Scholarlink

<p>Short-term load forecasting (STLF) plays a critical role in ensuring the reliable and economical operation of power systems, particularly under complex and dynamic meteorological conditions. While Deep Residual Networks (DRNs) have demonstrated strong capability in modeling nonlinear load patterns, existing studies have predominantly focused on architectural enhancements, with limited attention given to input-level feature processing. This study conducts a controlled empirical evaluation of existing feature processing strategies for STLF in Malaysia using a unified DRN framework, highlighting their impact on forecasting performance under tropical conditions. Five representative approaches are investigated, including raw feature input, Principal Component Analysis (PCA), Pearson correlation coefficient (PCC)-based feature selection, Random Forest (RF)-based importance ranking, and Autoencoder (AE)-based representation learning. Unlike prior Principal Component Analysis–Deep Residual Network (PCA–DRN) studies, this work systematically evaluates all these strategies within a unified DRN framework on a tropical dataset, providing new empirical insights into how input-level feature processing affects STLF performance under tropical conditions. Experimental results show that incorporating feature processing significantly improves forecasting performance. Among all evaluated methods, the PCA–DRN framework achieved the best results on the Malaysia Petaling Jaya dataset. The observed performance may be dataset-dependent, and further validation on other tropical or weather-sensitive datasets is suggested. Statistical significance analysis based on bootstrap resampling further confirms that the performance improvements of PCA–DRN are robust and not due to random variation. The results indicate that preserving the global variance structure of meteorological variables is more effective than feature selection or nonlinear representation learning, particularly in tropical power systems where load patterns are highly sensitive to short-term weather fluctuations. These findings highlight the critical role of feature representation in DRN-based STLF and provide practical insights for improving forecasting accuracy and generalization 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

Empirical evaluation of feature processing strategies in deep residual networks for short-term load forecasting: a Malaysian tropical power system case study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月26日

查看全文
  • Junchen Liu,
  • Faisul Arif Ahmad,
  • Khairulmizam Samsudin,
  • Fazirulhisyam Hashim,
  • Mohd Zainal Abidin Ab Kadir

摘要

Short-term load forecasting (STLF) plays a critical role in ensuring the reliable and economical operation of power systems, particularly under complex and dynamic meteorological conditions. While Deep Residual Networks (DRNs) have demonstrated strong capability in modeling nonlinear load patterns, existing studies have predominantly focused on architectural enhancements, with limited attention given to input-level feature processing. This study conducts a controlled empirical evaluation of existing feature processing strategies for STLF in Malaysia using a unified DRN framework, highlighting their impact on forecasting performance under tropical conditions. Five representative approaches are investigated, including raw feature input, Principal Component Analysis (PCA), Pearson correlation coefficient (PCC)-based feature selection, Random Forest (RF)-based importance ranking, and Autoencoder (AE)-based representation learning. Unlike prior Principal Component Analysis–Deep Residual Network (PCA–DRN) studies, this work systematically evaluates all these strategies within a unified DRN framework on a tropical dataset, providing new empirical insights into how input-level feature processing affects STLF performance under tropical conditions. Experimental results show that incorporating feature processing significantly improves forecasting performance. Among all evaluated methods, the PCA–DRN framework achieved the best results on the Malaysia Petaling Jaya dataset. The observed performance may be dataset-dependent, and further validation on other tropical or weather-sensitive datasets is suggested. Statistical significance analysis based on bootstrap resampling further confirms that the performance improvements of PCA–DRN are robust and not due to random variation. The results indicate that preserving the global variance structure of meteorological variables is more effective than feature selection or nonlinear representation learning, particularly in tropical power systems where load patterns are highly sensitive to short-term weather fluctuations. These findings highlight the critical role of feature representation in DRN-based STLF and provide practical insights for improving forecasting accuracy and generalization performance.