Besher Alhalabi

• Deep learning
• IoT
• Machine learning
• Algorithms

• Programming
• Machine learning
• Neural networks
• C#
• Python
• Data visualization
• Data analysis

• MSc Business Intelligence, Birmingham City University

Designing lightweight machine learning algorithms for IoT solutions for energy efficient predictive building control

The Internet of Things (IoT) are smart devices that sense objects or can be controlled remotely across existing network infrastructure. This has allowed more direct integration of the physical world into computer-based systems, and resulting in improved efficiency, accuracy and economic benefit in addition to reduced human intervention. In housing, IoT can perform various building controls for energy efficiency. For example, controlling lights, ventilation systems, cooling/heating systems. However, most control operations for these systems are performed reactively, for example, sensors waiting for the internal temperature to reach a particular set temperature to switch on/off heating/cooling systems, or using motion sensors to turn on/off building lights.

Reactive building control are slowly causing oversupply of energy, which acquire unnecessary cost. It can also cause undersupply for various services, which can cause customer dissatisfaction. For example, when these devices fail to maintain user's comfortable temperature, this is when the IoT requires longer time to run the heating system in response to a change in room temperature. This normally leads to lack of trust among users in IoT solutions, hence, prefer the traditional approaches using the manual control settings or the typical scheduling control mechanisms, which are inefficient and possibly worsen the issue of service oversupply/undersupply.

IoT devices can perform smart monitoring and building energy control, but, if they were able to learn about the building structure, tenant’s behaviour and relevant internal / external weather condition, they would generate smarter and faster control decisions. In this context, machine learning algorithms can be used on edge devices to proactively learn building behaviour to predict ahead the occurrence of various scenarios, hence, response faster to these conditions in real time. The benefits of this approach are no single point of failure, there will be no need for cloud connectivity and it eliminates any decision delays caused by high network latency or low network throughput. Running machine-learning algorithms on the edge devices, e.g. sensors, requires data acquisition, data processing, and decision making to be performed locally on the devices. However, due to IoT’s low processing power, the low memory available to them, and concerns about power consumption may all limit the ability to run machine learning at the edge device. Despite the continuous advances in IoT devices in terms memory, processing power and battery storage, running machine learning can easily exhaust its processor and quickly drain its battery power. In fact, IoT devices will sustain these constraints as their purpose is to be cheap in cost for large-scale deployments, can be deployed anywhere without the complication of cables and to “possibly” accommodate sensor mobility.

The aim of this project is to design lightweight machine learning algorithms that can run on IoT devices without exhausting its resources. It will also focus on performing multi-objective optimisation to identify optimal machine learning algorithms and data acquisition techniques that work the best for various types of IoT devices. In addition, to identify the optimal communication mode between IoT devices by using various collaboration models when processing the data. For example, IoT may share knowledge to generate adaptive accurate models for buildings to be efficiently monitored and controlled.

• Mohamed Gaber

• LinkedIn profile