CAREER: A Networking and Learning Co-Design Framework for Data-Efficient Resource Management

现代建筑需要通过收集各种传感器的数据，对供暖、通风和空调（HVAC）、照明、百叶窗和窗户进行复杂的管理。LoRaWAN（一种开放的LPWAN协议）与LoRa物理层技术（远程）已经成为大规模传感器网络的一个很好的选择，能够以低成本提供长距离通信。根据LoRa联盟（管理LoRaWAN的发展）的说法，LoRaWAN在智能建筑的许多方面都优于传统的无线网络（例如ZigBee和Wi-Fi），例如简单的网关可以轻松覆盖建筑物的几层楼，低功耗以及长达十年的电池寿命。然而，当前版本的LoRaWAN可以从提高的高传输精度和降低传感器节点的能耗中受益。为了解决当前用于建筑内气候控制系统的LPWAN和机器学习解决方案的上述限制，本项目旨在研究用于资源管理的整体网络和学习框架。特别是，该项目侧重于建筑能源管理作为一个例子应用。目标是设计一个建筑能源管理系统，在满足人体舒适度要求的同时，共同优化建筑内气候系统的能耗。提出的建筑能源管理系统旨在实现三个设计目标：1)最大限度地节约能源，同时保持居住者的舒适度；2)能够在多个系统的建筑物中部署；3)以数据效率寻找最优控制策略。拟议的研究活动将与加州大学默塞德分校的教育活动仔细结合起来，包括课程开发、跨学科教育和参与代表性不足的群体。加州大学默塞德分校是一所为西班牙裔服务的机构。该项目将建立一个整体的网络和学习框架，共同控制建筑中的气候。将实现三个设计目标：最大限度地节约能源，同时保持居住者的舒适度，易于部署在具有多个控制系统的建筑物中，并通过数据效率寻找最佳控制策略。为了实现这些设计目标，本项目分为三个研究重点：1)通过新颖的无速率数据传输机制和位级网络资源分配方案，构建可靠数据采集的室内低功耗广域网系统；2)通过解决训练数据不完整、表征学习缓解数据噪声、系统动力学模型偏差等基础研究问题，开发基于数据高效模型的资源管理强化学习系统；3)设计了一种网络与学习协同设计方案，该方案在统一的框架下考虑了一组优化目标，包括建筑节能、居住者舒适度、强化学习的数据效率和无线网络寿命。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。

Modern buildings require complex management of heating, ventilation, and air-conditioning (HVAC), lighting, blinds, and windows through collection of data from a variety of sensors. LoRaWAN (an open LPWAN protocol) along with LoRa physical layer technology (long range) has been a good choice for large-scale sensor networking with ability to offer long communication distance at low cost. According to LoRa Alliance (which manages the development of LoRaWAN), LoRaWAN outperforms conventional wireless networks (e.g., ZigBee and Wi-Fi) for smart buildings in many aspects, such as easy coverage of several floors in a building with a simple gateway, low power consumption, and long battery life of up to ten years. However, the current version of LoRaWAN can benefit from improved high transmission accuracies and lower energy consumption of sensor nodes. To tackle the above limitations of current LPWAN and machine learning solutions used for in-building climate control system, this project aims to investigate holistic networking and learning framework for resource management. In particular, the project focuses on building energy management as an example application. The goal is to design a building energy management system that optimizes the energy consumption of in-building climate systems jointly while meeting requirements of human comfort. The proposed building energy management system aims at achieving three design goals: 1) maximizing energy saving while maintaining occupants’ comfort, 2) being able to be deployed in buildings of multiple systems, and 3) searching for the optimal control policy with data efficiency. The proposed research activities will be carefully integrated with education activities at UC Merced, a Hispanic-serving institution, including curriculum development, interdisciplinary education, and engaging underrepresented groups.This project will develop a holistic networking and learning framework to jointly control climate in building. Three design goals will be achieved: maximizing energy saving while maintaining occupants’ comfort, being readily deployable in buildings with multiple control systems, and searching for the optimal control policy with data efficiency. TO achieve these design goals, the project is organized into three research thrusts: 1) building an indoor low-power wide area networking system for reliable data collection by a novel rateless-enabled data transmission mechanism and a bit-level network resource allocation scheme; 2) developing a data-efficient model-based reinforcement learning system for resource management by tackling fundamental research problems, such as incomplete training data, representation learning for mitigating data noise, and the bias problem of system dynamics models; 3) designing a networking and learning co-design scheme that considers a set of optimization goals in a unified framework, including building energy saving, occupants’ comfort, data efficiency of reinforcement learning, and wireless network lifetime.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.