A backpropagation neural network improved by a genetic algorithm for predicting the mean radiant temperature around buildings within the long-term period of the near future

• Published in: Building simulation 2022-03, Vol.15 (3), p.473-492
• Main Authors: Xie, Yuquan, Ishida, Yasuyuki, Hu, Jialong, Mochida, Akashi
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Atmospheric Protection/Air Quality Control/Air Pollution; Back propagation; Back propagation networks; Building Construction and Design; Buildings; Cluster analysis; Clustering; Datasets; Energy consumption; Engineering; Engineering Thermodynamics; Genetic algorithms; Heat and Mass Transfer; Meteorological parameters; Monitoring/Environmental Analysis; Neural networks; Performance prediction; Principal components analysis; Research Article; Thermal comfort; Thermal environments; Training; Vector quantization
• ISSN: 1996-3599; 1996-8744
• DOI: 10.1007/s12273-021-0823-6

This study aimed to develop a neural network (NN)-based method to predict the long-term mean radiant temperature (MRT) around buildings by using meteorological parameters as training data. The MRT dramatically impacts building energy consumption and significantly affects outdoor thermal comfort. In NN-based long-term MRT prediction, two main restrictions must be overcome to achieve precise results: first, the difficulty of preparing numerous training datasets; second, the challenge of developing an accurate NN model. To overcome these restrictions, a combination of principal component analysis (PCA) and K -means clustering was employed to reduce the training data while maintaining high prediction accuracy. Second, three widely used NN models (feedforward NN (FFNN), backpropagation NN (BPNN), and BPNN optimized using a genetic algorithm (GA-BPNN)) were compared to identify the NN with the best long-term MRT prediction performance. The performances of the tested NNs were evaluated using the mean absolute percentage error (MAPE), which was ≤ 3% in each case. The findings indicate that the training dataset was reduced effectively by the PCA and K -means. Among the three NNs, the GA-BPNN produced the most accurate results, with its MAPE being below 1%. This study will contribute to the development of fast and feasible outdoor thermal environment prediction.