Time Series Anomaly Detection in Smart Homes: A Deep Learning Approach Somayeh Zamani1,* , Hamed Talebi2 and Gunnar Stevens1 1 University of Siegen, Siegen, Germany 2 Amirkabir University of Technology, Tehran, Iran Abstract Fixing energy leakage caused by different anomalies can result in significant energy savings and extended appliance life. Further, it assists grid operators in scheduling their resources to meet the actual needs of end users, while helping end users reduce their energy costs. In this paper, we analyze the patterns pertaining to the power consumption of dishwashers used in two houses of the REFIT dataset. Then two autoencoder (AE) architectures with 1D-CNN and TCN as backbones are trained to differentiate the normal patterns from the abnormal ones. Our results indicate that TCN outperforms CNN1D in detecting anomalies in energy consumption. Finally, the data from the Fridge_Freezer and the Freezer of house No. 3 in REFIT is also used to evaluate our approach. Keywords Time series, Anomaly detection, Deep learning, Autoencoder, Temporal convolutional networks, Smart home, Sustainability 1. Introduction higher power consumption or damage in the most critical cases [8]. Throughout recent years, the energy demand has signifi- Thus, for optimization purposes in smart homes, via cantly gone up due to urban and industrial development implementing load monitoring systems and formulating alongside an increase in population [1]. Therefore, cli- smart anomaly detection models using machine learning mate change, global warming, and volatility in energy techniques, the abnormality can be mitigated [9]. To do prices have fuelled the interest in smart systems [2]. In so, it is essential to analyze the energy consumption of this regard, the huge potential increase of replacing tra- households in order to identify consumption patterns and ditional home appliances with new in-operation power- extract valuable information from smart homes [8][10]. consuming ones by 2040 has caused the residential sector In this paper, the power consumption patterns of dish- to account for roughly 60% over 2017-25 and 70% over washers used in houses No. 1 and 2 of the REFIT dataset 2025-40 of electricity demand increase of buildings. As are analyzed as examples of devices that are used based such, household appliances need to be operating effi- on the needs of users. Then, the data for each usage of ciently and used appropriately to achieve energy-saving the device is divided into different signals to properly goals [3]. train autoencoders with different backbones, including To this end, utilizing AI-based technologies and smart 1-dimensional CNN and TCN, to detect abnormal usage. homes as novel interventions to recognize abnormal For this purpose, any predicted value that is greater than power utilization and understand the reasons for each ab- twice the standard deviation of the electricity consump- normality could pave the way for end-consumers both to tion the day before is considered abnormal. renovate wasteful devices and adopt a more sustainable The rest of the paper is organized as follows. We pro- energy consumption behavior [4][5][6][7]. Moreover, it vide related work on anomaly detection in energy con- facilitates the prediction of end-users power demand as sumption in Section 2. In Section 3, our methodology is well as performing an optimal energy distribution by presented in details. Section 4 concludes the paper and grid operators depending on specific end-users’ needs. discusses future work. In addition, electrical anomalies are less likely to remain unnoticed for a long period of time which would result in 2. Related work AMLTS’22: Workshop on Applied Machine Learning Methods for Time Series Forecasting, co-located with the 31st ACM International Con- In the context of energy usage, anomalies are defined as ference on Information and Knowledge Management (CIKM), October deviations from expected behavior that occur when the 17-21, 2022, Atlanta, USA * Corresponding author. consumption of a household appliance does not corre- $ somayeh.zamani@uni-siegen.de (S. Zamani); spond with its normal pattern [3][11]. Among the key hamed.talebi.aut@gmail.com (H. Talebi); applications of anomaly detection by load monitoring, gunnar.stevens@uni-siegen.de (G. Stevens) are forecasting maintenance and energy efficiency [12]. Β© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). Thus, a smart plug, smart appliance, and other appliance- CEUR Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org) method, the feature representations are enforced to learn important regularities of the data so that reconstruction errors are minimized. Consequently, anomalies are dif- ficult to reconstruct from the resulting representations and are, therefore, subject to large reconstruction errors [17]. 3. Methodology 3.1. Dataset and preprocessing The REFIT Electrical Load Measurements dataset con- tain cleaned electrical consumption data in Watts for 20 households in the UK at both the aggregate and appli- Figure 1: The architecture of the CNN-based autoencoder ance level. The data is related to a period of two years (CNN-AE) comprising nine individual appliance measurements at 8-second intervals per house with 1,194,958,790 readings [16]. The models proposed in this paper are trained using level monitoring devices are needed to continuously mon- dishwasher data from houses No. 1 and 2. Furthermore, itor the power consumption of individual appliances in data from the Fridge_Freezer and the Freezer of house a house [8]. However, identifying anomalies, and their No. 3 is used to assess the effectiveness of our approach. nature of them should also be considered, which can be To begin with, it is necessary to resample the data categorized, based on different dimensions. In the data to convert it into equal time intervals r. Then using science world, anomalies are seen as either single points the following formula, the average sampling time, 𝑑 of that are not necessarily relevant to each other or a set the REFIT data is used to fill in a limited number of of data points that constitute a pattern and, therefore, signals, n with no data. The remaining empty intervals can be interpreted in relation to each other. The other are substituted with zero. dimension of anomaly detection that should be taken [οΈ‚ ]οΈ‚ 4*𝑑 into account is the context which refers to a deviation in 𝑛= (1) π‘Ÿ a particular context relating to the structure of the data [3][13][14]. For example, in the context of a warm sea- Additionally, for devices used according to users’ needs, son, a temperature report of -30 degrees Celsius can be consumption data must first be differentiated. The power anomalous; however, during a cold season, such a report consumption pattern may include turning the device on may be more common [3]. To this end, understanding and off several times per usage. The matching data is the available data will provide a solid foundation for im- therefore combined into relevant signals. Also, due to proving energy efficiency. For this purpose, there are the possibility of failure in some devices that can result thirty-one publicly available databases with several fea- in constant operation for an extended period, we assume tures, such as the geographical location, period of col- a maximum period for a device to operate. lection, number of monitored households, the sampling rate of collected data, and number of sub-metered appli- ances [15]. Regarding this, a valuable dataset is REFIT which includes cleaned electrical consumption in Watts for 20 households in the UK at both the aggregate and appliance levels [16]. On the other hand, Pang [11] has provided a comprehensive overview of current anomaly detection methods to gain an important understanding of their inherent capabilities and limitations in addressing some largely unsolved challenges in anomaly detection. Figure 2: The architecture of the TCN-based autoencoder According to his