Low-frequency datasets are more available to the public due to the lower cost and easier installation. The sampling rate and features vary from one dataset to another where the sampling rate in the datasets is between 0.2 Hz and 15 Hz. Smart [33], Indian ambient water and electricity sensing (iAWE) [34], REFIT [35], GREEND [36], office plug load dataset (OPLD) [37], rainforest automation energy (RAE) [38], electricity consumption and occupancy (ECO) [39], and Dataport [40] datasets measured the electrical consumption at resolution of 1 Hz. Whereas, Tracebase [41] and appliance consumption signature database (ACS-F1) [42] datasets recorded the consumption of the appliance at a sampling rate of 10 seconds as well commercial building energy dataset (COMBED) [43] dataset collected data at 20 seconds. There are datasets having resolution of 1 minutes such as almanac of minutely power dataset (AMPDs2) [44], Chulalongkorn university building energy management system (CU-BEMS) [45], Indian buildings energy consumption dataset (I-BLEND) [46], electrical-end-use dataset (EEUD) [47], and Dutch residential energy dataset (DRED) [48]. In addition, ENERTALK [49] was recorded at a sampling rate of 15 Hz.

Researchers tend to use a high sampling rate to study the patterns of loads and several features require a high frequency to extract them. High-frequency datasets are less available due to the high cost of installation. One of the datasets that is used by the public is reference energy disaggregation dataset (REDD) [50] which captured the signals of loads at 15 kHz as well can be used for a low rate at 1 Hz. Building-level fully labelled dataset for electricity disaggregation (BLUED) [51] contains the measurement recordings at 12 kHz. United Kingdom domestic appliance-level electricity dataset (UK-DALE) [52] and controlled on/off loads library (COOLL) dataset [53] measured the electricity consumption at 16 kHz and 100 kHz, respectively. Whereas, worldwide household and industry transient energy dataset (WHITED) [54] recorded the signatures of the total consumption at 44.1 kHz. Building level office environment dataset (BLOND) [55] collected the measurements in two categories, namely BLOND-50 had resolution at 50 kHz and BLOND-250 measured the electricity consumption at 250 kHz. HELD1 [56], LIT [57], and plug-level appliance identification dataset (PLAID) [58] also used a high sampling rate to record the energy usage at 4 kHz, 15 kHz, and 30 kHz respectively.

Running campaigns for actual datasets is a costly and time-consuming operation that requires a large amount of sensor infrastructure, money, and effort. Furthermore, training the network for recognising the appliances and running state-of-the-art algorithms of energy disaggregation and load identification is not possible without having precise consumption data. To address this problem, synthetic datasets are utilized to achieve cost savings and performance parity [32]. 5 synthetic datasets are published to the public including SmartSim [59], AMBAL [60], SHED [61], SynD [62], and NILM-Synth [63].

SmartSim [59] measured the appliances for one household at 1 Hz. Furthermore, SmartSim is integrated with NILM-TK, a publicly accessible NILM toolkit, and its synthetically produced traces are compared to traces from a real house to demonstrate that they give comparable quantitative and qualitative findings for representative energy analytics.

Automated model builder for appliance loads (AMBAL) [60] dataset contains a signal at a resolution of 1 Hz. AMBAL enables the generation of models using real-world device power consumption data acquired through smart plugs. Models are modelled as sequences of parametrized signatures, and complexity is kept low while achieving the required precision. Furthermore, a synthetic appliance trace generator is described, which allows for the recombination of appliance models to mimic user actions in houses of varying complexity. Simulated high-frequency energy disaggregation (SHED) dataset [61] consists of 8 buildings. SHED dataset covers the overall current consumption as well as specific consumption according to various categories for each building. Individual consumptions for buildings 1–6 are comprised of low-frequency power measurements, whereas those for buildings 7–8 are comprised of high-frequency current data. Every 30 seconds, one current waveform is captured, and 200 points are sampled for each current waveform. Power data are also taken at a frequency of 1/30 Hz.

