Employing electricity-consumption monitoring systems and integrative time-series analysis models: A case study in Bogor, Indonesia

Abstract

The Paris Agreement calls for maintaining a global temperature less than 2°C above the pre-industrial level and pursuing efforts to limit the temperature increase even further to 1.5°C. To realize this objective and promote a low-carbon society, and because energy production and use is the largest source of global greenhouse-gas (GHG) emissions, it is important to efficiently manage energy demand and supply systems. This, in turn, requires theoretical and practical research and innovation in smart energy monitoring technologies, the identification of appropriate methods for detailed time-series analysis, and the application of these technologies at urban and national scales. Further, because developing countries contribute increasing shares of domestic energy consumption, it is important to consider the application of such innovations in these areas. Motivated by the mandates set out in global agreements on climate change and low-carbon societies, this paper focuses on the development of a smart energy monitoring system (SEMS) and its deployment in households and public and commercial sectors in Bogor, Indonesia. An electricity demand prediction model is developed for each device using the Auto-Regression eXogenous model. The real-time SEMS data and time-series clustering to explore similarities in electricity consumption patterns between monitored units, such as residential, public, and commercial buildings, in Bogor is, then, used. These clusters are evaluated using peak demand and Ramadan term characteristics. The resulting energy-prediction models can be used for low-carbon planning.

Graphical abstract

Keywords

electricity monitoring / electricity demand prediction / multiple-variable time-series modeling / time-series cluster analysis / Indonesia

1 Introduction

1.1 Background and objectives

To attain a low-carbon society, it is necessary to transform the centralized energy system into distributed systems at city and regional scales. Because energy demand patterns vary spatially, more detailed data on energy demand provided by innovative Information Communication Technologies (ICTs) is expected to enable local energy demand and supply system optimization in which distributed renewable energy resources can be integrated with large-scale grid energy supply systems.

Energy information and data at local scales, particularly in developing countries, is persistently unavailable. However, there is enormous potential to reduce energy use in various sectors through the use of rapidly developing ICT systems in energy management. The innovative system described below would be feasible in applications such as hourly demand response and multiple energy supplementation.

ICT and Internet of Things (IoT) technological development have recently enabled real-time energy monitoring at multiple points, which will soon be possible to analyze mid- to long-term energy demand by assembling multi-sectoral and spatially variable data into shared knowledge databases. Social monitoring systems for energy consumption can also be used by local governments to verify urban and regional statistical data. The development of mid- to long-term energy demand prediction models using statistical methods and monitored data are also expected. In particular, it is important to consider electricity storage at non-peak times and differences in electric power demand over time in order to create models that can predict short-term (e.g., diurnal) changes and be applied to long-term data. In addition to developing predictive models for urban and regional level planning, the models are evaluated using time-series characteristics. It is also desirable to create a database of relationships between energy use and spatial characteristics, etc., using the validated models.

To predict long-term electricity consumption, which is influenced by various factors, the auto-regressive, multivariable Auto-regression eXogenous (ARX) model was developed to project changes in exogenous effects [1]. In this paper, a real-time electricity consumption monitoring system is developed which is useable in the Indonesian social environment and demonstrative monitoring systems suitable for a range of urban sectors, such as residential units, business offices, and retail shops in Bogor, Indonesia. The system consists of monitoring devices that collect sequential electricity demand data in the power distribution board.

Based on the real-time data obtained from the monitoring system, an electricity demand prediction model is developed for each unit using the auto-regressive exogenous model, which includes climate and social factors. This model features both multi-regression and auto-regression characteristics and is used to address energy demand data gaps in Indonesia via the following two methods. First, the data collection system involving the innovative energy monitoring system and developed prediction model is discussed more extensively than in Ref. [1]. Secondly, hourly and long-term electricity demand patterns are developed for the monitoring targets using time-series clustering. The objective is to expand the basic information in the database using the data collection system and data analysis methods developed herein so that the database can be used in future low carbon and climate scenario planning.

