Elsevier

Big Data Research

Volume 31, 28 February 2023, 100360
Big Data Research

Predicting Household Electric Power Consumption Using Multi-step Time Series with Convolutional LSTM

https://doi.org/10.1016/j.bdr.2022.100360Get rights and content

Abstract

Energy consumption prediction has become an integral part of a smart and sustainable environment. With future demand forecasts, energy production and distribution can be optimized to meet the needs of the growing population. However, forecasting the demand of individual households is a challenging task due to the diversity of energy consumption patterns. Recently, it has become popular with artificial intelligence-based smart energy-saving designs, smart grid planning and social Internet of Things (IoT) based smart homes. Despite existing approaches for energy demand forecast, predominantly, such systems are based on one-step forecasting and have a short forecasting period. For resolving this issue and obtain high prediction accuracy, this study follows the prediction of household appliances' power in two phases. In the first phase, a long short-term memory (LSTM) based model is used to predict total generative active power for the coming 500 hours. The second phase employs a hybrid deep learning model that combines convolutional characteristics of neural network with LSTM for household electrical energy consumption forecasting of the week ahead utilizing Social IoT-based smart meter readings. Experimental results reveal that the proposed convolutional LSTM (ConvLSTM) architecture outperforms other models with the lowest root mean square error value of 367 kilowatts for weekly household power consumption.

Introduction

In recent years, the topic of social Internet of things (IoT) has become an emerging IoT field of great interest. IoT is now an established paradigm that supports a variety of applications and services [19]. Social IoT is based on introducing a new vision of cooperation between smart objects in a similar way to what happens for humans when they communicate to achieve a common goal. So, several interconnected IoT devices can relate to each other and provide smart services in different contexts. The sectors that can benefit from this new paradigm and the connection of different smart objects to obtain increasingly pervasive and effective answers to their problems are those of transport, water, energy, etc. The potential of energy efficiency management and consumption forecasting has gradually been recognized by governments and energy research institutes as an important part of sustainable development. With the increase in population and living standards of citizens, household energy consumption is increasing steadily [63]. Power generation and distribution systems keep trying to maintain a balance between electricity demand and supply. According to the report of the annual energy outlook 2020, the annual growth of electricity demand averages about 1% over the 2019–2050 period [11]. A 90% of the time is spent indoors indicating the energy requirements of the buildings [48]. As a result, 80% to 90% of the buildings' life cycle is spent on people's occupational activities and comfort [59]. Consequently, the building sector consumes the most significant portion, i.e., 39% of the total energy consumption globally [40], and has 38% greenhouse gas emissions [49]. Electricity cannot be stored due to its physical characteristics, it must be consumed as it is generated in the plant [26]. With the continuous increase in the population and increased demand for energy, power consumption forecasting systems have become increasingly important recently.

Fluctuation of electric power consumption mainly depends upon the number of household electric items and the behavior of residents. Household electric consumption is derived from household occupancy. To implement household energy disaggregation based on occupancy aiming at automatic energy reduction is studied by utilizing IR ultra-wideband (UWB) radar and electrical signals [10]. Authors estimated the household occupancy using smart meter power consumption dataset [6]. They suggested an expandable and personalized energy efficiency method and also gave a new idea for the implementation of Social IoT-based smart metering infrastructure. However, consumer privacy concern is the most challenging issue for policymakers in smart metering implementation.

The performance of machine learning (ML) models mainly depends upon the data representation. While deep learning (DL) deals with the nonlinear transformation that provides high-level abstraction and eventually more profit [8]. DL techniques have been extensively used in various applications [44] and convolutional neural networks (CNN) surpass other existing methods in image classification [54] by preserving relationships between pixels. A recurrent neural network (RNN) is superior in natural language processing (NLP) tasks by storing sequential information. RNN also ensures that time-series information can be retained [17]. Long short-term memory (LSTM), a variant of RNN, is being used to extract spatial and temporal features in combination with CNN. The electrical energy consumption prediction problem is a time series problem. Kim et al. predicted energy consumption by combining CNN with LSTM [31]. They achieved higher performance than existing approaches by using a hybrid deep learning model. They also analyzed the variables influencing the prediction of power consumption. However, it is difficult for the traditional deep learning approaches to model spatial and temporal features of energy consumption.

Forecasting of electric energy consumption is a multivariate time series problem that predicts power consumption. However, these irregular seasonal trends of power consumption make it difficult for prediction methods to predict electric energy consumption [1]. Dataset provided by the UCI repository consists of seven variables and power consumption sampling for 2007–2011 and is considered a benchmark dataset in time series forecasting. These time series were collected from various Social IoT devices such as smart meter readings and used, in this study, as input for convolutional LSTM (ConvLSTM), a hybrid deep learning model that combines characteristics of CNN with LSTM. So, ConvLSTM architecture is used for electrical energy consumption prediction, where ConvLSTM-two dimensional (ConvLSTM-2D) finds a correlation of multivariate variables and LSTM layers build time series based temporal information to generate the final prediction. LSTM layer output is fed to the fully connected layer that finally predicts power demand. In this work, multi step-time series power consumption problem is explored that is to estimate the expected electric power consumption for the next week by using recent consumption. Then forecast the total active power of each day to the next week by the predictive model. This work aims at helping the planning expenditure and fulfilling the electricity demands of a single household on the supply side. The main contributions of this study are:

  • A hybrid deep learning framework is designed which combines CNN and LSTM models to predict the weekly power consumption for single households. It is built to predict total generative active power for the coming 500 hours.

  • Contrary to existing approaches that are single-tier, this study follows a two-phase approach of predicting generative active power in the first phase using LSTM and energy consumption prediction in the second phase using proposed ConvLSTM.

