NEURAL NETWORKS FOR PREDICTIVE HEATING VENTILATION AIR CONDITIONING CONTROLS

PREDICTIVE MODELS FOR HVAC CONTROL USING FEED-FORWARD NEURAL NETWORKS HAVE SHOWN GOOD PROMISE; HOWEVER, SUCH NETWORK TECHNIQUES ARE NOT DESIGNED FOR USE IN COMPLEX SPATIOTEMPORAL MODELING PROBLEMS SUCH AS PREDICTING TOTAL BUILDING CHILLED WATER DEMAND SINCE FEED-FORWARD MODELS HAVE NO INTERANL DYNAMICS. IN THIS PROJECT, WE PROPOSE TO EXTEND THE CURRENT WORK WITH NEURAL NETWORK PREDICTIVE MODELS THROUGH THE USE OF RECURRENT NETWORKS. SUCH NETWORKS HAVE SHOWN THE ABILITY TO ACCURATELY MODEL COMPLEX PROCESSES. WE WILL ALSO APPLY ADVANCED NEURAL NETWORK-BASED DATA ANALYSIS METHODS TO DETERMINE WHICH BUILDING PARAMETERS HAVE SIGNIFICANT PREDICTIVE POWER. THESE METHODS WILL BE APPLIED TO THE PROBLEM OF PREDICTING TOTAL CHILLED WATER DEMAND AND CHILLER POWER CONSUMPTION FOR SIMULATED BUILDING USING BLAST. THE BLAST PROGRAM WILL ALLOW EVALUATION OF THE NERUAL NETWORKS UNDER VARYING CONDITIONS, AND TO ACCESS THE ROBUSTNESS OF THESE METHODS TO CHANGES (SUDDEN AND GRADUAL) IN THE BUILDING CHARACTERISTICS. INCREMENTAL TRAINING METHODS WILL BE APPLIED TO ALLOW THE NETWORK TO ADJUST TO THESE CHANGES.