• Featured Projects

    View All Featured Projects

  • EFRI Representative Visit

Fourth Annual NSF EFRI Clemson-GTech-MST Workshop

Madren Conference Center, Clemson University, Clemson SC 29634

Program Details

Sunday Oct 28, 2012    
Reception 5 PM – 7 PM MC Patio
Monday Oct 29, 2012 Investigators’ Highlights & Keynote Day
Welcome & Opening Remarks Clemson ECE Dept. ChairDr. Darren Dawson
8:15 AM – 8:30 AM
Seminar Room II
Brain2Grid Project Overview and Accomplishments Dr. G. Kumar Venayagamoorthy – PI
8:30 AM – 9.30 AM
Seminar Room II
 EFRI-COPN: Neuroscience and Neural Networks for Engineering the Future Intelligent Electric Power Grid – Brain2Grid ProjectThe electric power grid is a complex system under semi-autonomous distributed control. It has spatial and/or temporal complexity, nonconvexity, nonlinearity, nonstationarity and uncertainty. Control algorithms do not yet exist at sufficient scale to guarantee stability over a wide range of nonlinear operating conditions, as witnessed by the August 2003 blackout.It has been shown that Adaptive Critic Design (ACD) based neurocontrollers can significantly improve the stability of a power system when they are applied in real time control loops. While these and other approaches have achieved ACD based adaptive optimal control, none scale up for controlling very large power systems in real-time. The Brain2Grid project combines neurobiology with engineering to develop new approaches to emulate brain-like capabilities for fast real-time control and decision making. These revolutionize nonlinear adaptive optimal control of very large systems, including the electric power grid. These advances are made by studying learning and adaptation for in-vitro neuronal networks interfaced to computing machines; applying new closed-loop hybrid neural architectures that are capable of learning to optimize, predict and control; improving the understanding of their simulation on new massively parallel systems; and validating these architectures on several power system testbeds, using complex, scalable simulations and real-time hardware. The overall goal is the utilization of biological network mechanisms to control of complex power systems, more brain-like.Neurocontrol offers nonlinear adaptive optimal control, which is particularly useful for nonlinear stochastic power networks, which would otherwise need different mathematical models and changing parameters for each network topology. Neural network architectures have worked well on small power systems, but do not replicate the powerful capabilities of the brain to learn, and compute in a massively parallel fashion, allowing for adaptive optimal decision making, prediction and pattern recognition among other things. Neurocontrol of the power grid with new architectures and algorithms increase real-time responsiveness to changing power loads and component failures, improve the dynamic and transient behavior of the power network, and improve grid reliability, ensuring local and wide area stability, reduce greenhouse gas emissions and assisting human experts in control rooms. The findings of this project provide a better understanding of multiple-time base system identifiers and controllers, and their interactions and scalability for large systems. The project outcomes have broad implications for all types of complex systems design. Advances in understanding the brain’s learning mechanisms at the network level could affect many areas of neuroscience and lead to new educational approaches. 
Thrust #1 – Neurobiology Highlights Dr. Steve Potter – co-PI
9:30 AM – 10:15 AM
Seminar Room II
Coffee Break 10.15 AM – 10:45 AM North Lobby
Thrust #2 – Learning Algorithms Highlights Dr. Donald C Wunsch II – co-PI
10:45 AM – 11.20 AM
Seminar Room II
