Datasets on South Korean manufacturing factories’ electricity consumption and demand response participation

This study describes the release of electricity consumption data of some manufacturing factories located in South Korea that participate in the demand response (DR) market. The data (in kilowatt) comprise individual factories’ total power usage details that were acquired using advanced metering infrastructures. They further contain details on the manufacture types, DR participation dates, mandatory reduction capacities, and response capacities of the factories. For data acquisition, 10 manufacturing companies are representatively selected according to the process regularity and company size standard of this study. Entire datasets are newly collected and available at one-minute intervals for seven months from 1 March to 30 September 2019. These datasets can be used in a variety of ways to contribute to the functioning of power systems and markets, including the conduction of industrial load characteristic analysis for load flexibility, estimation of demand-side considerations for virtual power plant design, and determination of energy markets and incentives to achieve carbon neutrality targets at the national level.


Background & Summary
Today, global energy and environmental conditions necessitate the widespread use of renewable energy sources for countries to achieve their carbon neutrality targets and, thereby, address climate change problems 1 . However, installing renewable energy resources without accounting for the power system reliability limitation causes system stress resulting from a supply-demand imbalance, such as from oversupply or excessive security 2 . This forces more ancillary generators in the system to stand by or promotes inefficient investment in power grid reinforcement 3 . To solve this problem, power system operators must understand the concept of load flexibility (LF). LF refers to the resources used to ensure the stable operation of the power system by facilitating dynamic changes, including increments and decrements, in demand. This includes implementing demand-side management (DSM), which changes power use patterns according to the time-series energy production characteristics of wind turbines or solar power sources to increase the application rate of renewable energy 4,5 .
However, although most of the DSM capacity for LF is met by industrial loads, there are quite a few obstacles to the acquisition of industrial demand data. In a competitive industrial environment, the data disclosure of industrial loads is prohibited since such data are considered a trade secret because a manufacturing plant's electricity consumption data can be used to infer the company's sales. To the best of the authors' knowledge, investigations on manufacturing factories' load data remain limited; only two studies require special mention in this respect: an investigation on the machine-level load data of a paper manufacturing factory in Brazil 17 and an examination of the normalized electricity consumption data of food and paper industries 18 .
In this study, the authors acquire data from volunteered industrial factories and analyze their characteristics to evaluate demand response (DR) availability of Korean industrial demands for securing power system and market flexibility. Furthermore, a market system is being designed to encourage factories to participate as LF resources.
The authors collect electricity consumption data from manufacturing factories in South Korea by using communication systems, including the advanced metering infrastructure (AMI). These factories participate in the DR market through DSM. Accordingly, the resulting dataset is unique and potentially a valuable consideration in several analyses, including.
• Expected locational DR capacity estimation by statistically estimating customer baseline load (CBL) and participation amount of each industrial sector. • Estimation of hourly LF by analyzing industrial demand consumption patterns.
• Consideration of demand-side utilization in virtual power plants.
• Design of the LF market and incentive price.

