Integrating Machine Learning in residential Building Management System
Towards self regulation by means of sensor-based modelling and data-based simulation. A case study.
This dissertation approaches the field of Internet of Things and adaptive Building Management Systems (BMS) in residential units. It aims to implement sensor-based modelling in conjunction with Machine Learning (ML) to generate a data-based model of the working cycles of a standard house. This way, several control policies regarding thermal comfort and energy consumption could be simulated, tested and optimized without the need of a complex physics-based model. The thesis also approaches the determination of the most relevant inputs for the estimation of several variables that have implications in energy demands and user’s comfort in dwellings, in an attempt to reduce the number of necessary sensors to instrument a house with an Intelligent BMS.
As an ideal study case, and drawing from previous research, the experiments were realized on a dataset gathered in a house equipped with over 100 sensors collecting data during a year in a 15 minute timestamp with simulated occupancy patterns. The determination of the most relevant inputs was performed by means of a Least-Squares Support Vector Machine model and Automatic Relevance Determination. The results showed that in most cases, interesting variables can be predicted with a reduced number of inputs, which in most cases are exclusively weather and temporal inputs. This implies that a heavy instrumentation is not necessary to implement an intelligent BMS.
The thesis also presents alternative ML models for specific variables, and proposes several practical applications prototyped in different pieces of software: A BMS able to evaluate behavioural patterns and program the activation of several devices according to expected usage (e.g. water heater), the estimation of energy generation in the next 12-24 hours by reading online weather forecasts to shift energy-demanding tasks to surplus times, and a simulation of an intelligent Thermostat to control internal temperature. The results show that a model based only on data and virtual human input can effectively bypass the need for physics-based models for management and optimization of residential unit’s control policies. Simpler and more affordable sensor installations are possible in new and already existing spaces to incorporate such a system. The presented models not only allow for energy savings and adaptiveness, but also fit in the SmartGrid framework, presenting in turn a potential for collective implementation.