Please use this identifier to cite or link to this item:
Title: Development of data mining-based big data analysis methodologies for building energy management
Authors: Fan, Cheng
Degree: Ph.D.
Issue Date: 2016
Abstract: Today's buildings are becoming not only energy intensive, but also information intensive. Building Automation Systems (BASs) are widely installed in modern buildings for automatic monitoring and control of the operation of various building services systems. BASs collect and store a huge number of sensor measurements and control signals at short time intervals. The effective utilization of the big BAS data can help to optimize and diagnose the performance of buildings so as to improve their operational performance. However, the big BAS data are not fully utilized due to the lack of advanced data analysis techniques and tools. BASs can only perform simple data analysis, such as historical data tracking, moving averages and benchmarking. Data mining (DM) is a promising solution for the knowledge discovery from massive data sets. However, it is extremely challenging for building automation professionals to keep up with the constantly emerging sophisticated DM techniques. Meanwhile, there is a knowledge gap between building professionals and advanced data analytics. DM itself cannot tell the value or the significance of the knowledge discovered, and domain knowledge in the building field is therefore still needed to interpret and apply the knowledge discovered. This research aims to develop generic DM-based methodologies for discovering knowledge in big BAS data and applying the knowledge to building energy management, such as identifying typical and atypical operation patterns, energy performance analysis, diagnosis and optimization. Based on a comprehensive exploration of the state-of-the-art DM techniques using case studies on the BAS data of a high-rising building in Hong Kong, the strengths and restrictions of a variety of advanced DM techniques taking into account of the characteristics of BAS data and the building operations are understood. This dissertation first presents a generic DM-based framework for knowledge discovery in massive BAS data and applications of the knowledge for building energy management. The framework consists of five phases, i.e., data pre-processing, data partitioning, knowledge discovery, post-mining and applications. The framework and the DM techniques involved at each phase are deliberately designed considering the characteristics of BAS data and the type of knowledge to be discovered. Based on the framework developed, the methodologies for discovering and applying three different types of knowledge, including cross-sectional knowledge, temporal knowledge and graph-based knowledge, are developed, tested and evaluated using BAS data retrieved from real buildings.
BAS data are usually stored in a single two-dimensional data table, where each column represents a variable and each row is an observation consisting of the values of different variables. Cross-sectional knowledge refers to the relationships and associations between variables (i.e., different columns) without taking into account the temporal dependency. A number of DM techniques, including clustering analysis, association rule mining and decision trees, are adopted to discover cross-sectional knowledge and to improve the reliability of the knowledge discovered. Post-mining methods, which bridge the knowledge discovered by DM techniques and domain expertise, are developed to enhance the efficiency and effectiveness in knowledge selection and application. Valuable knowledge has been discovered to understand building operation behaviors and spot energy conservation opportunities. Different from cross-sectional knowledge, temporal knowledge discovery focuses on discovering the temporal relationships between observations (i.e., different rows). In this case, the observations are considered as multivariate time series. The symbolic aggregation approximation (SAX), motif discovery and temporal association rule mining are applied as the main DM techniques. Two post-mining methods are developed to effectively utilize the knowledge discovered. The knowledge discovered can be used to characterize the dynamics in building operations and facilitate fault diagnosis and control optimization. A graph-based DM technique is developed for mining BAS data with potentially complex structures, rather than just a single two-dimensional data table. It ensures the knowledge discovery efficiency when the BAS data structure is complex, e.g., data are stored in multi-relational databases and cannot be easily merged into a single data table. With the population of building information modelling, a huge amount of valuable information related to building design and operations is becoming available for analysis, such as the text data for building construction and maintenance, and the spatial information of system components. Graphs provide great flexibility in integrating and representing various types of information; and the knowledge discovered using graph-based DM is highly interpretable. The frequent subgraph mining (FSM) and graph-based anomaly detection (GBAD) are selected as the primary mining techniques. Two problems are specifically addressed, i.e., graph generation based on BAS data and efficiency enhancement in knowledge selection and application. The graph-based mining methodology has been applied to represent different types of information, based on which frequent and atypical building operation patterns are detected. BAS data retrieved from the tallest building in Hong Kong and the Zero-Carbon Building are used to test and evaluate the methodologies. The results show that the knowledge discovered is valuable to identify dynamics, patterns and anomalies in building operations, assess building system performance and spot opportunities in energy conservation. The framework and the methodologies can contribute to develop more powerful and sophisticated BAS tools for effective utilizing the big BAS data for building energy management.
Subjects: Buildings -- Energy conservation.
Buildings -- Energy conservation -- Data processing.
Data mining.
Hong Kong Polytechnic University -- Dissertations
Pages: xvii, 215, 23 pages : illustrations (some color)
Appears in Collections:Thesis

Show full item record

Page views

Last Week
Last month
Citations as of May 28, 2023

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.