Artificial Neural Networks-Based Models for an Internally Cooled Liquid Desiccant Dehumidifier

Abstract

Advanced internally cooled liquid desiccant air dehumidifiers enhance overall air-conditioning efficiency by incorporating internal cooling mechanisms. Their existing simulation models require detailed dehumidifier information, are computationally expensive, and pose challenges in convergence, which makes them unsuitable for integration into Building Energy Simulation software for air-conditioning system simulation. This study aims to investigate a method to generate models with less computational demands during simulation using artificial neural networks to represent the operation of an internally cooled dehumidifier. A comprehensive finite difference model was used to generate a data set representative of the expected operation conditions of the device for training the neural network. Various network configurations were explored to assess their impact on prediction precision, following a trial-and-error approach. During the training process, the neural networks reached R values between 0.96 and 0.98 for the different variables. Then, the networks were implemented in a stand-alone code, independent from training, and using basic programming methods. In this implementation, the trained networks underwent a secondary evaluation for prediction accuracy using a distinct data set from the training stage, proving accurate to simulate the internally cooled dehumidifier, reaching values well below 10% for the five predicted outlet variables in comparison to the validated, detailed finite differences model. The overall best performance was found for a network comprising two hidden layers of ten neurons each. This is the initial step towards incorporating neural network models into specialized Building Energy Simulation software, as the foundation for conducting system-level, transient, and long-term simulations.