Biclustering has been proven to be a more powerful method than conventional clustering algorithms for analyzing high-dimensional data, such as gene microarray samples. It involves finding a partition of the vectors and a subset of the dimensions such that the correlations among the biclusters are determined and automatically associated. Biclustering ARTMAP (BARTMAP) is a recently introduced algorithm that enables high-quality clustering by modifying the ARTMAP structure, and it outperforms most previousbiclustering approaches. Hierarchical BARTMAP (HBARTMAP), introduced here, offers a biclustering solution to problems in which the degree of attribute-sample association varies. We also have developed a hierarchical version of Iterative Two-Way Clustering for comparison purposes and have compared these results with other methods as well. Experimental results on multiple genetic datasets reveal that HBARTMAP can offer indepth interpretation of microarrays, which other conventional biclustering or clustering algorithms cannot achieve. Thus, this paper contributes two hierarchical extensions of biclustering or co-clustering algorithms and comparatively analyzes their performance in the context of microarray data.
Thrust #2 – Power Electronics Highlights Dr. Keith Corzine – Senior Personnel
11:20 AM – 11:55 AM
Seminar Room II
Lunch 12 PM – 1:00 PM Seasons by the Lake
Thrust #2 – Intelligent Power Systems I Highlights Dr. Ronald Harley- co-PI
1:00 PM – 1:35 PM
Seminar Room II
Thrust #2 – Intelligent Power Systems II Highlights Dr. Kumar Venayagamoorthy- PI
1:35 PM – 2.25 PM
Seminar Room II
Thrust #2 – Cellular Neural Networks & Learning Highlights Dr. Bipul Luitel – Senior Personnel
2:25 PM – 2.55 PM
Seminar Room II
The information obtained from monitoring systems enables sensemaking leading to situational awareness. In a smart grid, monitoring of system variables such as voltages and speed deviations of generators is important for assessing its stability, and making proper control decisions. Development of wide area monitoring system is, hence, important for situational awareness; especially in a smart grid where integration of renewable resources, distributed generation and bidirectional power flows can lead to instabilities if proper control action is not taken at the right time, place and context. In this study, cellular neural networks (CNN) have been used to predict speed deviation of generators and bus voltages in a smart grid. Since CNN can be exactly mapped to represent components (generators and buses) of a smart grid, a CNN based wide area monitoring system has been developed. The main focus of this paper is to use an integrated CNN for predicting two parameters (generator speed deviations and bus voltages). The results show that the CNN method is potentially scalable and can be a useful tool for situational awareness in smart grids. 
Coffee Break 2.55 PM – 3:20 PM North Lobby
Keynote IImproving Situational Awareness of Power Grids Dr. N Zhou – DOE PNNL
3.20 PM – 4.20 PM
Seminar Room II
 Improving Situational Awareness of Power GridsSituational awareness (SA) is “understanding the current environment and being able to accurately anticipate future problems to enable effective actions” [from PNNL Web]. Lack of SA often leads to conservative operation and reduces asset utilization. In addition, lack of SA has been identified as one of core causes of several large scale power outages in North America (e.g., Southwest America in September 2011, Northeast blackout in August 2003, Western interconnection in August 1996). This presentation will review the current practice and challenges in SA. The Pacific Northwest National Laboratory’s on-going research into improving situational awareness will be overviewed in the following areas:Fast state estimation,Massive contingency analysis,Fast dynamic simulation,Real time path rating,

