Scholarlink

<p>In today’s interconnected world, energy consumption forecasting faces challenges due to client-side anomalies in time-series data. Federated Learning (FL) offers a decentralized solution by forecasting without directly accessing user data. However, the effectiveness of the global model can decline if local anomalies are not properly managed. We propose our lightweight framework EIF-FL: Elliptic envelope and Isolation Forest with FL. It employs unsupervised ensemble Anomaly Detection (AD) before applying the FL process. EIF-FL consists of two layers. The first is on client-side and employs Isolation Forest and Elliptic Envelope with majority voting for AD. The second utilizes Long Short-Term Memory to forecast using FL on server-side. Simulations on energy consumption datasets show that EIF-FL improves AD metrics with accuracy, precision, and recall, having 0.94, 0.91, 0.92 simultaneously compared to the literature. It also enhances FL forecasting performances with test loss improvement (5.74%), SMAPE (13%), MAE (17%) compared to FedAvg without AD.</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

Combining client-based anomaly detection and federated learning for energy forecasting in smart buildings

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Bouchra Fakher,
  • Mohamed el Amine Brahmia,
  • Ismail Bennis,
  • Abdelhafid Abouaissa

摘要

In today’s interconnected world, energy consumption forecasting faces challenges due to client-side anomalies in time-series data. Federated Learning (FL) offers a decentralized solution by forecasting without directly accessing user data. However, the effectiveness of the global model can decline if local anomalies are not properly managed. We propose our lightweight framework EIF-FL: Elliptic envelope and Isolation Forest with FL. It employs unsupervised ensemble Anomaly Detection (AD) before applying the FL process. EIF-FL consists of two layers. The first is on client-side and employs Isolation Forest and Elliptic Envelope with majority voting for AD. The second utilizes Long Short-Term Memory to forecast using FL on server-side. Simulations on energy consumption datasets show that EIF-FL improves AD metrics with accuracy, precision, and recall, having 0.94, 0.91, 0.92 simultaneously compared to the literature. It also enhances FL forecasting performances with test loss improvement (5.74%), SMAPE (13%), MAE (17%) compared to FedAvg without AD.