Short-Term Electricity Demand Forecasting for DanceSport Activities

Abstract

This paper introduces a novel hybrid deep learning-based approach for short-term electricity demand forecasting in dance sport activities. Traditional deep learning methods often overlook important spatial dependencies and key features like trend and seasonal patterns. To address these limitations, we propose a model that combines Transformer for temporal feature extraction and Graph Neural Networks for spatial feature extraction, enabling prediction based on spatial-temporal features. Additionally, we employ the decomposition techniques to extract seasonal and trend features from dance sports data. By integrating early fusion (feature-level fusion) and late fusion (score-level fusion) strategies, our model achieves superior performance, outperforming baseline methods by over 4% on benchmark datasets. Additionally, we conduct the ablation study to comprehensively analyze the impact of each module on prediction accuracy, providing valuable insights into the contribution of spatial, temporal, seasonal and trend features to the overall forecasting performance.

Keywords

• Short-term demand forecasting
• graph neural networks
• DanceSport
• hybrid fusion