Faisal Saeed , Abdul Rehman , Hasnain Ali Shah , Muhammad Diyan , Jie Chen , Jae-Mo Kang
{"title":"SmartFormer: Graph-based transformer model for energy load forecasting","authors":"Faisal Saeed , Abdul Rehman , Hasnain Ali Shah , Muhammad Diyan , Jie Chen , Jae-Mo Kang","doi":"10.1016/j.seta.2024.104133","DOIUrl":null,"url":null,"abstract":"<div><div>Electric load forecasting is a pivotal component in the power industry, providing essential intelligence for optimizing smart grid operations. Energy load data, inherently characterized as a multivariate time series, is influenced by various interdependent factors such as weather conditions, economic activity, and seasonal variations, all of which significantly impact the overall load dynamics. Though deep learning techniques, particularly with transformer-based models, have achieved significant progress in forecasting time series data, a gap exists in adequately acknowledging the importance of inter-series dependencies in multi-series load data. This paper proposes a novel graph-nested transformer model to effectively capture inter-series dependencies and forecast the load using a graph structure. The proposed Transformer model addresses two primary challenges: efficiently representing various temporal patterns and reducing redundant information within the series. In the proposed model, the graph neural network components are seamlessly integrated into the Transformer layers, allowing for the fusion of sequence encoding and graph aggregation in an iterative workflow. Evaluations across four distinct datasets demonstrate the superiority of the proposed model over state-of-the-art techniques in power load forecasting.</div></div>","PeriodicalId":56019,"journal":{"name":"Sustainable Energy Technologies and Assessments","volume":"73 ","pages":"Article 104133"},"PeriodicalIF":7.1000,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Sustainable Energy Technologies and Assessments","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2213138824005290","RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"ENERGY & FUELS","Score":null,"Total":0}
引用次数: 0
Abstract
Electric load forecasting is a pivotal component in the power industry, providing essential intelligence for optimizing smart grid operations. Energy load data, inherently characterized as a multivariate time series, is influenced by various interdependent factors such as weather conditions, economic activity, and seasonal variations, all of which significantly impact the overall load dynamics. Though deep learning techniques, particularly with transformer-based models, have achieved significant progress in forecasting time series data, a gap exists in adequately acknowledging the importance of inter-series dependencies in multi-series load data. This paper proposes a novel graph-nested transformer model to effectively capture inter-series dependencies and forecast the load using a graph structure. The proposed Transformer model addresses two primary challenges: efficiently representing various temporal patterns and reducing redundant information within the series. In the proposed model, the graph neural network components are seamlessly integrated into the Transformer layers, allowing for the fusion of sequence encoding and graph aggregation in an iterative workflow. Evaluations across four distinct datasets demonstrate the superiority of the proposed model over state-of-the-art techniques in power load forecasting.
期刊介绍:
Encouraging a transition to a sustainable energy future is imperative for our world. Technologies that enable this shift in various sectors like transportation, heating, and power systems are of utmost importance. Sustainable Energy Technologies and Assessments welcomes papers focusing on a range of aspects and levels of technological advancements in energy generation and utilization. The aim is to reduce the negative environmental impact associated with energy production and consumption, spanning from laboratory experiments to real-world applications in the commercial sector.