Elsevier

Energy and Buildings

Volume 141, 15 April 2017, Pages 96-113
Energy and Buildings

Artificial neural network (ANN) based model predictive control (MPC) and optimization of HVAC systems: A state of the art review and case study of a residential HVAC system

https://doi.org/10.1016/j.enbuild.2017.02.012Get rights and content

Abstract

In this paper, a comprehensive review of the artificial neural network (ANN) based model predictive control (MPC) system design is carried out followed by a case study in which ANN models of a residential house located in Ontario, Canada are developed and calibrated with the data measured from site. A new algorithm called best network after multiple iterations (BNMI) is introduced to help in determining the appropriate ANN architecture. The prediction performance of the developed models using BNMI algorithm was significantly better (between 6% and 59% better goodness of fit for various models) when compared to a previous study carried out by the authors which used the default single iteration ANN training algorithm of MATLAB®. The ANN models were further used to design the supervisory MPC for the residential HVAC system. The MPC generated the dynamic temperature set-point profiles of the zone air and buffer tank water which resulted in the operating cost reduction of the equipment without violating the thermal comfort constraints. When compared to the fixed set-point (FSP), MPC was able to save operating cost between 6% and 73% depending on the season.

Introduction

Building sector consumes about 40% and 30% of total energy in U.S. and Canada respectively [1]. Space heating can consume up to 60% of total sector energy in countries with extreme weather conditions such as Canada [2]. Therefore, it is essential to investigate the methods for reducing energy consumption and operating cost of these systems. Heating, ventilation and air-conditioning (HVAC) systems are very complex and nonlinear systems due to the interaction of a large number of subsystems (e.g., chillers, boilers, heat pumps, pipes, ducts, fans, pumps and heat exchangers) and thermal inertia of the buildings. In order to calculate the total energy consumed by HVAC system, each of these subsystems need to be modeled precisely taking into account all the mass and energy transfer across each subsystem [3]. Using conventional forward or physics-based modeling methods, developing precise dynamic models of each subsystem is very difficult and a significant effort is required for a detailed understanding of system physics [4]. On top of that, these models have a large number of parameters e.g., thermal capacitance and thermal conductance for heat transfer component in each subsystem. These parameters need to be either extracted from the manufacturer supplied data or estimated using the parameter estimation techniques. Using manufacturer supplied data, the estimated parameters are less precise if this data was recorded under different set of operating conditions. Using parameter estimation techniques require measurements of system performance data further complicating the development of forward models. Limitations of these forward modeling methods become very evident when models for a variety of HVAC system configurations need to be developed and tuned impeding the deployment of such methods in real world. Forward modeling methods are useful for simpler systems and research based analysis of complex systems but fail to satisfy industrial users. Traditionally, industry has been very reluctant to adapt complicated methods in modeling and control of HVAC systems and due to this reason majority of HVAC systems are still using very simple on/off or proportional-integral-derivative (PID) controllers instead of more modern controllers such as model predictive control (MPC). Lack of an advanced controller results in many performance and economic penalties such as higher energy consumption, higher operating cost, higher thermal discomfort and higher equipment wear and tear [5]. Researchers have shown that by adding a supervisory MPC controller to HVAC systems could result anywhere from 7% to more than 50% reduction in energy consumption and operating cost reduction [6], [7], [8], [9], [10], [11]. On some so called advanced HVAC systems, rule-based supervisory controllers [12] are implemented which use the operator knowledge and apply it to reduce the energy and operating cost such as appropriate start/stop time of HVAC, night set-back, precooling and preheating etc. These rule-based controllers are simple to implement and require no modeling and design effort and can provide significant savings when applied correctly. Problem with these systems is that operator has to constantly monitor and adjust the HVAC operation to meet the objectives of reducing energy consumption while maintaining thermal comfort. Rule-based controllers generally are not anticipatory controllers i.e., they cannot look into the future and decide the course of action appropriately and rather work based on the current state of system. Maintaining a rule-based controller is a tedious process and advanced HVAC control task is better suited for an intelligent controller such as supervisory MPC which can automatically take into account the variations in weather parameters over a future horizon and control the appropriate settings of HVAC systems at present [13]. Furthermore, supervisory MPC can take into account the dynamic electricity price and adjust the set-points of local level controllers for active and passive thermal energy storage to offset the peak load to off-peak hours.

