1 Introduction

Overall, households account for 20–30% of energy consumption in Europe, North America and Asia. Heating and cooling, lighting, and electric appliances are the three major contributors of this consumption [1, 2]. Moreover, the recent regulation in the EU and China has required buildings to utilize less energy and, at the same time, utilize more renewable energy [3, 4]. To promote efficient and comfortable energy consumption in smart buildings, there is a need to develop and deploy machine learning (ML) applications, neural network (NN) applications and other AI applications coupled with systematic data. However, despite the potential significance of AI in this area, few surveys or reviews have systematically categorized machine learning applications for energy utilization in smart buildings.

The purpose of this study is to review the application segments of the smart energy of buildings since 2009 using a mapping study to scope the topic, and, then, to concentrate on the review of the applications and techniques within that scope. We present our method in Sect. 2. We present our mapping study results and answer the related research questions in Sect. 3. In Sect. 4 we present our literature review results and in Sect. 5 answer the related research questions. We present the open issues and future work in Sect. 6 and our conclusions in Sect. 7.

2 Method

This study has been undertaken as a mixed study starting with a mapping study to scope the study topic and continuing with a related systematic literature review [5, 6]. The first goal of this study is to assess the status of application segments for the smart energy of buildings. This part of the study can be categorized as a mapping study. Then, the second goal of this study is to review the applications and the machine learning techniques identified through the mapping study. This part of the study can be categorized as a systematic literature review. The steps of the mapping and systematic literature review are presented below.

2.1 Research Questions

The research questions addressed by this study are:

  • RQ1: How much smart building energy activity utilizing machine learning has there been in 2009-2021?

  • RQ2: Who leads smart building energy research?

  • RQ3: What research topics are addressed?

  • RQ4: What are the open questions and limitations of the current research?

With respect to the research question concerning the amount (RQ1) of the research, the mapping study addresses this. We limited the search to the period 2009–2021. We recognize that the number of review papers prior to 2009 was scarce. However, there were relevant studies also prior to 2009, and we looked in few of them to have an idea of the prior research [7, 8]. To answer the RQ1, we identified the number of papers on smart building energy and the journals that published them. With respect to the research questions concerning the origins (RQ2) of the research, the mapping study addresses this. We considered the fields of science, to which researchers were affiliated, and the country, in which the organization was situated.

With respect to the research topics (RQ3), they were addressed both in the mapping study and the literature review. In the mapping study we identified the candidates for the taxonomy. The taxonomy was then proposed based on the mapping study and the literature review results. This was an iterative process. Furthermore, in the literature review we considered the scope of each article (i.e., whether it addressed taxonomy, whether it considered applications, and whether it looked at a machine learning centered research question). Consequently, we considered several issues:

  • RQ3.1: What are the discussed topics for smart building energy utilization?- This research question is answered in the mapping study.

  • RQ3.2: What are the existing taxonomies utilized for smart building energy utilization and what are their characteristics?- This research question is answered in the literature review.

  • RQ3.3: What is the taxonomy proposed for the applications for smart building energy utilization?- This research question is answered in the literature review.

  • RQ3.4: What are the characteristics of the applications and the parameters to compare the application areas?- This research question is answered in the literature review.

  • RQ3.5: Which are the techniques utilized for the applications of smart energy for buildings?- This research question is answered in the literature review.

  • RQ3.6: How are the techniques manifested in electric grid, building control, maintenance and security, and personalization? What are the issues?- This research question is answered in the literature review.

With respect to the limitations of this study and future directions of the research (RQ4), the Sect. 6 addresses this question. We considered both the applications and the machine learning techniques.

2.2 Search Process

The search process was initiated by comparing Web of Science and Scopus databases. We selected Scopus as it yielded more results when we proceeded with the first steps of the search process.

With respect to the mapping study, the search process was an automated search of reviews in 2009–2021 with broadly defined search keyword from 34 publications in Scopus including Renewable And Sustainable Energy Reviews, Energy And Buildings, Energies, IEEE Access, Building And Environment, IEEE Internet Of Things Journal, Sustainable Cities And Society, ACM Computing Surveys, and Journal Of Network And Computer Applications. The journals were selected because they were known to include either empirical studies or literature surveys, and to contain machine learning topics. Furthermore, three bibliographic analysis programs were compared (Bibliometrix, CiteSpace, and VOSviewer). We selected VOSviewer as it was the most stable one and had relevant references for the methods utilized [9].

With respect to the literature review, to extract any details about the taxonomies proposed so far, the search was a manual search of specific survey and review papers. The journals selected were the same as above. Each review publication was reviewed by the main author and the papers that addressed taxonomies for smart energy of buildings of any type were identified as potentially relevant. Then, the main author applied the detailed inclusion and exclusion criteria to the relevant papers (refer to Sect. 2.3), and another round of reading the paper abstracts, conclusions, problem statements and taxonomies was done.

With respect to the literature review, to extract more details about the applications and machine learning techniques, the search was a manual search of specific conference proceedings, journal papers, and research papers. The journals selected were the same as above. Each review publication was reviewed by the main author and the papers that addressed applications for smart energy of buildings of any type were identified as potentially relevant. Then, the main author applied the detailed inclusion and exclusion criteria to the relevant papers (refer to Sect. 2.3), and another round of reading the paper abstracts, conclusions and problem statements was done.

