Research Statement. Dong Chen

Transcription

1 Research Statement Dong Chen Cyber-Physical Systems (CPSs) and the Internet of Things (IoT) are an increasingly important class of systems that integrate computation and communication with physical processes. CPSs usually monitor and control physical processes in a feedback loop where physical elements affect computation and vice versa. Modern society now relies on CPSs to perform critical operations in sectors ranging from assisted living, power generation and distribution, building design and automation, transportation, manufacturing, health monitoring, and emergency management, all of which serve critical functions in our daily lives. Advances in CPSs promise to enable adaptability and scalability that will far exceed the current embedded engineering systems. The tight coupling between the cyber and physical worlds in CPS is enabling the accumulation of large amounts of data, which can be analyzed, interpreted, and appropriately leveraged for optimizations. Especially, when multiple CPSs are interacting and cooperating with each other, this big data analysis becomes a critical requirement to improve CPSs performance. Due to the widespread deployment of CPSs, our physical environment is constantly monitored in detail by millions of Internet-connected sensors, including smart meters, satellites, and radars, with much of the data made publicly available. Thus, sensor data is growing at a faster rate than ever before. As one example, utilities have deployed 70 millions smart meters, which record building energy usage at fine-grained intervals. The state-of-the-art meters are able to read energy data every second, and thus every million meters generate 86.4 billion readings per day. Recent studies estimate that data processed annually by CPSs will reach 44 zettabytes, or 44 trillion gigabytes by This increase will become overwhelming if the big data is not properly managed. A parallel trend is the increasing concern of security and privacy issues that are stemming from integrating with cyber systems using a diverse set of communication techniques. The more complex a system becomes, the more vulnerabilities it has. When CPSs fail due to unintentional faults, cyber attacks, or privacy threats, it could affect public safety, trigger loss of life, cause enormous economic damage, and thwart the vital missions and interests of businesses, cities, states, and even the nation. For example, recently, Symantec reported a group of hackers broke into utilities and energy companies supplying electricity to the U.S. power grid, and may now have the ability to cause major blackouts in the U.S. Thus, novel proactive and adaptive approaches are needed to strengthen security and reliance for CPSs. The interplay between these two trends naturally raises three key research problems: (1) How can we collect, transmit, and manage the massive streams of data being generated by CPSs in an energy-efficient, secure and privacy-oriented manner? (2) What are the security and privacy implications of these big data collection and analytics? and (3) How can we design and build CPSs that are reliable, secure, and private? These challenges lie at the heart of my current and future research. As part of my research, I have designed a novel wide range of energy analytics, e.g., solar performance modeling, solar disaggregation, energy data localization, and their privacy implications, that combine advanced analytical and Machine Learning (ML) techniques with a deep understanding of the physical processes underlying energy generation and consumption. My research is unique in that it cuts across multiple areas, including CPSs and energy systems, computer systems, security/privacy, and data science and ML. In particularly, this broad approach puts me in an ideal position to seek funding from a diverse set of programs, including the CSR, STC, and CPS programs at NSF, as well as programs at the DoE, e.g., Renewable Energy Systems and Building Technology office. 1 Big Energy Data Analytics and their Privacy Implications Residential and commercial buildings account for over 75% of electricity consumption in the U.S. As a result, retrofitting existing buildings using renewable energy (primarily solar) has become an important research challenge of both CPS energy system and societal need. In parallel, the cost to deploy renewable energy is rapidly decreasing. For example, the installed cost per Watt(W) for residential photovoltaics (PVs) decreased by 3 from 2009 to 2016 (from $8/W to $2.9/W), resulting in the installed aggregate solar capacity in the U.S. increasing by 30 (from 435 MegaWatts to 14,762 MegaWatts) over the same period. This increasing solar capacity is imposing operational challenges on utilities in balancing electricity s real-time supply and demand, as solar generation, even when aggregated across many deployments over a large region, is still more stochastic Dong Chen 1 of 5

