Utilizing Deep Transfer Learning Techniques to Forecast Building Energy Consumption for Enhanced Sustainable Energy Practices

Introduction: In recent years, accurately predicting building energy consumption has gained significant momentum to support decarbonization and efficient energy management. However, traditional machine learning methods depend largely on historical data for training and work best with accessible data. Currently, forecasting energy demand for new buildings is challenging due to limited consumption data. To overcome this challenge, this study explores the use of deep transfer learning techniques to predict energy usage in buildings with limited historical data. Methods: The research uses a structured experimental method, combining baseline models with advanced deep learning techniques. A benchmark dataset with real-time, appliance-level energy consumption data was utilized to develop the model. An Artificial Neural Network (ANN) was initially developed as a baseline predictor for the energy use of the targeted building. A deep transfer learning framework was then introduced, employing Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models. The pre-trained base model was developed and fine-tuned, enabling knowledge transfer from a data-rich source domain to the targeted domain. Results: Experimental results demonstrate that the deep transfer learning method surpasses standalone LSTM and GRU models, achieving the lowest Root Mean Square Error (RMSE) of 0.859061 and decreasing training time by about 1.18406 seconds. Discussion: These findings demonstrate the effectiveness of deep transfer learning in overcoming data scarcity. This method offers a patentable and sustainable approach to enhancing energy efficiency strategies in real-world building applications. Conclusion: Deep transfer learning offers an effective way to predict building energy consumption with limited historical data, supporting efficient energy management.