Matt Wytock

Large-scale probabilistic forecasting in energy systems using sparse Gaussian conditional random fields

Short-term forecasting is a ubiquitous practice in a wide range of energy systems, including forecasting demand, renewable generation, and electricity pricing. Although it is known that probabilistic forecasts (which give a distribution over possible future outcomes) can improve planning and control, many forecasting systems in practice are just used as “point forecast” tools, as it is challenging to represent high-dimensional non-Gaussian distributions over multiple spatial and temporal points. In this paper, we apply a recently-proposed algorithm for modeling high-dimensional conditional Gaussian distributions to forecasting wind power and extend it to the non-Gaussian case using the copula transform. On a wind power forecasting task, we show that this probabilistic model greatly outperforms other methods on the task of accurately modeling potential distributions of power (as would be necessary in a stochastic dispatch problem, for example).

Preventing cascading failures in microgrids with one-sided support vector machines

Microgrids formed by a network of power sources and power consumers yield significant advantages over the conventional power grid including proximity of power consumption to power generation, distributed generation, resiliency against wide area blackouts and ease of incorporation of renewable energy sources. On the other hand, unlike the conventional grid, microgrids are compliant where a single load or a single generation unit can often form a significant fraction of the total generation capacity. Here large excursions from the nominal operating conditions are possible motivating the need for safety mechanisms which isolate power electronic equipment from damage. Breakers serve the purpose of protecting equipment from surge conditions by shutting off, for example, generation units. However in microgrids, a loss of a single generation unit can have catastrophic impact on the viability of the entire system. Here settings on breakers cannot be chosen too conservatively to protect the equipment at the expense of system viability or too liberally which will result in equipment damage. The ensuing problem of striking a suitable compromise tends to be combinatoric in nature due to numerous states of breakers which is further exacerbated by an uncertain load profile and nonlinear nature of system dynamics. In this article we provide a methodology to determine current thresholds and guard times, the time interval when current is allowed to exceed threshold value, for each inverter for fail-safe operation of microgrid. We employ a machine learning approach to address the problem where we first demonstrate that conventional support vector machine (SVM) methodology does not yield a satisfactory solution. We then develop a one-sided SVM method and generalize it to yield nonlinear support boundaries which captures the need for fail-safe operation against system blackouts while protecting equipment. A simulation engine is developed to model a real microgrid which is used to generate data for assessing and guiding our approach.

Course-specific search engines: semi-automated methods for identifying high quality topic-specific corpora

Web search is an important research tool for many high school courses. However, generic search engines have a number of problems that arise out of not understanding the context of search (the high school course), leading to results that are off-topic or inappropriate as reference material. In this paper, we introduce the concept of a course-specific search engine and build such a search engine for the Advanced Placement US History (APUSH) course; the results of which are preferred by subject matter experts (high school teachers) over existing search engines. This reference search engine for APUSH relies on a hand-curated set of sites picked specifically for this educational context. In order to automate this expensive process, we describe two algorithms for indentifying high quality topical sites using an authoritative source such as a textbook: one based on textual similarity and another using structured data from knowledge bases. Initial experimental results indicate that these algorithms can successfully classify high quality documents leading to the automatic creation of topic-specific corpora for any course.

Dynamic energy management with scenario-based robust MPC

We present a simple, practical method for managing the energy produced and consumed by a network of devices. Our method is based on (convex) model predictive control. We handle uncertainty using a robust model predictive control formulation that considers a finite number of possible scenarios. A key attribute of our formulation is the encapsulation of device details, an idea naturally implemented with object-oriented programming. We introduce an open-source Python library implementing our method and demonstrate its use in planning and control at various scales in the electrical grid: managing a smart home, shared charging of electric vehicles, and integrating a wind farm into the transmission network.