Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 16, No. 2, April 2026, pp. 675 ∼ 686 ISSN: 2088-8708, DOI: 10.11591/ijece.v16i2.pp675-686 ❒ 675 A r eal-time appliance monitoring appr oach with anomaly detection f or r esidential houses Nimantha Madhushan 1 , Rasanjalee Rathnayak e 2 , Dhanushika Darshani 1 , Ashmini J ee v a 1 , Uditha W ijewardhana 1 , Nishan Dharmaweera 1 1 Department of Electrical and Electronic Engineering, Uni v ersity of Sri Jaye w ardenepura, Nuge goda, Sri Lanka 2 Department of Computer Engineering, Uni v ersity of Sri Jaye w ardenepura, Nuge goda, Sri Lanka Article Inf o Article history: Recei v ed Jun 13, 2025 Re vised Dec 18, 2025 Accepted Jan 16, 2026 K eyw ords: Anomaly detection Appliance identication Demand side managment Ev ent detection Intrusi v e load monitoring Non-intrusi v e load monitoring ABSTRA CT Monitoring electrical appliances in residential b uildings is essential for mini- mizing ener gy w aste and enhancing safety through the early detection of ab- normal conditions. While researchers ha v e in v estig ated both intrusi v e and non- intrusi v e load monitoring approaches, the non-intrusi v e approac h has emer ged as preferred due to its cost-ef fecti v eness and nonin v asi v e implementation. De- spite considerable progress in appliance monitoring and f ault detection systems o v er the past tw o decades, critical challenges and limitations persist. This paper proposes a lo w-comple xity appliance identication and monitoring solution to o v ercome those issues. Furthermore, the proposed solution is inte grated with an abnormal condition detection mechanism for critical appliances, aiming to sa v e ener gy and ensure the safety of the po wer system. Furthermore, the solution incorporates user feedback via a dedicated mobile application, enhancing adapt- ability and performance. The proposed s olution has be en v alidated in real-time en vironments using both custom and publicly a v ailable datas ets, demonstrating impro v ed accurac y in ener gy monitoring and increased consumer safety . This is an open access article under the CC BY -SA license . Corresponding A uthor: Nimantha Madhushan Department of Electrical and Electronic Engineering, Uni v ersity of Sri Jaye w ardenepura Nuge goda, 10250, Sri Lanka Email: nimanthamk@sjp.ac.lk 1. INTR ODUCTION Residential electric ener gy monitoring systems ha v e become a trending research area in recent decades due to the global ener gy crisis, rising ener gy costs, and v arious en vironmental concerns [1]–[5]. T raditional ener gy meters pro vide only the total ener gy consumption of the house and, f ail to of fer i nsights into the usage of indi vidual appliances [6]. As a result, consumers are unable to ef fecti v ely manage specic appliances, leading to ener gy w aste, reduced appliance longe vity , and causing hazardous operational conditions. Ho we v er , an accurate load monitoring (LM) mechanism has the potential to address these mentioned issues, allo wing for better ener gy management and impro v ed protection of the po wer system in the residential premises [6], [7]. Furthermore, the y also play a role in mitig ating global climate change by reducing unnecessary electricity consumption [3]–[5], [8], [9]. An accurate LM system s ha v e been sho wn to signicantly c u t electricity usage, leading to sa vings of up to 20% for consumers [2], [6], [7]. By identifying ener gy-hungry appliances and suggesting optimizations, these systems empo wer users to control their ener gy consumption and reduce their o v erall ener gy bills. Addi- tionally , manuf acturers also benet from introducing ne w ener gy-ef cient appliances based on the analytical J ournal homepage: http://ijece .iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
676 ❒ ISSN: 2088-8708 data collected from LM systems [2]. Be yond ener gy sa vings, LM systems can contrib ute to a wide range of applications, such as anomaly detection of appliances for enhanced protection of po wer systems and ambient assisted li ving applications for elderly people [10]–[13]. Abnormal condition detecti on, or anomaly detection, is g aining importance due to its ability to pre v ent system f ailures and damage to appliances [11]–[13]. Identifying malfunctioning appliances early helps con- sumers a v oid costly repairs and potential safety hazards [11]. Anomalies in po wer consumption patterns can indicate de vices that are not operating as intended, allo wing users to tak e correcti v e actions swiftly . Moreo v er , inte grating anomaly detection into load monitoring systems can contrib ute to the o v erall safety and ef cienc y of the po wer grid [14]–[16]. Such systems also benet elderly indi viduals, who can use