Non-iterative voltage stability analysis,

Dynamic state estimation, and

Advanced optimization.

Utilization of high performance computing (HPC) is going to be highlighted. It is expected that improved situational awareness will increase grid asset utilization; facilitate the integration of the renewable generation, distributed generation and plug-in hybrid vehicles; help avoid large-scale blackouts; and help fully utilize the potential of smart grid appliances.

 Zhenyu Huang (PI), Ning Zhou (Co-PI), Yousu Chen, Shuangshuang Jin, Ruisheng Diao, Stephen T Elbert, Yuri V Makarov, Karan Kalsi, Maria Vlachopoulou, Mark J Rice, Kurt R Glaesemann, Craig H Allwardt, Marcelo A Elizondo, Di Wu, Bharat Vyakaranam, Zhangshuan Hou, Barry Lee, Chunlian Jin, Xinxin Guo


Keynote II (CU COES Seminar) – Energy Challenges Facing the Nation and NSF Opportunities to Address Them Dr. Paul Werbos – NSF
5 PM – 6 PM
Earle Hall 100
This talk discusses how NSF-funded research can have big impacts as part of a larger world effort to address two grand challenges:(1) How can the OECD become totally independent forever from the need to use fossil oil at the soonest time, at minimum cost?; (2) How can the entire world become able to meet its electricity needs from >80% renewable sources at <= 10 cents per kwh total cost to the consumer with maximum probability at soonest time? Special funding opportunities and success rates will also be discussed, such as the EPAS program, whose submission window is now Oct. 1 – Nov. 1 every year.
Dinner 6:45 PM 102 Country Walk Circle
Tuesday Oct 30, 2012 Students’ Day  
Brain2Grid Website Highlights Sirjana
8:00 AM – 8:15 AM
Seminar Room II
Communicating with Living Neuronal Networks on Microelectrode Arrays for Computation and Control Robert Ortman
8:15 AM – 8:50 AM
Seminar Room II
 We present advances in methods for communicating with small-scale in vitro living neuronal networks (LNNs) intended to facilitate the study of learning and application of LNNs to advanced computation and control tasks. We have developed a novel sensory coding method to translate network inputs into electrical stimulation patterns, which reliably evoke separable responses in cultures of approximately 20000 rat cortical neurons living on microelectrode arrays (MEAs). We evaluated candidate electrical stimulation patterns for effectiveness via a support vector machine (SVM)-based method and used a genetic algorithm to construct subsets of highly separable stimulation patterns. Our methods achieved input/output bit rates of 16-20 bits per second at a 10% symbol error rate over time periods of up to 10 hours. We stimulated the LNN with a sinusoidal test input encoded via stimulation patterns our algorithm identified, and recovered the original input with a normalized root-mean-square error of 20-30% using only the LNN responses and a trained SVM classifier. Preliminary data indicates that the LNNs exhibit signs of short-term memory and plasticity in the context of sensory input encoded via our method.
Computation using Latency dynamics in Living and Artificial Neural Networks Riley Zeller-Townson
8:50 AM – 9.25 AM
Seminar Room II
 Spiking Neural Networks use the precise timing of action potentials to convey meaning. The conduction delays between neurons are one set of parameters that can be tuned to improve network performance on computational tasks, however no biologically inspired delay learning rules have been adopted by the artificial neural network community. This work introduces simple delay update rules that are based on how delays change in living neural networks, and then shows how these update rules alter synaptic plasticity, spike pattern formation, and memory in artificial networks. Additionally, we demonstrate how delay changes seen in living neurons could be used to improve performance for a prediction task. 
Biologically Inspired Artificial Neural Networks Jing Dai
9.25 AM – 10:00 AM
Seminar Room II
 Inspired by living neuron networks (LNNs) in the brain, artificial neural networks (ANNs) have been broadly used in various applications as a computational intelligence tool. However, due to many fundamental differences between ANNs and LNNs, despite the mature training mechanisms for ANNs, it is often challenging to use LNNs as a computational intelligence tool. To bridge the gap between ANNs and LNNs, a novel type of artificial neural network, i.e. biologically-inspired artificial neural network (BIANN) is proposed in this paper. The BIANN, which is based on spiking neuron models of LNNs, processes information in a more “brain-like” fashion than conventional ANNs. A reservoir-computing-based training approach is also proposed for BIANNs to serve as a novel modeling and control tool for practical applications. The feasibility of the proposed BIANN is illustrated for the prediction of a synchronous generator’s speed and terminal voltage signals in a single machine infinite bus electric power system setup. The proposed BIANN model is able to provide an accurate prediction for online monitoring of a generator. 
Coffee Break 10:00 AM – 10.15 AM North Lobby
Modeling and Intelligent Control for an Electric Power Micro-Grid Yi Deng
10:15 AM – 10:50 AM
Seminar Room II