2.3 Inclusion and Exclusion Criteria

For the mapping study, all peer-reviewed surveys and literature reviews from 2009–2021 were included using the search string ("smart building" AND ("energy" OR "energy-efficiency" OR "comfort" OR "anomaly detection")) AND ("machine learning" OR "supervised learning" OR "unsupervised learning" OR "neural networks" OR "reinforcement learning"). Primary research articles were excluded. For the related bibliometric analysis, a term was included if it had a minimum of 25 occurrences and a membership in the 100 most relevant terms. The relevancy was based on the full counting method of the bibliometric tool. Too generic terms (problem, condition), repetitive terms (IoT, IoT system), and irrelevant terms (review, study) were excluded.

For the literature review, the above mentioned studies as well as research articles from 2009–2021 were included. We included some articles that were cited in the surveys or literature reviews if the topic had only few search results. Furthermore, a paper had to have over 8 citations and it had to have some reference to the data utilized. For the inclusion criteria, the emphasis was on energy applications for existing buildings and their dwellers. We excluded healthcare, IoT, and smart city articles, articles concerning the initial design and construction of buildings as well as articles concerning the final reuse of building materials. We also excluded smart energy articles that exceed the topic identified. Finally we excluded repetitive articles with only a minor difference from the perspective of this study if there were several articles concerning similar application or methodology.

2.4 Data Collection and Analysis

With respect to the mapping study, the data extracted from each review was:

  • The source journal

  • The author(s), their institution and the country where it is situated

  • Titles, abstracts and citations

  • The number of co-occurrence of terms in reviews

A co-occurrence is the number of publications in which two terms occur together [9]. The data was summarized in graphs and tables to show:

  • The number of reviews published per year (addressing RQ1).

  • Countries of the authors and the fields of the surveys (addressing RQ2).

  • The clusters of terms in the data, i.e., candidates for application areas in the taxonomy (addressing RQ3.1).

  • The proposed taxonomy (addressing RQ3.2).

With respect to the literature review, the data extracted from each article was the same as above as well as:

  • Conclusions of the articles

  • Problem and issue statements concerning the topic in the text body

  • Taxonomy proposed (surveys and reviews only)

The data was summarized in graphs, tables, and analyzed in text format to address the remaining questions:

  • The existing taxonomies for the applications and their characteristics (addressing RQ3.3).

  • The characteristics of the application areas and their parameters (addressing RQ3.4).

  • Methods utilized in applications for the smart energy of buildings (addressing RQ3.5).

  • Applications reporting the most issues and issues related to them (addressing RQ3.6).

3 Results and Discussion for the Mapping Search

This section summarizes the results and answers the research questions related to the mapping search.

Figure 1 presents the number of articles that were identified in the search process. This answers the research question: How much smart building energy activity utilizing machine learning has there been in 2009–2021? In total 483 surveys or literature review articles were identified. The trend for the number of articles is rising, especially in the years since 2018.

Table 1 shows the 12 top countries of origin for the authors of these articles with the respective number of documents and citations. Furthermore, Fig. 2 shows the fields of the surveys according to Scopus. These answer the research question: Who leads smart building energy research? In respect to the countries, the United States, China and the United Kingdom are the three leading origins of the articles. There are clear differences between the countries in the number of citations. In average, an article from Denmark gets more citations than an article from Spain, for example. For the fields of surveys, engineering, computer science and energy are the top three fields.

Fig. 1
figure 1

Publishing trend in the domain of Machine learning and neural network applications for smart building energy utilization (source: Scopus)

Full size image
Fig. 2
figure 2

Fields of the articles (Source: Scopus)

Full size image
Table 1 The rank (id) of the country, countries of origin for the authors, number of documents, and number of citations
Full size table
Fig. 3
figure 3

Publishing trend in the domain of Machine learning and neural network applications for smart building energy utilization. (Source: Scopus, VOSviewer)

Full size image

The results of the bibliographic analysis applied on the mapping are presented in Fig. 3. This answers the research question: What are the discussed topics for smart building energy utilization? There are three major topics or clusters of terms. Moreover, with respect to the applications, many terms specific to an application emerge, such as ’microgrid’, and ’smart grid’. With respect to the machine learning techniques utilized for the applications, three terms emerge: machine learning, deep learning, and blockchains. However, any other terms related to supervised, unsupervised and reinforcement learning (RL) methods, that were part of the search string, are missing.

For the articles identifying the clusters, we limited the search only to the reviews which still resulted in a large number of papers. This means that machine learning techniques, that are often described in more detail in research articles, may have been missed with this search process. However, this gives some balance to the results on the scoping of the application areas.

For the bibliographical study, there are some logical overlaps as the bibliographic analysis program does not combine related terms, such as heating and ventilation vs. HVAC (heating, ventilation and air conditioning) or energy system vs. energy management.

For the machine learning techniques, analysis of the techniques is not feasible based on the results of the mapping study, and, therefore, is left to the literature review.

4 Results for the Literature Review

This section summarizes the results for the literature review.