1.2 Literature review

Hargreaves et al. [2] introduced a monitor in the UK and examined the use of visualization in saving energy at the city level. However, they targeted only households. Kindaichi et al. [3] analyzed the relationships between building energy demand and time and climatic factors using multiple regression analysis in order to predict energy demand. Kawano et al. [4] developed an electricity demand prediction equation using multi-regression analysis, and Kondou et al. [5] developed an electricity demand prediction equation using auto-regression analysis. However, these predictive models cannot consider the data from multiple electric devices. Thus, these models are not sufficient for low-carbon planning. Using multi-variable time-series analysis, Ma et al. [6] developed energy demand patterns using the ARX model for a heating, ventilation, and air conditioning system and analyzed the economic aspects of effective energy saving controls. However, they involved only an imaginary system, not the real electricity demand.

In Indonesia specifically, Utma and Gheewala [7] investigated life-cycle electricity use in high-rise apartment buildings. Permana et al. [8] investigated electricity demand dependence on household sector using urban development patterns in Bandong. However, both studies involved estimations using energy bills and did not consider the hourly demand. Chou and Gusti Ayu Novi Yutami [9] developed a psychological model for household smart meter installation and considered smart meter introduction strategies. However, they also targeted only residential energy use. The monitoring data used for low-carbon planning must involve multi-sector data collection.

The selection of time-series data clustering methods do exist. The dynamic time warping (DTW) [10] clustering method can be used to measure similarities between two temporal sequences. Sakoe and Chiba [10] applied DTW to spoken word recognition analysis. Nakamoto et al. [11] applied time-series data compression. However, this method has not been applied to analyses of real buildings and monitoring data devices.

In this paper, innovative electricity demand monitoring systems were developed and installed in buildings including research offices, hotels, and shopping malls, and collected sequential electricity demand data for individual devices such as air conditioners, lighting, and refrigerators. Electricity demand prediction models were then developed using the monitoring data. Lastly, a time-series cluster analysis was applied using the DTW method to validate the electricity patterns and evaluate the potential for application at an urban scale.

2 Methodology

2.1 Analytical procedure

A prediction model was developed based on the real-time electricity demand monitoring information as shown in Fig. 1.

First, a real-time electricity demand monitoring system was developed and installed in selected public, commercial, and residential buildings in Bogor, Indonesia. Numerous public buildings and a summer refuge were included in the monitoring campaign. Second, influential social and climate factors were investigated using the monitoring results in a multi-regression analysis.

An electricity demand prediction formula was developed that uses factors with over 5% significance in a multi-regression analysis in the ARX model, which is a multi-variant time-series analytical method; this method considers multi-variable effects using multi-regression analysis and considers the effects of past electricity demand using auto-regression analysis. The analysis spans from March 1, 2015 to January 31, 2016 for 2015-start monitored buildings, and from April 1 to December 31, 2016 for 2016-start monitored buildings.

The prediction models were verified using error rate indexes and a 1-to-1 line. The time-series cluster analysis was also applied to confirm the similarity of hourly and long-term trends using the energy pattern of average number of 1 for the target buildings.

2.2 Case study area background

Figure 2 shows the final energy use and consumption by sector in Association of South-east Asian Nations (ASEAN) countries and Japan [12]. In 2013, the total energy use in Indonesia was about 1.61 × 10⁸ t-oil equivalent, the CO₂ emissions of which are relatively large compared to other ASEAN countries. The share of energy consumption of the residential, commercial, agriculture, etc., sectors is large, accounting for nearly 40%. The consumption of these sectors is higher in highly populous countries such as Vietnam and the Philippines than in other countries. Figure 3 displays expected CO₂ emissions from 1990 to 2040 in Association of South-east Asian Nations (ASEAN) and Japan [12]. In 1990 and 2013, Japanese CO₂ emissions are large which are several times that of any of the ASEAN. However, Indonesian CO₂ emissions are expected to surpass those of Japan in 2040 at an equivalent of 2% of world emissions. In addition, consumption is increasing more quickly in highly populous countries such as Vietnam and the Philippines than in other ASEAN. There is a great potential to reduce energy demand and CO₂ emissions, especially in the commercial, public, and residential sectors in Indonesia.

Bogor is located approximately 50 km south of Jakarta and has an area of 21.56 km² and a population of approximately 1 million. Bogor is home to developed small industries pertaining mainly to food and beverage production and a large number of employees that commute to the Jakarta metropolitan region. The service industry, which includes hotels and restaurants, has grown rapidly in recent years. Bogor was founded in the Dutch colonial period. It has typical public buildings such as the Presidential Palace, Botanical Garden, and Bogor Agricultural University (IPB).