2 and less predictable than aggregate demand. To address these issues, both academia and utilities have a strong interest in solar data analytics to accurately monitor, predict and react to variations in intermittent solar power. White-box and Black-box Solar Analytics. Prior solar data analytics are mostly white box approaches that assume detailed information (e.g., tilt, orientation, size, efficiency, number of modules, nominal operating cell temperature, wiring loss, inverter type) from a deployment to physically model all solar generation effects. Thus, these white box approaches require much less training data to calibrate the modeling and are highly accurate. Unfortunately, white-box physical modeling is impractical at large scales, as it requires manually gathering and recording detailed information from 70 millions of solar sites. Homeowners with small rooftop solar sites generally do not have the expertise or motivation to collect this information. In addition, these white box learnings may also require inputs that are difficult to accurately measure, e.g., dust build up, air velocity. In contrast, some recent solar analytics works using black-box approaches are mostly leveraging Machining Learning (ML) techniques that automatically learn unknown relationships from monitored data, and adapt as they change over time. However, this kind of black-box solar analytics usually does not incorporate fundamental well-known physical models of solar generation. In addition, these MLbased black-box approaches require a significant amount of historical solar data to train an accurate model. Unfortunately, these training data is generally not available for either new deployments or deployments that do not continuously monitor and store the data. Hybrid Solar Analytics. Instead, I present a hybrid black-box approach that enables a wide range of accurate solar analytics, including performance modeling, disaggregation, and localization, with limited training data and without knowledge of key system parameters by integrating black-box ML approaches with whitebox physical models. 1.1 Big Energy Data Analytics (Greenmetrics 17, eenergy 17) This section describes my work in energy data analytics, including solar performance modeling, solar disaggregation, solar forecasting, and solar performance debugging. Solar Performance Modeling. Solar modeling is useful for a variety of solar energy analytics, including indirect monitoring, forecasting, disaggregation, anonymous localization, and fault detection. Significant recent work focuses on ML models, which leverage only historical energy and weather data for training. Interestingly, these ML techniques are often off the shelf and do not incorporate well-known physical models of solar generation based on fundamental properties. Instead, prior work on physical modeling generally takes a white box approach that assumes detailed knowledge of a deployment, that utilities are not able to gather. I designed a configurable hybrid approach [1] that combines the benefits of both by enabling users to select the parameters they physically model versus learn via ML, and found that it significantly improves model accuracy. Solar Disaggregation. ML-based solar models require significant historical solar generation data for training. Unfortunately, pure solar generation data is often not available, as the vast majority of grid-tied solar deployments are behind the meter, such that utilities only have access to net meter data that represents the sum of each building s solar generation and its energy consumption. To address this problem, I designed SunDance [2], a black-box technique that accurately disaggregates solar generation from net meter data without access to a building s pure solar generation data for training. It leverages multiple insights into wellknown fundamental relationships between location, weather, solar irradiance, and physical characteristics and requires only a building s location and a minimal amount of historical net meter data, e.g., as few as two datapoints. I also identified a new fundamental relationship, which I call the Universal Weather-Solar effect [3], that has not been articulated in the past and is broadly applicable to other solar energy analytics. (In progress) Solar Forecasting and Performance Debugging. I use the solar performance model for solar forecasting by feeding it forecasted weather conditions, rather than current weather conditions. Then, I compare with state-of-the-art forecasting modes that do not use empirical physical weather models, to better understand the benefits of my hybrid approach. Solar performance debugging is the problem of determining what is causing a solar panel s output to be less than its maximum output, i.e., what fraction is due to weather conditions versus non-optimal installation characteristics (tilt, orientation) versus shade from obstructions. My hybrid model can differentiate these effects because it models them explicitly and independently. This enables solar-owners to understand how much they could improve their solar panels performance. Dong Chen 2 of 5

3 1.2 Energy Data Privacy Implications (BuildSys 13, Percom 14, TOSG 15, BuildSys 16, SmartGridComm 16, BigData 17) Recently, the Department of Energy (DoE) released a Voluntary Code of Conduct (VCC) policy which enables utilities to provide energy data to third-party analytics companies as long as it is anonymized by stripping it of associated account information, e.g., a name and address. In addition, the anonymous energy data is often not handled securely, and even made publicly available over the Internet. My research below shows that energy data is not anonymous, and it could be accurately localized. Solar-based Localization. My key insight is that solar energy data is not anonymous: since every location on Earth has a unique solar signature, it embeds detailed location information. I designed SunSpot [4] to localize anonymous solar-powered homes using their solar energy data. SunSpot is able to localize a solar-powered home to a small region of interest that is near the smallest possible area given the solar energy data resolution, e.g., within a 500m and 28km radius for per-second and per-minute resolution, respectively. Weather-based Localization. Another key insight is that every location on Earth not only has a unique solar signature but also a distinct weather signature that uniquely identifies it. This is because that energy consumption, wind, and solar data largely correlate with weather, e.g., temperature, wind speed, and cloud cover. I designed Weatherman [5] to localize public energy data using a time-series