Informs Annual Meeting 2017

MA69

INFORMS Houston – 2017

MA69

4 - Nonparametric Prognostics using Convolution Processes: An Alternative Approach Based on Transfer Learning Raed Al Kontar, University of Wisconsin-Madison, 1513 University Avenue, Room 3255, Madison, WI, 53706, United States, alkontar@wisc.edu, Shiyu Zhou In this paper, an alternative view on modeling condition monitoring (CM) signals is proposed. This view draws its roots from multitask learning and is based on treating each CM signal as an individual task. Each task is then expressed as a convolution of a latent function drawn from a Gaussian process (GP), and the transfer of knowledge is achieved through sharing these latent functions between historical and in-service CM signals. As a result, all task are represented through a functional graphical model. Aside from being non-parametric, the individualistic approach in our model is able to account for heterogeneity in the data and automatically infer the commonalities between different signals. 371C Joint Session QSR/ENRE Electricity: Data-driven Systems Engineering Approaches in Wind Power Systems Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Eunshin Byon, University of Michigan, 1205 Beal Avenue, 1205 Beal Avenue, Ann Arbor, MI, 48109, United States, ebyon@umich.edu Co-Chair: Hoon Hwangbo, Texas A&M University, College Station, TX, 77840, United States, hhwangbo@tamu.edu 1 - Contextual Bayesian Ascent Algorithm for Cooperative Wind Farm Control Jinkyoo Park, jinkyoo.park@kaist.ac.kr We discuss a probabilistic optimization method for identifying the joint optimal operational conditions of wind turbines in a wind farm for varying wind condition. We developed a Contextual Bayesian Ascent (CBA) algorithm, an extended version of Bayesian Optimization (BO) that can find an optimum of a target function through the iterative learning and sampling procedure. CBA takes into account context information that modifies the target function and restricts the searching space so that it can rapidly find an optimum. The results from simulation studies using an analytical wind farm power function show that the CBA algorithm can achieve an almost optimum control policy. 2 - Assessment of Wind Parameter Sensitivity on Ultimate and Fatigue Wind Turbine Loads Latha Sethuraman, National Renewable Energy Laboratory, National Wind Technology Center, 15013 Denver West Parkway, Golden, CO, 80401, United States, Latha.Sethuraman@nrel.gov, Amy Robertson Wind turbines are designed using a set of simulations to ascertain the structural loads that the turbine could encounter, but the only probabilistic variable considered is the mean wind speed. Other wind parameters, such as turbulence spectra, sheer, veer, spatial coherence, and component correlation are deterministic values that in reality could vary vastly in time and between sites and have a significant effect on the loads. This paper seeks to assess the sensitivity of different wind parameters on the resulting turbine ultimate and fatigue loads during normal operational conditions. Eighteen different wind parameters are screened using an Elementary Effects approach with radial points. 3 - Wake Effect Parameter Calibration in Wind Power Systems Bingjie Liu, bingjiel@umich.edu, Matthew Plumlee, Eunshin Byon In design and operations in wind power systems, physical interactions among wind turbines, called wake effects, is one of the significant factors that affect power generation performance. Among several wake modeling approaches, physics-based engineering models have been widely used due to its simplicity. The accuracy of engineering wake models, however, highly depends on the parameters. This talk proposes a calibration approach for estimating the relationship of wake parameters and environmental factors. 4 - Multivariate Power Curve Estimation an Adaptive Kernel Method Reflecting the Domain Knowledge Hoon Hwangbo, Texas A&M.University, 522 Southwest Parkway, Apt A, College Station, TX, 77840, United States, hhwangbo@tamu.edu, Yu Ding Accurate estimation of wind turbine power curve improves wind power prediction and hence economic assessment of existing or potential wind energy projects as well as operational controls and condition monitoring of turbines. Deviating from the predominant practice of univariate power curve modeling, a recent study introduced a multivariate power curve model utilizing additive multivariate kernels, which significantly improves the prediction accuracy. In this study, we further strengthen the multivariate power curve model by allowing adaptive bandwidths and enhancing the additive structure maintained between the multivariate kernels that ultimately reflect the domain knowledge better. MA68

371D Learning and Optimization for Resilient Power Grids Sponsored: Energy, Natural Res & the Environment Environment & Sustainability Sponsored Session Chair: Yury Dvorkin, New York University, Seattle, WA, 98115, United States, dvorkin@nyu.edu Co-Chair: Yu Zhang, University of California, Santa Cruz, Albany, CA, 94706, United States, zhangy@ucsc.edu 1 - T&D Integration: Value of Learning about Consumers Yury Dvorkin, Assistant Professor, New York University, 5 Metrotech Center, New York, NY, 11201, United States, dvorkin@nyu.edu We will present a modeling framework to analyze techno-economic interactions between transmission and distribution electrical grids. This framework intends to allocate flexibility of distributed energy resources between transmission and distribution grid support services. Since availability of some distributed energy resources is constrained by socio-economic behavior of electricity consumers, we will use a learning-compatible approach to account for behavioristic aspects that cannot be considered using formal models. The usefulness of the proposed framework will be illustrated on benchmark transmission and distribution systems. 2 - Learning through Power Distribution Grid Probing Vassilis Kekatos, Virginia Tech, Blacksburg, VA, United States, kekatos@vt.edu Although power distribution grids are challenged by limited observability, many nodes are equipped with smart meters collecting voltage and power injection readings. The key idea here is to engage the smart power electronics to probe the grid and infer non-metered loads. Probing can be accomplished by commanding inverters to fluctuate their (re)active power injections. Recording the incurred voltages at actuated and metered nodes can unveil injections at non-metered nodes. It is analytically shown that probing over two time instances is successful if and only if non-metered nodes can be matched to metered nodes through vertex- disjoint paths as numerically corroborated using actual data too. 3 - Modeling and Optimization of Complex Buildings with Deep Neural Networks Yuanyuan Shi, University of Washington, Seattle, WA, United States, yyshi@uw.edu, Yize Chen, Baosen Zhang Buildings are responsible for 40% of the global energy consumption and aroused serious economical and environmental concerns. Previous model-based optimization and control methods for building energy management either involves tedious modeling processes or cannot fully capture the exact dynamics of sophisticated buildings. Here we propose an alternative approach which is model- free and data-driven. By utilizing high volume of data coming from advanced sensors, we train a deep neural network which could accurately represent the operation dynamics of buildings. The trained network is then directly fitted into a constrained optimization problem to find the optimal control strategies. 4 - Black Start Allocation for Power System Restoration Georgios Patsakis, University of California Berkeley, 1460 Cedar St, Berkeley, CA, 94702, United States, gpatsakis@berkeley.edu, Deepak Rajan, Shmuel S.Oren Extended blackouts of the power system, even though rare, can have a very high cost impact. System operators use heuristically predefined sequences for restoring the grid back to its initial state, based on units with the ability to start autonomously, called Black Start (BS) units. Recent literature tries to optimize over such candidate sequences. In light of pending BS unit retirements, we instead aim to optimally allocate BS units on the grid, while also extending current optimization models to include grid considerations. Scenarios corresponding to partial outages of the grid are considered and a decomposition algorithm is utilized to solve the problem. A test case is simulated.

157