Synthetic energy dataset (SynD) [62] provides a simulated home that is used for a total of 21 appliances for 180 days. Measurements are taken from the aggregate and the submeters at a rate of 5 Hz. NILM-Synth [63] was recently published, as a synthetic dataset-generating tool that can generate complicated, realistic power waveforms that are ideal for training, testing, and benchmarking NILM algorithms. The program and load library are natively compatible with major numeric computing applications, such as NILMTK, a popular software framework for developing NILM algorithms. Table 1 shows the common public datasets which are accessible to the public and a comparison between the datasets, highlighting their main characteristics, such as region, sample resolution, duration, features and the number of sites.

Steady-state signatures are the load features extracted when the appliance is in its steady-state operation [8]. Because the fluctuations in signatures may be at their smallest during the steady-state operation of the appliance, steady-state features may not need a large number of sampled signals [24]. Sadeghianpourhamami et al. [83] proposed a review of feature groups and identified the efficient feature group by execution of methodical feature elimination. The fluctuations in patterns are likely to be at their lowest during the device's steady-state operation. As a result, the characteristics of a steady-state do not need the use of a high sampling rate. Moreover, this is to be observed that several appliance signatures can be overlapped due to their comparable characteristics.

Brito et al. [84] demonstrated a new load signature based on instantaneous real power-reactive power (P-Q). The findings demonstrate that instantaneous P-Q produces more graphically distinct patterns than traditional load characteristics, particularly when dealing with low-power loads, whereas the latter commonly overlap. The major disadvantage of this technique is that appliances with comparable power draw characteristics overlap in the P-Q signatures, making it impossible to discriminate between them only if it is based on power step differences. To address this challenge, V-I trajectories are applied to discriminate between appliance events and other appliances events which gives a unique shape for each load more than the active power feature [85]. Angle, length, maximum distance, variations, average slope, Trajectory Centroid, and overshoot are all characteristics determined by the form of the V-I trajectory [86]. Kholeif et al. [87] investigated the effect of changing the voltage on the features extracted in the V-I trajectory method. The findings showed that apparent power, current, and harmonics are affected by voltage variation. Iksan et al. [88] expand and evaluate appliance load signatures based on a hybrid method that uses the V-I trajectory. The identification findings not only identify between equipment that is resistive, inductive, and capacitive, but they can also distinguish between equipment that is linear/non-linear and electrical appliances that are low in power.

Further, the frequency domain exemplification of signals and features can be detected. Harmonic features are used for the recognition in contrast to the time domain of device-specific current harmonic components that efficiently classify device events [24]. In addition to having excellent anti-interference capabilities, the steady-state characteristics are useful in identifying high-power consumption equipment such as electric heaters. In the case of appliances with Type II, Type III, and Type IV loads, however, focusing only on the steady-state makes the identification of appliances with these types of loads more challenging [89]. Steady-state approaches are widely used due to their low cost and ease of obtaining power parameters such as active power, reactive power, apparent power, and power factor. After all, it is highly limited due to overlapping appliance signatures [17]. Table 4 lists the summary of features extraction based on steady-state including real power, reactive power, current harmonics, and V-I trajectory while discussing the advantages and disadvantages.

The features obtained during the transient state would disclose unique information about the device, which would assist in the more accurate identification of the events. Electrical signals collected at a higher frequency are required to capture the characteristics of transient states (in kHz). So, the hardware module for extracting transient characteristics from high-sampled input signals should be able to process high-sampled input signals [24]. The process of high sampling rate needs a NILM system to enable the identification of transient features. The most used analyses in the transient state are transient V-I trajectories, Inrush current, higher-order harmonics, and transient power [24], [86].

