MDL-Load: a power load forecasting method for small-sample scenarios

Abstract

Accurate power load forecasting is of great importance for maintaining the balance of power systems. Existing methods have difficulty in effectively integrating non-temporal features such as holiday information and are prone to overfitting under small-sample conditions. Meanwhile, large language models are limited in performance when directly applied to this task, as they do not inherently incorporate domain knowledge of power systems or the specialized characteristics of load sequences. To address these issues, this paper proposes a small-sample multimodal differential forecasting method based on large language model fine-tuning. The constructed structured contextual prompts integrate temporal numerical data with holiday semantics, enabling efficient alignment of multimodal information in a high-dimensional feature space. In addition, a differential mechanism and a low-rank adaptation fine-tuning strategy are introduced to transform absolute values into temporal increments, guiding the model to overcome numerical drift and capture the underlying evolution patterns of load series. Furthermore, a two-stage decoupled evaluation mechanism is designed to avoid the distortion trap of autoregressive evaluation metrics. Experimental results demonstrate that the proposed method effectively surpasses the accuracy limits of traditional models and significantly improves forecasting accuracy and generalization ability in small-sample scenarios.