Residential Building Heating Load Prediction using Deep Learning TabNet Model

Abstract

Efficient heating load (HL) predictions in residential buildings are vital in energy optimization and cost savings. Deep learning models with a high ability to solve complex tasks could be good tools for this purpose. Therefore, in this study, we proposed a new deep learning model for heating load prediction based on an attentive interpretable tabular learning model (TabNet). The in-hand dataset contains only 768 records which causes deep learning models’ weak performance in comparison with classical machine learning tools. To solve the problem and utilize the ability of deep learning models, a new hybrid model that combines TabNet and a tabular data augmentation module based on GAN and CGAN methods has been proposed. The data augmentation module increased the size of the training dataset 5 times. The performance results indicated that the BiLSTM outperforms other well-known deep learning models without data augmentation, including ResNet, Fully Connected Neural Networks (FCNN), TabNet, And LSTM. By utilizing the data augmentation module the TabNet-GAN model outperformed other deep learning models and brought comparable results with other classical machine learning HL prediction models. Therefore, the TabNet-GAN model could be used for residential building HL  prediction and help engineers select the best plan/design from the energy usage perspective.