Temporal Convolutional Network with Adaptive Diffusion Model for Generation–Load Probabilistic Forecasting

Accurate generation–load forecasting is essential for the stability and efficiency of modern power systems. However, large-scale renewable integration and diverse user demand introduce strong nonlinearity and uncertainty, making probabilistic forecasting challenging. To address this, we propose a Temporal Convolutional Network with Adaptive Diffusion for generation–load probabilistic forecasting, called TCN-AD. TCN-AD employs a temporal convolutional encoder to capture long-term dependencies and local variations. In addition, an adaptive diffusion mechanism dynamically adjusts noise intensity to model time-varying uncertainty. Notably, a multi-scale fusion module and periodic attention mechanism further enhance the perception of multi-scale and cyclical patterns. Finally, a TCN-based denoising decoder refines the reverse diffusion process to reconstruct temporal dependencies effectively. Experiments on real-world load, solar, and wind datasets show that TCN-AD consistently outperforms baselines in both deterministic and probabilistic forecasting.