Cross Domain Transfer Learning Techniques for Residential Energy Forecasting

Residential energy prediction has been accomplished by various techniques. For the improvement of residential energy forecasting in terms of accuracy and efficiency, a Graph Neural Network (GNN) model with transfer learning capacity and Temporal Fusion Transformer (TFT) model have been adopted and compared in this study. Group 1 utilizes the designed GNN model, and Group 2 utilizes the TFT model. Both models use a sample size of 20 datasets and analyze various time periods and trends in energy usage. The predictive accuracy, training time, computational complexity, and flexibility of both models are compared. With a 2.7% increase in prediction accuracy, 35% decrease in training time, 20% decrease in computational cost, and improved generalization across many residential energy datasets, the TFT model is superior to the GNN model. The results indicate that the suggested TFT model provides a more efficient and scalable residential energy forecasting method compared to the GNN model with better computing efficiency and flexibility.