It has been shown that transient real power and reactive power signatures may be utilized to monitor the functioning statuses of electrical appliances via the process of transient identification. This combination of characteristics is particularly well suited for identifying appliances that have very lengthy transient time, generate considerable spikes in their power use, and then exhibit slower changing variations. The edges and slopes of these transients or their power profiles are used to identify them [83]. Le et al. [98] analysed and extracted a new transient signal from the active power signal based on a low sampling rate. The state of the appliance is to obtain a transient signal to detect ON/OFF events of appliances. Athanasiadis et al. [99] proposed the turning-on of a target appliance by processing the measured active power transient response and estimating its consumption in real-time. The NILM system can identify in real-time when an appliance is turned on based on its information-rich transient response sampled at 100 Hz.

Because voltage has a significant degree of resemblance to other electrical devices, it is difficult to differentiate between them efficiently. It is possible to solve this limitation by using a harmonic analysis based on the Fourier transform to perform feature extraction, which may be used to reflect the variations between various equipment types. The Fourier transform technique and the Wavelet transform method are two of the most often used harmonic detection methods [100], [101]. Drouaz et al. [102] demonstrated new transient features based on the time-frequency domain by using the Stockwell transform to generate the shape of the frequency harmonic voices.

Hasan et al. [103] investigated the unique shape of a current at the starting phase (turn-on state) of a device. The adoption of a shapelet-based classification technique is time-consuming and may lead to computing difficulties. Nevertheless, by offering both analytical premises and experimental data, this study establishes simple computationally cost-efficient, and robust shapelet discovery and matching criteria that are both simple and resilient. Table 5 lists the summary of features extraction based on transient state including transient real power, transient reactive power, shape, size, and duration as well as the list of the advantages and disadvantages.

Non-traditional features have been of considerable interest in feature extraction approaches in recent times. This category contains all the characteristics that distinguish it from the two conventional groups. providing details of the detected appliance by using the environmental quantities such as starting and ending time, day of the week, temperature, and sensed light [24], [104]. Senave et al. [105] investigate the characterization of the heat loss coefficient of homes based on onboard monitoring data of their interior temperature and energy consumption, which were gathered throughout the heating season. Despite there is lesser trend in using non-traditional features, it can provide high performance for the NILM system, therefore needs to be considered.

Most of the researchers concentrate on machine learning to investigate the performance of load classification. Machine learning algorithms in NILM are divided into two categories; supervised learning and unsupervised learning.

In machine learning, supervised learning techniques are algorithms that learn to map input data to a target data attribute, which is usually a class label. To be able to identify the appliances that are operating at any given moment, supervised learning approaches need the training of individual appliance data [108]. Supervised learning techniques contain several approaches based on artificial neural networks (ANN) [109], [110], support vector machines (SVM) [95], k-nearest neighbours (k-NN) [111], [112], Hidden Markov models (HMM) [75], [113], genetic algorithms (GA), decision tree (DT) [114], random forest (RF) [115], naïve Bayes (NB) [116]. To illustrate, HMM can be used either supervised or unsupervised learning, whereas the genetic algorithm can solve optimization problems in machine learning. In supervised techniques, the labelling process involves complete monitoring of the loads to be detected. The training takes extracted events to obtain the appliance's signatures.

Figure 8.

Load disaggregation techniques are based on machine learning, optimization, deep learning, and graph signal processing

Recently, Gao et al. [117] suggested frequency spectrum in combination with time-domain analysis to identify loads based on SVM and particle swarm optimization (PSO). Kelati et al. [118] have proven that the accuracy of a k-NN classifier is dependent on the quantity of the training datasets for the appliances under consideration. Yang et al. [116] demonstrated that by using a small number of training samples with generic NB and the average power, it is possible to obtain good classification accuracy. The NB algorithm is preferable to be applied since it just takes a little quantity of the training data to forecast the class of the appliance. Himeur et al. [119] presented an efficient NILM framework that comprises a unique fusion of various time-domain characteristics offered to extract appliance patterns from a large number of data sources. In addition, Choksi et al. [120] have introduced a novel approach based on a power pattern matrix and DT for properly recognizing the patterns of household appliances. This is further affirmed by Nalmpantis et al. [104] who proposed a comprehensive analysis of machine learning approaches from qualitative to quantitative. Therefore, neural networks have shown that they are capable of meeting the condition of generalization with a high degree of confidence.