2.3 Energy consumption monitoring system

An innovative monitoring system was developed that included a 3G line system and recovery of the hard collection time. A system overview is demonstrated in Fig. 4. The system, which monitored sequential electricity demand, was installed on the power panel board in each building using aggregation devices and a Fuji Multifunction Protector and Controller (FMPC, © Fuji Electric FA Components & Systems), a compact digital multipurpose relay device that offers protection, operation, measurement, monitoring, and transmission functionality for high-voltage power distribution facilities. At the time of installation, the auditory and breaker OFF/ON testing was used and the connected devices were verified. The monitored electricity demand was aggregated on the cloud server. The monitoring began in 8 buildings on March 1, 2015 and in 7 additional buildings in February 2016.

Households as well as commercial and public sectors were focused on. The residential sector constitutes approximately 36% of the electricity demand in Bogor and thus has a great potential for energy savings. Figure 5 depicts the spatial locations of target buildings in this paper. Power outages were necessary during equipment installation. During the initial stage of research, systems were preferentially installed in buildings willing to cooperate with this research, as they might be more apt to properly manage the equipment after installation.

Systems were installed in a total of 15 buildings, including 3 public offices, a research office that shared space with a business office, a hotel, a café, a shopping mall, and 7 residences with different family configurations. High-demand devices and areas were monitored such as air conditioning (AC), lighting, server rooms, guest rooms, etc. Table 1 lists target building information and monitoring periods.

2.4 Analytical models

An electricity demand prediction formula was developed using the ARX model, which is a multivariate time-series analysis method. Equation (1) shows the mathematical structure of this model. This model calculates the electricity demand at time t using the electricity demand at some past time i and exogenous factors.

(1) {{\displaystyle y}}_t={\displaystyle \sum _i^p{{\displaystyle M}}_{t-i}}{{\displaystyle X}}_{t-i}+b+{{\displaystyle \epsilon }}_t,

where y _t is the electricity demand at time t, M _t is the coefficient vector of the external variable at time t, X _t is the external variable including y _t at time t, ε _t is the error term at time t, b is a constant, and p represents the duration of the calculation. Because long-term effects were considered, a maximum p value of 10 was applied.

The most appropriate model was selected using the Akaike Information Index (AIC) and the time-series cluster analysis, the DTW, and the furthest neighbor method were applied for validation among the monitoring target buildings. Prediction results from March 1 to December 31, 2016 were used for the time-series clustering. The normalized (to 1) rate of hourly variability in a given day for hourly time-series clustering and the normalized (to the average monthly electricity demand) long-term variability rate in a given month for long-term time-series clustering were calculated.

Using the monitoring data as the dependent variable and social and climate factors as explanatory variables, the multiple regression analysis was applied to analyze the factors that affected electricity demand. A holiday dummy (1: Saturday, Sunday, and national holidays; 0: weekday) was applied to account for the influence of holidays on behavior, as well as an office hour dummy (1: 8:00–18:00 LST; 0: other times) and Islamic Ramadan dummy (1: June 18–July 16, 2015 and June 6–July 4, 2016; 0: other Day). These factors affect human behavior and device usage. Dummies for climate factors including temperature, humidity, and wind speed were also applied.

Climate factors were obtained from hourly open data from Open Weather Map in Bogor. However, these data contain anomalies, such temperatures as over 200°C; thus, the anomalous data, such as temperatures over 50°C, wind speeds over 30 m/s, and missing data, during analysis were omitted. As a result, the maximum temperature is 37.78°C and the maximum wind speed is 9.7 m/s. Figure 6 depicts the supplementation of missing temperature data. However, because the monitoring system had a long system failure time before June 2015, this term could not be complemented. Besides, because the monitoring data had a lot of data lack time in June 2015, complemented data in this term is lower in accuracy than those in the other term. However, it is considered that the other term was well complemented.

Serial correlations and homoscedasticity were explored using adjusted statistical methods. Normal regression analysis cannot be applied to monitoring data displaying heteroscedasticity [13]. Thus, the monitoring data found to display heteroscedasticity were adjusted using the Newey-West method [14]. Time-series analyses must use continuous data. However, the monitoring data feature some missing data points. Therefore, linear interpolation was used to generate missing data. These adjusted data were applied in the time-series analysis.

