Long-term scenario generation for distribution network loads based on interpretable diffusion models

• Interpretable diffusion model enhances long-term load scenario generation, overcoming GAN limitations. • High-credibility annual load scenarios support hydrogen storage capacity planning. • A physically interpretable generation mechanism is developed through a temporal signal decoupling framework. • The proposed method achieves 31% average improvement over benchmarks in key metrics. The increasing complexity of modern power systems reveals significant limitations in traditional distribution network planning methods when addressing load evolution patterns under multi-factor coupling effects. Generating long-term, physically interpretable scenario data has become crucial for enhancing planning rationality. This paper proposes an interpretable diffusion model-based method for long-term distribution network load scenario generation. The approach enhances the diffusion model by integrating a Transformer architecture and a temporal decomposition mechanism to explicitly model multi-scale load characteristics. A spectral domain loss is introduced to ensure frequency-domain fidelity. Experimental results demonstrate that the proposed method significantly outperforms existing benchmarks, achieving an average improvement of 31% in key metrics including Context-FID and Correlational Score. The generated scenarios exhibit high fidelity in both temporal dynamics and frequency-domain characteristics, effectively supporting downstream planning tasks. Ablation experiments further validate the contributions of key model components. This work provides a credible and interpretable solution for generating long-term load scenarios, offering robust data support for distribution network planning involving emerging elements like hydrogen storage.