Several research works have been carried out on different intelligent methods for BEMS, and the most used are categorized as: (i) learning based methods of the AI domain, including a support vector machine (SVM) [128,129], K-nearest neighbor (K-NN) [128], artificial neural network (ANN) [130], reinforce learning (RL) [131], etc.; and (ii) model-based methods, such as the model predictive control (MPC) technique [132].

A learning-based control model where self-scheduling loads and the ESS of the building ensure the maximum usage of PV power, curtailment load profiles, and a reduced energy cost. The optimization algorithm optimizes the building cost and minimizes the fluctuation between PV generation and energy consumption, relying on predicted control information using a machine-learning algorithm [11]. In addition, it demonstrates the benefit of combining model-based and learning-based control methods into a single management framework for controlling multiple aspects of building performance [11,133].

The predictive algorithm model of energy consumption in smart buildings with a Microsoft Azure cloud-based machine learning platform wherein three methodologies namely SVM, K-NN, and ANN have been adopted for comparison to find the most accurate prediction model as compared to actual demand. Notably, the collected data is partitioned between 70% training data and 30% testing data. Though it took the most time to train, SVM was able to predict more accurately than the other two. The K-NN approach, on the other hand, was faster at completing the training, and ANN was faster than SVM, but it took many hours. However, a high training time is required for the SVM model to achieve a higher performance [134]. Most researchers focus on maintaining the indoor comfort of the buildings by implementing complexity prediction algorithms with higher cost and computational time, and power consumption. The model is carried out and investigated on five standard supervised machine learning models: polynomial regression (PR), support vector regression (SVR), random forest (PF), extreme gradient boost (XGB), and neural network (NN). In contrast, the best result output is shown for the SVR model with a tiny prediction error. A trustworthy NN model needs extensive data training set input and more processing time, while the amount of data is relatively small in the proposed system application, so it is rational to avoid the NN [135].

A novel deep neural-network-based algorithm for the day-ahead hourly energy consumption profile prediction of residential and commercial buildings in terms of occupancy rate and seasonality was investigated. A survey was conducted by a business manager, an electrical engineer, and a data scientist about all machine learning techniques, in which deep ANN achieved the highest score (multi-criteria analysis = 189) with a high accuracy prediction rate of about 98%. Firstly, training data was generated using synthetic load generation, where 100,000 data set points (load profiles) were produced for each occupancy rate type. The proposed forecasting model introduced these data to train an eight-layer deep neural network-based model and made decisions based on a limited number of inputs, evaluated using root mean square error (RMSE) and the coefficient of determination metrics [136].

Q-learning-based peak load reduction (QL-PLR) uses RL to present the optimal residential energy management (REM), which can decline only peak load demand and is associated with the dissatisfaction of consumers [137]. The designation of a dueling deep Q network (DDQN) captures the real-time state of the grid, demand-side strategy for interruptible loads (IL), along with the safe limit of regulating the system voltage and a reduction in peak demand loads, and the operation cost of distribution system operators (DSO). It is noteworthy that deep reinforcement learning (DRL) is adopted with DDQNs to optimize the DR management of IL under the TOU tariff and variable patterns of electricity consumption. To obtain the long-term profit for the DR management problem of IL, the Markov decision process (MDP) was formulated and solved using a DDQN-based DRL algorithm [133]. In addition, in comparison to a rule-based control (RBC) policy, DRL may also be utilized to control thermal storage in commercial buildings and can lower system operating costs by more than 50% [133]^[1].

MPC has a high level of efficiency in a building’s energy management. On the other hand, finding a good control-oriented model for MPC is a difficult task. To address this issue, data-driven models are applied to MPC tasks that have universal approximation capabilities [138]. To improve a building’s performance and take advantage of the TES’s operating flexibility, an MPC technique was devised. The findings revealed that having a TES in a commercial building allows for more flexibility in participating in DR programs, resulting in lower energy costs and demand charges while maintaining occupants’ comfort [139]. In the presence of an ESS, MPC was utilized to control the functioning of a commercial building in DR programs for bi-directional power flow with the main grid. The results of this study demonstrated that by exploiting the flexibility of the HVAC system, MPC resulted in a reduction in building running costs as well as an improvement in power grid performance [140].