3 Results and discussion

3.1 Monitoring results

Figure 7 shows the monitoring results for each building from the beginning of monitoring to December 31, 2016. Explanations for confirmed events and anomalies are also described. However, because the university office includes corridors in the “other” category, and because the target range was not clarified, this target was omitted from the subsequent analysis. Monitoring failures at the city office, local council, city office branch, and shopping mall in the initial monitoring period prevented data collection. Frequent missing data and anomalies also occurred in initial monitoring period in other buildings, preventing the use of this data in the analysis. The research office, hotel, and shopping mall had no significant changes or periodic peaks in electricity demand except during the electric line construction period and the initial failure period.

Conversely, the city office, local council, and city office branch have large differences between peak and non-peak periods. The café features decreasing electricity demand after 2016. However, monitoring data collection was interrupted around May 2016 due to line cutting by mice. Due to problems in management, it was necessary to conduct surveys to explore energy saving effects. All of the residential locations have relatively stable electricity demand. Buildings with 2015-start monitoring have low accuracy terms, but the accuracy terms change between February 2016 and 1 to 2 months after. Residence 1 has periods during which consumption increases by several times the surrounding values. It is also possible that these periods represent data abnormalities. Residence 4 features reduced electricity demand. However, this residence underwent electricity line construction in 2016, and some monitoring devices were unable to monitor the electricity demand during that time. Residence 6 has periods during which consumption increases by a factor of two. Residence 7 has an unexplained decreased electricity demand from September 2016. The influence of variability and events during this period will be investigated in the future.

3.2 Results of multi-regression analysis

Table 2 tabulates the results of the multi-regression analysis, in which all of the external variables were adjusted using the Newey-West method. A significance of X% means that the given variable affects energy demand with a probability of X% over a random variable. The t value represents the magnitude of the influence of the explanatory variable on the dependent variable, where larger values denote larger influence. The coefficient of determination is an index of the similarity between the measured and predicted values. In general, it is represented by R ² and has a value of 0 to 1, where higher similarity produces values closer to 1.

The hotel and some of the residences, such as Residence 2, did not show significant differences between the office hour dummy and total electricity demand. Many buildings and devices do show significant differences between the office hour dummy, holiday dummy, and electricity demand. These results reflect the fact that electricity demand is affected by human activity. Thus electricity demand varies with the level of human activity.

Excluding the local council building, AC is significant at the 5% level, with large t values in many cases. Thus, AC is greatly affected by human activity much of the time. The AC electricity demand in commercial and public buildings features a large t value with temperature. However, residences with a low AC consumption are not significantly influenced by temperature. AC is a large part of the electricity demand of many buildings. Therefore, temperature also greatly affects the total electricity demand.

The t value of the intercept is large for continuously used devices such as servers and refrigerators, and the t values of temperature and time period were lower than that for AC. However, refrigerator utility is different in each building. The t value of residential refrigerators is lower than those for commercial and public refrigerators, and the refrigerators in Residences 2 and 3 do not have significant intercepts. Furthermore, the Ramadan and temperature t values in Residence 4 are larger than the intercept t value.

Similar to results found in past research [3,4], electricity demand is affected by human behavior and the thermal environment. The results confirm fixed electricity demand in devices such as refrigerators through engineered electricity demand calculations [15]. However, these devices are significantly affected by human activity and climate factors.

Using the results of the multi-regression analysis discussed above, the influential factors were checked. However, not all of the multi-regression analysis results reached R² = 0.6. These values are not sufficiently large. Thus, the multi-regression model is not appropriate for electricity demand prediction.

3.3 Results of multivariable time-series analysis

Here, the observed data were compared with the results of the ARX model time-series regression. The period from June 12, 2015 to January 31, 2016 were selected for analysis as the 2015 Start Group and from April 1 to December 31, 2016 as the 2016 Start Group. Using the minimum AIC model, the p number for all buildings was set to 10, excluding Residence 6. The p of Residence 6 was set to 3.

Figure 8 exhibits prediction results for 1 week (August 1 to 7, 2015) for the 2015 Start Group and 1 week (May 1 to 7, 2016) for the 2016 Start Group. For commercial and public buildings, the model R ² values are>0.7. Thus, these models are sufficiently significant for use. Hourly prediction results are different for each building in Fig. 8.

