**1. Introduction**

Demand Response (DR) provides an opportunity for consumers to play a significant role in the operation of the electric grid by reducing or shifting their electricity usage from peak time to off peak and/or altering their usage pattern in response to time-based tariffs or other forms of financial incentives to improve Power Quality (PQ). DR also plays an important role in a smart grid in helping consumers plan their consumption pattern and optimize electricity usage without compromising on their needs [1–3], and is made possible through (i) identification of unnecessary consumption of electricity at an individual appliance level, (ii) alerting consumers with timely information that helps to balance the load between appliances and (iii) leading to reduced bills. DR is approached through the following four steps:


#### *1.1. Identification of Load Features*

In the past 30 years, researchers have tried several diverse electrical and non-electrical features to uniquely identify all types of home appliances with different operation modes [4]. Event-based techniques have been employed to identify turn-on and turn-off events.

The identification of electrical appliances is possible through their electrical behavior (active power (P) reactive power (Q), voltage (V), current (I), harmonics, power factor (pf), phase angle, etc.). Moreover, PQ and VI trajectories were used as described in [5–7]. These parameters can be at a steady state or transient (turned on). There is some research using non-electrical behaviors such as light emitted (lumens), heat generated (joules), vibration and sound (noise), EMF (Electromagntic Field) produced, etc. If we employ non-electrical behavior to recognize appliances, then we need respective sensors to collect data for analysis. While this method can be considered as non-intrusive from an electricity point of view, it is an expensive proposition and may be invasive from a personal privacy point of view.

A detailed literature review was presented by Antonio Ruano et al. in [4]. As per the review article, the observations of various load features for disaggregation are 1. active power and reactive power, 2. voltage, current and fundamental phase angle, 3. average displacement power factor, 4. VI trajectory and current vectors, 5. steady-state current transients, 6. PQ disturbance trajectories, 7. third and fifth harmonic amplitudes, 8. average Total Harmonic Distortion (THD), 9. the maximum magnitude of the first eight harmonics 10. The rate of change of transient signals, 11. Shannon and Reini entropies, 12. the spectral band energy of the current spectrum, 13. wavelet transform coefficients, 14. features of power spectrum density, 15. occupancy data, 16. usage patterns and 17. Electro-Magnetic Interference (EMI) signals.

It is quite evident that, initially, researchers tried the active power of the load and its steady-state active power (P) features that are readily available in an energy meter for load disaggregation, which requires a low sampling rate at the time of data acquisition. Later, it was found that, due to the additive nature of active power, it is only useful and deterministic when the power ratings of appliances are distinct, and the sum of the power ratings of various combinations of appliances in operation is also distinct. Therefore, the active power feature is suitable mostly for high-power appliances and is not helpful to discern the simultaneous operation of appliances with the same ratings. Later, researchers employed transient power features [8] (ON slope at the start and step changes) to detect appliance status at a higher sampling frequency (high to very high). Load features such as steady-state power characteristics versus transient power characteristics, capturing events like turn-on step changes, are quite cumbersome. Some researchers used multiple harmonics [9] and the harmonic phase [10,11], since they are induced by nonlinear loads, which are small in amplitude.

#### *1.2. Load Disaggregation*

Once the features are decided as discussed above, they are used to discern the start time and duration of operation of the appliances. There have been several attempts by various researchers to disaggregate loads using multiple electrical and non-electrical features [4]. Usually, a power supplier provides a single energy meter outside the premises and records the cumulative energy consumption since its installation. Periodic billing (bill cycles) is calculated by subtracting the current reading from the reading taken at the previous bill cycle and applying a tariff. This procedure does not allow the consumer (i) to realize opportunities for saving energy at an appliance level and (ii) to prevent unnecessary consumption before it happens. The most effective method for load disaggregation is to meter the power at every appliance, termed appliance-level monitoring (ALM) [12]. G.W. Hart [13] recognized that ALM is highly prohibitive due to (a) the high number of appliances in use, (b) the various types of appliances and (c) inconsistent times and durations of operation of appliances at home. He proposed, for the first time, in 1992 that there is a need for identifying individual loads without sub-metering [13,14] and termed this Non-Intrusive Appliance Load Monitoring (NIALM), which is now referred to as Non-Intrusive Load Monitoring (NILM).

There are various ways to discern appliances using digital signal processing (DSP), wavelet transform (WT) and artificial intelligence (AI). Recently, machine learning (ML) [15] and deep learning (DL) techniques have been trialed for this purpose. In particular, two types of ML, namely supervised learning (e.g., classification) and unsupervised learning (e.g., clustering), have been widely used for NILM [16–24]. Load features and the selected disaggregation technique decide the sampling rate for data acquisition for the desired accuracy.

A concise and updated review of the various features reported in the literature for NILM and a comprehensive feature selection from a benchmarked dataset are reported in [25]. A multi-objective evolutionary algorithm is proposed in [26], where five objective functions using active power, apparent power, reactive power, current waveform, and harmonics as load signatures are established to identify several electrical appliances. Antonio Ruano et al. [4] carried out an exhaustive review of disaggregation approaches, including the most recent ones, namely machine learning and deep learning, and they are 1. the Hidden Markov Model (HMM), 2. Deep Neural Networks (D-NN), 3. k-NN (K-nearest neighbors) classifiers, 4. Naïve Bayes Classifiers, 5. Tensor and Matrix Factorization, 6. NN auto encoders, 7. Graphic Signal Processing, 8. Multi-Layer Perceptron, 9. linear searches of databases, 10. Maximum a Posteriori probability, 11. fuzzy "C" means, 12. Discriminative Disaggregation Sparse Coding, 13. The Factorial Hidden Markov Model, 14. hierarchical HMM, 15. variants of HMM, 16. Viterbi Decoding, 17. density-based spatial clustering of applications with noise, 18. quadratic discriminant analysis, 19. Modified Combinatorial Optimization, 20. Long- and Short-Term Memory Recurrent Neural Networks, 21. Karhunen–Loève Spectral Decomposition, 22. iterative subsequence dynamic time warping, 23. particle filtering, 24. Maximum a Posteriori (MAP) criteria, 25. Gaussian Process Classifiers, 26. rule-based classifiers, 27. decision trees, 28. Adaboost classifiers, 29. several supervised classifiers, 30. Self-Optimizing Mapping (SOM), 31. Particle Swarm Optimization, 32. Ant Colony Optimization, 33. Siamese Artificial NN, 34. PCA (Principal component analysis), 35. location-aware energy disaggregation frameworks, and microscopic power features and pattern recognition (reported for NILM in [27]).

It is noted that all of these methods are non-deterministic and hence non-repeatable, which are useful for predictions with some certainty, but not to the extent that the inferences and insights of these methods are beneficial for their deployment in production for consumers to respond in real time and realize the benefits then and there. Furthermore, these methods draw correlations, not causations, to effectively convince consumers to change their behavior.

A comparison between traditional non-deterministic NILM methods versus proposed deterministic methods is illustrated in Table 1, and gives reasons for the limitations of machine learning and deep learning techniques over the proposed deterministic experimental approach.


**Table 1.** Comparison between traditional non-deterministic Non-Intrusive Load Monitoring (NILM) methods versus proposed deterministic method.

The realization of benefits and the value of DR lie in the accurate measurement of a load feature that could be used by employing a suitable load feature that can determine, with certainty, the consumer demand and suggest alternate propositions for a better response, either from an energy conservation or PQ perspective.