4.1 Results for the Taxonomy

Table 2 summarizes the search results of the previous studies. Figure 4 shows the taxonomy proposed here. It is based on results of the mapping search and literature review (taxonomy) search. It was iterated during the consequent steps of the search process.

Fig. 4
figure 4

Taxonomy for smart energy applications for the built environment

Full size image
Table 2 Identified keysurveys for building applications with a taxonomy
Full size table

For the articles investigated for the taxonomy, we systematically went from the highest cited surveys to lowest cited surveys. However, for the emerging studies (e.g., personalization), we had to select articles with fewer citations, or had fewer articles in our disposal, than for well-established areas (e.g. smart grid).

4.2 Results for the Applications

Table 3 summarizes the search results for applications. The columns Application domain, Application area, Objectives, and Related articles are the relevant ones for this section.

In respect to energy management systems, a further result was that 49% of the EMS applications are related to a whole building’s energy load, while the rest 51% to its other loads, such as heating or cooling. Adjacently, the time horizon is less than 24 h for 41%, one day for 35%, and longer for 24% of the applications [31].

For the research articles and conference papers investigated for the applications, we started from 5120 articles and systematically went from the highest cited articles to lowest cited surveys. However, for the emerging areas (e.g., blockchain contracts), we had to select articles with fewer citations, or had fewer articles in our disposal, than for well-established areas (e.g. microgrids). This is a bit disappointing and resulted in missing some possibilities to quantitative analysis for some application areas, but a quantitative analysis is difficult to do for emerging topics. The applications are discussed in more detail in the Sect. 5.2.

4.3 Results for the Machine Learning Techniques

This section summarizes the results for the machine learning techniques.

Table 3 presents the results of the search for the techniques utilized for the applications. We identified 85 articles by the search process. Potentially relevant studies were excluded according to the inclusion and exclusion criteria. Furthermore, from these articles, we identified the following techniques: Artificial Neural Network (ANN), Autoregressive Moving Average (ARMA), Bayes, Bee Colony, Blockchains, Classification and Regression trees (CART), Clustering, Convolutional Neural Networks (CNN), data analysis, Deep Belief Network (DBN), Deep Neural Network (DNN), hybrid models (e.g., Deep Neural Network and Long-Short-Term Memory (DNN-LSTM)), Deep Reinforcement Learning (DRL), ensemble methods (e.g., severals ANNs), Extreme Gradient Boosting (XGB), Firefly, Fuzzy Logic, Generative Adversarial Networks (GAN), Genetic Algorithm (GA), Gradient Boosting (GB), K-Nearest Neighbors (kNN), Nash Equilibrium, Particle Swarm Optimization (PSO), Random Forest (RF), Recurrent Neural Networks (RNN), Reinforcement Learning (RL), Support Vector Machines (SVM), and Support Vector Machines with Dragonfly. Noteworthy, several algorithms have been applied to energy management systems and microgrids. Undoubtedly, this reflects the bigger amount of research done in these areas.

For the application domains, we tried to identify a relationship with the machine learning techniques utilized. Overall, out of all 85 applications, 33% utilize RL or DRL, 22% other neural networks and 45% other methods. Applications concerning both load and money objectives utilize prominently reinforcement learning method in both smart grid and energy management domains. Out of the 27 related solutions, 48% utilize RL or DRL, 22% other neural networks and 30% other machine learning or analytical methods. Applications concerning personal comfort utilize to a noticeable extent reinforcement learning method. Out of the 12 related solutions, 33% utilize RL or DRL, 8% other neural networks and 58% other machine learning or analytical methods.

Table 4 presents the results for the issues identified for the applications and techniques. We identified 19 topic types that were divided into 34 topic areas based on an analysis of the search results. We tried to identify a relationship with the issues and either the applications or the techniques. The reported issues for the techniques and related models were predominant in comparison to the issues for the application (design) as such.

Table 3 Overview on applications and the methods utilized
Full size table
Table 4 Identified issues in the applications and techniques
Full size table

In respect to energy management systems, a further result is that the selected time-horizon affects the yielded results. For example, [129] compares deep reinforcement learning and common supervised models with a shorter or longer time-horizon concluding that there was a trend of rising model error with multi-step forecasts and with longer time-horizons. Finally, to make more data, [63] proposes a way of synthesizing rule-based controllers with simple rules and with simulated data utilizing support vector machines and AdaBoost; however, the practical applicability of this was not proven at that point of time (2014).

5 Discussion on the Literature Review

5.1 Discussion on the Taxonomy

In this section, we discuss the answers to our research questions related to the taxonomy.

5.1.1 What are the Existing Taxonomies Utilized for Smart Building Energy Utilization and What Are Their Characteristics?

Overall, we identified 21 relevant sources during the search (Table 2). The applications constitute various application domains, such as smart grid, several types of building energy controls and managements, personalization, and maintenance and security. The issues and challenges faced are heterogeneous. We have analyzed them in more detail in the context of the techniques in Sect. 4.3 as many of them turned out to be technique related. The applications were arranged into taxonomies, classifications or frameworks.

