Collaborative Research: CISE-MSI: RCBP-RF: CPS: Develop Scalable and Reliable Deep Learning-driven Embedded Control Applied in Renewable Energy Integration

该奖项全部或部分由《2021年美国救援计划法案》（公法117-2）资助。近年来，深度学习主要在图像处理、语言处理等领域取得了成功。然而，在实时控制领域，深度学习刚刚开始挑战比例-积分-导数控制器在工业应用中的主导地位，例如用于可再生能源集成的电源转换器的实时控制。许多亟待解决的问题，包括训练难度、在嵌入式设备上实施的挑战，都制约了深度学习在嵌入式控制环境中的发展和实施。为了克服这些困难，本项目旨在开发新颖的可扩展训练算法和新颖的深度神经网络控制器架构，以适应嵌入式控制设置的严格要求。该跨学科项目将开发可扩展且可靠的深度学习驱动的实时电源转换器嵌入式控制，用于集成太阳能等可再生能源。具体来说，该项目旨在(a)开发可扩展的、并行的、快速的训练算法，用于高采样频率和使用高性能计算或云平台的长时间轨迹学习，这将显著减少训练时间，从几天，甚至几周到几个小时；(b)开发可在嵌入式设备中实现的新型深度神经网络架构，例如：数字信号处理器/现场可编程门阵列，而不影响神经网络的通用性和额外的计算能力和存储要求。该项目将建立并加强两所少数民族服务机构之间的跨学科和机构间合作：德克萨斯农工大学金斯维尔分校和北卡罗来纳州农工州立大学。该项目将吸引、留住和教育更多的少数民族，特别是西班牙裔、非裔美国人和女性学生参加博士课程。开发的嵌入式控制新训练算法和新架构可以扩展到其他领域，如生物信息学、图像、机器人等。所开发的技术将为可再生能源并网提供深度学习驱动的智能控制，有助于解决美国将更多可再生能源并网的迫切需求。项目产生的研究存储库（数据、代码、模拟等）将存储在德克萨斯州农工大学金斯维尔分校和北卡罗来纳州农工大学的数字存储库中，并确保更广泛的计算机科学和可持续能源研究社区能够长期访问国家科学基金会规定的至少三年的访问权限。公共使用的数据文件可以通过项目网站（https://sites.google.com/view/dr-xingang-fu/home和https://sites.google.com/view/letuqingge/home）通过两个校区的数字存储库直接访问。限制使用的数据文件是在删除可能严重损害数据分析潜力的潜在识别信息后分发的。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).Recently deep learning has succeeded mainly in image processing, language processing fields. However, in the real-time control field, deep learning has just started to challenge the dominant role of proportional-integral-derivative controllers in industrial applications, e.g., real-time control of power converters for renewable energy integration. Many urgent problems including training difficulty, the implementation challenges on embedded devices, are curbing deep learning from the development and implementation in embedded control settings. To overcome these difficulties, this project aims to develop novel scalable training algorithms and novel deep neural network controller architectures to fit the strict requirement of embedded control settings.The interdisciplinary project will develop scalable and reliable deep learning-driven embedded control of power converters in real-time for integrating renewable energy such as solar power. Specifically, this project aims (a) to develop scalable, parallel, fast training algorithms for high sampling frequency, and long-time duration trajectory learning using an high performance computing or cloud platform that will significantly reduce training time from several days, even weeks to several hours, (b) to develop novel deep neural network architectures that can be implemented in embedded devices, e.g., Digital Signal Processors / Field-programmable Gate Arrays without compromising the neural network generalizability and extra computing power and storage requirements.The project will build and enhance interdisciplinary and inter-institution collaborations between two Minority Serving Institutions: Texas A&M University-Kingsville and North Carolina A&T State University. The project will attract, retain, and educate more minorities particularly Hispanic, African-American, and female students to attend Ph.D. programs. The developed new training algorithm and new architectures for embedded control can be extended to other fields, e.g., bioinformatics, image, robotics, etc. The developed technologies will result in deep learning-driven intelligent control for grid integration of renewable resources and help solve the urgent need to integrate more renewable energy into the power grid in the United States. The research repository (data, code, simulations, etc.) generated from the project will be deposited with the digital repository at Texas A&M University-Kingsville and North Carolina A&T State University and ensure that the broader computer science and sustainable energy research community have long-term access for a minimum of three years prescribed by the National Science Foundation. Public-use data files can be accessed directly through the project websites ( https://sites.google.com/view/dr-xingang-fu/home and https://sites.google.com/view/letuqingge/home ) via the digital repository on both campuses. Restricted-use data files are distributed after removing potentially identifying information that would significantly impair the analytic potential of the data.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.