A CONTROLLER FOR HVAC SYSTEMS WITH FAULT DETECTION CAPABILITIES BASED ON SIMULATION MODELS

This paper describes a control scheme with fault detection capabilities suitable for application to HVAC systems. The scheme uses static simulation models of the system under control to generate feedforward control action, which acts to supplement a conventional PI(D) feedback loop. The feedforward action reduces the effect of plant non-linearity on control performance and provides more consistent disturbance rejection as operating conditions change. In addition to generating feedforward control action, the models act as a reference of correct operation. Faults that occur in the HVAC system under control cause the PI(D) controller to provide a greater than normal control action to compensate for faultinduced inaccuracies in the feedforward models. The controller monitors the level of feedback compensation and generates alarms when thresholds are exceeded. The paper presents results from testing the controller with a dual-duct air-handling unit. INTRODUCTION Heating, ventilating, and air-conditioning (HVAC) systems are typically controlled using proportional plus integral (and sometimes plus derivative) PI(D) control law. In practice, HVAC systems exhibit nonlinear operating characteristics, which causes control performance to vary when operating conditions change. Poor control performance can lead to occupant discomfort in a building, greater energy consumption, and increased wear on controlled elements, such as actuators, valves, and dampers. In a conventional PI(D) feedback loop, the controller does not contain much information about the process it is controlling. Faults that lead to performance deterioration or a change in system behavior are often masked within a feedback loop. The control scheme described in this paper uses a model of the correctly operating system to supplement a conventional PI(D) feedback loop. The model is used as part of a feedforward control regime in order to reduce the effects of plant non-linearity on control performance. In addition, the model acts as a reference of correct behavior, which facilitates the detection of faults that develop in the controlled system. Several researchers (e.g. Gertler, 1998; Glass et al., 1994; Isermann, 1995; Patton et al., 1995) have proposed fault detection and diagnosis schemes based on the use of models. The main trade-off with model-based schemes is configuration effort versus model accuracy. Generally, the greater the potential accuracy of the models, the greater the effort required to configure the models for operation. The models in the proposed controller have therefore been selected to be configurable from performance information typically available during a system life cycle. Although model accuracy and fault sensitivity are sacrificed to a certain extent, the paper demonstrates that the proposed scheme is capable of detecting two important faults in the air-handling unit tested. DESCRIPTION OF THE CONTROLLER Figure 1 shows the control and fault detection scheme. A conventional PI(D) feedback loop generates control action ( uPI) based on the error between the setpoint and the controlled variable. This feedback control action is then supplemented by a control signal ( uFF) generated by a simulation model, which is an inverse representation of the system. The model is in static form and produces a control action appropriate for the current setpoint and measured disturbances. The control scheme is similar to that proposed by Hepworth and Dexter (1994), which used an adaptive neural network as the inverse system model. PI Contro l ler Inverse Sys tem M o d e l u FF u PI uPI+FF var iable cont ro l led + + + Sys tem dis turbances