Scholarlink

<p>Accurate energy load forecasting is essential for optimizing power grid management, reducing operational costs, and enhancing energy efficiency. Traditional forecasting models rely heavily on real-world historical data, which often contains noise, irregular fluctuations, and privacy constraints, limiting their generalization and robustness. This study explores the impact of synthetic data augmentation on load forecasting models by comparing real data, synthetic data (generated by utilizing an autoencoder-based generative adversarial networks (GAN)), and blended-data training approaches. A comprehensive quantitative and qualitative evaluation was conducted across multiple forecasting models, including autoregressive integrated moving average (ARIMA), convolutional neural network-long short-term memory (CNN-LSTM), gated recurrent units (GRU), neural basis expansion analysis for time series (N-BEATS), and temporal fusion transformer (TFT), with a particular focus on the 70% real + 30% synthetic (70R+30S) blended model. Statistical analysis using ANOVA, Kruskal–Wallis, and Tukey’s HSD tests validates that blended models significantly outperform real-data-only models in predictive accuracy and error stability. The 70R+30S model achieved 40.16% reduction in mean absolute percentage error (MAPE), 39.95% reduction in mean absolute error (MAE), and 39.72% reduction in root mean square error (RMSE) compared to real-data models, confirming its superiority in balancing adaptability and noise reduction. Furthermore, qualitative findings highlight that synthetic data effectively enhances model robustness, mitigates privacy concerns, and ensures stable forecasting performance even in the presence of missing or highly volatile real-world data.</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

Synthetic data-augmented AI models for load forecasting: Enhancing accuracy, robustness, and generalization through optimized blending strategies

  • 期刊论文

  • 发表日期：

    2026年4月30日

查看全文
  • Mohit Choubey,
  • Rahul Kumar Chaurasiya,
  • J S Yadav

摘要

Accurate energy load forecasting is essential for optimizing power grid management, reducing operational costs, and enhancing energy efficiency. Traditional forecasting models rely heavily on real-world historical data, which often contains noise, irregular fluctuations, and privacy constraints, limiting their generalization and robustness. This study explores the impact of synthetic data augmentation on load forecasting models by comparing real data, synthetic data (generated by utilizing an autoencoder-based generative adversarial networks (GAN)), and blended-data training approaches. A comprehensive quantitative and qualitative evaluation was conducted across multiple forecasting models, including autoregressive integrated moving average (ARIMA), convolutional neural network-long short-term memory (CNN-LSTM), gated recurrent units (GRU), neural basis expansion analysis for time series (N-BEATS), and temporal fusion transformer (TFT), with a particular focus on the 70% real + 30% synthetic (70R+30S) blended model. Statistical analysis using ANOVA, Kruskal–Wallis, and Tukey’s HSD tests validates that blended models significantly outperform real-data-only models in predictive accuracy and error stability. The 70R+30S model achieved 40.16% reduction in mean absolute percentage error (MAPE), 39.95% reduction in mean absolute error (MAE), and 39.72% reduction in root mean square error (RMSE) compared to real-data models, confirming its superiority in balancing adaptability and noise reduction. Furthermore, qualitative findings highlight that synthetic data effectively enhances model robustness, mitigates privacy concerns, and ensures stable forecasting performance even in the presence of missing or highly volatile real-world data.