This paper provides some new techniques to predict the electric load using Multiple Linear Regression (MLR) model, which adopts a statistical approach assuming that the past load and weather data have an information for forecasting the target load. Since the conventional general MLR prediction performance can be degraded by seasonal effects, we propose some new MLR techniques to improve the prediction performance. We have found that the performance of the proposed MLR can be further improved by solving weighted least squares problem and clustering the training set. Also, we try to compare the prediction performances for such techniques to find out the best. Our argument will be demonstrated by the two case studies with real electric and weather data.

A self-organizing network refers to a computer network that can configure, manage, and optimize itself. To achieve this goal, it first collects the network information for performance analysis at some control entities. After the control entities determine the optimized parameters, they push the settings back to the network again using the same channel they use to collect the information. Network optimization has been extensively studied over the last two decades. Most works ignore actual mechanisms for data collection and parameter update. This is partially true because device vendors, such as Cisco or Juniper, have their own control interfaces. Unfortunately, these interfaces are generally not compatible with each other, and so a unified control protocol, OpenFlow, emerged.

OpenFlow defines a set of commands for data collection and parameter update for wired networks. However, wireless networks, or the most richly existing Wi-Fi, were not the primary targets of OpenFlow. In this dissertation, we first showed OpenFlow protocol could be migrated from wired networks to wireless networks with minor software modification and use to build a software-defined mobile ad hoc network (SD MANET) prototype.

Seeing the tight constraints in our SD MANET, we moved on to loosen these limitations. We came up with an approach to support most nowadays smart devices using the standard IEEE 802.11v/r protocol to replace the last mile connection between the devices and their associated access points. Our solution did not require any hardware or software modification on users' devices but only need the built-in IEEE 802.11v/r support, which can be found in many mainstream smart devices.

In addition to smart devices having rich network capability. We noticed another group of wireless connectivity devices but did not support advanced protocols such as IEEE 802.11v/r due to their simplified architectures. These devices include the ever-popular internet-of-things devices built on simple, low-power and low-cost Wi-Fi system-on-chip solutions. Because Wi-Fi connection consumes lots of power, today's Wi-Fi IoT devices generally require external power support, and we can hardly see battery-powered Wi-Fi IoT devices. Seeing the demand, we design a non-coherent wake-up receiver that takes over the channel monitoring task of a power-consuming Wi-Fi interface so that the Wi-Fi interface can be completely turned off. Our wake-up receiver consumes only 20-40 $\mu$W when monitoring the channel, whereas a general Wi-Fi interface can easily drain more than 100 mW. We also came up with an extendable finite-state-machine design that supports multicast wake-up. Multiple receivers can wake up through a single, carefully selected wake-up signal.

Practicality is the main idea of this dissertation. We have seen solutions with amazing performance but required either huge investment or complicated hardware design. In this dissertation, we chose the other way around by first analyzing the capabilities of existing frameworks and design solutions installed as overlays. Through this process, practicality is guaranteed.