this technology to ensure their appliances are functioning optimally , enhancing their quality of life in assisted li ving en vironments [10]–[13]. Non-intrusi v e load monitoring (NILM) and intrusi v e load monitoring (ILM) are the tw o concepts used to address both ener gy management and anomaly detection. The NILM techniques allo w for the monitoring of indi vidual appliances by analyzing the aggre g ated po wer signal from a single measurement point, usually at the main ener gy entry point of the house [6], [17], [18]. The rst concept w as introduced by G.W . Hart in 1992 using acti v e and reacti v e po wer v alues of appliances [19]. This method is cost-ef fecti v e compared to ILM, which requires installing measuring apparatus on each de vice [2], [17], [18], [20]. The Sense [21] and Emporia [22] are tw o commercially a v ailable products that rely on LM approach, where Sense ener gy meter utilizes NILM approach and Emporia follo ws ILM approach. NILM has pro v en ef fecti v e in identifying operating appliances and detecting abnormal operations in real-time, making it ideal for anomaly detection applications [11]. Furthermore, e v ent-based NILM approaches, which track the switching on and of f of appliances, are particularly suited for real-time anomaly detection. Despite t h e promise of appliance identi cation solutions and anomaly detection solutions proposed during the past tw o decades [2], [15], [23], se v eral challenges remain unresolv ed [2], [5], [15], [23], [24]. Most of the proposed solutions used e xpensi v e and comple x hardw are arrangements to acquire data from the po wer system [5], [6], [25]–[28], tar geted only identication and monitoring of selected high-po wer - consuming appliances [3], [29]–[34], required labeled datasets for both appliance identication and abnormal operation detection [2], [20], [35]–[40], and used high computing po wer to run the appliance identication solutions. In response to these challenges, a lo w-comple xity appliance identication and monitoring system that combines both LM approaches to enhance the monitoring and management of residential electric appliances. Our approach focuses on impro ving real-time detection capabilities to ensure timely identication of abnormal conditions. The system is tested using both a custom dataset and publicly a v ailable datasets, yielding results that demonstrate its ef fecti v eness in pro viding accurate, real-time appliance monitoring and anomaly detec- tion. A comparison between the proposed solution and s tate-of-the-art m ethods is presented in T able 1. K e y contrib utions are as follo ws. a. Real time appliance identication and monitoring system w as proposed for residential houses co v ering all the electric appliances in the premises. Lo w comple xity machine learning solution w as proposed for appliance identication to reduce the computational po wer requirement when it is deplo yed in real w orld settings. b . Lo w-comple xity data acquisition (D A Q) system w as de v eloped to acquire data from po wer system and it is easy to install in the house. Since, the cost of the solution is primary concern of consumers, e xpensi v e hardw are component is not suitable for real w orld settings. c. An accurate e v ent detection methodology w as proposed to identify e v ents of both lo w and high-po wer consuming appliances. By identifying lo w po wer consuming appliances such as light b ulbs, f ans, consumer can switch OFF unw anted appliances and acti v ely contrib ute to the ener gy sa vings strate gies. d. A consumer feedback mechanism w as introduced to enhances system accurac y and user e xperience by de v eloping a mobile application. The de v eloped application can be used for vie w consumption data in real time and gi v e correct appliance names to the system. e. Real time anomaly detection methodology also proposed in p a rallel to the appliance identication to detect abnormal operations of critical appliances in the house. The outline of the paper as follo ws. Section 2 e xplain the prototype de v elopment and proposed ap- pliance identication and monitoring solution. T est ed results are discussed in sections 3 and 4 describes the conclusion of the research w ork. Int J Elec & Comp Eng, V ol. 16, No. 2, April 2026: 675-686 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 677 T able 1. Comparison with e xisting appliance identication solutions Real time Can identify lo w Lo w cost Can detect Ability to Can be Ref. monitoring po wer consuming appliances D A Q system ne w appliances detect anomalies Generalized [5] X X X X X X [12] X ✓ ✓ X X X [17] ✓ ✓ ✓ X X X [26] X X X X X X [27] X X X X X X [28] X ✓ X X X X [30] X ✓ X X X X [34] X ✓ X X X X Proposed ✓ ✓ ✓ ✓ ✓ ✓ 2. METHOD The proposed appliance identicati on and monitori ng solution follo ws both intrusi v e and non-intrusi v e monitoring approaches to reduce the comple xity of the total system. Appliances can be cate gorized into tw o groups based on their a v erage po wer consumption patterns: linear and non-linear . Non-linear po wer -consuming appliances include w ashing machines, high-performance laptops, and computers. Therefore, tw o types of D A Q units are included, such as the main unit (MU) to acquire total ener gy consumption and auxiliary units (A Us) to acquire indi vidual ener gy consumption of non-linear po wer -consuming appliances. Figure 1 illustrates the o v erall block diagram of the proposed approach. Figure 1. Basic block diagram of the proposed h ybrid appliance identication and monitoring solution 2.1. Data acquisition system The ESP-8266 (i.e., Node MCU) microcontroller board w as used as the main processing part of de v el- oped measurement units and it is used in most of Internet of Things (IoT)-based projects. It has b uilt-in wireless delity (W i-Fi) f acility enables the transfer of recorded data to a cloud service. Google Firebase online database w as used in this w ork as the cloud service. The data transfer rate is congured for 10-second interv als based on the e xperimental data [6]. The PZEM-004T ener gy meter module w as used to read the acti v e po wer , root-mean square (RMS) current, RMS v oltage, po wer f actor (PF), and frequenc y of the po wer system. Figure 2 illustrates the o v erall process of the de v eloped solution. 2.2. A ppliance identication and monitoring pr ocess The NILM approach w as used to identify linear po wer -consuming appliances. The consumption of those appliances can be calculated by subtracting A U data from MU data. Consumption of non-linear po wer - consuming appliances can be directly vie wed from A U data and it follo ws ILM approach. NILM concept consist with four steps: data acquisition, e v ent detection, feature e xtraction, and appliance identication. The steps are discussed as follo ws. A r eal-time appliance monitoring appr oac h with anomaly detection for ... (Nimantha Madhushan) Evaluation Warning : The document was created with Spire.PDF for Python.
678 ❒ ISSN: 2088-8708 Figure 2. Basic block diagram of the o v erall process 2.2.1. Ev ent detection The e v ent detection methodology proposed here is only used to identify state transitions (i.e., e v ents) of linear po wer -consuming appliances. Acti v e po wer v ariation w as selected as the best feature to detect e v ents in the po wer system. The e v ent detection methodology proposed in our pre vious w ork w as used with a 50- second o v erlapping sliding windo w [6]. Since the data recording interv al is x ed at 10 seconds, sliding windo ws consist of 5 data points. 2.2.2. F eatur e extraction Since the research w ork tar gets a lo w-comple xity appliance identication solution, po wer v alues were selected for appliance identication. Based on the e xperimental data, a threshold v alue of ( P thr eshol d ) 350 W w as dened to di vide high and lo w-po wer -consuming appliances. Figure 3 sho ws the scatter plot of acti v e po wer (P) and reacti v e po wer (Q) of commonly used dif ferent lo w-po wer -consuming appliances. According to that, P and Q v alues can be used for appliance identication. In Figure 3, the black-colored curv e represents a ne wly observ ed appliance, which closely resembles the prole of a pedestal f an. Consequently , the ne w appliance is identied as a pedestal f an. F or high-po wer -consuming appliances such as electric k ettles and rice cook ers, their distinct operat- ing durations pro vide an additional discriminatory feature not typically present in lo w-po wer appliances [6]. Therefore, both the po wer v alues and operational time are used to correctly identify those appliances. Equation (1) is used to calculate the acti v e po wer v alue ( P new ) of an appliance that e xperiences a state transition. P new = | P af ter − P bef or e | (1) Int J Elec & Comp Eng, V ol. 16, No. 2, April 2026: 675-686 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 679 The po wer v alues of the appliance for 1 minute are recorded after detecting an ”ON” e v ent as per the procedure sho wn in Figure 2. The system then calculates the a v erage po wer consumption v alue ( P af ter ) from the recorded v alues. The po wer v alues before the e v ent ( P bef or e ) are used to obtain the a v erage v alues of the 1 st and 2 nd points of the e v ent windo w [6]. In the case of an OFF e v ent, the P af ter v alues are calculated by getting the a v erage of the 4 th and 5 th points. The P bef or e v alues are calculated from