• Energy storage modeling and optimization

    An illustration of how renewable energy sources can be combined with different energy system components to improve reliability.

    Storage allows us to compensate for generation variability, especially from renewables. However, storage is expensive and complex to manage. Our work studies sizing of storage, drawing on its analogy to buffers in computer networks, and operation of storage, that is, when to charge and when to discharge storage.

    • Battery sizing: Although solar energy is varying throughout the day, if we could spread out the total daily solar energy throughout the day, we can have a useful constant power source. We aim to minimally size the batteries to guarantee a certain fixed power from the sun either for the whole day (firm-up solar PV power) or for a time interval during the day (day-ahead dispatch problem).
    • Battery operation: The application of the renewable energy and battery system plays a role on how the battery should be operated. For example, a home owner might want to use the system to minimize payments to their electricity provider (utility). The optimal operation would differ substantially depending on the pricing scheme (eg. time-of-use pricing) that is in effect. We aim to find solutions to the question of how to optimally operate the system in a given environment, and propose practical online operating strategies that result in near-optimal system performance.

    Spec-Based Lithium-Ion Storage Models

    This work looks at mathematical models for Lithium-ion batteries that are tractable enough to be included as part of mathematical optimization programs. We compare the state-of-the-art model with real Lithium-ion batteries, using experiments conducted at the Technology Center for Energy in Ruhstorf, Germany. Our experiments show that the accuracy of the state-of-the-art model is very limited with respect to a wide range of charging and discharging rates. Using the insights from our experiments, we derive and validate two new models that represent different trade-offs between accuracy and tractability. We evaluate the error of our models compared to a real battery, and make a favourable comparison with the state-of-the-art model. This work has been published in the proceedings of ACM eEnergy 2016, and was featured in a poster presented at the 2016 Cheriton Symposium and was a runner-up for the Best Poster award.

    Energy system control with deep neural networks (in progress)

    We are working on a project to effectively control a simulated residential energy system using deep neural networks. The goal of the control algorithm is to minimize the amount of money spent on electricity in order to power a building by efficiently utilizing local photovoltaic generation and a large energy storage device. For the given system, we explore three different approaches: fully-connected networks, DDPG reinforcement learning, and convolutional neural networks. We compare the effectiveness of our trained networks with four benchmarks: a simple control algorithm as an upper bound on grid cost, two refined control algorithms developed in prior work, and the offline optimal oracle which gives a lower bound on grid cost. Once an effective approach for a simple energy system has been established, we will apply it to a complex system for which no effective control strategies have been developed.