Finally, at this step of our survey, we identified a research gap: we did not discover a taxonomy that covers the domain of energy utilization (electric grid, smart building energy management and control, personalization, and maintenance and security) for smart buildings.

5.1.2 What is the Taxonomy Proposed for the Applications for Smart Building Energy Utilization?

Figure 4 shows the proposed taxonomy that was based on the analysis of the mapping study and the literature survey. It consists of the electric grid, the smart building energy management and control, the personalization, and maintenance, the security, and the data application domains with several applications types for each of them. The taxonomy was iterated during the literature review to show more details.

Notice that the taxonomy uses terms as they appear in the articles, albeit often in abbreviated form. The exception is the term ’cost-aware scheduling’ that this study adopted. As the energy load prediction as such has evolved to scheduling of devices and to taking into account the cost factors, there is a need for defining terminology. Even though some of the identified articles propose applications to these current aims, they mostly do not articulate this in precise terms.

This taxonomy is different from the existing ones, as the taxonomy of some surveys concentrate on a type of machine learning, for instance, neural networks [24], or on an application area of a smart building, for instance building load [30]. Moreover, [27] considers operation, control and retrofit; however, it does not include the human-in-the-loop as such in its taxonomy. Even though, for example, [130] presents solutions that incorporate building occupants in sensing and control frameworks, it does not include other smart building components. [2] concentrates on energy management systems, but not on the related energy components.

5.2 Discussion on the Applications

In this section, we discuss the answers to our research questions related to the applications.

5.2.1 What are the Characteristics of the Applications and the Parameters to Compare the Application Areas?

Overall, for the four main application domains of the taxonomy, each of them contains many applications (Table 3). In the following we present the applications for the smart grid, the building energy control and management, the personalization, and the maintenance and security domains.

The first strand of literature presents smart grid (electric grid) related solutions in smart buildings. In contrast to traditional grids, the smart grids with demand side management can deploy tasks to manage electricity load or improve energy efficiency [131]. According to the surveys [132,133,134] the main components of a smart grid are advanced control to utilize data, integrated appliances, integrated renewable energy sources, integrated energy storages, optimization of the grid use, and safety measures. The main objectives for the applications of the smart grid are to flatten the load peaks either by reduction of consumption during high demand or by shifting of load to low demand periods [25, 135]. The broader objectives include also lower greenhouse gas emissions and design of adequate interfaces to AI applications, for instance.

Reducing peak load demand that increases the grid capacity and reliability is called demand side management [102]. The smart building owners are involved in participating in the demand side management through different demand-response (DR) programs of a smart grid. As stated by the United States Department of Energy, DR refers to “a tariff or program [...] to induce lower electricity usage at times of high market prices or when grid reliability is jeopardized” [136]. The demand-response programs are broadly divided into incentive-based schemes and price-based schemes. The incentive-based schemes are further divided to direct control and indirect control schemes where the former refers to a mechanism, by which a utility remotely adjusts or disrupts energy consumption of controllable appliances in a short notice (1 s–1 min). The indirect control refers to a mechanism, by which a utility notifies end-users and asks them voluntarily to reduce the peak load with prior notification (up to some hours). Both of these incentive-based schemes offer monetized rewards for the end-users and can include different kinds of bidding mechanisms or commitments [137].

In turn, the pricing schemes are further divided into time-of-use, real-time pricing, and critical peak pricing schemes. Time-of-use pricing scheme settles a group of prices in advance and then applies them to different predefined intervals of a calendar day [138, 139]. Critical peak pricing is according to the United States Federal Energy Commission a scheme that “encourages reduced consumption during periods of high wholesale market prices or system contingencies by imposing a pre-specified high rate or price for a limited number of days or hours” [140]. For instance, critical peak pricing can be used together with pre-set prices, but the trigger to change the price is dynamic, such as a change in wind-power generation [104]. Real-time pricing responds continually to the energy spot market, because of which it changes constantly. The utility may reveal the price a day ahead or an hour ahead so the end-users can adjust their energy consumption accordingly [133, 141]. An overview of the demand-response framework is in Fig. 5.

Other techniques to enable load reallocation can include, for example, smart meters that can be utilized, on the other hand, by the grid operator for clustering the energy consumption pattern, and on the other, by the local producer-consumer (also known as prosumer) for trading the energy based on the current state of the renewables, market prices and energy consumption [16, 142]. That is, a smart meter is a component that enables communications between the grid and a smart building which is utilized for forecasting, clustering, classification and optimization [16, 143]. However, there were relatively few papers concerning smart meters in the smart building area; the focus was more on the utility price-setting.

Fig. 5
figure 5

Demand-response framework overview. Source [137]

Full size image

For pricing and incentive schemes, to address the lack of motivation of the end-users, the grid operators have run demand-response trials with customers, such as in New Zealand in 2007-2020, which concluded that raising awareness of the system was one of the main goals [144]. However, if the personal monetary incentives are small, sacrificing the comfort level in exchange for financial incentives is not accepted by the majority of the residential dwellers [117, 118]. (See also Table 4 for more issues).

