Transfer-Learning Based BiLSTM for Building Energy Consumption Prediction Considering Seasonal Features

Accurately predicting building energy consumption is essential for power grids to efficiently plan energy scheduling and ensure stable operation. However, challenges arise due to limited energy consumption data for new buildings, making precise forecasting difficult. Additionally, energy consumption patterns vary across seasons, further complicating prediction accuracy. To address these issues, this paper proposes a Transfer Learning-based BiLSTM model that incorporates seasonal features for building energy consumption forecasting. First, the source domain data are clustered into a new source dataset using the k-means clustering algorithm based on relevant features. Next, the target domain is divided into four seasonal subsets—spring, summer, autumn, and winter. For each season, the optimal source domain is selected using the Maximum Mean Discrepancy (MMD) criterion. The Bidirectional Long Short-Term Memory (BiLSTM) model is initially pre-trained on the selected source domain, and the resulting model is saved. Finally, transfer learning with fine-tuning is applied to the target domains to improve the predictions. This approach enhances the accuracy of energy consumption forecasting by effectively leveraging seasonal characteristics.