In designing low-energy buildings, mathematical optimization can be used as a powerful tool to minimize the consumption of energy, while continuous and discrete issues are the two primary types of optimization challenges [141,142]. These problems can be formulated as binary, integer, or mixed integer optimization problems. Optimization algorithms are applied to solve these types of problems in various engineering fields [143].

A brief summary of the most selected optimization algorithms applied to BEMS problems is as follows:

ACO was motivated by observations of ant behavior [144]. This approach was originally created to handle discrete optimization problems before being expanded to include continuous variables. ACO for the continuous domain is the name of the extension (ACOR). ACOR showed better performance in finding the best solution as compared to other benchmark algorithms: particle swarm optimization with inertia weight (PSOIW), hybrid particle swarm optimization and Hooke–Jeeves (PSO-HJ), and the Nelder–Mead (NM) algorithm [142]. An ACO load scheduling strategy for a smart home was proposed. The goal was to achieve the best possible use of integrated renewable energy sources. This is accomplished by concentrating on the total electricity bill, TOU, and the overall improved quality of life (QoL) [145].

ABC is an optimization algorithm based on honey bee foraging behavior [146]. In [147], a HEMS for household appliances was proposed by implementing DR schemes for residential consumers with facilitating renewable energy integration. This framework was solved based on an improved ABC algorithm. In [148], the authors proposed a new approach of ABC with Knowledge Base (ABC-KB) for the management of power and the occupant’s preferred environment inside a residential building. ABC-KB uses less power than GA and PSO.

PSO is an optimization technique based on the social behavior of bird flocking or fish schooling [149]. The conventional continuous PSO algorithm was modified to binary spaces, while BPSO is a binary variant of PSO. It is a bird-inspired optimization technique based on flocks of birds looking for food. Birds move in certain locations and velocities when foraging for food [150]. In [151], they adopted and developed cooperative particle swarm optimization (CPSO) to optimize user comfort and the electricity bills of individual homes as well as avoiding peak loads and peak rebounds on the grid. In [152], they provided a regularized PSO algorithm for optimally controlling battery energy in a grid-connected microgrid, lowering power costs.

GA is an iterative optimization technique inspired by live creatures’ genetic processes. New genes are created that inherit the characteristics of their parents [153]. Chromosome representation and algorithmic flows are the two main components of a GA. An algorithmic flow is an iterative technique for generating and evolutionarily selecting chromosomes to obtain high-quality solutions, while a chromosomal representation is a scheme for modeling a solution [154]. GA was used to optimize the scheduling of ESS and plug-in electric vehicle (PEV) operations in a home energy management strategy in order to reduce daily electrical energy expenditure for the user [155]. Residential load management solutions may necessitate appliance scheduling to achieve specific goals such as load factor reduction, a PAR ratio reduction, or energy cost reduction. This problem is solved using GA [156].

In [157], the proposed hybrid algorithms have better performance and faster convergence compared to a single-based algorithm. The hybrid-based algorithms such as bacterial foraging optimization (BFO), gray wolf optimization (GWO), wind-driven optimization (WDO), enhanced differential evolution (EDE), and the harmonic search (HS) algorithm can solve the DSM optimization problems in SG. A detailed summary of other optimization algorithms is illustrated in Table 1.

A single subject owns a number of the neighborhood’s buildings in the single-owner scenario. The architecture for energy optimization is centralized in design. The neighborhood buildings are owned by various subjects in the multi-owner scenario, each of whom seeks to reduce their individual energy costs. The building owners additionally consent to provide flexibility to the entire neighborhood and run their own local generation and storage units in a coordinated manner to achieve neighborhood-level goals. The architecture for energy optimization is hierarchical. A centralized optimization for a single-owner neighborhood with a high level of transparency and a hierarchical two-level optimization for a neighborhood with multiple owners and a reduced level of transparency are two separate optimization algorithms. Additionally, in order to lower the neighborhood net load as viewed from the grid side and maintain the neighborhood pollution emissions below a predetermined threshold, the neighborhood energy optimization algorithms schedule the generation and storage equipment based on energy prices. It has been demonstrated that the use of flexible resources, such as thermal storage (related to thermal comfort levels) or electrical storage, enables one to pursue an economic goal while utilizing the flexibility of the local energy supply.