A second prominent strand of literature presents energy management systems (EMS) and other control systems for smart buildings. An energy management system consists of hardware (sensors, actuators, information technology components) as well as software for operating logic (controls and alarms). These systems are called with several resembling terms (building energy management, home energy management, building automation and control systems); however, these terms are broadly considered to be close synonyms to each other [134]. From a functional perspective, energy management systems are tools to shift, and constrain energy consumption and production profiles automatically in a smart building at defined times. An EMS usually creates an optimal consumption and production schedule by considering multiple objectives such as energy costs, environmental concerns, load profiles, and consumer comfort [12]. That is, the objectives of an EMS of a smart building include typically energy cost reduction, user-comfort maximization, energy load (and generation) profiling, and emissions cut [12]. However, depending on the EMS, the system can utilize a different set of objectives, such as user-comfort, energy savings, or indoor air quality omitting, for instance, emissions [10, 11].

In turn, occupancy is a significant aspect that is related to a considerable amount of applications for energy management and personal comfort [31, 122, 145]. Some stand-alone aspects of a building’s energy control, such as energy load as such, or lighting, can also be a focus of an application.

In turn, microgrids are a prominent topic in the literature of energy management [13, 146]. Microgrids are defined by the US Department of Energy as: “A group of interconnected loads and distributed energy resources (DERs) with clearly defined electrical boundaries that acts as a single controllable entity with respect to the grid and can connect and disconnect from the grid to enable it to operate in both grid-connected or island modes” [147]. In other words, microgrids for smart buildings consist of electric appliances, local (renewable) energy sources, utility energy sources, and energy storage systems as well as advanced controls (Fig. 6). Moreover, scheduling the charging-discharging of electric vehicles (EVs) can be part of building energy management [14]. The main objectives of a microgrid are energy-efficiency and cost-efficiency for a smart building and its owner. Moreover, the utilization of microgrids contributes to the smart grid stability and reliability [15, 133]. The concept is close to the energy management systems, but with a more focused definition than the definitions of various energy management systems.

Fig. 6
figure 6

Overview of a microgrid in a smart building. Source [24]

Full size image

A third strand of literature presents the personalization in smart buildings which relates to comfortable utilization of buildings, or, some other personal aspects, such as (personal) energy market participation that the smart grid concept has enabled recently.

Figure 7 presents the main components to manage personal comfort indoors. The models have evolved in the last ten years from classification of human activity, to construct more explicit models utilizing human signals and activity to predict the comfort experienced [148]. These personal comfort applications save energy with the reasoning that they increase the utilization rates of the buildings. That is, people tend to prefer comfortable environments [149]. Furthermore, in practice the control models seldom interact directly with building users. However, there are ideas on buildings that ’talks to me’, or takes into account the personalities of the users [126, 150].

Fig. 7
figure 7

Challenges of machine learning application to achieve user-comfort for building users

Full size image

Being the most studied domain of comfort, the human thermal comfort comprises several factors impacting it as depicted in Fig. 8 [77]. Moreover, balancing between the thermal comfort and energy costs is a common dual target for comfort modeling [128]. Noteworthy, use-comfort evaluation is predominantly based on a majority opinion as defined in standards. For example, ASHRAE or ISO15251 standards specify the combinations of indoor thermal and personal parameter values that will produce thermal environmental conditions acceptable to a majority of the occupants within the space [151, 152].

Fig. 8
figure 8

A graph with environmental and human factors impacting thermal comfort for making a model and predicting the thermal preference. Source: [77]

Full size image

As for the benefits, [22] yielded a result based on field-studies that occupant voting and preference profiling in comfort-aware HVAC operations illustrated a median of 20% of energy savings. However, in practice, personalization is often limited to lighting. An exapmle of this is, at least according to the information available, a so-called smart building in Amsterdam opened in 2014 [153].

In turn, smart building measurements can be utilized for formulating recommendations. For instance, [28, 142] present applications for anomaly detection of energy consumption in buildings. Some solutions can detect anomalies in specific hours in the day, or specific days on building level; hence they can provide end-users with a personalized feedback to reduce wasted energy. Furthermore, [19, 22] present anomaly detection to identify occupancy and comfort, which can be utilized for user-centric operations of HVAC systems.

This kind of information can be extracted from energy control and management systems to evaluate personal thermal comfort as well. For example, occupancy behavior and possible recommendations can be gained from the system that [17] presents. This application for energy demand flexibility has a range of information on subjects, such as photovoltaic cells and wind, as well as heating, ventilation, air conditioning (HVAC) systems, energy storage, building thermal mass, appliances, and mechanisms to learn occupant behaviors. Buildings can become more flexible in terms of power demand from the power grid if all of these measures are considered from the beginning in modeling.

In turn, energy market participation can be regarded as part of the personalization of smart building energy utilization. For instance, blockchains that can be utilized for decentralized energy trading and that fulfill privacy, security and trust objectives [154] (See Fig. 9). Also machine learning based solutions may be utilized for energy market participation [17].

Fig. 9
figure 9

Overview of blockchain technology in energy peer-to-peer trading. Source: [155]

Full size image

A fourth major strand of literature concerns instabilities, faults and intrusion of building energy systems, that is, maintenance and security. For instance, [156] presents solutions for integration of power systems that include instabilities and security concerns that machine learning solutions can solve. Recently, blockchains have been introduced to ensure together data security and privacy for smart building applications [157]. Moreover, at building level, [18] presents several unsupervised learning methods for analysis of performance that benefits maintenance, such as clustering, novelty detection, motif and discord detection, rule extraction, and visual analysis. Information retrieved can be used for both maintenance and security applications.