Algorithms of supervised learning can effectively classify the loads due to the training that they need and the labelling of the data. However, several limitations include intense focus on the training time and the unseen data. Supervised methods may take a longer time to train based on the size of the data as well as failure to identify the unseen data which needs to be updated when there is a new appliance or changes in the power pattern.

Unsupervised methods do not require any training before classification. This is an important advantage since, in this way, minimum effort is required from the user and the intrusiveness involved in building a database is reduced [121]. Furthermore, the unsupervised approaches are not reliant on the number of appliances that make up the aggregated load and are capable of dynamically adapting to changes in the power system over time [122]. Moreover, researchers have used unsupervised learning approaches in NILM systems such as HMM [113], k-means clustering [123], and Gaussian mixture models (GMM).

The use of HMM is to recognize the load state and power consumed when the hidden variables can be estimated (appliance states). The transition of the matrix can be used to observe appliance state changes. Instead of taking into account the appliance's actual functioning states, traditional HMM-based models instead rely only on basic empirical data. To solve this problem, a new model suggested by Wang et al. [124] is called the Adaptive Factorial HMM model, and it provides for a more accurate search for the operating condition of each appliance and subsequent accurate modelling using an adaptive search. A new infinite Factorial Hidden Markov Model constrained on Contextual features (iFHMMCC) has been proposed by Hasan et al. [103]. Furthermore, it is utilized to reduce the computational cost of inference and makes the model more tractable. Mostafavi et al. [125] proposed a new approach that can remove the need for human training and potentially expedite the adoption of low-cost power monitoring based on Bayesian networks. The adaptive Density Peak Clustering-Factorial Hidden Markov Model (ADPC-FHMM) is a model that Wu et al. [126] suggested to automatically detect the functioning condition of any appliance based on its power consumption. The findings show that the suggested model is indeed effective and superior. Wu et al. [127] also presented Time-Efficient Factorial Hidden Semi-Markov Model (TE-FHSMM) to boost computing performance. Two publicly accessible datasets and a real-world scenario indicate that the suggested model delivers excellent results in terms of both computing performance and efficiency. Ponrak et al. [128] proposed method uses a semi-supervised learning approach, with improved K-Means and Hierarchical methods. It involves determining the best number of clusters, selecting features, and managing the data. The findings present the suggested technique that precisely displays the energy consumption of each electric device. Papageorgiou et al. [129] presented an unsupervised NILM approach under the consideration of odd harmonic currents. To create unlabelled data clusters, the suggested technique first uses a k-means clustering algorithm. The findings show that when the harmonic content is taken into account, the F-Score index improves considerably for all analysed combinations.

Algorithms of unsupervised learning can learn without any labels of data which reduces the time for labelling manually and can be inserted after the classification stage. Moreover, they can reduce the dimensionality of the appliances' patterns. On the other hand, the classification of loads can be less accurate due to no training as well as may take a longer time to obtain the whole possibilities based on the number of features.

Optimization approaches address the load disaggregation challenge and optimize the problem. Recognizing load by this method matches the identified unknown appliance feature with the known appliance feature in the database and attempting to minimize the error of mismatch as well as detecting the nearest similarity. Researchers have proposed various optimization methods including aided linear integer programming (ALIP), mixed-integer linear programming (MILP), integer quadratic constraint programming (IQCP), integer programming (IP) [130], and combinatorial optimization methods which include artificial bee colony (ABC), PSO, and GA [131].