Building optimization problems are considered MILP problems that have been solved using MPC [117]. Intelligent controlling has been used to manage loads efficiently and an ANN strategy has been adopted to maintain a comfort zone in the building for the occupant, with an MILP scheduling technique to decay the peak demand of consumer [166]. An energy management agent (EMA) consists of an ANN and MPC for the modeling and optimization of building flexibility. The Monte Carlo Tree search-based planning and control was used to find the optimal policies with an ANN. The system can predict the demand for a day ahead and has a tiny prediction error [167]. The proposed system consisted of an ANN and MPC. In contrast, an ANN was used for renewable energy (i.e., solar and wind) forecasting and ensured the optimized usages of generated energy, and MPC is adopted for intelligent home control [168].

The ANN is used to accurately predict power consumption and indoor temperature selection by given weather, occupancy, and temperature setpoints as input. At the same time, a GA has been taken to adapt to the ANN to minimize energy consumption, and an optimization control strategy was assessed in case of the day ahead and MPC [169].

The uncertainty of environmental variables and users’ preferences has been tackled using a data-driven machine-learning approach. Furthermore, a lifelong multitask framework was adopted to exploit structural similarities in control policies as there were different room sizes in buildings. Kernel-based learning was pursued as well to mitigate the non-linearity policy. Finally, a dual decomposition method was employed to cope with DR constraints across the spaces, transforming the overall problem into a series of unconstrained stochastic optimization problems for individual rooms. The method was verified via numerical experimental based on semi-real data sets [170].

Three artificial intelligence techniques were used to solve the problem of energy demand planning in smart homes. First, the modification of the elitist none-dominated sorting genetic algorithm II to demand-side management was applied and accounted for electricity fluctuations over time, priority in the use of equipment, operating cycles, and a battery bank. Second, the forecast of demand-side consumers, distribution generation, renewable energies, and weather for a day ahead from the nearest meteorology office was considered for demand-side management by employing the SVR technique. Third, the k determined user comfort levels through the cluster technique [171]. Table 2 shows a full explanation of each optimization control technique, along with its benefits and drawbacks.

Neighborhoods or districts are not frequently included in the application of optimum control ideas to achieve energy efficiency in buildings and the optimal exploitation of regional resources. In [171]^[2], they consider a neighborhood with several buildings that have agreed to coordinate how they use their energy loads and resources in order to achieve some overall objectives, while still allowing for the pursuit of individual optimization goals. When this occurs, a building’s local resources should work together with a top-level optimization engine to balance the accomplishment of local optimization goals with neighborhood-level goals. A hierarchical optimization algorithm was introduced to divide the optimization at the building level and the neighborhood level in such a way that the bottom level managed the individual building objectives and the top level addressed the neighborhood-level objective, in order to address the most general case of multiple ownership neighborhoods. In particular, the building-level energy management would give the top-level optimizer flexibility and provide recommendations on control measures to implement to move the neighborhood closer to achieving its objectives [172]^[2]. The optimistic assumption in bilevel optimization (i.e., the two layers of optimization tasks are nested one inside the other) states that the consumers choose the best option that benefits the retailer the most. The pessimistic variant, on the other hand, deals with the scenario in which the consumers give the retailer their least preferred optimal response, protecting against potential losses brought on by an unexpected choice. However, the work in [173]^[3] shows the formal properties of the best solutions to a bilevel tariff optimization issue for both the computationally challenging general case with an arbitrary number of consumers and the particular case with an easily tractable single consumer. The key relevance of these findings is that by perturbing the issue data and the optimal price vector, the pessimistic variant may be reduced to the optimistic one, which also yields the first effective solution technique for the pessimistic version. On the other hand, a numerical case study was offered to show that, if consumers do not select their optimal solution as expected, addressing the optimistic problem could directly result in a significant loss of profit for the retailer. In [174]^[4], they emphasize the distinction between the optimistic and pessimistic versions of the bilevel optimization problem with regard to energy management.