A building cluster load forecasting method based on time-series-aware multi-task learning

Building cluster load forecasting is a critical technology for achieving supply-demand balance and stable operation of energy systems. However, data noise and feature redundancy hinder prediction accuracy. To address this, a time-series-aware multi-task learning (MTL) method for load forecasting is proposed and validated on a real-world building cluster dataset. Firstly, an adaptive denoising framework combining Variational Mode Decomposition (VMD) and Stationary Wavelet Transform (SWT) is employed to enhance data quality. Secondly, a dynamic feature selection mechanism, integrating Pearson Correlation Coefficient (PCC) and Gradient Boosting Decision Tree (GBDT), identifies high-contribution feature subsets to improve model adaptability. Finally, an MTL model, incorporating a Convolutional Neural Network(CNN), Bidirectional Long Short-Term Memory Network(BiLSTM), and time-series clustering, is developed to capture shared and individual temporal patterns across buildings. Results show that the proposed method significantly improves the prediction accuracy, with a weighted root mean square error of 24.6% higher than the baseline model.