weather database that includes data from over 35,000 locations. Weatherman localizes coarse (one-hour resolution) energy consumption, wind, and solar data to within 16.68km, 9.84km, and 5.12km, respectively. Weatherman is more accurate using much coarser resolution data than prior work on localizing only anonymous solar data using solar signatures. (In progress) Net Meter Localization. I develop a technique to extract an accurate location from net meter data, which combines energy consumption and solar data. I first extended my prior work on SunDance to also learn a rough location that is accurate enough for disaggregation. To further hone in on a location, I leveraged my prior work on Weatherman to localize a small region of interest. I plan to train a ML classifier that is able to identify visible solar sites from satellite imagery. I will also train an additional ML classifier that can then identify the rough size and orientation of the panels for each site. Then, I can filter the list of visible solar sites to those that best match the solar data based on their size and orientation. Privacy Implication (Localization). SunSpot and Weatherman could also leverage similar ML techniques in the work of net meter localization to localize the source of energy data towards a specific site. Thus, energy data that includes solar generation, wind generation, energy consumption, and net meter data is not anonymous and can be accurately localized. The location of energy data can provide important contextual information to interpret big data analytics, but it can also enable third-parties to link private behavior derived from energy data with a particular home. Therefore, SunSpot and Weatherman expose a new kind of privacy threat that has never been discussed before. These localization techniques not only explore the privacy threats of energy data, but also are public useful tools for researchers that need to recover the locations of energy data to learn insights behind energy data. For the home owners that are unwilling to reveal their privacy, how to prevent this privacy exposition becomes an important challenge. In order to address this problem, I plan to design an advanced encryption protocol that hides solar and weather signature in net meter data transmission. Privacy Implication (Occupancy). My NIOM work [6] shows that a home s pattern of electricity usage generally changes when occupants are present due to their interaction with electrical loads, thus energy consumption data indirectly leaks sensitive information about a home s occupancy, which is easy to detect because it highly correlates with simple statistical metrics, such as power s mean, variance, and range. By combining NIOM and Weatherman, utilities and third-parties can easily detect the occupants personal behaviors, e.g., daily routines, eating habits, vacation schedules, etc. in a specific building. Privacy Preserving. In order to prevent the above occupancy privacy threat, I designed a CHPr-enabled water heater [7] and [8] that regulates its energy usage to thwart a variety of occupancy detection attacks without violating its objective to provide hot water on demand, and my results show that a standard 50- gallon CHPr-enabled water heater prevents a wide range of state-of-the-art occupancy detection attacks.. Data and Software Release. I collected 10 billions of fine-grained interval energy data and weather data from 35,000 sites. I released these datasets, and they have been cited/requested 110 times by other researchers. I also developed SmartSim [9], a publicly-available device-accurate energy trace generator that enables researchers to mine net meter data to learn insights into home energy usage and occupants behavior. Dong Chen 3 of 5

4 1.3 Cybersecurity Issues and Countermeasures (ICGEC 10, CSSS 11, CSIS 11, TIIS 12, IJWMC 12, CSEE 12) In addition to big energy data analytics and their privacy implications, my research also focuses on secure architecture and cybersecurity strengthening techniques, e.g., lightweight key management, and malicious device detection, to build secure CPSs. Secure Architecture. As different wireless communication techniques and edge networking infrastructures are continually integrated into CPSs, the interacting and cooperating between multiple CPSs in a secure manner is becoming more challenging. To address this problem, I investigated security and privacy issues in CPSs, and presented a wide range of countermeasures that solve or relieve these security and privacy challenges in [10] and [11]. I then designed a secure architecture [12] to connect multiple diverse CPSs. Key Management. Key management is an effective approach to ensure authentication and the prerequisite for all security operations in a system. Unlike traditional cyber systems, CPSs devices could be highly heterogeneous. Most recent works that do not consider the cost and overhead on diverse nodes cannot be directly applied to CPSs. To address this issue, I designed a lightweight n-collusion resistant secure key management scheme [13] to strengthen the authentication in CPSs. Malicious Device Detection. CPSs/IoT typically leverage wireless sensor and actuator networks to monitor and control physical environment. Due to its wireless nature, a device could be easily captured by an adversary, which may result in its non-cooperative behavior or misbehavior with the rest of the nodes in the network (a.k.a. network partitioning). I designed TRM-IoT in [14] and [15], a behavior-based trust and reputation management model. My results show that it could not only detect the malicious nodes but also improve the system performance by avoiding these malicious nodes. TRM-IoT has been cited 155 times. 