A novel approach for residential load appliance identification

Highlights

• A statistical approach for residential load identification is proposed.

• The features obtained from only current signal are used.

• The loads active on grid are identified on any time without waiting any change.

• Considerably accurate results (over 98 %) are obtained with this approach.

Abstract

The analysis of the energy consumption of a house is important for its energy management. With the expansion of smart homes, energy management of a house gained more importance. To manage this expansion, loads should be identified. In this study, a novel load appliance identification approach is proposed. This approach utilizes from only current waveform while extracting the features. In the proposed approach, firstly a data preprocessing is performed to extract one period signal from the measurement. Then Fast Fourier Transform (FFT) of the current signal is calculated and the real and imaginary parts of the transform are evaluated separately. Statistical features such as maximum, minimum and standard deviation of the real and imaginary parts are extracted. After the feature extraction procedure, the boundaries of each load appliance in terms of extracted features are determined to build a rule table and the load appliance is identified using these rules. In this study, the identification of both individual appliances and different combinations of appliances are performed. The results show that this new approach provides successful identification performance with over 98 % identification rate. Furthermore, it is demonstrated that the separately evaluation of real and imaginary parts of the Fourier transform provides around 4.7 % improvement.

Introduction

A great effort continues increasingly in the area of Smart Grids research including energy production, distribution, and consumption issues since they play a critical role in the sustainability (Jurado, Nebot, Mugica, & Avellana, 2015). Generation and consumption are two critical subjects of the energy management in a grid. In consumption side, it is crucial to identify the loads and to understand their behavior. Figuring out household electricity consumption characteristic is a key research issue for sustainable consumption (Sakah, de la Rue du Can, Diawuo, Sedzro, & Kuhn, 2019; Sun, Zhou, & Yang, 2020). Analyses in this area are necessary to acquire the detailed information about household energy efficiency, determine the household electric energy consumption structure, and understand the impact of user behaviors on their household energy efficiency (Wang, Sun, & Liu, 2015; Wu, Han, Liu, & Qi, 2018; Zhang, Zhang, Liu, & Guo, 2015). Therefore, energy planning of a residential building requires information about the used loads and this issue is starting point of this study. In studies on energy planning of a home, the loads are organized according to some parameters such as tariff used, the load characteristics, energy consumption of the load etc. to obtain better efficiency. To organize such a study, the loads in use in a certain time should be known and therefore, load identification is selected as research topic in this study. Load monitoring is an essential part of load identification. Intrusive load monitoring is an option for this purpose and in this approach separate sensors are employed for each load to be monitored. This process provides accurate measurements with smart plugs; however, there are disadvantages such as high hardware costs and additional complexity of the communication network (Lin et al., 2016). The other approach is non-intrusive load monitoring proposed by Hart in 1980s (Hart, 1992). In Non-intrusive load monitoring (NILM), an analysis is performed to determine changes in the signals coming from appliances, and their individual energy consumption are deduced without entering the building. Since the procedure does not require intruding into the house when measuring the power consumption, it is called non-intrusive (Aladesanmi & Folly, 2015). After monitoring the signal (intrusive or non-intrusive), the load used in the house should be identified. Electrical loads (appliances) generally present unique characteristic in the electrical signals such as current, voltage and power (Aladesanmi & Folly, 2015; Zoha, Gluhak, Imran, & Rajasegarar, 2012). Such features can be employed to identify the load appliance. Basically, a load identification system includes monitoring, feature extraction and identification procedures.

Several methods have been used previously in the literature for load identification. Among them C-means clustering (Guillén-García et al., 2019), Karhunen-Loève expansion (Welikala, Thelasingha et al., 2019), Maximum likelihood classifier (Henao, Agbossou, Kelouwani, Dube, & Fournier, 2015), Fuzzy logic (Welikala, Dinesh, Ekanayake, Godaliyadda, & Ekanayake, 2019), Artificial Neural Networks (Chang, Lian, Su, & Lee, 2014), Artificial Immune Algorithm (Tsai & Lin, 2012), Decision Tree (Buddhahai, Wongseree, & Rakkwamsuk, 2018), etc. can be listed as foremost ones. Machine learning techniques provide well identification performance; however, these techniques have high computational complexity. Furthermore, an increase in the number of appliances that will be identified will also cause an increase in computational complexity. Additionally, in such methods, the samples used in the training are critical for load identification performance. Unlike such methods, rule-based methods generally have less computational complexity, good identification performance and easy to implement especially in embedded systems. Although rule based methods have limitations in explaining every scenario as a rule, by considering advantages above mentioned, a rule based approach is employed in this study.