A simpler alternative to forward models is inverse or data-driven models [4], [14]. Inverse models can be developed comparatively easily since they do not require the understanding of system physics. In order to train the inverse models, a comprehensive set of input-output data of system is needed under all possible working conditions. Therefore, the ease of development of inverse models comes at the cost of reduced generalization capability compared to the forward models. Accuracy of inverse models decreases when training data deviates from testing data. Therefore, it is critical to train the inverse models with a training data that covers all the operating conditions which could be challenging especially for large scale systems such as HVAC systems which operate under a wide range of weather conditions throughout the year. For such systems, models trained under one set of conditions may not be accurate enough under different set of test conditions therefore the adaptive models are sometimes used. Alternatively, many researchers use different models in heating and cooling seasons for HVAC systems. Researchers have developed many types of inverse models such as frequency domain models (first and second order over-damped process with dead time) [15], data mining algorithms (artificial neural network-ANN [16], support vector machine-SVM [17]), fuzzy logic models (fuzzy adaptive network [18], Takagi-Sugeno fuzzy models [19], adaptive network based fuzzy inference system [20]), statistical models (regression [21], auto regression exogenous [22], auto regression moving average exogenous [23], auto-regressive integrated moving average [24]), state-space models (sub-space state space identification [25]), geometric models (thin plate spline approximation [26]), case-based reasoning (topological case base modeling [26]), stochastic models (probability density function approximation [27]) and instantaneous models (just in time models [28]). Out of all these modeling methods, ANN is the most popular method due to its high accuracy to model nonlinear systems compared to other methods. ANN mimics the human brain by using several neurons in multiple layers. The weights of these neurons are generally trained by using supervised learning methods. Appropriately trained ANN can approximate any nonlinear process to a high degree of accuracy. There are many different types of ANN structures but multi-layer perceptron (MLP) feedforward structure is the most popular one. Other types of ANN structures include radial basis function neural network, recurrent neural network and feedforward neural network with dynamic neutrons etc.

Since models developed with ANN are nonlinear, therefore, MPC approaches based on ANN are nonlinear. Linear MPC problems have guaranteed solution [29]. Minima of a linear optimization problem can be efficiently found by using optimization approaches such as active set method or interior-point method. On the other hand, nonlinear optimization problem may be non-convex and it may have many local minima. An algorithm that guarantees global minima of a nonlinear optimization problem does not exist in the literature yet. Another problem with nonlinear MPC is that the optimization solution might take a long time to converge which could be critical for fast moving processes though in the case of supervisory MPC for HVAC systems, this may not be an issue since supervisory MPC output only updates once every few minutes to every few hours. Global optimization methods such as evolutionary algorithms and simulated annealing can work with nonlinear optimization problems but present convergence and computational complexity issues.

The following are the main objectives and contributions of this paper:

  • i

    Highlighting the current research trends in ANN based MPC and its applications to HVAC control systems;

  • ii

    Development of a new ANN training algorithm to appropriately tune the network weights and aid in the selection of network architecture;

  • iii

    Comparison of developed ANN models with a previous research paper from the authors to highlight the ANN prediction performance improvements as a result of appropriate training algorithm;

  • iv

    Simulation of the ANN based MPC controller on the accurately calibrated residential HVAC system model; and

  • v

    Analysis of potential energy and cost savings using MPC compared to the fixed set-points on a residential HVAC system.

Section snippets

Review of ANN based MPC and optimization of HVAC systems

This section presents an in-depth review of ANN based MPC and optimization research for various types of HVAC systems. These approaches differ based on the types of buildings and HVAC systems, control objectives, modeling data generation, ANN architecture and optimization method selection. Following subsections discuss each of these factors in detail with specific examples from literature. At the end of this section, the energy and cost savings potential of ANN-MPC approaches is presented.

