and their advantages and disadvantages are compared. At the same time, the technical direction of the possible use of artificial intelligence technology for emission reduction is also discussed. (3) The paper discuss carbon emission accounting for building sectors, including regional and individual building accounting. The article lists the current accounting models and introduces the CBCEM model suitable for our country. Methods of building life cycle accounting are also presented. (4) Building a big data platform for building operation data supervision is recommended to solve the current problems in carbon emission accounting. Other important issues, e.g., privacy protection and random people positioning, are also promising research directions in the future.