Preventive maintenance includes several applications. A maintenance solution can combine several domains, so that anomaly detection can be utilized not only in cases of power transmission failures but also in cases of intrusion[21, 156].

A fifth strand of the literature presents big data (issues) for smart building energy utilization [26, 119, 120]. All solutions based on machine models require applicable data. Hence, an underlying data model for buildings and people is necessary. Moreover, the topic of data models is a pervasive theme in the smart energy application papers surveyed here; therefore, it is discussed. The big data includes data on monitoring, control, maintenance, automation, and personalization of energy utilization in smart buildings. [26] classifies data to the following seven classes including, but not restricted to, energy savings, appliance recognition, occupancy detection, user preference detection, anomaly detection, energy disaggregation and energy demand prediction. Moreover, typically, the system components communicate over a centralized hub via a common interface [158]. For instance, a typical interface in the European Union is the REST API [159] that specifies the data transfer between an appliance and a machine learning algorithm.

For human related data, [160] summarizes a disparity in human modeling studies: some focus on occupant behavior in order to model that, while, on the other hand, some focus on occupants’ interaction with given buildings and devices in order to develop interfaces. This disparity raises the question how to adequately combine these human data models for applications developed. One noteworthy development is that mobile and portable technologies have made it possible to provide personal recommendations. The present smart building equipment controls can be connected to interactive (mobile) applications, such as thermostats [161].

Even though not included in this study, we recognize that Internet of Things and multi-agent systems provide data and tools to management systems, for example, by providing data from sensors [162, 163]. Moreover, there are several ways to achieve the objectives, not necessarily related to an application utilizing machine learning. For example, for personalization of the immediate environment can be achieved with personal gadgets, such as personal fan or heating device. [164] proposes using a heated and cooled chair to increase comfort. However, while providing the answer to the personalizing, in a strict sense, this proposal does not belong to the domain of learning models for buildings that we scrutinize here. At the same time; in spite of the above, these kinds of gadgets could be integrated as part of a recommendation (model) for personal comfort.

Finally, as a generic summary for building operations, the United Nations has defined some broad goals for sustainable cities and communities [165, 166].

5.2.2 Juxtaposing the Definitions of Some Terms

As the mapping study suggested, the number of topics for articles has grown, especially since 2018. The literature study indicated that the solutions have become more specialized, and perhaps even more overlapped. To make an analysis on ‘smart building’, we need to scrutinize this concept. We noted that it is close to ‘smart energy, ‘smart grid’, and ’smart city’ according to the literature review; therefore, we juxtapose all of them. The definition of smart building covers energy conservation, safety, comfort, lighting, heating, and appliances. Machine learning is part of smart building concept as models used to control smart buildings learn from data gathered from sensors and smart devices in contrast to buildings relying on a pre-programmed models [167,168,169]. More recent definitions of smart buildings consist of heterogeneous artificial intelligence (AI) models to solve problems related not only to buildings directly but also to personalization, electric grid, and data security according to [134, 170]. Utilizing these models requires that the building has advanced control mechanisms that can ensure the reliability and operating points for the models. This control system is often referred to as (building) energy management systems, but some other system, such as average voltage control, can be deployed as well [133, 133]. In this paper, we have discussed energy applications for smart buildings having the advanced controls in the broad definition of the term.

In turn, the basis for present-day energy systems in most countries is a simple scheme, where energy resources are converged to meet the demand; moreover, increased demand is met with increased utilization of resources [171]. Notwithstanding this simple scheme, ‘smart energy’, ‘smart energy systems’, or sometimes ‘intelligent energy’, is defined as sustainable energy scheme within the electricity, gas, building, and industrial sectors where a few of these solutions utilize renewable energy sources, energy storages, and national or local energy systems that are energy-efficient in a flexible manner [24, 171, 172]. In turn, smart grid refers to advanced electric power grid infrastructure for improved efficiency, enhanced reliability and safety involving the electricity generation and utilization [133, 134]. Finally, ’smart city’ can be defined as networked infrastructure coupled with high technologies, creative social and environmental industries, that focuses on achieving sustainability. ’smart city’ incorporates a wide range of intelligent systems from education to waste management [133, 134].

In this study, we have concentrated on the ’smart building’ and the ’smart grid’ concepts.

5.3 Discussion on the Literature Survey for the Machine Learning Techniques

In this section, we discuss the answers to our research questions related to the machine learning techniques.