the a v erage of the 1 st and 2 nd points. The same process is then repeated with reacti v e po wer v alues. Figure 3. Scatter plot of P and Q v alues of dif ferent set of residential appliances 2.2.3. A ppliance identication Since e v ery house has lighting loads, at the be ginning, the databas e i s on l y fe d wit h dat a abou t l ight- emitting diode (LED) b ulbs. Other appliance data will be stored in the database automatically after installing the system. The proposed appliance identication process follo ws a self-supervised learning approach, and a training period of one month (i.e., 30 days) from the system installation date w as dened based on the e xperimental results. During that period, it is assumed that each appliance will be operating. The k-Nearest Neighbors (k-NN) model w as utilized for the prediction process. If a ”switching-on” e v ent is detected by the system, the a v erage po wer v alues (i.e., acti v e and react i v e v alues) of the ne wly switched-on appliance are calculated. It then determines whether it is a lo w or high-po wer - consuming appliance. The k-NN model uses the updated database to predict a lo w-po wer -consuming appliance. The system then calculates the po wer dif ferences ( P dif f ) between the predicted and actual appliances. Based on the e xperimental results, a 10 % mar gin of P dif f is allo wed for all appliances [6]. If the calculated P dif f surpasses the threshold, it can be an indication of the presence of a ”ne w appliance” and a temporary database w as used to store appliance details during the training per iod. The system checks the temporary database for an y matched appliances before creating a ne w entry . A match indicates that the appliance has been in operation before, and therefore, the details of the appliance are permanently stored in the main database. If an y high-po wer -consuming appliance is found, the identication process will be run in the switching- of f state. Until then, it is sho wn as an appliance, indicating consumption without assigning an y name. Both the po wer v alues and operational time are considered for the prediction of these appliances. At the switching-of f state, the s ame 10 % of po wer le v el ( P dif f ) and 5-minute time mar gin were used for the v erication pro- cess. The database stores the data from those appliances along with their operational times. The same process discussed abo v e will be follo wed to store ne w appliance data. 2.3. Mobile application A mobile application w as de v eloped to sho w consumption data to consumers and g ather feedback. Figure 4(a) illustrates the main screen of the app, and the total acti v e po wer consumption, system v oltage, total current, and frequenc y of the system are visible to the consumer . By clicking the ”Switched ON Appli- ance” b utton in the main screen, consumers can see operating appliances in the system sho wn in Figure 4(b). Consumers can enter correct names using the screen sho wn in Figure 4(c). Before entering the correct name, the appliance name is sho wn as ”appliance n”. The v ariable n can tak e an y inte ger v alue from 1 to 10. The A r eal-time appliance monitoring appr oac h with anomaly detection for ... (Nimantha Madhushan) Evaluation Warning : The document was created with Spire.PDF for Python.
680 ❒ ISSN: 2088-8708 solution allo ws for the storage of 10 ne w appliances. Additionally , the same screen allo ws for the entry of A U names. The consumption details of non-linear po wer -consuming appliances can be seen from the screen sho wn in Figure 4(d). Figure 4. Screens of the mobile application, (a) main screen, (b) screen of switching ON appliances screen, (c) names of appliances, and (d) consumption of non-linear appliances 2.4. Abnormal conditions detection pr ocess The proposed non-intrusi v e appliance identication method does not ha v e an y supervised parameters (i.e., labeled data). Therefore, the abnormal condition detection system cannot operate during the training period. After the training period of the system ends, the detection process will start automatically . A statistical approach w as used to detect abnormal condit ions of a selected set of appliances [11]. In statistics, a v alue in a normal distrib ution greater than ( µ + 2 σ ) or less than ( µ − 2 σ ) is dened as an anomaly v alue [41]. The mean v alue of the distrib ution is dened as µ , and σ is the standard de viation of it. Most of the pre viously published w orks [15], [23], [39], [42] used labeled data. Therefore, the y follo wed the 3 σ rule to dene whether it w as an anomaly or not. That rule consisted of 99.7 % data from the distrib ution [11], [43]. Ho we v er , that rule is not suitable because the proposed system