Graph Model, Multi-head Attention Mechanism, Ultra-Short-Term Load Forecasting, User-Level Load Forecasting Abstract: Against the backdrop of strong volatility and randomness in loads within the new power system, power load forecasting confronts challenges like limited feature information and difficulties in capturing peak demand. In this research, we propose a transformer-based deep graph model that integrates multi-scale periodic features, edge type-aware attention mechanisms, and multi-objective loss optimization for high-accuracy electricity time series forecasting. First, we construct multi-scale graph structure data, mapping time points to nodes and modeling continuity and periodicity dependencies through horizontal edges (adjacent time series), periodic edges (the same time period across days), and selfcyclic edges. Second, we introduce learnable edge embedding vectors into the encoder and improve the multi-head attention mechanism of the Transformer to enable the model to adaptively distinguish between local conduction and cross-periodic association patterns. Finally, we design a hybrid loss function to collaboratively optimize overall accuracy and peak tracking ability. Experiments based on enterprise-level 15-minute granular load data demonstrate that our research method significantly outperforms benchmark models such as LSTM and GRU in terms of MAE, RMSE, and MAPE. Additionally, ablation experiments verify the effectiveness of each component's design. https://mc.manuscriptcentral.com/jcmse Journal of Computational Methods in Science and Engineering For Peer Review Page 1 of 12 https://mc.manuscriptcentral.com/jcmse Journal of Computational Methods in Science and Engineering 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 For Peer Review Transformer-based Deep Graph Model for Electricity Time Series Modelling Yixin Li1, Yiwen Jiang1*, Hanyu Feng1, Lin Wang2 and Xu Cheng2 1. State Grid Information & Telecommunication Center (Big Data Center), Beijing, China. 2.School of Computer Science and Technology, Tongji University, Shanghai, China. * Correspondence to: 1064131401@qq.com Abstract: Against the backdrop of strong volatility and randomness in loads within the new power system, power load forecasting confronts challenges like limited feature information and difficulties in capturing peak demand. In this research, we propose a transformer-based deep graph model that integrates multi-scale periodic features, edge type-aware attention mechanisms, and multi-objective loss optimization for high-accuracy electricity time series forecasting. First, we construct multi-scale graph structure data, mapping time points to nodes and modeling continuity and periodicity dependencies through horizontal edges (adjacent time series), periodic edges (the same time period across days), and self-cyclic edges. Second, we introduce learnable edge embedding vectors into the encoder and improve the multi-head attention mechanism of the Transformer to enable the model to adaptively distinguish between local conduction and cross-periodic association patterns. Finally, we design a hybrid loss function to collaboratively optimize overall accuracy and peak tracking ability. Experiments based on enterprise-level 15-minute granular load data demonstrate that our research method significantly outperforms benchmark models such as LSTM and GRU in terms of MAE, RMSE, and MAPE. Additionally, ablation experiments verify the effectiveness of each component's design. Keywords: Graph Model, Multi-head Attention Mechanism, Ultra-Short-Term Load Forecasting, User-Level Load Forecasting 1.