Mohamed Ibrahim Abdi 國立台灣科技大學 營建工程系(指導教授 :Min-Yuan Cheng)
Predicting building thermal loads is critical for optimizing building energy management systems and enhancing energy efficiency. The high-dimensional characteristics of HVAC loads necessitate the selection of an optimal number of features to improve model accuracy. In this research, a novel approach is introduced that utilizes weighted feature selection methods. Four different techniques, including Pearson, Spearman, and Kendall correlations, as well as a dimensionality reduction method, Principal Component Analysis, are used to determine four distinct sets of features based on the dimensionality of feature numbers. After that, the combination of Neural Networks (NN) and Bidirectional Gated Recurrent Units (BiGRU) is used with the Optical Microscope Algorithm (OMA), an optimization algorithm inspired by microscopic mechanisms, to fine-tune the NN-BiGRU framework. This Opticalinspired Bidirectional Machine Learning model (OMA-NN-BiGRU) predicts the outcomes from the four data methods. Additionally, OMA is applied to optimize the weight combinations of the predictions from the individual feature selection methods, based on their contributions to the final prediction. The results demonstrate that the proposed model achieves significantly higher accuracy than the single methods. The weighted feature selection model demonstrates superior performance, with a Root Mean Squared Error
(RMSE) of 0.046, a Mean Absolute Error (MAE) of 0.032, a Mean Absolute Percentage Error (MAPE) of 12.6%, and R² values of 0.95 and 0.90, respectively. The OMA-NNBiGRU model outperforms other compared models in building thermal load prediction. This study offers fresh insights into creating an integrated approach based on different input
variable dimensions.
Keywords: Thermal Load Prediction; Weighted Feature Selection Approach; NeuralNetwork; Bidirectional Gated Recurrent Unit; Optical Microscope Algorithm