EAGER:Incremental and Distributed Learning in Nonstationary Environments with Applications to Wind Forecasting

项目概述本提案研究了可再生能源中应用的新型在线核学习算法。PI在机器学习和信号处理领域拥有经验和以前资助的NSF研究。PI现在正在进入可再生能源(特别是风能预报)的新领域，试图通过应用机器学习和信号处理知识在这一领域进行基础和应用研究。这一热切的提议提供了一个进入可再生能源和可持续发展这一领域的机会，这一领域对国家和夏威夷都具有关键意义(夏威夷清洁能源倡议(目标是到2030年70%的能源来自可再生能源和节能实践))。这项建议还将帮助PI成为可再生能源领域的技术和管理领导者，因为他与夏威夷大学的其他研究人员和行业领导者合作，寻求更多的研究和教育机会。智力优势：PI以前曾研究过最小二乘核算法，但对其实现仍有担忧。该建议通过考虑平衡性能、空间复杂性和计算复杂性的算法来解决其中的许多问题。我们开发了一套在线核算法，从子空间最小二乘算法到核LMS算法的变体。在线核算法工作在对偶空间，系统的维度随着训练样本的增加而增加。必须对数据进行窗口处理。对于子空间算法，通过控制支持向量的个数和矩阵的维度，来平衡性能和复杂度。我们还结合信号处理和机器学习的结果开发了分布式在线核算法。分布式算法是集成学习算法，非常适合以分布式方式收集信息的传感器网络中的操作。传感器网络可能有相关的物理限制(通信和能源成本)，这使得分布式处理和学习更具吸引力。为了理解这些不同的算法，我们分析它们的性能并确定不同算法的复杂性。采用系统辨识的方法分析了算法的均方误差性能。我们还研究了自适应在线学习算法的收敛性质。研究和分析了分布式学习算法的性能、复杂度和物理代价。一些分布式学习算法类似于Boosting算法，我们研究了这些算法的收敛行为。我们还考察了学习算法在非平稳环境中的性能。重点研究在线核学习算法在漂移环境中的性能。随着在线核学习算法的发展和分析，我们将更好地将学习算法应用于各种应用中。特别是，我们考虑将这些算法应用到风预测中。我们在这方面的一些初步结果表明，在线核算法可以准确地做出短期风速预测。我们将考虑使用分布式核心算法从多个传感器进行预测和长期预测。分布式在线核算法也非常适合于从传感器网络中提取其他信息。更广泛的影响：该提案是一项综合研究和教育努力，可能会对机器学习、信号处理和可再生能源产生重大影响。该方案的研究将加深对不同应用领域中自适应在线非线性核算法的理解和采用。随着美国和夏威夷走向更多的清洁能源解决方案(可再生能源和能源效率)，在做出关于能源储存和使用的决策时，将需要部署更多的智能。这份提案展示了如何将机器学习算法与信号处理相结合，用于准确的风速预测。这项拟议的研究还将对夏威夷大学的本科生和研究生产生重大的教育效益。将特别努力与工程学院的夏威夷原住民科学与工程导师计划(NHSEMP)合作，鼓励夏威夷原住民、太平洋岛民和女学生进入研究生研究计划。这项提案将补充小组最近在可再生能源领域撰写的其他提案。目标是让夏威夷大学工程学院成为利用信息技术开发清洁能源解决方案的领先者。PI还将研究与帝国理工学院、伦敦和日本的国际研究人员的合作。

Project SummaryThis proposal examines novel on-line kernel learning algorithms for applications in renewable energy. The PI has experience and previous funded NSF research in the machine learning and signal processing areas. The PI is now moving into new areas of renewable energy (specifically wind forecasting) attempting to do both basic and applied research in this area by applying machine learning and signal processing knowledge. This EAGER proposal provides an opportunity to move into this area of renewable energy and sustainability that is of key importance both nationally and in the state of Hawaii (with the Hawaii Clean Energy Initiative (goal of having 70% of energy generated from renewable sources and energy efficient practices by 2030). This proposal will also assist the PI in becoming a technical and administrative leader in the renewable energy area as he collaborates with other University of Hawaii researchers and industry leaders in pursuing more research and educational opportunities. Intellectual Merit: The PI has previously studied least squares kernel algorithms, but there are still concerns about their implementation. This proposal addresses many of these issues by considering algorithms that balance performance, space complexity, and computational complexity. We develop a suite of on-line kernel algorithms varying from subspace least squares algorithms to variants of kernel LMS algorithms. The on-line kernel algorithms work in the dual space where dimensionality of systems increase as we add training examples. Data must be windowed. For subspace algorithms the number of support vectors, and the dimensionality of the matrices, LS is controlled to balance performance and complexity. We also develop distributed on-line kernel algorithms combining results from signal processing and machine learning. The distributed algorithms are ensemble learning algorithms that are well suited for operation in sensor networks where information is gathered in a distributed manner. The sensor networks may have physical constraints (communication and energy costs) associated which make distributed processing and learning more attractive. To understand these different algorithms we analyze their performance and determine the complexity of the different algorithms. A system identification approach is used to analyze the mean and mean squared error performance of algorithms. We also study the convergence properties of the adaptive on-line learning algorithms. Distributed learning algorithms performance, complexity, and physical costs are studied and analyzed. Some of the distributed learning algorithms are similar to boosting and we study the convergence behavior of these algorithms. We also look at the performance of learning algorithms in nonstationary environments. A focus will be on studying the performance of on-line kernel learning algorithms in drifting environments. With a development and analysis of the on-line kernel learning algorithms we will be in a better position to apply the learning algorithms in a variety of applications. In particular we look at applying these algorithm to wind forecasting. We have some preliminary results in this area that show that on-line kernel algorithms can accurately make short term wind prediction. We will consider prediction from multiple sensors using distributed kernel algorithms and longer term prediction. Distributed on-line kernel algorithms are also well suited to extract other information from sensor networks. Broader Impacts: The proposal is an integrated research and educational effort that can have major impacts to machine learning, signal processing, and renewable energy. Research from this proposal will add to the understanding and adoption of adaptive on-line nonlinear kernel algorithms in different application areas. As the United States and Hawaii move towards more clean energy solutions (renewable energy and energy efficiency) more intelligence will need to be deployed in making decisions about energy storage and usage. This proposal shows how machine learning algorithms combined with signal processing can be used for accurate wind prediction. The proposed research will also have major educational benefits to both undergraduate and graduate students at the University of Hawaii. Special effort will be given to work with the Native Hawaiian Science and Engineering Mentorship Program (NHSEMP) in the College of Engineering to encourage Native Hawaiian, Pacific Islander, and women students to enter graduate research programs. This proposal will complement other group proposals that the PI has recently written in the renewable energy area. A goal is to have the University of Hawaii, College of Engineering being a leader player in using information technology in the development of clean energy solutions. The PI will also look at collaborations with international researchers in Imperial College, London and Japan.