Conditional Diffusion with Dual-Domain Constraints: Resolving Holiday-Workday Imbalance via Fourier-Consistent Load Synthesis

To address the imbalance between working-day and holiday load distributions caused by scarce holiday samples—which degrades short-term load-forecasting accuracy and timeliness—this paper presents a conditional diffusion–forecasting unified framework. A TA-Diffusion model with few-step sampling, conditioned on working-day/holiday labels, generates high-fidelity holiday load sequences; these sequences are combined with a multi-scale CNN-BiLSTM-attention network to achieve fast and accurate prediction. Empirical results on the Panama City holiday subset show that the framework raises the coefficient of determination (R2) from 0.785 to 0.948, lowers the normalised root-mean-square error (nRMSE) and normalised mean absolute error (NMAE) by 22.4% and 25.7%, respectively, and reduces inference time by 58% compared with a standard diffusion approach. The method markedly enhances accuracy and deployability, providing an example of applying generative deep learning in power-system practice.