The electricity demand in the research office was dominated by AC and the server, and energy consumption prediction for these two devices is highly reproducible. Thus, the total energy use prediction is also highly reproducible. However, the predicted value is larger than the observed value immediately before the peak period. The total R ² value for the hotel is approximately 0.7, which is the smallest R ² value among the commercial and public buildings. This may arise from the multitude of peaks in a given week and the fact that the electricity patterns are not stable throughout a given day. As a whole, both peak time and peak electricity demand can be reproduced. Most of the electricity demand in the café arose from AC use. This demand is well reproduced when the peak time AC electricity demand is large. However, the reproducibility of the start time of increases or decreases in the AC electricity demand during peak times is low. The electricity demand patterns of the city office and local council are quite similar. Both models reproduce peak electricity demand. However, these models forecast excess demand on non-peak days and also produce a negative electricity demand number during non-peak times.

The city office branch has a total R ² value of approximately 0.9, and this model performs well, as seen in Fig. 8. Low electricity demand devices, such as receptacles, feature low reproducibility and departures at peak times. Thus, it is necessary to account for such deviations when using the model for energy saving evaluations. The shopping mall has a total R ² value of approximately 0.9. However, reduction in electricity use is predicted to occur sooner than observed, and these models feature low reproducibility without office hours, sometime calculating a negative value. These trends show large differences between model performance in peak and non-peak times. Thus, the prediction model for this building may not be suitable with only 1 equation.

The residential R ² values are lower than those of commercial and public values at R ² = 0.32–0.77. This may be caused by the fact that residential electricity demands were low, but measured using the same monitoring system installed in commercial and public buildings. Besides, these results may be influenced by unstable peaks and consumption patterns. The total R ² values for Residences 2 and 6 are very low because of multiple peaks in a day. On the other hand, Residences 1, 5, and 7 have total R ² values over 0.7 because they feature 1 peak in a day. The peak time reproducibility is high. However, non-peak time reproducibility is low. The predicted Residence 1 non-peak demand is larger than the observed demand. In Residence 3, AC use is not stable. The predicted results are larger than observations when the AC is not in use and lower than observations when the AC is being used. Thus, it is necessary to generate a prediction expression for AC use. Residence 4 has a pool, which produces unexpected consumption. Electricity demand was predicted even when the pool was not in use. Thus, the total predicted electricity demand is larger than the observed demand when the pool is not in use and lower than observations when the pool is being used. Thus, it is necessary to generate a prediction expression for pool use in Residence 3.

Table 3 shows the observed and predicted daily electricity demand per area and error rate index without the university office. The university office will be surveyed and calculated using the same method. The daily electricity demand per area is less than 3% without the shopping mall. Because the shopping mall electricity demand per area is approximately 8.5%, it is necessary to adjust these data before use. However, these predictions did not cause large deviations in daily data.

Table 3 shows the mean absolute percentage error (MAPE), the peak mean absolute percentage error (PMAPE), and the relative root mean square error (RRMSE) values. The MAPE values for commercial and public buildings are 6.9% – 50.1%, while residential MAPE values are 18.7% – 35.8%. As large as the city office MAPE is, these values are similar to past research results (12% – 42%) in Japan [5]. The large MAPE values of buildings with low PMAPE values indicate a large difference in electricity patterns between different time periods, such as peak and non-peak.

In the Agency for Natural Resources and Energy Guidelines [16], the relative root mean square error (RRMSE) is used as an accuracy index for deal demand response, where an index value of 20% or less is preferable. In this study, this index measures 10.5%–47.6%. The RRMSE values for the research office, hotel, and city office branch are 20% and under. Thus, in order to determine a deal demand response method, the prediction model must be improved using methods such as including a state equation.

Models can predict mid- and long-term electricity demand with a high accuracy, and can also predicted peaks. However, some of the model accuracies are insufficient for Negawatt deal demand response methods, and could be used only in electricity fee demand response unless the models are improved to include state variables.