This paper reports the related skills learnt during the last year and the ongoing work of the project, which is part of the Emerging Frontiers in Research and Innovation (EFRI) project. For the ongoing work, the details of a micro-grid are presented. The micro-grid contains the typical elements in present and promising future power systems, e.g., synchronous machines, wind power plant, photovoltaic energy, and energy storage. The work plan for the next year is also presented. 
Dynamic Performance Models of Wind Turbine Generators Prajwal Gautam
10:50 AM – 11:25 AM
Seminar Room II
This paper presents the dynamic performance model of different type of wind turbines using neural networks. A Doubly-Fed Induction Generator (DFIG) simulation model developed in Real Time Digital Simulation platform is used to establish wind power output curve based on wind speed profile recorded at National Renewable Energy Laboratory’s (NREL’s) National Wind Technology Center (NWTC). The neural network model is trained and tested with different wind data sets to verify to learning dynamics of wind turbine for power estimation. With the promising results on the simulation model, a 3.2 kW permanent magnet synchronous machine type wind turbine installed at Real-Time Power and Intelligent Systems (RTPIS) Laboratory was chosen for testing. In absence of the exact and detailed electro-mechanical parameters of wind turbine generators from manufacturers, neural networks are used as intelligent methods to develop accurate short-term planning and operational model for wind turbines. Moreover, individual wind turbines with their own performance characteristics can be used for power estimation or forecasting of the wind farms. In this study, dynamic neural network is used for learning the system dynamics and estimating power output of Skystream 3.7 wind turbine. Based on actual wind speed and wind power outputs, a neural network based wind turbine model is developed and integrated into a micro-grid simulation model consisting of photovoltaic, energy storage and loads. 
Damping Electromechanical Oscillations in Large-Scale Power Systems Using Intelligent Aggregated Control Diogenes Molina
11:25 AM – 12.00 PM
Seminar Room II
This paper presents the development and evaluation of damping controls for large-scale power systems using approximate dynamic programming and aggregation techniques. Intelligent controllers trained via approximate dynamic programming have been shown to be able to deal with the complex, non-linear, and time varying nature of small power systems. However, the algorithms used for training the neural networks used to implement these controllers can be difficult to scale up to the proportions needed for controlling realistically sized power systems. Aggregation techniques are commonly used to generate simplified representations of large power systems for offline simulation studies. As measurement networks based on PMU technology become commonplace in the smart grids of the future, it will be possible to generate these simplifications online and use them for closed loop control. These representations significantly reduce the size and complexity of the system “seen” by the controller, and alleviate the intelligent controller scalability issue. They also preserve the features needed for effective system-wide damping control. Simulation results for a benchmark 16-machine system show that the proposed controllers can help power systems maintain stability even after conventional stabilizing technology begins to fail. Preliminary simulation results for a realistic model of a large portion of the Brazilian power system illustrate the benefits of aggregation for simplifying the control task, and highlight the importance of the robustness provided by intelligent controls. 
Scalable Monitoring and DSOPF Control in Smart Grids Karthikeyan Balasubramaniam
12:00 AM – 12:35 PM
Seminar Room II
Economic dispatch, i.e. set points for machines in electric grid is set every 15 minutes. Variations within this period are handled by linear controllers with little to no system wide information. With the advent of wind and solar energy, conditions of high uncertainty and high variability are introduced. Currently, controllers such as AGC and regional voltage controller are linear controllers which operate with little to no system wide information. An adaptive, optimal real-time controller based on adaptive critics design called dynamic stochastic optimal power flow controller (DSOPF) is proposed. Although power system has become more observable with the use of synchrophasor systems, real-time control is hindered by power system communication delays which can range from a few milliseconds to several seconds depending on distance and communication media. Hence a scalable wide area monitoring system is design using cellular computational networks which predicts state variables for one or more time steps ahead of time and hence essentially compensates for communication delays. The DSOPF controller works in tandem with wide area monitoring system and hence real-time control action can be taken with system wide information. The work would realize a scalable non-linear controller working with system wide information in real-time to control a nonlinear, non-stationary system – power grid, under conditions of high uncertainty and variability and would result in monetary and security benefits. 
Lunch 12:40 PM – 1:45 PM Seasons by the Lake
Student Posters & Demos 2.15 PM – 5.45 PM SB2 and SB3, Riggs Hall
Banquet Dinner 6:30 PM – 8.30 PM Meeting Room IV