Given the problems of gradual oil depletion and global warming, energy consumption has become a critical factor for energy-intensive sectors, especially the semiconductor, manufacturing, iron and steel, and aluminum industries. In turn, reducing energy consumption for sustainability and both tracking and managing energy efficiently have become critical challenges. In response, we analyzed electricity consumption from the perspective of load profiling, which charts variation in electrical load during a specified period in order to track energy consumption. As a result, we proposed a time series data mining and analytic framework for electricity consumption analysis and pattern extraction by streaming data mining and machine learning techniques. We identified key factors to predict the state of the annealing furnace and detect abnormal patterns of the load profile of their electricity consumption. Our experimental results show that the dimension reduction method known as piecewise aggregate approximation can help to detect the state of the annealing furnace.
You have full access to this open access chapter, Download conference paper PDF
As a cornerstone of modern civilization and economic growth, electricity is critical for industrial and economic advancement, as well as a driving force for sustainable development. Indeed, social development correlates positively with power consumption, which in Taiwan, especially the consumption of electricity, has risen rapidly due to economic, industrial, and commercial growth.
In relation to total exports, Taiwan’s manufacturing-oriented economy exports a considerable share of manufactured goods. Currently, most industries in Taiwan have replaced manual operation with machine operation during fabrication, which requires a sufficient but not excessive supply of stable electricity. In fact, too much or too little electricity can cause mechanical malfunctions and thereby reduce the efficiency of both production and electricity. As Table 1 shows, Taiwan Power Company’s statistics from 2015 reveal that the industrial sector consumes an exceptionally large proportion of electricity—even up to more than 50% of the total consumed in Taiwan.
All tasks of analysis involved using the load profiles of the electricity consumption of the targeted machine. For the proposed framework, we preliminarily deploy the data warehousing framework to observe and analyze the load profiles of electricity consumption and the relationships among various attributes (e.g., electric power, temperature, and product weight). Next, we select and confirm key attributes to identify the state of the annealing furnace based on the results of analysis and consulted with domain experts. We confirm that either the electric power or temperature information of the operating machine can help to identify the entire machine operational process, which is 1,440 min on average. We use the temperature information of the operating machine to identify three states: warm-up, heat retention, and cooling.
We apply the PAA method to discretize streaming data into n segments with timestamps in order to build the prediction model. We will refine the SAX algorithm, which is a symbolic representation of time series for dimensionality reduction and indexing with a lower-bounding distance measure to further extract subsequent patterns. It can help the system to detect abnormal energy patterns and machine operational states by symbolizing energy load profiles to make further energy-optimization decisions in real time. We will apply an agglomerative hierarchical clustering approach to discriminate normal and abnormal electric patterns—that is, to group the electric patterns for further analytical and prediction tasks. We plan to next conduct a series of experiments to construct a prediction model in order to identify their operational states (i.e., warm-up, heat retention, and cooling), the target annealing furnace, and abnormal energy patterns. We also included associated experiments of parameter selection of the PAA method in our experiments.
Ultimately, the goal of our series of studies is to deploy a visualized decision support system and propose actionable energy-saving strategies for co-operating iron and steel plants to solve real-world problems. We present the entire framework for electricity consumption analysis and detail some of the modules in the following sections.
Table 3 presents all of the attributes of the annealing furnaces related to electricity consumption analysis in our research. We adopted a data mart to visualize and observe the initial load profiles of electricity consumption. In general, data warehousing is fundamental to business intelligence, and data collection, data management, and data analysis techniques (e.g., data mart design with extraction, transformation, and loading tools) can help business analytics use data intelligently. Accordingly, we deployed the data warehousing framework to observe the load profiles of electricity consumption (Fig. 2) and analyzed the relationships among various attributes (e.g., electric power, temperature, and product weight. Figure 3 presents the fact table of our research. The data warehousing platform had two chief goals: to analyze the load profiles of each annealing process and to define annealing states based on the selected attributes of load profiles.
Data warehousing helped us to confirm the load profiles of each annealing process in order to preliminarily identify the normal or abnormal state of the machines. We confirmed that either the electric active power or temperature information of the operating machine can help to identify the entire machine operational process, which is 1,440 min on average. We used the data of annealing process from April 1, 2014, to December 31, 2014 to train and construct the prediction model to detect each machine’s state and condition.
After selecting the attributes that were useful for periodical data analysis, we adopted the star schema to build the data mart (Fig. 2). The three dimension tables are the machine information table, the product information table with time information with different granularity table, and a fact table that shows the load profiles of current and temperature, among other things. Based on the analytical results of load profile, we used the temperature information of the operating machine to identify three states: warm-up, heat retention, and cooling. By extension, we could further identify the normal or abnormal states of each annealing process. We show one load profile of active power and temperature of one annealing furnace in Fig. 3 (Table 2).
Based on the methods, we defined time series data and related notations (Table 3). We denoted time series data of an attribute i as S = (s1, s2,….sn), with the length of a time series in n and w as the dimensionality of the space to index the time series data. Put differently, a time series of length n can be represented in w dimensional space and each feature point by a feature frame of fix length (i.e., n/w). For PAA, the result is \( \overline = (\overline >> ,\overline >> , \ldots ,\overline >> ) \) – that is, w-dimensional space by vector \( \overline \) . The ith feature point of \( \overline \) can be derived from Eq. (1).