3.4 Prediction equation relevance analysis

Figure 9 shows comparisons between ARX predicted results and observed valued. The black line is the 1-to-1 line, on which the predicted value is equal to the observed one. Much of the predicted data with large R ² values lie along the 1-to-1 line. Most of the research office data lie on the 1-to-1 line and are distributed in a single group. Thus, this model is adequate. However, outliers from the 1-to-1 line are more frequent than in other buildings. If the hourly demand control is to be used, these outliers must be analyzed. The hotel data also lie on the 1-to-1 line. However, these data are more disperse than the research office data. This is consistent with the small R ² value for the hotel data. The café, city office, local council, and city office branch are less disperse than the research office and hotel data. Many of these data fall along the 1-to-1 line. However, these buildings have electricity demands near 0 kWh/h, and the predicted results contain some negative values. In addition, the distribution near 0 kWh/h is more disperse than other parts of the distribution in the café and city office data. Thus, it is appropriate to make separate prediction models for the low electricity and high electricity groups. While the big shopping mall data also lie on the 1-to-1 line, this distribution is divided into 3 groups:<1000 kWh/h, 1000 – 1700 kWh/h, and>1700 kWh/h. It is thought that these categories represent groups with different determinants of electricity demand, and that the MAPE of this building is lower than the PMAPE. The influence of holidays and office hours is large, as evidenced by large t values in the multi-regression analysis.

Residential data are more disperse than are commercial and public building data, and the range of the predicted results tends to be smaller than the observations. This tendency is conspicuous in small-R² buildings like Residences 2 and 6. However, it is also seen in large-R² buildings like Residences 5 and 7. Thus, these data may not be consistently dispersed between the upper and lower groups. Most of the Residence 1 data are<1 kWh/h, but some data are>3 kWh/h. Thus, while the R² value is not bad, some groups may have different determinants of electricity demand. Some of the predicted Residences 1 and 5 results are negative. Thus, it is appropriate to make separate prediction models for groups such as the café and city office in order to increase model accuracy. Predictive models must have a given minimum predictive ability. Thus, the few models that produce disperse results must be improved before use in energy demand control and response.

3.5 Time-series cluster analysis results

Figure 10 shows the results from the DTW furthest neighbor method clustering analysis using predicted electricity demand trends from April to December 2016. The hourly variability is divided into 4 clusters by peak and non-peak electricity demand patterns. In Cluster 1, the peak electricity demand is about 1.6 times the average consumption and the minimum consumption is about half of the average. This group includes buildings with daytime peaks, such as the shopping mall and café, and nighttime peaks, such as residences. These types were not separated by statistical analysis. However, categories such as residential or commercial buildings can be separated. Cluster 2 features a smaller range than Cluster 1. The variability of Cluster 3 covers only a small range. This cluster changes only 10%–20%, in terms of increases and decreases, over the course of a day. Cluster 3 has little potential for energy savings through demand response. Cluster 4 features very large variability over the course of a day. In this cluster, the peak electricity demand roughly doubles the average consumption, and the minimum consumption is approximately 10%–20% of the average. This group differs from Cluster 3, as it has a great potential for energy savings through demand response.

In addition, long-term variability can be divided into 4 clusters by Ramadan consumption patterns. Cluster 1 features an infinitesimally increasing electricity demand during Ramadan. Cluster 2 includes only the shopping mall and features larger increases in Ramadan rate than Cluster 1 because many people go to restaurants for dinner during Ramadan. Conversely, Clusters 3 and 4 decrease during Ramadan. It is thought that human activities decreases during daytime and that electricity demand is interconnected with these activity trends.

The results above allow us to classify short, medium-, and long-term power consumption patterns. All of the buildings in hourly variability Cluster 3 are included in long-term variability Cluster 1. Thus, the hourly variability clusters share common characteristics with the long-term variability clusters. Furthermore, many of the public buildings are included in hourly variability Cluster 4, and all of these are included in long-term variability Cluster 4, suggesting that hourly and long-term variability can be estimated without monitoring data.

4 Conclusions

An innovative electricity demand monitoring system was installed in selected commercial, public, and residential buildings. The aim was to create high-precision electricity demand prediction models and investigate energy demand trend clustering, which might help in future low-carbon planning in Indonesia. A real-time electricity demand monitoring system was developed and motioning devices were installed in the public, commercial, and residential buildings. Then, factors influencing energy consumption using multi-regression analysis were investigated and hourly electricity prediction models were developed using ARX models and the significant factors found earlier. The developed models have high R ² values and similar MAPE values to those found in previous research in Japan. Thus, these models can be used to produce highly accurate predictions, which may provide basic data for electricity management techniques, such as ADR and supply-side planning, which can be used to realize low-carbon societies.

