Random Neural Network Based Smart Controller for Heating, Cooling and Ventilation in Domestic and Non-Domestic Buildings

  • Abbas Javed

Student thesis: Doctoral ThesisDoctor of Philosophy (PhD)

Abstract

Buildings currently consume 40% of the total energy in most developed countries. Nowadays,people spend more than 90% of their lives in buildings therefore indoor comfort is crucial.However, smart control of heating, ventilating and air conditioning (HVAC) systems while maintaining adequate indoor comfort remains problematic. Training algorithms (such as gradient descent - GD) often lead to sub-optimal performance. Furthermore, intelligent sensors to control the HVAC system require high processing power and memory, both of which are difficult to solve.
In this work, three controllers (model-free random neural network (RNN), RNN model based controller, and hybrid RNN with model predictive control (RNN-MPC)) were developed and tested for maintaining comfortable environment in residential building by controlling the motorized thermostatic radiator valves (TRVs) mounted on radiators. Global optimization techniques (particle swarm optimization (PSO), differential evolution (DE), and artificial bee colony (ABC)) and local optimization techniques (sequential quadratic programming(SQP)) were used for training the RNN. The hybrid training algorithms were developed by integrating the global optimization technique (ABC, PSO) with local optimization technique (SQP) were implemented for the first time to train the RNN. The integration of cloud computing with wireless sensor networks (WSN) was proposed for training of RNN models,data storage and data representation. The three different architectures of smart controller(Case 1- smart controller with cloud based control algorithm implementation, Case 2- RNN based centralized smart controller. Case 3- cloud enabled decentralized smart controller )were tested and compared in terms of control accuracy, battery consumption, control decision delay and memory consumption.
The results show the accuracy of the RNN controller was 26% more than artificial neural networks (ANN) controller and conserves 2.75% more energy than ANN controller at predicted mean vote (PMV) based temperature set points. The performance of MPC was slightly better than RNN controller in terms of control accuracy but MPC was computationally expensive as it required physical model of the building and optimization technique to determine the optimal parameters. The ANN controller was reliable only for the set points included in the training dataset. The RNN controller was robust and performed accurately even for the set points not covered in the training dataset.
Thus, the RNN model based controller was further developed for optimizing the energy consumption of the residential heating system. The integral squared error (ISE) of MPC for maintaining PMV based setpoints were 64.9% higher than RNN model based optimization techniques. The integral absolute error (IAE) of MPC for PMV based set points were 57.78% less than RNN model based controller. The hybrid RNN-MPC model was trained with dataset of MPC for maintaining PMV based setpoints and dual setpoints. However hybrid RNN-MPC controller was not robust.
In terms of training algorithms ABC-SQP outperformed other training algorithms as evidenced by lower MSE and NRMSE. A further problem confronting smart controllers isthe accurate estimation of room occupancy. The present study therefore used CO2 and temperature from the HVAC duct and indoor air for estimating the occupancy. The results showed that accuracy of occupancy estimation was up to 83% and energy consumption was reduced between 27% to 48%. The decentralized architecture for smart controller was also implemented in which occupancy estimation algorithms were implemented on sensor nodes while RNN models for PMV based setpoint estimation and HVAC control were implemented on base station. The occupancy estimation algorithm was further improved (up to 88%) in decentralized architecture with hybrid RNN occupancy estimation algorithm (which combines RNN based occupancy estimation model and occupancy estimation with passive infrared sensor/magnetic reed switches). Of the three architectures for smart controllers, Case 3 was found to the most efficient in terms of power consumption and control decision delay. Thus,the decentralized architecture of the RNN based smart controller with hybrid occupancy estimation algorithm trained with ABC-SQP algorithm is the best architecture in terms of battery consumption, control decision delay, and control accuracy.
Date of Award2016
Original languageEnglish
Awarding Institution
  • Glasgow Caledonian University
SupervisorHadi Larijani (Supervisor)

Cite this

'