  • Multi-step forecasting is used to facilitate application usage in the real world, such as social IoT-based smart grid planning. To prove the significance of the proposed model, extensive experiments are performed on the power consumption of a single household dataset.

The rest of the paper is organized as follows. Section 2 discusses a few pieces of research related to the current study. Section 3 presents an overview of the methodology adopted for the current research as well as a detailed description of the tweet dataset used for the experiment. Results are presented in Section 4. Section 5 describes other possible Social IoT applications of the sensors used when they are integrated into a larger Social IoT architecture. Finally, Section 6 discusses conclusion and future work.

Section snippets

Related work

The selection of the time series forecasting method depends upon many factors, like the availability of relevant datasets, desired accuracy, and so forth. Conventional statistical-based approaches have been used by many researchers to solve time series problems. With the immense progress in ML and artificial intelligence, researchers are now utilizing deep learning models to forecast electric load problems in Social IoT platforms. In recent years, Social IoT has become a hot topic, of great

Proposed methodology

The proposed hybrid neural network model is based on CNN and LSTM layers. Fig. 1 depicts the suggested architecture and the two phases implemented for predicting global active power and weekly residential energy consumption. Phase I uses an LSTM-based model to forecast the upcoming 500 hours of global active power. While in phase II, the ConvLSTM model predicts weekly electricity usage using time series data collected from Social IoT smart meter readings. ConvLSTM-2D layer captures spatial

Experimental results and discussions

In this section, the electric power consumption dataset used in the experiment, experimental setup, and results of both phases of the proposed approach are discussed. In phase I, global active power consumption is predicted using LSTM. In phase II, single household power consumption is forecasted for a week by leveraging the information made available by Social IoT smart devices. To achieve the goal of energy prediction for the week ahead, experiments are performed in five scenarios: (i) LSTM

A possible application scenario in the social IoT

The smart environment analyzed in this work is equipped with an ecosystem of smart meter readings that collect data on the electricity consumption of a home. However, the smart system described can be easily integrated into a wider Social IoT architecture. This new environment can associate with the devices used in other heterogeneous smart sensors and objects, both in terms of various components and characteristics but also in terms of appropriate programming languages to interact with them.

Conclusion

In an ecosystem of smart devices able to communicate with each other to solve similar problems, the concept of Social IoT is born and developed. In this paradigm, the smart objects are autonomous in their communication with other ones for their need. Nowadays, there is a growing interest in using this concept in different sectors such as transportation, utilities, home automation, and so on. In this study, deep learning models as five different scenarios are utilized to forecast electrical

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (65)

  • P. Nejat et al.

    A global review of energy consumption, CO2 emissions and policy in the residential sector (with an overview of the top ten CO2 emitting countries)

    Renew. Sustain. Energy Rev.

    (2015)
  • N. Somu et al.

    A hybrid model for building energy consumption forecasting using long short term memory networks

    Appl. Energy

    (2020)
  • C. Spandagos et al.

    Equivalent full-load hours for assessing climate change impact on building cooling and heating energy consumption in large Asian cities

    Appl. Energy

    (2017)
  • H. Takeda et al.

    Using the ensemble Kalman filter for electricity load forecasting and analysis

    Energy

    (2016)
  • D. Wang et al.

    DeepSTCL: a deep spatio-temporal ConvLSTM for travel demand prediction

  • Z. Wang et al.

    Random forest based hourly building energy prediction

    Energy Build.

    (2018)
  • G. Zhao et al.

    Energy consumption in machining: classification, prediction, and reduction strategy

    Energy

    (2017)
  • K. Zhou et al.

    Understanding household energy consumption behavior: the contribution of energy big data analytics

    Renew. Sustain. Energy Rev.

    (2016)
  • M. Ahmad

    Seasonal decomposition of electricity consumption data

    Rev. Integr. Bus. Econ. Res.

    (2017)
  • R. Ahmad et al.

    A machine-learning based ConvLSTM architecture for NDVI forecasting

    Int. Trans. Oper. Res.

    (2020)
  • M. Alhussein et al.

    Hybrid CNN-LSTM model for short-term individual household load forecasting

    IEEE Access

    (2020)
  • I. Ashraf et al.

    Application of deep convolutional neural networks and smartphone sensors for indoor localization

    Appl. Sci.

    (2019)
  • S.N. Azzah et al.

    Resolving energy consumption issues and spectrum allocation for future broadband networks

    IEEE Access

    (2021)
  • G. Bedi et al.

    Development of an IoT driven building environment for prediction of electric energy consumption

    IEEE Int. Things J.

    (2020)
  • Y. Bengio et al.

    Representation learning: a review and new perspectives

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2013)
  • A. Bogomolov et al.

    Energy consumption prediction using people dynamics derived from cellular network data

    EPJ Data Sci.

    (2016)
  • B.P. Center, Annual energy outlook 2020,...
  • T. Chai et al.

    Root mean square error (RMSE) or mean absolute error (MAE)? – Arguments against avoiding RMSE in the literature

    Geosci. Model Dev.

    (2014)
  • Y. Chakhchoukh et al.

    Electric load forecasting based on statistical robust methods

    IEEE Trans. Power Syst.

    (2010)
  • K.T. Chui et al.

    Predicting students' performance with school and family tutoring using generative adversarial network-based deep support vector machine

    IEEE Access

    (2020)
  • F. Divina et al.

    Hybridizing deep learning and neuroevolution: application to the Spanish short-term electric energy consumption forecasting

    Appl. Sci.

    (2020)
  • K. Greff et al.

    LSTM: a search space odyssey

    IEEE Trans. Neural Netw. Learn. Syst.

    (2016)
  • Cited by (14)

    View all citing articles on Scopus
    View full text