Case study: residential HVAC system

Following section describes the simulation of ANN-MPC on a residential HVAC system of Toronto and Region Conservation Authority (TRCA) Archetype Sustainable House (ASH) located in Vaughan, Ontario, Canada. There are two identical semi-detached houses called House A (ASHA) and House B (ASHB). ASHA has the HVAC system found in many Canadian households comprising of heat recovery ventilator, AHU and air source heat pump. Whereas, ASHB has the futuristic HVAC system focused on reducing the energy

Conclusions

In this paper, a comprehensive review of existing ANN based MPC approaches was carried out. Existing ANN-MPC approaches have been applied by HVAC researchers to various buildings including the university, airport, residential complex and office building but they have not been used for the most common residential houses found in all parts of Canada and North America. The researchers have generally considered the minimization of energy consumption while ignoring the operating cost of HVAC

Acknowledgements

The measured data was supplied by the Toronto and Region Conservation Authority team at Archetype Sustainable House. This research would have not been possible without their active collaboration.

References (59)

  • Q. Bi et al.

    Advanced controller auto-tuning and its application in HVAC systems

    Control Eng. Pract.

    (2000)
  • A. Kusiak et al.

    A data-driven approach for steam load prediction in buildings

    Appl. Energy

    (2010)
  • K. Chen et al.

    Fuzzy adaptive networks in thermal comfort

    Appl. Math. Lett.

    (2006)
  • S. Soyguder et al.

    Predicting of fan speed for energy saving in HVAC system based on adaptive network based fuzzy inference system

    Expert Syst. Appl.

    (2009)
  • L. Ferkl et al.

    Ceiling radiant cooling: comparison of ARMAX and subspace identification modelling methods

    Build. Environ.

    (2010)
  • I. Zlatanovic et al.

    Energy-saving estimation model for hypermarket HVAC systems applications

    Energy Build.

    (2011)
  • A.E. Ruano et al.

    The IMBPC HVAC system: a complete MBPC solution for existing HVAC systems

    Energy Build.

    (2016)
  • X. He et al.

    Performance optimization of HVAC systems with computational intelligence algorithms

    Energy Build.

    (2014)
  • H. Huang et al.

    A neural network-based multi-zone modelling approach for predictive control system design in commercial buildings

    Energy Build.

    (2015)
  • A. Kusiak et al.

    Optimization of an HVAC system with a strength multi-objective particle-swarm algorithm

    Energy

    (2011)
  • A. Kusiak et al.

    Minimization of energy consumption in HVAC systems with data-driven models and an interior-point method

    Energy Convers. Manag.

    (2014)
  • X. Wei et al.

    Multi-objective optimization of the HVAC (heating, ventilation, and air conditioning) system performance

    Energy

    (2015)
  • Y. Zeng et al.

    Predictive modeling and optimization of a multi-zone HVAC system with data mining and firefly algorithms

    Energy

    (2015)
  • B.C. Ng et al.

    Application of adaptive neural predictive control for an automotive air conditioning system

    Appl. Therm. Eng.

    (2014)
  • E. Asadi et al.

    Multi-objective optimization for building retrofit: a model using genetic algorithm and artificial neural network and an application

    Energy Build.

    (2014)
  • L. Lu et al.

    HVAC system optimization – in-building section

    Energy Build.

    (2005)
  • M. Ning et al.

    Neuro-optimal operation of a variable air volume HVAC&R system

    Appl. Therm. Eng.

    (2010)
  • K. Li et al.

    Forecasting building energy consumption using neural networks and hybrid neuro-fuzzy system: a comparative study

    Energy Build.

    (2011)
  • N. Alibabaei et al.

    Development of Matlab-TRNSYS co-simulator for applying predictive strategy plannin models on residential house HVAC system

    Energy Build.

    (2016)
  • Cited by (480)

    View all citing articles on Scopus
    View full text