Detailed Analysis of Energy Consumption for an Office Building

Modern buildings consume a significant share of global power production. The primary sources of the power consumption are heating, ventilation and air conditioning (HVAC), light, cooking equipment, technical equipment like elevators, office facilities. Optimization of the buildings power consumption requires real-life data both for strategic planing and operational control. Here we analyze detailed power consumption of a research institute campus which is a medium-sized office buildings. Our data contain detailed electric measurements of 100 three-phase power lines which power all facilities inside the building. Each line is associated with several power consumers like lights, HVAC and office equipment. We analyze weekly and monthly trends, typical patterns in power consumption, and provide rough analysis of what classes of power consumers are most active in specific periods of time.


Introduction
Energy systems all over the world now have to address two competing goals.First, they must to increase the electricity production to meet the rising demand.Second, the ultimate goal of almost all developed countries is to reduce the emission of CO2 and the consumption of fossil fuel.The latter challenge leaded to recent growth in alternative energy sources like windmills and solar stations.These energy sources actually allowed to reduce emission of greenhouse gases and need in fossil fuel, but also poses new challenges for energy infrastructure.
Another approach is to improve the efficiency of existing energy networks by modern control techniques like machine learning [1], [2].These methods rely on relevant data at each level of grid operation.Both researchers and engineers need publicly available datasets to develop and test novel approaches to energy management and control.The important part of energy studies is optimization of buildings energy consumption since buildings are responsible for about 30% of global energy consumption [3].To the best of our knowledge current datasets on building consumption are mostly focused on residential buildings.Less data is available for office buildings while they consume lots of energy and its consumption patterns are obviously different from residential buildings.
Our project is aimed to provide a dataset on the energy consumption of the campus of Trapeznikov Institute of Control Sciences.This is a medium-sized building where most power consumers are similar to office buildings.The dataset was first presented in [4].The data include recordings from one hundred electric meters installed at the grid powering the building.Electric recordings are supplemented with local meteorological data and indoor climate variables collected by several wireless climate sensors.Here we study power consumption patterns observed in the data.First, we provide the descriptive statistics on the power consumption.Second, we explore seasonal and daily power load patterns.We show significant differences in seasonal consumption patterns which is expected because of the two factors: the location and the working cycle typical for the research institute.This information may be used to plan and optimize power supply, make infrastructure upgrade decisions etc.

Related research
Modern research of smart buildings is centered around data-driven approaches like machine learning.These methods highly depend on the relevant data, so many research datasets have been published in the past decade.In [dataset survey] the survey and analysis of datasets available is provided.The authors consider the following potential applications for them: energy saving (A1), appliance recognition (A2), occupancy detection (A3), preference detection (A4), energy disaggregation (A5), demand prediction (A6), anomaly detection (A7).The number of electric meters used for a single building varies from 1 to 158, the from sampling rate varies from 1 hour to 100 kHz.However, in most datasets the sampling period is 1-60 sec.The authors highlight the four factors influencing the building energy consumption: occupancy patterns, user behavior, weather data and energy cost.
Our dataset may potentially be used for energy disaggregation also known as Non-Intrusive Load Monitoring (NILM) [5].This is the problem of separating the overall power load into records generated by individual appliances.The datasets intended for NILM usually include applicance-level recordings [6] or synthetic data generated by deep neural networks trained on real applicances [7].We provide recordings for separate power lines which also may be interpreted as individual power consumers.Hence, our data may serve as a testbed for energy disaggregation algorithms.
Another branch of research explores patterns of energy consumption.A New York office building data are explored in [8].The authors use machine learning methods to clusterize power load patterns.They also use features like outdoor temperature, month, day of week etc. to predict the power load at the specific hour.In this data the power load is higher during summer months which is directly opposite to our data showing the higher consumption in winter.