5.3.1 Which are the Techniques Utilized for the Applications of Smart Energy for Buildings?

According to the search results in Table 3, the application to solve problems related to the smart building energy utilization consisted of machine learning and neural network applications, and to a lesser extent, pure data analysis based solutions. The machine learning and neural networks are utilized to develop classification and regression models where classification refers to a method that predicts a discrete class label output for the given task, while regression predicts a continuous quantity output for the given task. Both of these two common methods are part of supervised learning [173]. However, other methods were also applied. For example, unsupervised learning methods were utilized for some problems, mostly to discover more about the task in hand. Unsupervised learning refers to a method that uses machine learning algorithms to analyze unlabeled data and cluster it to smaller sets. Moreover, unsupervised learning can extract associations between the pieces of data [174]. In turn, a combination of supervised and unsupervised methods, so-called semi-supervised learning, was also utilized by algorithms that, for example, actively querying users for labels, such as personal comfort experience [175]. Finally, reinforcement learning methods were applied to some problems, such as complex real-time learning tasks with rewards and punishments that try to find an optimal policy for the model.

Choosing a machine learning technique matters as a method has some inherent issues associated with it. Overall, a desired outcome is that a model created in one environment (building) can be used in another environment without significant degradation in its results. This is called model generalization capability.

For design criteria, the following article raises application design criteria unlike many other articles [64]. These six major design criteria for a smart energy application are implementation feasibility, usability, computational time, accuracy, randomization, and adaptability. An example of a definition for one of these criteria is the implementation feasibility that is defined as the level of ease to implement the technique in the restricted amount of time and resources.

In respect to the energy management systems, the current heating, ventilation and air conditioning (HVAC) systems are still operated by simple feedback controls, such as on-off control or proportional-integral-derivative (PID) controls to a large extent. These simple control strategies are proven; however they do not consider predictive information, which affects their energy performance. Optimal control strategies such as Model Predictive Control (MPC) address these drawbacks by iterative optimization of an objective function over a receding time horizon. However, the MPC models do not adapt well to different buildings. Recently, several reinforcement learning models to predict the load combined with cost function models to predict the costs have been introduced for energy management systems to enhance the energy and cost efficiency of smart buildings [103, 176].

In respect to renewable energy generation, the sun related phenomena (solar radiation, wind) supply energy intermittently due to changes of weather. Therefore, renewable energy sources need to be integrated by a suitable solution design. Time-series forecasting is typically applied to energy management [121] but other machine learning methods are applied as well [30].

In respect to occupancy applications, implementation and utilization of applications vary; hence, doing a comparison for the models derived is arduous as concluded in the comparison of 19 papers for forecasting occupancy [177].

In respect to personalization applications, there are mixed results on their use. The results of [82] indicate, only around 30% of the occupants’ thermal comfort improved, and it decreased in 70% cases, when the user-driven control strategy was implemented in a disheveled manner. The problem identified was that performance of the demand-driven control strategy was mainly influenced by insufficient and poorly distributed data, as well as the effect of thermal expectations on occupants’ thermal responses.

In respect to maintenance and security, blockchains have been introduced to ensure together data security and privacy for smart building applications [157].

5.3.2 How are the Techniques Manifested in Electric Grid, Building Control, Maintenance and Security, and Personalization? What are the Issues?

Overall, machine learning applications are suitable for non-linear modeling and are often more accurate and faster than traditional deterministic models in the domain of energy utilization; hence there will be more of them in the field in the future [121].

According to the results in Table 4, there are several issues related to the techniques ranging from generic ones to more method related ones. Here we raise prominent topics to further discussion.

In respect to the generic issues identified, there is a tendency for machine learning research to claim to deliver better performance compared with previous studies as a survey of 144 recent research papers states [27]. The results of this literature survey agree with this piece of findings. This is partly due to the lack of common metrics. Moreover, evaluation of the quality of the results or the applicability of the results are often limited or missing. This is due to deficient reporting of the issues; moreover, for a proportionally large number of studies, these aspects were absent in the abstract part so that a systematic literature review utilizing database tools was not able to retrieve them. However, there are some exceptions, such as [59] that explicitly evaluate the applicability of the model in a critical manner.

In respect to techniques, reinforcement learning has recently been applied to a proportionally large number of the advanced building control solutions. It is probably utilized because it provides a way to avoid the tedious work of developing and calibrating a detailed model, as yielded by, for example, traditional Model Predictive Control (MPC) of buildings that has been in use since the 1970 s [122, 128, 176]. There have been some claimed problems with the reinforcement model generalizability (Table 4). However, at the same time some other literature states the opposite is true as some aspects of the modeled applications are general. An adequate policy gradient (a mapping from state to action) generalizes well; for example, turning on the heating when the indoor temperature is low remains the same for almost every building regardless of the other possible goals. Therefore, a rising trend is to study using policy gradient [103] or some other technique claimed to generalize well, such as actor-critic [178] technique and apply them to so-called transfer learning [179]. Moreover, an additional benefit here is programming efficiency. Rather than training a RL controller for every individual building, it can be more efficient to train RL controllers on a small number of buildings and then apply them to a stock of buildings.

In respect to the time-series, they are long-proven and still utilized for volatile phenomena, such as natural phenomena (e.g., wind, sun radiation) and human behavior (e.g., electric vehicle charging patterns, lighting, occupancy).

In respect to data analysis, a prominent application field for these is the human thermal comfort approximation. There is a comprehensive effort to collect enough statistical data to form credible analytical or machine learning models to this end (more about the databases in question further below).

