Participation in CAI 2026
Our lab member and PhD candidate, Stelio Bompai has successfuly submitted “Diffusion-based Data Augmentation for Short-Term Multivariate Energy Prediction in Data-Scarce Scenarios” on CAI 2026(https://www.ieeesmc.org/cai-2026/). This work explores diffusion-based generative models as a data augmentation strategy using the ETTh1 multivariate energy dataset, evaluating point and quantile forecasting across XGBoost, LSTM, and BiLSTM architectures, with synthetic data selectively added to neural models. Results show that BiLSTM benefits significantly from diffusion-based augmentation, achieving up to 15.3% reduction in RMSE and 8.1% reduction in MAE, along with consistent improvements in quantile forecasting, while LSTM performance deteriorates across all synthetic-to-real data ratios. Bias–variance analysis suggests that the improvements mainly stem from variance reduction at moderate levels of augmentation, whereas excessive synthetic data can introduce bias and degrade performance.