Data
We explore energy consumption data collected in 2021 from January to December.The campus is located in Moscow at 55 • 39' N, 37 • 32' E. The primary focus was two main buildings: administrative (A) and laboratories (L).Each building is powered with multiple 3-phase 380V lines called feeders.Most feeders are divided into three independent 220V lines which further branch into multiple connections to separate rooms and facilities.We installed electric meters into each 3-phase feeder line.The measurements are collected with period varying from 1 to 15 sec.Electric variables include total power, active and reactive power, phase voltages and currents, grid frequency and harmonics.In this paper we are interested only in total power.Fig. 1 and 2 and illustrates the extent of the survey and the amount of data collected.In February, 7 meters were added to the existing 55, in May the number of installed meters reached 81, and in September-December -95.As the number of meters increased, so did the number of observations collected: from 500,000 measurements per month at the beginning of the year, more than 3 million measurements per month were collected by December.The number of measurements collected could depend on both the activation/deactivation of individual electrical lines in the building and the performance of the meters.
The electric measurements are supplemented with indoor climate and local weather measurements collected by two weather stations installed at the top of the building.The weather variables are temperature, wind direction, humidity and atmospheric pressure.Data overview  is provided in Table 1.The more detailed description of the dataset is provided in [4].The dataset itself is available at https://energy.ipu.ru/en/data-sets/.

Energy consumption patterns
Here we explore the patterns of energy consumption which reflects the lifecycles of the building.In Fig. 3 the dynamics of overall power load is shown.While the data contains some gaps in the start of the year, the overall weekly and seasonal cycles is visible.The energy consumption is low in the summer and high in the winter.The main reason is that most employees take vacation during summer, so the activity is low.The second reason is that during winter many employees use personal heaters which generate additional load.Fig. 4 shows the distribution of averaged hourly power load during working days and holidays.Both distributions are right-skewed, but the workday distribution is much more centered.The mean hourly load is 171 kWt for working days and 127 kWt for holidays.Surprisingly, the mean hourly load for holidays only 25% less than for working days.The maximal hourly load is 477 kWt for working days and 293 kWt for holidays, which is 38% less.
The difference in consumption patterns by day of the week is confirmed by Fig. 5.At night, consumption remains at the same level for both weekdays (Monday to Friday) and weekends (Saturday and Sunday).However, during the working day, consumption increases and peaks around midday.The observations for Friday are slightly different from those for Monday-Thursday, which is explained by the shortened working day -the decline in consumption starts  a few hours earlier than on other working days.It is worth noting that the window of growth of electricity consumption is wider than the window of working hours (from 9.00 to 18.00): the early morning growth is explained by the beginning of the work of the Institute's services, while the late evening decline can be explained by the frequent evening activities of the Instituteseminars, events related to the work of thematic workshops and others.
The next aspect of interest is the dependence of energy consumption on weather conditions.As shown at Fig. 6, the outside temperature can vary significantly depending on the time of year.Table 3 summarises the overall statistics of the dataset from the weather station located on the roof of the building.Fig. 7 and Fig. 8, show the average hourly and daily energy consumption data by season.
The highest energy consumption, as well as the greatest variation in consumption, is observed in winter, and the lowest in summer.Of particular interest is the comparison between spring and autumn: the average temperature (Fig. 6) and wind speed (Fig. 9) of these seasons do not differ significantly (less than 1%), while the average energy consumption in autumn is 20% higher.A possible explanation could be that more observations are available for the autumn.
Another explanations are the significant difference in the length of the day or relative humidity between these seasons (differences in average values of around 30%, Fig. 10).

Conclusion
Modern studies of energy grids and smart buildings need publicly available and relevant data.We collected a dataset on energy consumption of a research institute, which is a medium-sized office building (available at https://energy.ipu.ru/en/data-sets/).The data include recordings from the 100 meters installed at the electric grid of the building, indoor climate and outside weather.Sampling period for electric recordings is between 1 and 15 seconds.

Figure 1 .
Figure 1.Number of electric meters installed into feeders.

Figure 2 .
Figure 2. Number of observations obtained from electric meters.

Figure 3 .
Figure 3. Dynamics of the building power load during 2021 year.The power is averaged by 2-hour intervals.

Figure 5 .
Figure 5. Hourly electricity consumption by day of week.Figure 6. Weather station temperature data

Figure 6 .
Figure 5. Hourly electricity consumption by day of week.Figure 6. Weather station temperature data

Figure 8 .
Figure 8. Energy consumption by day of week and season.
Currently the data are available for 2021 year.The sampling rate and total duration are suitable for a range of research applications: energy disaggregation, short-term and long-term prediction of power consumption, study of seasonal power consumption patterns.We analyze some patterns of energy consumption in the collected data.The data show stable daily and weekly consumption patterns.Total energy consumption is lower in summer IC-MSQUARE-2023 Journal of Physics: Conference Series 2701 (2024) 012145

Table 2 .
Summary statistics of temperature change data obtained from the weather station, • C.

Table 3 .
Summary statistics of energy consumption depending on the time of year, kWt.