Generative modeling for mid-term probabilistic load forecasting based on latent diffusion

Accurate mid-term load forecasting is essential for effective energy storage operational planning and benefit realization for various grid and end-user services. However, it remains challenging due to the limited accuracy of weather predictions beyond the short-term horizon. This paper presents an innovative mid-term probabilistic load forecasting method, specifically targeting month-ahead hourly load and daily peak predictions. We develop a generative modeling framework that jointly learns the statistical relationships between load and temperature using a latent diffusion model. By capturing these dependencies in a data-driven latent space, the approach can generate realistic mid-term scenarios that maintain load–temperature correlations even without reliable weather forecasts. The methodology first converts daily load and temperature profiles into compact image representations to capture monthly patterns. A vector-quantized variational autoencoder embeds these images into a latent space, where a conditional latent diffusion model is trained using historical and calendar information. During inference, new latent samples are generated and decoded to produce month-ahead hourly and daily peak load scenarios. We validate the proposed method using multiple real-world datasets with varying aggregation levels and demonstrate its superior performance compared to existing methods with around 15 % numerical improvements, particularly in terms of various probabilistic and deterministic forecasting measures.