2 Future Directions I will continue exploring challenges in big data analytics, cybersecurity and privacy threats in CPSs. The research approaches in my big energy data analytics are quite general. They can be used in any situation where domain knowledge can be used to identify key physical parameters that can be estimated before residuals are estimated using ML. My expertise and experience with big data analytics and cybersecurity enhancing provides me with an excellent background to work in new areas. Electric Vehicle Grid Integration. Due to the widespread deployment of plug-in electric vehicles (PEVs), the power grid is being transformed from a one-way electricity delivery system to a two-way intelligent transmission and distribution system that connects many smart devices and infrastructures components. PEVs promise to reduce carbon emissions by exploiting renewable energy sources for battery recharge, and could potentially serve as electricity storage bank to flatten the fluctuations in power generation caused by the intermittent nature of renewable energy sources. However, this Vehicle-to-Grid (V2G) integration has raised many grid management issues. In addition, plugging the vehicles in the recharging infrastructures may expose private information regarding the user s locations and traveling habits. I am interested in: (1) big data analytics in PEV charging/discharging systems to monitor and enhance power quality and grid stability with high penetration of solar energy; (2) exploring approaches enabling the export of PEVs power to assist in grid outages; (3) designing advanced PEV charging systems that can reduce peak power demands and access smart grid value potential; and (4) investigating cybersecurity issues and privacy threads in this electric vehicle and grid integration and the approaches to prevent them. Personal Health Monitoring. Health care systems are facing a new challenge. According to the U.S. Bureau of the Census, the number of elder people is doubled from 35 million to nearly 70 million in In parallel, total health care spending in the U.S. has reached $1.8 trillion in 2014 with almost 45 million people uninsured. These data statistics suggest that heath care in the U.S. needs a major changing towards more scalable, more proactive, and more affordable solutions. Personal health monitoring is the key to helping this transition. I am interested in: (1) big data management in medical informatics; (2) usability of health data to benefit patient and public health safety, privacy and security; (3) user-driven privacy expectations from hospital information systems; (4) privacy in hospital information systems; and (5) data aggregation and visualization technologies for population-based reporting. Dong Chen 4 of 5

5 References [1] Dong Chen and David Irwin. Black-box solar performance modeling: Comparing physical, machine learning, and hybrid approaches. In the 2017 ACM Greenmetrics (Greenmetrics 17), Urbana-Champaign, IL, [2] Dong Chen and David Irwin. Sundance: Black-box behind-the-meter solar disaggregation. In the eighth ACM International Conference on Future Energy Systems (e-energy 17), Hong Kong, [3] Dong Chen and David Irwin. One model to rule them all: Black-box physical solar performance modeling using (almost) no training data. In In submission, [4] Dong Chen, Srinivasan Iyengar, David Irwin, and Prashant Shenoy. Sunspot: Exposing the location of anonymous solarpowered homes. In the 2016 ACM International Conference on Systems for Energy-Efficient Built Environments (BuildSys 16), Stanford, CA, USA, [5] Dong Chen and David Irwin. Weatherman: Weather-based localization of anonymous energy meter data. In 2017 IEEE International Conference on Big Data (BigData 17), [6] Dong Chen, Sean Barker, Adarsh Subbaswamy, David Irwin, and Prashant Shenoy. Non-intrusive occupancy monitoring using smart meters. In the Fifth ACM Workshop On Embedded Sensing Systems For Energy-Efficiency In Buildings (BuildSys 13), Roma, Italy, [7] Dong Chen, Sandeep Kalra, David Irwin, Prashant Shenoy, and Jeannie Albrecht. Preventing occupancy detection from smart meters. IEEE Transactions on Smart Grid, 6(5): , Feb [8] Dong Chen, David Irwin, Prashant Shenoy, and Jeannie Albrecht. Combined heat and privacy: Preventing occupancy detection from smart meters. In the 12th IEEE International Conference on Pervasive Computing and Communications (PerCom 14), Budapest, Hungary, [9] Dong Chen, David Irwin, and Prashant Shenoy. Smartsim: A device-accurate smart home simulator for energy analytics. In the 2016 IEEE International Conference on Smart Grid Communications (SmartGridComm 16), Sydney, Australia, [10] Dong Chen, Guiran Chang, and Jie Jia. Study on the interconnection architecture and access technology for internet of things. In International Conference on Computer Science and Service System (CSSS 11), [11] Dong Chen and Guiran Chang. A survey on security issues of m2m communications in cyber-physical systems. KSII Transactions on Internet and Information Systems, 6(1):24 45, [12] Dong Chen, Guiran Chang, Lizhong Jin, and Xiaodong Ren. A novel secure architecture for the internet of things. In the Fifth International Conference on Genetic and Evolutionary Computing (ICGEC 10), [13] Dong Chen, Guiran Chang, Dawei Sun, Jie Jia, and Xingwei Wang. Lightweight key management scheme to enhance the security of internet of things. International Journal of Wireless and Mobile Computing, 5(2): , [14] Dong Chen, Guiran Chang, Dawei Sun, Jie Jia, and Xingwei Wang. Trm-iot: A trust management model based on fuzzy reputation for internet of things. Journal of Computer Science and Information Systems, 8(4): , [15] Dong Chen, Guiran Chang, Dawei Sun, Jie Jia, and Xingwei Wang. Modeling access control for cyber-physical systems using reputation. Journal of Computers Science and Electrical Engineering, 38(5): , September Dong Chen 5 of 5