does not ha v e labeled data. Therefore, the 2 σ rule w as used for the proposed method. F our commonly used appliances were selected for the abnormal condition detection process: rice cook ers, electric k ettles, refrigerators, and light b ulbs [11]. Light b ulbs normally operate during the night and early morning hours. Some houses, such as bathrooms and storerooms, may also operate during the daytime [11]. Those operations are mainly based on consumer beha viors. During the training period of the system, the operating times are recorded automatically by the system. If an y lighting load operates dif ferently from the recorded times, it can be dened as an abnormal operation. Based on the e xperimental data, a threshold v alue of 2 hours w as used. Electric heaters and ri ce cook ers are the main heating appli ances in most residential houses in South Asian re gion. F or them, both the operational time and po wer v alues are considered for anomaly detection. A cont inuous monitoring mechanism w as proposed for these appliances. The maximum allo w able time is calculated from pre vious data using the ( µ ± 2 σ ) condition, and the s ystem continuously checks whether the appliance is operated for more than that time. If an y v ariation is detected, the system automatically sends an alert to the consumer . Further , at the switchi ng-of f state, the a v erage po wer consumption is also check ed with pre vious data, and if an y v ariation is detected, an alert is sent to the consumer . The refrigerator is the main cooling appliances in Sri Lankan residential houses and it sho ws a grad- ually decreasing po wer consumption. Therefore, po wer consumption in the ON and OFF states is dif ferent [15], [11]. Furthermore, the refrigerator will operate se v eral times per day . Both the operational period and time between the tw o operations mainly depend on consumer beha viors and the load of the refrigerator . The proposed method checks four parameters [11]: i) The a v erage po wer at the switching ON state, ii) The a v erage Int J Elec & Comp Eng, V ol. 16, No. 2, April 2026: 675-686 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 681 po wer at the switching OFF state, iii) Operational time, and i v) Duration between tw o consecuti v e operations. All the parameters are check ed at the switching-of f state of the appliances. Allo w able v alues are calculated from the pre vious data using the ( µ ± 2 σ ) condition. 3. RESUL TS AND DISCUSSION The proposed solution w as tested in both real time and simulation based e xperiments. First, the non- intrusi v e appliance identication process w as tested across three data sets. Second, the proposed abnormal condition detection process w as tested by adding articial anomalies into the original operations (i.e., without anomalies) of selected appliances. The standard accurac y matrix w as used to analyze the performance of the proposed solution [44]. The Accur acy v alues of each test can be observ ed using (2). Accur acy = T P + T N T P + T N + F P + F N (2) Where TP , FP , TN , and FN are the true positi v e, f alse positi v e, true ne g ati v e, and f alse ne g ati v e cases, respec- ti v ely . T rue positi v e is an outcome where the implemented model correctly predicts the positi v e class; f alse positi v e is an outcome where it predicts the positi v e class incorrectly; true ne g ati v e is where the model predicts the ne g ati v e class correctly; and f alse ne g ati v e is where the model predicts the ne g ati v e class incorrectly . 3.1. Real-time appliance identication A custom appliance list w as used as the rst data set to v erify the v al idity of the proposed sol ution. The w ork in v olv ed twelv e dif ferent types of linear -po wer -consuming appliances. Since light b ulbs are a v ailable at e v ery house, LED light b ulbs with three dif ferent w attage v alues were used. Before applying the appliance identication process, the e v ent detection process w as tested on o v er 300 pre-recorded e v ents. The proposed process achie v es a 98 % accurac y le v el in simulation-based testing. After that, the total solution w as tested in a real-w orld house with a s et of selected appliances. The time tak en for the detect e v ent w as analyzed, and on a v erage, 0.000223 seconds (0.2 milliseconds) were observ ed in real-time operations. The summary of the identication result of the custom data set is sho wn in T able 2. The identication of the satellite tele vision decoder and 7W LED b ulb sho wed the lo west accurac y compared with other appliances due to the similarity in a v erage po wer consumption v alues. T able 2. Identication results of appliances in the custom dataset Appliance Rated w attage (W) Identication