Scholarlink

<p>The inherent intermittency and temporal changes in solar irradiance create the need for developing the strong predictive model for reducing system uncertainty, boosting grid resilience, and assisting with the best energy-management practices. In this regard, machine learning techniques, especially deep learning architectures, have proven to be highly effective in capturing non-linear and non-stationary dynamics in time series of photovoltaic power. As a result, this study suggests a novel framework for solar PV power forecasting that combines metaheuristic-driven hyperparameter optimization with a Transformer-based prediction model. In order to methodically determine the best tuning strategy, this study compares three sophisticated optimization algorithms including VSO, SCA, and HLO in improving the Transformer’s predictive accuracy and generalization performance. The findings show that the Transformer–VSO model compared to others indicates strong and balanced generalization performance with high R2 equal to 97.88% and low RMSE of 11.71 and MAE of 6.27 during testing. Furthermore, its low MAPE of 22.20% verifies dependable accuracy under various operating situations, and low NRMSE value of 0.07 indicates consistent performance over the normalized data range. According to the convergence behavior analysis, the Transformer-VSO model exhibits superior convergence behavior, as indicated by a quick and steady decrease in MSE from 0.03724 to less than 0.03714 within a small range, which confirms the model’s robustness, stability, and efficacy in producing dependable predictive performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust Renewable Energy Prediction Using Transformer Networks Coupled With Evolutionary Metaheuristic Search Strategies

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Renping Yu,
  • Tianhe Yang,
  • Song Hu

摘要

The inherent intermittency and temporal changes in solar irradiance create the need for developing the strong predictive model for reducing system uncertainty, boosting grid resilience, and assisting with the best energy-management practices. In this regard, machine learning techniques, especially deep learning architectures, have proven to be highly effective in capturing non-linear and non-stationary dynamics in time series of photovoltaic power. As a result, this study suggests a novel framework for solar PV power forecasting that combines metaheuristic-driven hyperparameter optimization with a Transformer-based prediction model. In order to methodically determine the best tuning strategy, this study compares three sophisticated optimization algorithms including VSO, SCA, and HLO in improving the Transformer’s predictive accuracy and generalization performance. The findings show that the Transformer–VSO model compared to others indicates strong and balanced generalization performance with high R2 equal to 97.88% and low RMSE of 11.71 and MAE of 6.27 during testing. Furthermore, its low MAPE of 22.20% verifies dependable accuracy under various operating situations, and low NRMSE value of 0.07 indicates consistent performance over the normalized data range. According to the convergence behavior analysis, the Transformer-VSO model exhibits superior convergence behavior, as indicated by a quick and steady decrease in MSE from 0.03724 to less than 0.03714 within a small range, which confirms the model’s robustness, stability, and efficacy in producing dependable predictive performance.