Success of a load identification algorithm depends on features, and there are several features used in different studies as summarized below. Andrean et al. (2018) use active and reactive power, harmonics, current waveform and non-active current as features where the combination of cascade filtering approach with the Committee Decision Mechanism is used as identification algorithm (Andrean, Zhao, Teshome, Huang, & Lian, 2018). Guillén-García et al. (2019) present a non-intrusive method to identify the industrial loads. Energy Power Quantities and current values are employed as features and C-means algorithm is used for classification (Guillén-García et al., 2019). Liu et al. (2018) propose an admittance-based load signature to improve the identification accuracy of the appliance with on/off states and called it Resolution-Enhanced Admittance (REA). The current and voltage signals are used as features and the aggregate of ON/OFF appliance is interpreted as the transfer function of a linear time-invariant system. Then the phase information of the transfer function is computed with fundamental and harmonic frequencies (Liu, Wang, Zhao, & Liu, 2018). Welikala et al. (2019) focus on supply voltage variability issue for Real-Time Non-Intrusive Load Monitoring (RT-NILM). To validate the method, the active power and voltage measurements collected with 1 Hz sampling rate from a real house are used and over 92.5 % accuracy is obtained. The method provides good accuracy however an extended learning period is required on capturing power and voltage variations for appliances showing load dependent characteristics (Welikala, Thelasingha et al., 2019). Wu et al. (2019) uses an algorithm based on I-V characteristics of high-frequency data. They utilize from the principle of constant capacitive and inductive characteristic of load under same voltage conditions and the periodic current of previous switching appliance is calculated by using steady periodic current. Furthermore, the behavior of the habits in the residence is considered to determine the possible combination of devices used (Wu, Jiao, Liang, & Han, 2019). Buddhahai et al. (2018) present a non-intrusive load monitoring system which uses some electrical features such as current, real power, reactive power, and Power Factor (PF) collected in one-minute period. By using Decision Tree as the multi-label classification algorithm utilized from a multi-label classification scheme, promising results are obtained (Buddhahai et al., 2018). Bouhouras et al. (2017) propose using 1st, 3rd and 5th harmonic orders in the load signatures formulation. A data processing methodology s employed in order to keep the load signature short and simple. The analysis indicates that the developed load signatures improve the efficiency of the NILM algorithm and the higher harmonic currents improve the identification performance (Bouhouras et al., 2017). Lin et al. (2016) use some load signatures such as current waveform, current harmonic, active-reactive power and geometric properties of I-V curves are employed for identification. An experimental platform is implemented includes a fan, a desk light, a water fountain, a microwave oven and an electric heater as loads. It is indicated that over 90 % recognition rate is achieved in the experiments (Lin et al., 2016). Wu et al. (2019) propose an event-based non-intrusive load identification method for residential loads that monitors the circuit to determine a switching event. Characteristic filtering is used then to filter the decomposition current according to the unique harmonic components of each load (Wu et al., 2019). Bouhouras et al. (2019) propose the use of harmonic current vectors to improve the performance of the NILM algorithm. The harmonic current vectors are more distinctive features than the current amplitudes in using simultaneously operating appliances for load identification. Current amplitude, phase angles in the form of vector projection on the x-axis and harmonic current vectors are used as features for load identification and their effects on identification performance are investigated. The stand-alone operations of appliances and the different combinations of appliances are considered, and the results show that usage of harmonic current vectors has significant impact on identification performance (Bouhouras, Gkaidatzis, Panagiotou, Poulakis, & Christoforidis, 2019).

When the studies aforementioned above are investigated, it can be seen that some signals such as current, voltage, power and harmonics are prominent for load identification. One of the discriminative signals is current, and most of the studies utilize from Fourier transform to obtain more distinctive features and therefore, in this study current signal is selected as study area. In conventional approach, the magnitude of the current signal’s Fourier transform is employed as feature. Different from such studies, the imaginary and real parts of this transform are evaluated separately. Five different features obtained from standard deviation, maximum and minimum values of these parts are used in this study. A rule based non-intrusive load identification approach is developed by using extracted features to identify the load appliances. Six different load appliances include, Kettle, Monitor, Laptop Charger, Halogen Lamp, Heater and Vacuum Cleaner which are widely used in a house are selected for testing the proposed approach. After current signal measurement, firstly zero-crossing one period current signal is separated from measured two period signal, in data preprocessing phase. Then, the features are extracted from this signal and a feature vector consist of five features is obtained. A possible matching between feature vector and appliance or appliance combination in rule table is searched. The aim of this study is to determine the loads on the line at any time and therefore tracking a change in load status (to be activated or deactivated) is not necessary for identification. The organization of the paper is as follow; the load appliances used in this study are described in Section 2 while the developed approach is explained in Section 3. The experimental results are presented in Section 4, and finally the conclusions are given in Section 5.