Boosting short term electric load forecasting of high & medium voltage substations with visibility graphs and graph neural networks

• Published in: Sustainable Energy, Grids and Networks Vol. 38; p. 101304
• Main Authors: Giamarelos, Nikolaos, Zois, Elias N.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.06.2024
• Subjects: Graph Neural Networks; Short-term Load Forecasting; Visibility Graph; Visibility Graph; Short-term Load Forecasting; Graph Neural Networks
• ISSN: 2352-4677; 2352-4677
• DOI: 10.1016/j.segan.2024.101304

Modern power grids are faced with a series of challenges, such as the ever-increasing demand for renewable energy sources, extensive urbanization, climate and energy crisis, which set new standards and requirements in terms of their stability, sustainability and security, paving the way to transition to smart grids. Energy management, planning and control systems that include state-of-the-art technologies such as fifth-generation technology standard for cellular networks, Artificial Intelligence, Machine Learning and Internet of Things are considered to be the driving force of this transition, with short-term load forecasting constituting a crucial component in any such system. In this work, a novel graph neural networks model is proposed, which is boosted by the properties of visibility graphs for day-ahead mixed load forecasting. Specifically, the electric load time-series are transformed into an undirected graph by employing the visibility graph approach and then, trains a graph neural network with inputs the obtained graph along with the corresponding adjacency matrix. The efficiency of the so-called visibility graph neural networks is assessed by conducting a plethora of experiments using real data from a high voltage/medium voltage substation. Extensive results that presented confirm the dominance of the proposed method over widely used machine learning methods. •A novel graph neural networks model boosted by visibility graphs is proposed for STLF.•The proposed VGNN is tested for day-ahead predictions upon real mixed power-load data.•Not standard load conditions are considered as prediction targets.•The VGNN outperforms all competitive methods in terms of MAE and R2.•The computational complexity of GNN is significantly decreased by introducing VG.