In respect to blockchains, the results indicated a new trend: recently blockchains have been utilized for safe contracts and their applicability to the other application fields is also being scrutinized.

In respect to many other machine learning techniques utilized, they are more or less evenly distributed among the applications.

Finally, in respect to the data, the quality of data has an impact on the model results. [121] yielded a result that models utilizing field-collected data and taking the time-aspect into account achieve more accurate predictions than simulation-based models. Many surveys emphasize the importance of relevant (recent) data; furthermore, some emphasize model explainability, for example, through so-called grey box models, where machine learning models are combined with the physics-based equations and exogenous data for the model. However, to form a common ground and ameliorate the research preconditions, some preliminary efforts exist to provide more open data. These include, for instance, the ASHRAE Global Thermal Comfort Database [180] on personal behavior and comfort data, and the Building Data Genome Project [181] on building data.

5.4 Limitations

The procedures used in this study have deviated from the advice presented in Kitchenham’s guidelines in several ways:

  • The search related to the literature review was organized as a manual search process of a specific set of journals and conference proceedings and not as an automated search process. This was consistent with the other research aiming at identifying research trends and wider topics as opposed to a pure quantitative technology evaluation.

  • A single researcher was responsible for selecting the candidate studies.

The adapted process implies that we may have missed some relevant studies, and thus underestimate the extent of the related application areas or technologies. However, in this study, there was a systematic tertiary survey part on literature surveys, as there were relatively few primary studies for certain application areas (personalization and blockchain contracts), and the data extracted from the selected survey articles are considered to be relatively objective, so we do not expect several data extraction errors.

6 Open Questions

We discovered several unsolved issues and challenges while conducting this study. Several issues are common for the majority of solutions while each method or application may have specific problems of its own. This section answers the research question: What are the open questions and limitations of the current research?

  1. 1.

    Smart energy solutions for buildings are based on data from several sources, including building variables (energy load, for example). Even though there are some databases that have this data, the lack of applicable data is perhaps the primary issue when making applications. What properties of buildings to report? How many buildings to include? What kind of appliance and energy load model to utilize? What parameters of energy are needed (voltages, currents, phase shifts)? What is an applicable measurement frequency? How long measurement periods are adequate?

  2. 2.

    Moreover, the models include data on occupancy information and consumer preferences. How to define the control system for occupants (what is controlled and how is it controlled?)? What comprises a full predictive control? How to define and measure consumer preferences? Are preferences pre-set, or is there human-in-the-loop directly or indirectly?

  3. 3.

    Applications for smart energy for buildings utilize models created to solve tasks related to not only to energy utilization directly, but also to building users and possible anomalies that may affect the energy utilization. How to apply the collected data to train and validate models? How to adequately label data for different uses? For example, how to create a model, which allows a model trained with a data-rich building to be used in another building with limited data (and is this even feasible)? How to develop a robust model for one building or buildings that can be generalized to others (assuming the same level of data richness)?

  4. 4.

    Each new piece of machine learning research tends to claim to deliver better performance compared with previous studies. How to define a common metric to compare the created models?

  5. 5.

    The applications connect the buildings into the communications networks. How to detect and resolve intrusions? How to integrate fault detection and preventive maintenance into building operations?

  6. 6.

    Energy management solutions may require quick response due to rapid changes in the energy needs. How to adapt to real-time data to enable quick response?

  7. 7.

    The decentralized database of blockchains has consensus mechanisms that can be applied to data security, for example when making contracts. How to apply blockchains to solutions so that they are safe and keep privacy? How do they work in smart grids? Is there a need for more standards?

These open issues also turn into summary of future challenges to be researched in the field. One current challenge includes short-term energy savings by controlling the building energy usage to mitigate the peak loads in the grid as the energy shortages have affected, especially, the European Union. This challenge is related to the above open issue 6. Finally, this section has answered the last research question: what open questions remain?

7 Conclusions

This paper has presented a comprehensive study about prevailing machine learning applications for smart building energy utilization. As the application areas are heterogeneous, various methods and techniques have been proposed to solve the questions on smart grid, smart building energy management and control, personalization and maintenance and security solutions. Furthermore, this paper has proposed a taxonomy for machine learning applications for smart buildings covering the application domains identified. The predominant application domain includes energy management systems and microgrids. This study has highlighted some energy management systems with different architectures; noteworthy, one that includes building energy load, local renewable energy sources, local energy storage system, and electric vehicles. Satisfying the two major and, to some extent, conflicting objectives for an energy management system, that are user-comfort and energy-costs, has resulted in models that are sometimes complex. Moreover, intrusion, faults, and personal involvement are apparent challenges and require smart building applications of their own as smart buildings are required, by definition, to interact with their environment. In turn, a comparative analysis on machine learning approaches and pros and cons in each application domain has been presented. Although solutions are frequently developed using machine learning methods, sometimes pure data analysis approaches are applied at least as a part of a solution. The overall open issues for the smart building energy utilization have also been presented toward the end of this survey, as similar challenges occurred in many of the papers. They are often related to lack of labeled data or other data quality. However, the open issues also relate to the overall objectives, data privacy, model generalization, and some issues of local vs. distributed applications. These open issues also turn into summary of future challenges to be researched in the field.