A study of machine learning-based model predictive control on building energy system

• Main Author: Zhang, Qindong
• Other Authors: Wan Man Pun
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Mechanical engineering::Energy conservation; Engineering::Mechanical engineering::Control engineering
• Online Access: https://hdl.handle.net/10356/152491
• Institution: Nanyang Technological University

At present, most of the building automation and control (BAC) are using the conventional PID control or simple feedback control. However, due to the simplicity and absence of knowledge of the controlled system, the control performance of the conventional control is poor for dynamic systems. As a result, the BAC system using traditional feedback control to deal with building dynamics makes a big challenging in terms of managing building energy and indoor thermal comfort. This dissertation tries to propose a machine learning-based model predictive control (ML-based MPC controller). The MPC controller is able to solve the optimization and control problem of nonlinear dynamic system. Machine learning is one of the most advanced and popular data driven technologies to be the reliable alternative of traditional physical building energy model. To build the MPC controller, four types of ML model, which are support vector machine (SVM), random forest (RF), nonlinear autoregressive exogenous model with external inputs (NARX) and deep network long-short-term-memory (LSTM) are analyzed by using generated building climate and thermal comfort data. The data is obtained though building thermal simulation on IES VE software. The results show that the four models have similar prediction accuracy, but the computing cost is different. Generally, among the four models, NARX and SVM are acceptable to work as the prediction model of MPC controller. After ML model analysis, a MPC controller is developed by merging MPC and ML technologies. SVM model was selected as the building thermal model and LSTM neural network was used as controlled emulator in Matlab Simulink. This MPC controller aims to control room air handling unit (AHU) in a target office in Singapore. The MPC combined with ML to capture building thermodynamics and comfort, is used to predict the thermal comfort index, namely predicted mean vote (PMV), in room space. Subsequently, the cost function with multi-objectives is used to optimize energy consumption and thermal comfort. At last, the performance of developed MPC controller is evaluated according to simulation results.