accurac y (%) LED b ulb 7 94.50 LED b ulb 9 100.00 LED b ulb 12 100.00 Electric heater 1000 100.00 Electric k ettle 1800 100.00 T oaster 800 100.00 Rice cook er 900 100.00 Pedestal f an 55 96.00 Hair dryer 750 100.00 LED T ele vision 59 100.00 Refrigerator 70 100.00 Satellite tele vision decoder 12 90.00 Aquarium pump 20 100.00 Air conditioner 1800 100.00 3.2. Simulation-based test r esults Secondly , tw o publicly a v ailable data sets w as used to v erify the generalization of our proposed methodology . The T racebase data set w as made in German y using 40 dif ferent appliances in more than 10 households and of ce en vironments [45]. In that data set, there are appliance-based acti v e po wer recordings with tw o dif ferent sampling rates: 1 Hz and 1/8 Hz. The proposed methodology necessit ates that appliances ha v e reacti v e po wer v alues. Therefore, these v alues are calculated using a v erage PF v alues [46]. The selected appliances from the T racebase dataset, as well as the identication accurac y are sho wn in T able 3. A r eal-time appliance monitoring appr oac h with anomaly detection for ... (Nimantha Madhushan) Evaluation Warning : The document was created with Spire.PDF for Python.
682 ❒ ISSN: 2088-8708 The v acuum cleaner and the hair dryer are tw o c omparable ener gy-intensi v e de vices. Ne v ert heless, their operational durations v ary . The hair dryer functioned for around 5 minutes, while the v acuum cleaner operated for a duration of 12 minutes. Consequently , the operational period of the appliance can enhance the system’ s accurac y in identifying things. In certain states, the LCD tele vision’ s po wer consumption w as comparable to that of a refrigerator , resulting in the lo west accurac y . A total of 10 dif ferent appliances were tested in the T racebase dataset, and an a v erage of 89.47 % accurac y le v el w as obtained. The second data set is the iA WE [47] and three appliances were tested with the proposed non-intrusi v e identication solution. The achie v ed accurac y v alues are sho wn in T able 4. Since the number of appliances used for the test is less, a higher accurac y le v el w as observ ed. Most of the propose d appliance monitoring solutions achie v ed o v er 90% accurac y le v el for only a se- lected set of appliances [3], [30]–[34]. Ho we v er , those require labeled data, rely on a selected set of appliances, and are tested only for simulations. The solution proposed in this research does not rely on labeled data, and it is a generalized solution. Further , the solution can w ork in real w ork settings in real-time. Therefore, the solution is more reliable for monitoring all the appliances in a residential house. T able 3. Identication results of appliances in the T racebase dataset Appliance A v erage po wer consumption (W) Identication accurac y (%) Electric k ettle 2164.87 100.00 Lamp 41.63 100.00 V acuum cleaner 1131.88 100.00 T oaster 717.17 100.00 Hair dryer 1190.09 100.00 Cooking sto v e 828.05 91.67 Laundry dryer 2523.08 83.00 Liquid crystal display (LCD) tele vision 45.44 30.00 Refrigerator 190.71 90.00 Freezer 73.89 100.00 T able 4. Identication results of appliances in the iA WE dataset Appliance A v erage po wer consumption (W) Identication accurac y (%) Air conditioner 1694.31 100.00 W ater motor 593.46 100.00 Refrigerator 128.80 – 108.50 100.00 3.3. Abnormal condition detection test r esults The ne xt step of the solution in v olv ed testing the proposed anom aly detection approach on identied appliances across the three datasets. F our appliances from the custom dataset, three appliances from the T race- base dataset, and one appliance from the iA WE dataset were used in this testing phase. The selected datasets were augmented by introducing anomalies manually as pre viously done in [11]. The obtained results are sho wn in T able 5. F or refrigerators and freezers, anomalies ar e check ed at the switching OFF s tate. Both the operational time and consumption v alues are check ed. F or heating appliances (i.e., k ett les and rice cook ers) and light b ulbs, real-time monitoring w as considered. On a v erage the proposed anomaly detection methodology achie v es, a 99% accurac y le v el for real-time testing [11]. T able 5. Abnormal condition detection accurac y le v els Dataset Appliance Operational time - Detection ratio operational time - Po wer -based anomalies - at the Real-time monitoring at the switching OFF state switching OFF state Custom Refrigerator - 12/12 10/10 Custom Rice cook er 10/10 10/10 10/10 Custom Electric k ettle 11/12 11/12 9/9 Custom Light b ulbs 12/12 - - iA WE Refrigerator - 10/10 10/10 T racebase Refrigerator - 10/10 12/12 T racebase Freezer - 11/12 10/10 T racebase