Conditional diffusion models for probabilistic day-ahead load forecasting in modern power systems

Accurate probabilistic load forecasting is crucial for the safe and stable operation of power systems. However, mainstream forecasting paradigms, such as sequential multi-step prediction, are often challenged by error accumulation and practical lag-like effects (e.g., temporally smoothed trajectories during sharp ramps), which degrade their reliability. To address this knowledge gap, we shift the problem from direct prediction to conditional generation. We introduce DALNet, a novel denoising diffusion probabilistic model specifically engineered for probabilistic day-ahead load forecasting (PDALF). Instead of predicting a sequence of values, DALNet learns the complex conditional distribution of entire load curves and generates them as holistic samples, which mitigates error accumulation and reduces lag effects associated with sequential forecasting. To enhance the model’s ability to capture the intricate dynamics of load data, we propose a novel temporal multi-scale attention block (TMSAB) that integrates both positional and temporal dependencies. Furthermore, we utilize kernel density estimation (KDE) to reconstruct the distribution of generated load curves and employ Kullback–Leibler (KL) divergence to compare them with the actual data distribution. Experimental results demonstrate that DALNet excels in load forecasting accuracy and establishes a new and effective generative framework for tackling predictive tasks in power systems.