Usman Gani Joy , Shahadat Kabir , A.F.M. Farhad , Asraful Islam
{"title":"基于自适应多尺度注意力的跨异构数据集鲁棒能源和资源预测框架","authors":"Usman Gani Joy , Shahadat Kabir , A.F.M. Farhad , Asraful Islam","doi":"10.1016/j.prime.2025.101070","DOIUrl":null,"url":null,"abstract":"<div><div>Accurate time series forecasting is crucial for energy management and resource allocation, aligning with the United Nations Sustainable Development Goals (SDGs) for affordable and clean energy (SDG 7). However, dataset complexity and heterogeneity challenge existing models, such as Principal Component Analysis-Transformer (PCA-Transformer), which rely on static techniques and struggle with feature extraction, scalability, and adaptability to evolving patterns. This study introduces a framework integrating a Wavelet-Inspired Multi-Scale Transform (WIMST), Adaptive Transform (AT), Residual Blocks (RB), and Attention Mechanism (AM). The WIMST dynamically extracts multi-scale features, AT models complex interactions, RB stabilize deep training, and AM captures long-range dependencies, synergistically addressing non-linear patterns in dynamic datasets. Evaluated on the University of California, Irvine (UCI) Appliances Energy and U.S. Energy Information Administration (EIA) Renewable Energy datasets, the model achieves a Root Mean Square Error (RMSE) of 0.1899, Mean Absolute Error (MAE) of 0.1478, and coefficient of determination (<span><math><msup><mrow><mi>R</mi></mrow><mrow><mn>2</mn></mrow></msup></math></span>) of 0.9998 on UCI (versus baselines’ RMSE of 0.350–14.39), and RMSE values of 0.4393 (Hydro), 0.3093 (Waste), 2.9767 (Solar), and 0.1081 (Geothermal) on EIA, reducing errors by 70%–97% compared to Autoregressive Integrated Moving Average (ARIMA). These results highlight superior accuracy and robustness for energy forecasting across diverse applications.</div></div>","PeriodicalId":100488,"journal":{"name":"e-Prime - Advances in Electrical Engineering, Electronics and Energy","volume":"13 ","pages":"Article 101070"},"PeriodicalIF":0.0000,"publicationDate":"2025-07-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"An adaptive multi-scale attention-based framework for robust energy and resource forecasting across heterogeneous datasets\",\"authors\":\"Usman Gani Joy , Shahadat Kabir , A.F.M. Farhad , Asraful Islam\",\"doi\":\"10.1016/j.prime.2025.101070\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>Accurate time series forecasting is crucial for energy management and resource allocation, aligning with the United Nations Sustainable Development Goals (SDGs) for affordable and clean energy (SDG 7). However, dataset complexity and heterogeneity challenge existing models, such as Principal Component Analysis-Transformer (PCA-Transformer), which rely on static techniques and struggle with feature extraction, scalability, and adaptability to evolving patterns. This study introduces a framework integrating a Wavelet-Inspired Multi-Scale Transform (WIMST), Adaptive Transform (AT), Residual Blocks (RB), and Attention Mechanism (AM). The WIMST dynamically extracts multi-scale features, AT models complex interactions, RB stabilize deep training, and AM captures long-range dependencies, synergistically addressing non-linear patterns in dynamic datasets. Evaluated on the University of California, Irvine (UCI) Appliances Energy and U.S. Energy Information Administration (EIA) Renewable Energy datasets, the model achieves a Root Mean Square Error (RMSE) of 0.1899, Mean Absolute Error (MAE) of 0.1478, and coefficient of determination (<span><math><msup><mrow><mi>R</mi></mrow><mrow><mn>2</mn></mrow></msup></math></span>) of 0.9998 on UCI (versus baselines’ RMSE of 0.350–14.39), and RMSE values of 0.4393 (Hydro), 0.3093 (Waste), 2.9767 (Solar), and 0.1081 (Geothermal) on EIA, reducing errors by 70%–97% compared to Autoregressive Integrated Moving Average (ARIMA). These results highlight superior accuracy and robustness for energy forecasting across diverse applications.</div></div>\",\"PeriodicalId\":100488,\"journal\":{\"name\":\"e-Prime - Advances in Electrical Engineering, Electronics and Energy\",\"volume\":\"13 \",\"pages\":\"Article 101070\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2025-07-18\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"e-Prime - Advances in Electrical Engineering, Electronics and Energy\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S2772671125001779\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"e-Prime - Advances in Electrical Engineering, Electronics and Energy","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2772671125001779","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
An adaptive multi-scale attention-based framework for robust energy and resource forecasting across heterogeneous datasets
Accurate time series forecasting is crucial for energy management and resource allocation, aligning with the United Nations Sustainable Development Goals (SDGs) for affordable and clean energy (SDG 7). However, dataset complexity and heterogeneity challenge existing models, such as Principal Component Analysis-Transformer (PCA-Transformer), which rely on static techniques and struggle with feature extraction, scalability, and adaptability to evolving patterns. This study introduces a framework integrating a Wavelet-Inspired Multi-Scale Transform (WIMST), Adaptive Transform (AT), Residual Blocks (RB), and Attention Mechanism (AM). The WIMST dynamically extracts multi-scale features, AT models complex interactions, RB stabilize deep training, and AM captures long-range dependencies, synergistically addressing non-linear patterns in dynamic datasets. Evaluated on the University of California, Irvine (UCI) Appliances Energy and U.S. Energy Information Administration (EIA) Renewable Energy datasets, the model achieves a Root Mean Square Error (RMSE) of 0.1899, Mean Absolute Error (MAE) of 0.1478, and coefficient of determination () of 0.9998 on UCI (versus baselines’ RMSE of 0.350–14.39), and RMSE values of 0.4393 (Hydro), 0.3093 (Waste), 2.9767 (Solar), and 0.1081 (Geothermal) on EIA, reducing errors by 70%–97% compared to Autoregressive Integrated Moving Average (ARIMA). These results highlight superior accuracy and robustness for energy forecasting across diverse applications.