A number of models featured RRMSE indexes over 20%. Only the research office, hotel, and city office branch had RRMSE values under 20%. In addition, these models output some negative predicted values. Accordingly, these predictive models must be improved if they are to be used in demand response deal methods. The DTW time-series cluster analysis results were also compared to the total electricity prediction data in order to verify electricity pattern similarity from April to December 2016. The results indicated 4 different hourly electricity demand trend clusters and 4 different long-term electricity demand trend clusters. Clustering results also suggested that hourly variability and long-term variability trends could be estimated without monitoring data.

Future research targets will include improved analysis, such as the development of innovative predication models and anomaly analysis for considering Negawatt deal demand response, and the development of hourly energy evaluation models by demand response. In addition, the relationships between electricity demand trends using both monitoring data and non-monitoring data such as questionnaire and field surveys pertaining to monitoring sites will be investigated.

References

[1] Maki S, Ashina S, Fujii M, Installing energy monitoring system for consumer sector in Indonesia and energy use prediction by multiple-time series modelling. Journal of Japan Society of Civil Engineers, Ser.G (Environmental Research), 2017, 73(6): II_35 – II_43 (in Japanese)

[2] Hargreaves T, Nye M, Burgess J. Keeping energy visible? Exploring how householders interact with feedback from smart energy monitors in the longer term. Energy Policy, 2013, 52: 126 –134

[3] Kindaichi S, Nishina D, Murakawa S, Tanaka T, Horioka K. Energy demand and the factor analysis in a cafe in the campus. Journal of Environmental Engineering, 2014, 79(696): 191 –199

[4] Kawano H, Yamada S, Abe H, An electricity demand forecasting method for office buildings using a small data approach. Information Processing Society of Japan Consumer Device & System, 2014, 4(2): 1 –9 (in Japanese)

[5] Kondo S, Nobayasi M, Hokoi S. Forecasting model for electricity demand in residential house based on time series analysis. Journal of Japan Society of Energy and Resources, 2016, 37(1): 34 –42 (in Japanese)

[6] Ma J, Qin J, Salsbury T, Xu P. Demand reduction in building energy systems based on economic model predictive control. Chemical Engineering Science, 2012, 67(1): 92 –100

[7] Utma A, Gheewala S H. Indonesian residential high rise buildings: a life cycle energy assessment. Energy and Buildings, 2009, 41(11): 1263 – 1268

[8] Permana A S, Perera R, Kumar S. Understanding energy consumption pattern of households in different urban development forms: a comparative study in Bandung City, Indonesia. Energy Policy, 2008, 36(11): 4287 –4297

[9] Chou J S, Gusti Ayu Novi Yutami I. Smart meter adoption and development strategy for residential buildings in Indonesia. Applied Energy, 2014, 128: 336 –349

[10] Sakoe H, Chiba S. Dynamic programming algorithm optimization for spoken word recognition. IEEE Transactions on Acoustics, Speech, and Signal Processing, 1978, 26(1): 43 –49

[11] Nakamoto K, Yamada Y, Suzuki E. Fast clustering for time-series data with average-time-sequence-vector generation based on dynamic time warping. Transactions of the Japanese Society for Artificial Intelligence, 2003, 18(3C): 144 –152 (in Japanese)

[12] The Energy Data and Modelling Center, Japan. EDMC Handbook of Japan’s & World Energy & Economic Statistics 2016. The Energy Conservation Center, Japan, 2016 (in Japanese)

[13] Iwata G. Statistical Method for Economic Analysis, 2nd ed. Toyo Keizai Inc., 1983 (in Japanese)

[14] Newey W, West K. A simple, positive semi-definite, heteroscedasticity and autocorrelation consistent covariance matrix. Econometrica, 1987, 55(3): 703 –708

[15] Tomikoshi D, Ikaga T, Kawakubo S, Development of a suggestion tool for energy-saving actions based on the analysis of residents’ behaviors and energy demand. AIJ Journal of Technology and Design, 2013, 19(42): 655 – 660 (in Japanese)

[16] Agency for Natural Resources and Energy, Japan. Guidelines for trading negawatts. 2017–07

RIGHTS & PERMISSIONS

Higher Education Press and Springer-Verlag GmbH Germany, part of Springer Nature