Methods
The load aggregators performing brokerage transactions in the DR market are authorized to collect electricity usage information from the system operator through the AMI for DSM. In this study, the authors first introduce international and Korean demand response programs in detail. Subsequently, they describe a novel communication system in which a load aggregator collects relevant data through the AMI and finally classify the industrial demand data collected from the factories participating in DR programs by manufacture type.
Demand response programs. DR is defined as a tariff or a program established to motivate changes in electric use by end-use customers in response to changes in the price of electricity over time or to give incentive payments designed to induce lower electricity use at times of high market prices or when grid reliability is jeopardized 19 . It is classified into price-based DR for economic operational purposes and intensive-based DR for system security purposes. Figure 1 illustrates DR programs included in the planning and operation of power  www.nature.com/scientificdata www.nature.com/scientificdata/ system in detail. In DR programs, the participation performance of resources is evaluated based on CBL estimation 19 . In general, the average demand usage of past days without participating in DR is used in calculating CBL. Table 2 describes DR services of independent system operators (ISOs) in the US, which are internationally benchemarked [20][21][22][23][24][25] .
Korean DR market consists of six programs depending on the purpose as shown in Table 3 26 . In recent years, along with traditional DR programs, they expanded to mitigate environmental issues, including fine dust problems and supply/demand balance due to rapid renewable energy penetration. Participants are restricted from entering the market depending on the type and capacity of resources they have. Table 4 describes Korean ISO's DR services in detail 26 .
Monitoring set-up. In the proposed communication system, watthour pulse (WP) and end-of-interval (EOI) signals are received in one-minute units through the AMI's photocoupler, which is installed to charge electricity bills to the manufacturing company. The WP-based wattage data are synchronized with the EOI signal and delivered to the server in real-time. Further, the system involves storing the process of monitoring data for a short period to improve data acquisition quality. When data delivery fails, the communication system performs a resending the stored data to the server. After a certain number of retries fail, the data is extinguished by storage period expiration. The well-collected data are backed-up every 30 days. To upload the data to the server, one can select the interface from among Ethernet, RS-232, and RS-482 ports according to the communication environment. Figure 2 illustrates the overall hardware communication network design.
Industrial demand data classification. In Korea, the manufacturing industry is classified into 40 industries. Among them, 10 industries, namely petrochemical, fine chemical, cement, steel, forging, food, paper, metal, electricity/electronics, and textile, mainly participate in the DR market and function as ancillary service resources. The number of their companies account for 44.92% of all industries. The authors selected five representative types which account for 48.36% of the aforementioned 10 manufacturing factories: cement, forge,   www.nature.com/scientificdata www.nature.com/scientificdata/ metal, paper, and steel. Only 11.59% of the companies included in the types are actually participating in the DR program. Therefore, it is expected that they still have high potential that can be utilized as LF resources 27 .
Data from 20 volunteer factories with data disclosure agreements were obtained. Finally, 10 factories with regular manufacturing processes and their company sizes (e.g., number of employees, sales, and manufacturing scales) were selected in this study. Figures 3-7 illustrate the five representative manufacturing processes. To maintain information security, the company name and factory location are not disclosed in this paper, and net power consumptions without normalization are mentioned to preserve data originality. This study presents the data measured for seven months from 1 March 2019 to 30 September 2019. During the measurement period, a DR was issued twice; Table 5 depicts the date and time of DR participation, mandatory reduction capacity, and response capacity of each factory for the load aggregator's transaction. Average power consumption for 4 days out of the past 6 days (holidays and weekends), excluding highest and lowest consumption days Table 4. Summary of the details of CBL evaluation methods in South Korea.  www.nature.com/scientificdata www.nature.com/scientificdata/   www.nature.com/scientificdata www.nature.com/scientificdata/

Data Records
The entire dataset comprises 10 comma-separated value (CSV) files 28 , summarised in Table 6. As mentioned earlier, the total electricity consumption (kW) of each factory was measured in this study. The CSV files of each factory have 308160 rows, including N/A spaces and outliers, which indicate one-minute-interval data (1440 data points/day) for 214 days during the 7-month data collection period in 2019. Since the method of preprocessing data is selected and applied according to various research purposes, the authors provided raw data for reuse without preprocessing. Each file has two columns: one indicates time information (in the YYYY-MM-DD hh:mm format), while the other indicates the factory's real-time electricity consumption. For better reuse, the    www.nature.com/scientificdata www.nature.com/scientificdata/ Korean system load data file of the same period is provided together 28 . The dataset has been made publicly available under the creative commons license CC BY 4.0 hosted on the figshare repository.