In recent works, Dash et al. [130] presented a two-stage NILM technique that is both simple and effective. It is presented with a standard deviation-based automated state identification technique (stage 1) and a penalty-based improved IP-based load disaggregation approach (stage 2). Several test results show that the suggested improved IP-based technique is an effective and less complicated solution to the load disaggregation issue. However, concerns with transient signals and circumstances where certain states are binary combinations of other states have been demonstrated to cause problems with IP-based disaggregation on actual data. To overcome this problem, Bhotto et al. [132] presented an ALIP disaggregation method that addresses many of the drawbacks of the IP-based approach. The experimental findings show that ALIP has a considerable accuracy improvement over the IP-based disaggregation approach. Wittmann et al. [133] introduced the mixed-integer linear constraints (MILC) technique to effectively predict the behaviour of complicated load signatures. The suggested technique is also evaluated in the comparison to IP and ALIP. The results show that the suggested technique is capable of reliably identifying and allocating individual energy signatures in a computationally efficient way, as well as outperforming existing methods. Kong et al. [134] presented NILM system based on segmented IQCP. The findings show that the method may be used with current smart meters to build a device-level load model for future smart grid research and applications.

In recent works based on combinatorial optimization methods, Lin et al. [135] demonstrated the PSO-based NILM technique and verified it in an actual housing situation. According to the experimental findings of the suggested technique, the PSO-based NILM presented in this study has a load recognition accuracy of 64.06%. The accuracy of the model considers low where state-of-the-art NILM models achieved from 85% to over 95% in disaggregating household energy consumption as shown in Table 6. The accuracy can vary widely depending on the specific dataset, the complexity of the household, the types of appliances present, the quality of the data, and the techniques used for load disaggregation. Guo et al. [136] suggested a novel combinatorial optimization-based load disaggregation approach. First, the PSO algorithm is enhanced by including temporal probability in the fitness function. Second, investigate the multi-model combination load disaggregation approach, which parallelizes the enhanced PSO, IP algorithm, and delta feature (DF) model. The experimental findings indicate that the approach performs well for load disaggregation. To increase the accuracy of load classification, Ghosh et al. [137] suggested the ABC optimization technique by reducing the computational complexity of the processor which requires less training of appliance profiles. The suggested approach provides a higher rate of classification accuracy than the current techniques in use.

Optimization approaches can highly achieve identity precision and able to solve the problems of machine learning. However, the limitations of optimization methods are costly to be computed, also it difficult to discern the loads with the same or overlapping signatures, mostly if the database does not extract anonymous appliances will tend to have a lower trend for research. Moreover, the larger computational overhead of the disaggregation method may result in inaccurate findings, and delayed operation.

Analytical tools such as GSP are useful for visualizing complicated correlations between data samples [138]. GSP has been commonly applied, in both supervised and unsupervised learning. Unsupervised GSP involves filtering and clustering to disaggregate and classify loads without any previous knowledge of specific load shapes, while supervised GSP is trained using labelled samples of individual appliance data [139].

Recently, Bochao Zhao et al. [140] applied semi-supervised GSP to tackle the difficult challenge of reducing the influence of measurement noise and unknown loads on load disaggregation performance. As a result, the method improved NILM performance when applied to three events based on NILM approaches and two real-world datasets with multiple residences and varied noise levels. The disaggregation of water heaters and HVAC system components have been investigated by Green et al. [139]. A supervised GSP method was used; and the findings demonstrate that the values of disaggregation sequencing, graph scaling factor, and graph classifier threshold have a substantial influence on supervised GSP NILM performance. He et al. [141] presented two NILM algorithms based on the emerging concept of GSP. In the first place, it reduces the overall variance of the graph. Second, simulated annealing is used to further fine-tune the entire graph-variation solution. The suggested approach, which does not need any extra hardware installation, might operate with traditional smart meters, such as recording data every 10 seconds through the Consumer access device. Consumer access devices empower individuals to take control of their energy consumption, reduce wastage, and potentially save on energy costs as well obtaining information on power consumption in several seconds to the consumer. Zhang et al. [138] showed a two-step GSP-based method proposed for low-rate NILM problems. The suggested technique treats disaggregation as an optimization and classification problem with a regularisation term defined by the overall smoothness of the graph signal. The disaggregation findings based on the REDD dataset indicate the exceptional quality of the proposed method.