W ater k ettle 7/8 8/8 9/9 Int J Elec & Comp Eng, V ol. 16, No. 2, April 2026: 675-686 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 683 Based on the observ ed results, the proposed non-intrusi v e appliance identication method can accu- rately identify linear -po wer -consuming appliances in real time. Unlik e other approaches, it does not depend on labeled data and demonstrates strong generalization capability . Additi onally , since the solution emplo ys con v entional machine learning algorithms, its comple xity is signicantly lo wer than neural-netw ork-based ap- pliance identication methods. The proposed system also requires fe wer A Us than systems using an ILM approach, such as the Emporia ener gy monitoring system [22]. As a result, it is more cost-ef fecti v e than ILM solutions. Furthermore, the anomaly detection mechanism does not rely on labeled data and can adapt to appliances with v arying consumption patterns. 3.4. Futur e w orks The proposed solution w orks on the cloud service, and there may be pri v ac y concern issues by con- sumers. Therefore, implementation of the total system in an edge computing de vice will increase the security . Future w ork includes the de v elopment of an edge computing de vice to identify and monitor all the appliances. It can be used to inte grate rene w able sources and can be used as a single dashboard to manage all the po wer sources. Further , propose an appliance name v erication process to enhance the accurac y , which also includes future w orks. 4. CONCLUSION This study presents a h ybrid appliance identication and monitoring system that syner gizes the strengths of NILM and ILM approaches to address criti cal challenges in residential ener gy management. The proposed solution achie v es real-time identication of linear -po wer -consuming appliances usi ng a lo w-comple xity , self- supervised k-NN model, eliminating dependenc y on labeled datasets and reducing computational o v erhead compared to neural-netw ork-based methods. By inte grating A Us only for non-linear appliances, the system signicantly lo wers hardw are costs relati v e to full ILM s y s tems lik e Emporia, while maintaining high accurac y of 89% on a v erage across datasets. V alidation across custom and public datasets conrmed the system’ s generalizabil ity and ef fecti v enes s in di v erse residential settings. Future w ork will focus on edge computing inte gration to address pri v ac y con- cerns and rene w able ener gy inte gration, further adv ancing the system’ s utility in smart grids and demand-side management. By bridging g aps in af fordability , real-tim e capability , and unsupervised learning, this research contrib utes to sustainable ener gy consumption, enhanced safet y , and consumer empo werment in residential ener gy ecosystems. A CKNO WLEDGMENTS This research w as supported by the Science and T echnology Human Resource De v elopment Projec t, Ministry of Higher Education, Sri Lanka, funded by the Asian De v elopment Bank (Grant No: R1/SJ/02). FUNDING INFORMA TION This research w as supported by the Science and T echnology Human Resource De v elopment Projec t, Ministry of Higher Education, Sri Lanka, funded by the Asian De v elopment Bank (Grant No: R1/SJ/02). A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contri b ut or Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Nimantha Madhushan ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Rasanjalee Rathnayak e ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Dhanushika Darshani ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Ashmini Jee v a ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Uditha W ije w ardhana ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Nishan Dharma weera ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ A r eal-time appliance monitoring appr oac h with anomaly detection for ... (Nimantha Madhushan) Evaluation Warning : The document was created with Spire.PDF for Python.
684 ❒ ISSN: 2088-8708 C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY − The supporting data of this study are openly a v ailable in ”iA WE” at https://ia we.github .io, [47]. − The supporting data of this study are openly a v ailable in ”T racebase” at https://github .com/areinhardt/ trace- base, [45]. − The data that support the ndings of this study will be a v ailable in Github repository named ”appli- ance data” at https://github .com/NimanthaMadhushan/appliance data. REFERENCES [1] M . Herath, C. J. Ang ammana, and M. Liyanage, “ A study of the ef fects of appli ance ener gy signatures on dif ferent neural net- w ork types in nonintrusi v e load monit oring, ” IEEE T ransactions on Instrumentation and Measurement , v ol. 72, pp. 