APPLICATION OF NEURAL NETWORKS TO PREDICT THE PERFORMANCE OF A RUN-AROUND MEMBRANE ENERGY EXCHANGER (RAMEE)
The objective of this thesis is to develop a fast model to predict the performance of a Run-Around Membrane Energy Exchanger (RAMEE). The RAMEE is a novel air-to-air energy recovery system that is capable of transferring both heat and moisture between two air streams. The utilization of a properly controlled RAMEE in the HVAC system of a building can significantly reduce the required energy to condition the ventilation air of the buildings. A neural network (NN) approach is applied to model the steady-state and transient performance of the RAMEE under a wide range of operating and system parameters. In order to approximate the underlying function of a physical system using a NN approach, a set of examples (data points) that describes the behavior of the system is required to train NNs. In this study, two separate numerical models that predict the steady-state and transient effectiveness of the RAMEE are utilized to generate the required training data sets. The Back-Propagation algorithm is applied to feed-forward NNs of different architectures to minimize the errors between the predicted effectivenesses by the numerical models and the NN models. Root Mean Squared Error (RMSE) between the results of steady-state NN models and the steady-state simulations are 0.05 °C for the sensible NN and 0.02 gv/kga for the latent NN. The accuracy of the transient NNs are reported in terms of Mean Absolute Difference (MAD) which are 0.5 °C for the sensible model and 0.2 gv/kga for the latent NN. The steady-state NN models show an excellent accuracy and the accuracy of transient NN models are quite acceptable for energy transfer calculation purposes, which is the main application of the NN models. The main advantage of NNs over numerical models is the non-iterative nature of the NN models that provides a very fast feed-forward model that can generalize and predict the RAMEE effectiveness for any practical operating condition in a fraction of second. This simplicity of predictions allowed the steady-state NN models to be used with a simple optimization algorithm to find the optimal performance of RAMEE during each operational hour. The TRNSYS computer simulations used the output of the optimized NNs to predict the annual energy savings caused by an optimally controlled RAMEE for an office building as well as a hospital. The results for an office building show up to 43% heating energy saving in cold climates, and up to 15% cooling energy saving in hot climates. The same analysis for the application of an optimally controlled RAMEE in the HVAC system of a hospital shows even more energy savings. The optimized RAMEE reduces the annual heating energy by 58 ‐ 66% in cold climates, and the annual cooling energy by 10 ‐ 18% in hot climates. The RAMEE allows the heating system to be downsized by 45% in cold climates, and the cooling system to be downsized by 25% in hot climates.
DegreeMaster of Science (M.Sc.)
SupervisorSimonson, Carey J.; Besant, Robert W.
CommitteeSchoenau, Greg J.; Gupta, Madan M.; Gokaraju, Ramakrishna
Copyright DateMay 2012
run-around membrane energy exchanger, neural network, heat and moisture exchange, steady-state, transient, sensible effectiveness, latent effectiveness