study, Autoencoders, which are a sub- (TCN-AE) set of the generic normality feature learning category, aim to learn some low-dimensional feature representa- tion space on which the given data instances can be well 3.2. The proposed models reconstructed. While this is a widely used method for data compression or dimension reduction, by using this The development of time series anomaly detection al- gorithms has recently received considerable attention. Autoencoder-based approaches are often used to identify by a factor of 𝑠. To do so, groups of size 𝑠 are averaged anomalous behavior by analyzing the reconstruction er- along the time axis. ror of the data [18][19][20]. Having learned a nonlinear In the decoder module, the downsampled sequence transformation of the input data into a compressed rep- is returned to its original length by performing a near- resentation, latent variables are used to reconstruct the est neighbor interpolation on the upsampled sequence. original input. On the other hand, utilizing the convolu- Upsampled sequences are passed through a second TCN tion mechanism in sequential models is computationally with independent weights parameterized similarly to the optimal [21]. Also, due to CNNs’ equivariance proper- encoder-TCN. As a final step, the input sequence is re- ties and sparse interactions, they are translated from constructed with a Conv1D layer that ensures that the computer vision into the time domain using temporal dimensionality of the input is matched (by setting π‘˜ = 1 convolutional networks (TCN)[20]. and 𝑛𝑓 π‘–π‘™π‘‘π‘’π‘Ÿπ‘  = 𝑑) [20]. As described in the next section, In the following sections, we will describe how we the input sequence and its reconstruction will be used used autoencoders (AEs) for time series data that utilize for detecting anomalies after TCN-AE has been trained. 1-dimensional CNNs and TCNs as building blocks to detect energy anomalies in the REFIT dataset. 3.2.1. CNN-based autoencoder (CNN-AE) We used TensorFlow to implement the architecture con- sisting of two smaller sequential models, an encoder and a decoder. Also, considering the speed of the model con- vergence, our CNN-based autoencoder is comprised of 3 layers of Conv1D using the data of the households’ dishwashers. Furthermore, a nonlinear ReLu activation function is used in each convolution layer. In this model, a standard rate of 0.2 is considered for the dropout layer to randomly remove 20% of the upper layer during learn- ing. Figure 1. shows the layers and the number of input and output parameters of each. The input layer is 320 Γ— 1 Figure 3: (a) Examples of the normal energy consumption (3200 seconds), calculated according to the maximum of the dishwasher (b) Examples of the abnormal energy con- operation time of the device. sumption of the dishwasher 3.2.2. TCN-based autoencoder (TCN-AE) The temporal convolutional network (TCN) combines 3.3. Experimental results simplicity with auto-regressive prediction, residual 3.3.1. Anomaly detection blocks, and a very long memory. In general, a TCN can be broken down into three components: a list of dilation We compute a threshold value of 2𝜎 above the predicted rates 𝐷 = {π‘ž1 , π‘ž2 , ..., π‘žπ‘›π‘Ÿ }, the number of 𝑛𝑓 π‘–π‘™π‘‘π‘’π‘Ÿπ‘  , and value to measure the trend in electricity consumption the kernel size π‘˜, which is the same for all filters in a TCN over time, where 2𝜎 is the standard deviation on the [20][22]. Inspired by a classical (deep) autoencoder, the day before the actual moment [23]. An abnormal state is TCN autoencoder encodes sequences, along the temporal defined as a value exceeding the threshold for predicted axis, of length 𝑇 into a compressed representation of electricity consumption at the actual moment. Equations length 𝑇 /𝑠 (where 𝑠 ∈ Z+ ) and then tries to reconstruct (2) and (3) show the calculation of 𝜎 and π›Ύπ‘‘β„Žπ‘Ÿπ‘’π‘ β„Žπ‘œπ‘™π‘‘ : the original sequence[20]. βˆšοΈ‚ βˆ‘οΈ€π‘› 𝑖=1 (π‘₯𝑖 βˆ’ π‘₯) 2 In Figure 2, each layer of the TCN-AE is