Using supervised techniques can outperform unsupervised techniques due to the provided samples of complicated operational cycles to the disaggregation algorithm, which may not be taken into account in unsupervised methods. Furthermore, supervised techniques enable the algorithm to be tailored to a specific set of load types, allowing for higher precision in classification [139].

Deep learning is a machine learning concept based on artificial neural networks. For many applications, deep learning models outperform shallow machine learning models and traditional data analysis approaches [142]. In NILM, researchers have found some deep learning approaches such as convolutional neural network (CNN) [143], autoencoder (AE) [144], recurrent neural network (RNN) [145], and long-short term memory (LSTM) to improve load disaggregation by extracting features automatically [146], [147]. Wang et al. [148] proposed reconstructing the load signals from collected data and load identification by using denoising autoencoder (dAE) and LSTM. The approach dAE and LSTM outperformed HMM and dAE. However, LSTM was able to store and bring out the information through the data of time sequence. With the use of high-dimensional sparsity hourly energy data, Jia et al. [149] have developed a tree-structured CNN classifier for the study of temporal patterns. Osathanunkul [150] investigated if decreasing the sample rate would shorten the time required to disaggregate an aggregated reading while still producing an acceptable result as a disaggregation algorithm, RNN with LSTM is used. Load data for small office appliances such as laptops, computer monitors, etc. have been proposed by [151]. It is impossible to overlook the potential energy savings that each tiny item has, especially in major enterprises and institutions. It is shown that the RNN that has been presented obtains superior outcomes. Zehua et al. [152] showed a new method of NILM based on CNN. Although CNN-based load classification requires a huge quantity of data to properly fit the network, the benefit is significant. Chen et al. [153] suggested a new NILM model, which is a hybrid model combining the CNN and LSTM networks. The CNN network is used to extract the V–I image characteristics, while the LSTM network is used to recover the electrical equipment time characteristics. Tests show that the model is better than the k-NN approach and the SVM, CNN, and LSTM methods.

Recently, researchers have focused on transferability to decrease data gathering, which is a key strategy for addressing the issue of inadequate deep learning training data. This method seeks to transfer knowledge from the source domain to the target domain by assuming that the training and test data must be distributed independently and identically [154]. A V-I trajectory-enabled transfer learning approach was presented by Liu et al. [155]. Different electric loads may be distinguished by the shape and colour information of their V-I trajectories using correct trajectory pre-processing and colour encoding methods by transferring a convolutional neural network called AlexNet that has already been trained to the new NILM tasks. However, the technique can detect type-I loads, but type-II loads remain a difficult challenge. D’Incecco et al. [156] introduced seq2point learning for NILM, which is based on appliance transfer learning and cross-domain transfer learning. The findings demonstrated that CNN layers trained on washing machines may be applied to other appliances. This suggests that all appliances utilize the same signatures for NILM, which might simplify the data collection process. The federated learning and TL that Li et al. [157] presented maintains high performance while protecting the privacy of the users. The findings are similar to state-of-the-art NILM algorithms with robust privacy protection, which is useful for deploying the distributed algorithm to home Internet of Things devices.

The approaches based on deep learning algorithms may consider several factors in order to have an efficient model where the size of data depends on the similarity of loads and number of appliances which require a huge data to identify the loads easily. Large data may not lead to an increase in the accuracy of the model until it reaches a steady point. However, normalizing data can improve neural networks and be more accurate. Table 6 demonstrates the summary of recent NILM models. The relation between the features and the classifiers is described with various approaches in the remarks. F1-score or F1-measure is calculated by obtaining the data test's precision and recall measuring the test's accuracy and to demonstrate the model's performance as shown in the table below.