Scholarlink

<p>This study introduces and evaluates the Optimal Feature Selection Convolutional LSTM (OFSCL) model, a hybrid deep learning architecture designed for accurate short-term energy demand forecasting at distribution substations using real-world data. Unlike traditional statistical models that struggle with non-linear and high-dimensional patterns, OFSCL integrates automatic feature selection within convolutional layers and employs LSTM units to effectively capture temporal dependencies. This architecture achieves superior predictive accuracy, robustness to noise, and adaptability across varying demand patterns, outperforming classical machine learning approaches and advanced deep learning models such as LSTM and CNN. After extensive hyperparameter tuning, OFSCL achieved an R² of 90%, an MAE of 0.55 MWh, and an RMSE of 0.74 MWh, demonstrating strong forecasting performance. The model captures both spatial and temporal dynamics in load data. Additionally, gradient-based sensitivity analysis identifies air temperature, month, and relative humidity as key contributors to forecasting accuracy, enhancing robustness against environmental variability and supporting informed feature prioritization.</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

A hybrid optimal feature selection and Conv-LSTM model (OFSCL) for short-term energy demand forecasting in distribution substations of Ahvaz, Iran

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月2日

查看全文
  • Mehdi Mohammadian Mehr,
  • Hossein Farzin,
  • Elaheh Mashhour

摘要

This study introduces and evaluates the Optimal Feature Selection Convolutional LSTM (OFSCL) model, a hybrid deep learning architecture designed for accurate short-term energy demand forecasting at distribution substations using real-world data. Unlike traditional statistical models that struggle with non-linear and high-dimensional patterns, OFSCL integrates automatic feature selection within convolutional layers and employs LSTM units to effectively capture temporal dependencies. This architecture achieves superior predictive accuracy, robustness to noise, and adaptability across varying demand patterns, outperforming classical machine learning approaches and advanced deep learning models such as LSTM and CNN. After extensive hyperparameter tuning, OFSCL achieved an R² of 90%, an MAE of 0.55 MWh, and an RMSE of 0.74 MWh, demonstrating strong forecasting performance. The model captures both spatial and temporal dynamics in load data. Additionally, gradient-based sensitivity analysis identifies air temperature, month, and relative humidity as key contributors to forecasting accuracy, enhancing robustness against environmental variability and supporting informed feature prioritization.