described 𝜎= (2) by its parameters within the box. TCN-AE receives a 𝑛 sequence π‘₯[𝑛] of length 𝑇 and dimensionality 𝑑 as its π›Ύπ‘‘β„Žπ‘Ÿπ‘’π‘ β„Žπ‘œπ‘™π‘‘ = 𝛾 Μ‚οΈ€ + 2𝜎 (3) input. Using a TCN, the encoder first processes input sequence π‘₯[𝑛] of length 𝑇 and dimension 𝑑. Afterward, a where 𝜎 is the standard deviation, π›Ύπ‘‘β„Žπ‘Ÿπ‘’π‘ β„Žπ‘œπ‘™π‘‘ is the one-dimensional convolutional layer with π‘ž = 1, π‘˜ = 1, threshold, 𝛾 Μ‚οΈ€ is the predicted value, π‘₯𝑖 is the electricity and 𝑛𝑓 π‘–π‘™π‘‘π‘’π‘Ÿπ‘  = 8 is used to reduce the dimensionality of consumption, π‘₯ is the average electricity consumption, the TCN’s output. As the last layer in the encoder, the and 𝑛 is the number of samples. temporal average pooling layer downsamples the series A normal electricity usage pattern detection for the dishwasher is shown in Figure 3 (a), where the real-time Table 1 Evaluation of CNN1D-AE performance with a different data division ratio Model Division Ratio MAE MAPE CNN1D-AE 9:1 0.1781 %22.55 CNN1D-AE 8:2 0.1570 %20.07 CNN1D-AE 7:3 0.1702 %21.99 Table 2 Evaluation of TCN-AE performance with a different data division ratio Model Division Ratio MAE MAPE TCN-AE 9:1 0.1527 %21.39 TCN-AE 8:2 0.1371 %17.52 TCN-AE 7:3 0.1412 %19.88 threshold curve follows the sequence trend, indicating tection techniques, abnormal behaviors can be mitigated. the model depicts the dishwasher’s normal electricity us- This is possible, especially if user-centric explainable age. As can be seen in the figure, the real power consump- recommender systems are combined with anomaly de- tion curve does not exceed the threshold range, which tection modules. However, there is no proper labeled indicates a normal level of electricity consumption. As dataset available to develop accurate algorithms or de- shown in Figure 3 (b), anomalous consumption patterns tect different types of anomalies. Accordingly, we plan to occur when actual values exceed the threshold. Conse- run a laboratory to build the first appropriately labeled quently, the method can distinguish between normal and energy anomaly dataset. abnormal consumption behavior. 3.3.2. Evaluation Acknowledgments As Table 1. and Table 2. show, for both architectures, the This research has been funded by the EU Horizon 2020 best performance is obtained with the data division ratio Marie SkΕ‚odowska-Curie International Training Network of 8:2, and clearly, TCN-AE is more efficient than CNN- GECKO, Grant number 955422. AE. Our unsupervised approach has also been evaluated using the data from the Fridge_Freezer and the Freezer of house No. 3 in REFIT. The results in Table 3. confirm References the best division ratio of 8:2 and the higher performance [1] M. Pothitou, A. J. Kolios, L. Varga, S. Gu, A of TCN compared to CNN1D. framework for targeting household energy savings through habitual behavioural change, International 4. Conclusion and future work Journal of Sustainable Energy 35 (2016) 686–700. [2] D. Marikyan, S. Papagiannidis, E. Alamanos, A This paper presents the starting point of our work on systematic review of the smart home literature: A studying how would applying deep learning algorithms, user perspective, Technological Forecasting and and explainability improve energy efficiency, environ- Social Change 138 (2019) 139–154. mental sustainability, and user adoption. In this regard, [3] S. S. Hosseini, K. Agbossou, S. Kelouwani, A. Carde- first, we preprocessed our data by resampling and differ- nas, N. Henao, A practical approach to residential entiating each usage. Next, the extracted patterns of dish- appliances on-line anomaly detection: A case study washer usage in houses No. 1 and 2 of the REFIT dataset of standard and smart refrigerators, IEEE Access 8 were analyzed. Two deep learning models, CNN-AE and (2020) 57905–57922. TCN-AE were then trained to detect abnormalities. While [4] T. Schwartz, S. Denef, G. Stevens, L. Ramirez, the TCN backbone performed better, we evaluated our V. Wulf, Cultivating energy literacy: results from a models using the data from the refrigerators of house No. longitudinal living lab study of a home energy man- 3 in REFIT as well. agement system, in: Proceedings of the SIGCHI Through the implementation of energy monitoring Conference on Human Factors in Computing Sys- systems and the formulation of intelligent anomaly de- tems, 2013, pp. 1193–1202. Table 3 Evaluation of the model’s performance using the Fridge_Freezer and the Freezer data of house #3 Model Division Ratio MAE MAPE CNN1D-AE 9:1 0.1678 %21.15 CNN1D-AE 8:2 0.1486 %19.17 CNN1D-AE 7:3 0.1649 %20.29 TCN-AE 9:1 0.1422 %18.21 TCN-AE 8:2 0.1264 %16.33 TCN-AE 7:3 0.1353 %17.70 [5] Y. Himeur, K. Ghanem, A. Alsalemi, F. Bensaali, [16] D. Murray, L. Stankovic, V. Stankovic, An electri- A. Amira, Artificial intelligence based anomaly cal load measurements dataset of United Kingdom detection of energy consumption in buildings: A re- households from a two-year longitudinal study, Sci- view, current trends and new perspectives, Applied entific data 4 (2017) 1–12. Energy 287 (2021) 116601. [17] M. Hasan, J. Choi, J. Neumann, A. K. Roy- [6] S. Chen, T. Liu, F. Gao, J. Ji, Z. Xu, B. Qian, H. Wu, Chowdhury, L. S. Davis, Learning temporal reg- X. Guan, Butler, not servant: A human-centric ularity in video sequences, in: Proceedings of the smart home energy management system, IEEE IEEE conference on computer vision and pattern Communications Magazine 55 (2017) 27–33. recognition, 2016, pp. 733–742. [7] B. Zhou, W. Li, K. W. Chan, Y. Cao, Y. Kuang, [18] S. Hawkins, H. He, G. Williams, R. Baxter, Outlier X. Liu, X. Wang, Smart home energy management detection using replicator neural networks, in: In- systems: Concept, configurations, and scheduling ternational Conference on Data Warehousing and strategies, Renewable and Sustainable Energy Re- Knowledge Discovery, Springer, 2002, pp. 170–180. views 61 (2016) 30–40. [19] P. Malhotra, A. Ramakrishnan, G. Anand, L. Vig, [8] M. Castangia, R. Sappa, A. A. Girmay, C. Camarda, P. Agarwal, G. Shroff, LSTM-based encoder- E. Macii, E. Patti, Detection of Anomalies in House- decoder for multi-sensor anomaly detection, arXiv hold Appliances from Disaggregated Load Con- preprint arXiv:1607.00148 (2016). sumption, in: 2021 International Conference on [20] M. Thill, W. Konen, T. BΓ€ck, Time series encodings Smart Energy Systems and Technologies (SEST), with temporal convolutional networks, in: Inter- IEEE, 2021, pp. 1–6. national Conference on Bioinspired Methods and [9] A. Alsalemi, Y. Himeur, F. Bensaali, A. Amira, Their Applications, Springer, 2020, pp. 161–173. Appliance-level monitoring with micro-moment [21] Y. Kim, Convolutional neural networks for sen- smart plugs, arXiv preprint arXiv:2012.05787 (2020). tence classification, arXiv preprint arXiv:1408.5882 [10] K. Das, Detecting patterns of anomalies, Carnegie (2014). Mellon University, 2009. [22] S. Bai, J. Z. Kolter, V. Koltun, An empirical evalua- [11] G. Pang, C. Shen, L. Cao, A. V. D. Hengel, Deep tion of generic convolutional and recurrent net- learning for anomaly detection: A review, ACM works for sequence modeling, arXiv preprint Computing Surveys (CSUR) 54 (2021) 1–38. arXiv:1803.01271 (2018). [12] I. Ullah, F. Yang, R. Khan, L. Liu, H. Yang, B. Gao, [23] H. Pan, Z. Yin, X. Jiang, High-Dimensional En- K. Sun, Predictive maintenance of power substa- ergy Consumption Anomaly Detection: A Deep tion equipment by infrared thermography using Learning-Based Method for Detecting Anomalies, a machine-learning approach, Energies 10 (2017) Energies 15 (2022) 6139. 1987. [13] V. Chandola, A. Banerjee, V. Kumar, Anomaly De- tection: A Survey. ACM Computing Surveys, vol 41 (2009) 15. [14] X. Song, M. Wu, C. Jermaine, S. Ranka, Conditional anomaly detection, IEEE Transactions on knowl- edge and Data Engineering 19 (2007) 631–645. [15] Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, Build- ing power consumption datasets: Survey, taxonomy and future directions, Energy and Buildings (2020) 110404.