technical Validation
This section discusses the visualization of data to clarify the quality of the dataset, which includes missing data, outliers, and weekly pattern plots. The missing data plot and outlier information indicate the availability of minute details on the electricity consumption of each factory, whereas the weekly pattern plots provide the characteristic insights into power consumption according to the manufacturing type and working/non-working date conditions. The summary of manufacturing factories' dataset statistics is described as shown in Table 7.
Missing data. Figure 8 illustrates the missing electricity consumption data of 10 factories. The missing data plot for the entire data collection period (where the missing data are indicated using black lines) is shown on the left side of the figure. Further, the horizontal bars on the right visually represent the percentage of missing data over the study period. The manufacturing factories have an average data availability of 98.7%. An exception is the Metal 2 factory, whose missing data rate is more than 10% due to data collection errors in April 2019. Data with a 20% or less missing rate guarantees quality through missing data imputation 29 . The approach for time-series missing data imputation provided in this study is classified mainly into five categories: deletion, neighbor-based, regression-based, multi-layer-perceptron-based, and deep-learning-based approaches. The description and practical methods of each approach were reviewed in detail as shown in Table 8 30-40 . outliers. Figure 9 illustrates the 10 factories' daily electricity consumption profiles during data collection periods. As an index for outlier detection, the interquartile range (IQR) of the box plot was considered. As a result of extracting data located outside the range of 3 sigma of the normal distribution from each demand data, 4, 38, and 1 outlier were detected in Cement 1, Cement 2, and Paper, respectively. The approach for time-series outlier   Table 7. Summary of manufacturing factories' dataset statistics. www.nature.com/scientificdata www.nature.com/scientificdata/ data detection provided in this study is classified into four categories: statistical, unsupervised discriminative, unsupervised parametric, and supervised approaches. The description and practical methods of each approach were reviewed in detail as shown in Table 9 [41][42][43][44][45][46][47][48][49][50] . Accordingly, the authors propose to scale and utilize the raw data according to the research purpose.
Weekly patterns. Figure 10 shows the 10 factories' weekly electricity consumption patterns, obtained by averaging the electricity consumption during the data collection period by day of the week. Each factory reveals approximate periodicity according to its own manufacturing process. The factories that implemented automated

Approach Description Method
Deletion 30,31 Elimination of observations with missing values in raw data Listwise deletion and pairwise deletion Neighbour based 32,33 Missing data imputation through neighbours identified by the clustering method KNN and DBSCAN Regression based 34,35 Missing data prediction by modelling correlations between a dependent variable and independent variables based on historical data AR, ARX, and ARIMA Multi-layer perceptron based 36,37 Missing data estimation by designing a model minimizing the loss function of fully connected network NLP and ANN Deep learning based [38][39][40] Missing data prediction by designing network including information over time RNN and GRU Table 8. Summary of missing data imputation methods for time series data.

Fig. 9
Electricity consumption daily profiles of 10 manufacturing factories during data collection periods.

Approach Description Method
Statistical approach [41][42][43] Outlier detection through a function describing the relationships between a dependent variable and independent variables based on historical data ARMA, ARIMA, VARIMA, and EWMA Unsupervised discriminative approach 44,45 Outlier detection through similarity measurement based on clustering method K-means, SOM Unsupervised parametric approach [46][47][48] Outlier detection through probabilistic model about state or value over time HMMs Supervised approach 49,50 Outlier detection through a model trained with labelled data SVM Table 9. Summary of outlier detection methods for time series data.
www.nature.com/scientificdata www.nature.com/scientificdata/ processes (Steel 2, Cement 1, and Cement 2) recorded a steady electricity use even on non-working days. The factories' electricity consumption varied according to their size; for example, employees, sales, and production scale. In particular, factories with high electricity usage (Metal 2, Steel 2, and Cement 1) tended to avoid operating on  www.nature.com/scientificdata www.nature.com/scientificdata/ time intervals with high electricity rates. Despite the limitation of the 7-month acquisition period, the characteristics of weekly demand usage were strongly confirmed. Figure 11 provides the factories' electricity consumption profiles at the DR participation day (13 June 2019), which confirm the factories' responded capacities. The capacity is calculated as the difference between the CBL (denoted using cyan lines in Fig. 11) and the actual load (denoted using red lines). The CBL is a general standard used for settlement in national DR markets. In this study, the factories' average power consumption in the same time for four out of the past five days, excluding holidays, is considered the CBL. As additional information,

code availability
The code implementation was done in R 4.0.5 using R studio. The scripts to perform data visualization are available in 28 .