Development of an energy cost prediction model for a VRF heating system

• Published in: Applied thermal engineering 2018-07, Vol.140, p.476-486
• Main Authors: Park, Bo Rang, Choi, Eun Ji, Hong, Jongin, Lee, Je Hyeon, Moon, Jin Woo
• Format: Article
• Language: English
• Subjects: Air temperature; Algorithms; Artificial neural network; Artificial neural networks; Coefficient of variation; Control model and algorithm; Control theory; Energy consumption; Energy cost; Error analysis; Fluid pressure; Heating; Heating system; Indoor environments; Learning theory; Mathematical models; Model accuracy; Model testing; Neural networks; Neurons; Optimization; Performance evaluation; Performance tests; Predictive and adaptive controls; Predictive control; Refrigerants; Thermal environments
• ISSN: 1359-4311; 1873-5606
• DOI: 10.1016/j.applthermaleng.2018.05.068

•A predictive and adaptive ANN model was developed for controlling heating system.•The model predicted heating energy cost for the different variable settings.•Model optimization was conducted for the accurate and stable prediction.•The optimized model demonstrated its prediction accuracy within the recommended level. This study developed a predictive model using artificial neural network (ANN) to forecast the energy cost for a variable refrigerant flow (VRF) heating system. The energy cost is predicted with the ANN model by considering the set-points for the refrigerant condensation temperature, condenser fluid temperature, condenser fluid pressure, and air handling unit supply air temperature together with past operational data and other climatic data. The predicted energy cost was used as a determinant for the control algorithm to optimize the heating system operation in terms of cost. The study consisted of three steps: initial model development, model optimization, and performance evaluation. The neural network toolbox in the Matrix laboratory was used to develop the model and conduct the performance tests. For the model training and performance evaluation, data sets were collected in the winter from a test building. Initial model consisted of a structure that included ten input neurons and a learning method. Then, the optimization process was used to find the optimal structure of the ANN model, which was 1 hidden layer with 15 hidden neurons, while the optimal learning method had a 0.5 learning rate and 0.4 momentum. In the performance evaluation, the optimized model demonstrated its prediction accuracy to be within the recommended level, with 0.8417 r2 and 4.87% coefficient of variation root mean squared error between the measured and the predicted costs, thus proving its applicability in the control algorithm to supply a comfortable indoor thermal environment in a cost-efficient manner.