Scholarlink

<p>Residential electrical load forecasting is considered essential for the planning and operation of electricity distribution systems. However, due to the unpredictable and fluctuating behavior of households and the complexity of datasets in this field, accurate load forecasting is often challenging. Additionally, the task of meeting electricity demand during household peak loads has been identified as a major challenge for modern power systems. This article uses a class-based approach for load forecasting to differentiate between on-peak, mid-peak, and off-peak areas of electrical load consumption. For each class, short-term electric load forecasting is performed using an ensemble model based on statistical, Machine Learning (ML), and Deep Learning (DL) methods. For these three methods, feature engineering uses encoding cyclical features to capture periodic patterns, incorporating the eXtreme Gradient Boosting (XGB) model to identify and rank influential features affecting load consumption. Within each class, the most accurate method is selected based on the performance metrics. Analysis of Variance (ANOVA) test is also employed to assess the reliability of the results through multiple runs. The results indicate that the DL method achieves higher accuracy in the on-peak class, the ML method performs better in the mid-peak class, and the statistical method yields the best results in the off-peak class. In the final forecast of electric load, it has been seen Mean Absolute Percentage Error (MAPE) of the proposed model is 2.4, 1, and 0.6% lower than statistical, ML, and DL methods respectively.</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

Very Short-Term Class-Based Residential Load Forecasting: An Ensemble of Statistical, ML and DL Models with Cyclical Feature Engineering

  • 期刊论文

  • 发表日期：

    2025年7月5日

查看全文
  • Abdolsalam Rasouli,
  • Mohammad Rastegar,
  • Mostafa Fakhrahmad

摘要

Residential electrical load forecasting is considered essential for the planning and operation of electricity distribution systems. However, due to the unpredictable and fluctuating behavior of households and the complexity of datasets in this field, accurate load forecasting is often challenging. Additionally, the task of meeting electricity demand during household peak loads has been identified as a major challenge for modern power systems. This article uses a class-based approach for load forecasting to differentiate between on-peak, mid-peak, and off-peak areas of electrical load consumption. For each class, short-term electric load forecasting is performed using an ensemble model based on statistical, Machine Learning (ML), and Deep Learning (DL) methods. For these three methods, feature engineering uses encoding cyclical features to capture periodic patterns, incorporating the eXtreme Gradient Boosting (XGB) model to identify and rank influential features affecting load consumption. Within each class, the most accurate method is selected based on the performance metrics. Analysis of Variance (ANOVA) test is also employed to assess the reliability of the results through multiple runs. The results indicate that the DL method achieves higher accuracy in the on-peak class, the ML method performs better in the mid-peak class, and the statistical method yields the best results in the off-peak class. In the final forecast of electric load, it has been seen Mean Absolute Percentage Error (MAPE) of the proposed model is 2.4, 1, and 0.6% lower than statistical, ML, and DL methods respectively.