Dynamic learning and resource management under uncertainties for smart grid and cognitive radio networks

Abstract

The importance of timely applications and decisions in dynamic environments, has led to the integration of intelligent networks to increase efficiency and end-user satisfaction in various application domains including telecommunication and power grid networks. Contemporary intelligent networks require advanced statistical signal processing and optimization tools to learn, infer and control their operation. This integration poses new challenges and has witnessed the emergence of novel resource management and learning techniques to cope with dynamics. In addition, in order to have implementable resource management algorithms, it is crucial to model the underlying sources of uncertainty in the optimization framework. This thesis develops algorithms for resource allocation under channel uncertainty in cognitive radio (CR) communication networks and contributes to demand coordination under uncertainty in power networks.Demand coordination through real-time pricing is addressed first by capitalizing on the uncertainty involved in the consumption behavior of consumers. Prerequisite to the demand coordination task is learning the uncertainty present in power consumption data. The dependency of consumers' consumption behavior on the announced prices and their neighbors' behavior, is modeled through graphical models. In particular, the electric vehicle (EV) consumers are considered and the adopted model also captures dynamics of EV consumers' time-varying charging decisions. Leveraging the online convex optimization (OCO) framework, an online algorithm for tracking the model is devised. With minimal assumptions on the structure of the temporal dynamics, and while accounting for the possibly adversarial consumption behavior of consumers, the proposed online algorithm provides performance guarantees. The probability distributions obtained through the tracking algorithm are then deployed as input to stochastic economic profit maximization for real-time price setting.Learning in the presence of missing data is a pervasive problem in statistical data analysis. Next, attention is turned to tracking the dynamic charging behavior of EV consumers, when at each time slot some of the consumers' consumption decisions are possibly missing. The problem amounts to online classification with missing labels. An online algorithm is proposed to wed real-time estimation of the missing data with learning of complete data in the OCO framework.As regards CR networks, this thesis introduces novel resource allocation algorithms for orthogonal frequency-division multiple access (OFDMA) CR under channel uncertainty where the unique approaches can be fitted to a class of large-scale robust mixed-integer problems. Due to the lack of cooperation of the licensed system, CRs must resort to less efficient channel estimation techniques thus incurring an inevitable channel estimation error. It is shown that CR interference constraints under channel uncertainty can be cast as chance constraints. On the other hand, instead of just modeling the user rates by logarithmic functions of transmit-powers, justified under ideal Gaussian coding, practical finite-alphabet constellations are adopted which leads to an optimization objective of a weighted sum of mutual information. When multiple users are present, due to the combinatorial search for optimal subcarrier assignment, the problem is non-convex and hard to solve, as the optimization variables are coupled across all subcarriers. To circumvent the resulting computational hurdle, tight and conservative approximations of the chance constraint are introduced to break the coupling and enforce separability per subcarrier. The <italic>separable</italic>problem across subcarriers opens the door to the dual decomposition approach, which leads to a near-optimal and computationally efficient solution.