1–10, 2023, doi: 10.1109/TIM.2023.3305664. [2] Y . Himeur , A. Alsalemi, F . Bensaali, A. Amira, and A. Al-Kababji, “Recent trends of smart nonintrusi v e load m onitoring in b uild- ings: A re vie w , open challenges, and future directions, ” International Journal of Intelligent Systems , v ol. 37, no. 10, pp. 7124–7179, 2022, doi: 10.1002/int.22876. [3] M . D. Silv a, Q. Liu, and O. F . Darteh, “ A recent re vie w of NILM frame w ork: De v elopment and challenges, ” in 2022 IEEE Intl Conf on Dependable, Autonomic and Sec ure Computing, Intl Conf on Perv asi v e Intelligence and Computing, Intl Conf on Cloud and Big Data Computing, Intl Conf on Cyber Science and T echnology Congress (D ASC/PiCom/CBDCom/CyberSciT ech) , 2022, pp. 1–7, doi: 10.1109/D ASC/PiCom/CBDCom/Cy55231.2022.9927902. [4] T . L. Y oung, J. Gopsill, M. V alero, S. Eikvag, and B. Hicks, “Comparing four machine learning algorithms for household non- intrusi v e load monitoring, ” Ener gy and AI , p. 100384, 2024, doi: 10.1016/j.e gyai.2024.100384. [5] R. Liaqat, I. A. Sajja d, M. W aseem, and H. H. Alhelou, “ Appliance le v el ener gy characterization of reside ntial electricity demand: Prospects, challenges and recommendations, ” IEEE Access , v ol. 9, pp. 148676-148697, 2021, doi: 10.1109/A C- CESS.2021.3123196. [6] N. Madhushan, R. Rathnayak e, A. Udh yani, U. W ije w ardhana, and N. Dharma weera, “Rule-based self-supervised learning non- intrusi v e load monitoring system for residential houses, ” in 2023 IEEE 8th International Conference on Recent Adv ances and Inno v ations in Engineering (ICRAIE) , 2023, pp. 1–6, doi: 10.1109/ICRAIE59459.2023.10468116. [7] P . A. Schirmer and I. Mporas, “Non-intrusi v e load monitoring: A re vie w , ” IEEE T ransactions on Smart Grid , v ol. 14, no. 1, pp. 769–784, 2023, doi: 10.1109/TSG.2022.3189598. [8] R. Nayak and C. D. Jaidhar , “Enhancing anomaly detection in critical systems using household appliance po wer consumption data, ” IEEE Sensors Journal , v ol. 24, no. 17, pp. 27 677–27 686, 2024, doi: 10.1109/JSEN.2024.3426090. [9] Y . Himeur , A. Alsalemi, F . Bensaali, and A. Amira, “ A no v el approach for detecting anomalous ener gy consumption based on micro-moments and deep neural netw orks, ” Cogniti v e Computation , v ol. 12, no. 6, 2020, doi: 10.1007/s12559-020-09764-y . [10] C. L. Athanasiadis et al. , “ A smart ener gy management system for elderly households, ” in 2022 57th International Uni v ersities Po wer Engineering Conference (UPEC) , 2022, pp. 1–6, doi: 10.1109/UPEC55022.2022.9917856. [11] N. Madhushan, R. Rathnayak e, D. Darshani, U. W ije w ardhana, and N. Dharma weera, “ Abnormal operations detection of residential electric appliances using non-intrusi v e load monitoring, ” in 2024 IEEE 9th International Conference for Con v er gence in T echnology (I2CT) , 2024, pp. 1–6, doi: 10.1109/I2CT61223.2024.10544325. [12] C. Chen, G. Geng, H. Y u, Z. Liu, and Q. Jiang, “ An end-c loud collaborated frame w ork for transferable non-intrusi v e load monitor - ing, ” IEEE T ransactions on Cloud Computing , v ol. 11, no. 2, pp. 1157–1169, 2023, doi: 10.1109/TCC.2021.3132929. [13] D. Gonzalez, M. A. P atricio, A. Berlang a, and J. M. Molina, “V ariational autoencoders for anomaly detection in the beha viour of the elderly using electricity consumption data, ” Expert Systems , v ol. 39, no. 4, p. e12744, 2022, doi: 10.1111/e xsy .12744. [14] H. Apaydin- Ozkan, “ Appliance-le v el anomaly detection by using control charts and articial neural netw orks with po wer proles, ” Sensors , v ol. 22, no. 17, p. 6639, 2022, doi: 10.3390/s22176639. [15] S. S. Hosseini, K. Agbossou, S. K elouw ani, A. Cardenas, and N. Henao, “ A practical approach to residential appliances on- line anomaly detection: A case study of standard and smart refrigerators, ” IEEE Access , v ol. 8, pp. 57 905–57 922, 2020, doi: 10.1109/A CCESS.2020.2982398. [16] D. A. M. Lemes et al. , “Lo w runtime approach for f ault detection for refrigeration systems in smart homes using w a v elet transform, ” IEEE T ransactions on Consumer Electronics , v ol. 70, no. 1, pp. 4447–4456, 2024, doi: 10.1109/TCE.2023.3328147. [17] M. T . Ahammed et al. , “Real-time non-intrusi v e electrical load classication o v er iot using machine learning, ” IEEE Access , v ol. 9, pp. 115 053–115 067, 2021, doi: 10.1109/A CCESS.2021.3104263. [18] A. Ruano, A. Hernandez, J. Ure˜na, M. Ruano, and J. Garcia, “NILM techniques for intelligent home ener gy management and ambient assisted li ving: A re vie w , ” Ener gies , v ol. 12, no. 11, p. 2203, 2019, doi: 10.3390/en12112203. Int J Elec & Comp Eng, V ol. 16, No. 2, April 2026: 675-686 Evaluation Warning : The document was created with Spire.PDF for Python.