智能电表
米
计算机科学
能源消耗
消费(社会学)
适应性学习
智能电网
实时计算
人工智能
工程类
电气工程
社会学
天文
社会科学
物理
作者
Nawaf Abdulla,Mehmet Demirci,Suat Özdemır
标识
DOI:10.1016/j.segan.2024.101342
摘要
Forecasting short-term residential energy consumption is critical in modern decentralized power systems. Deep learning-based prediction methods that can handle the high variability of residential electrical loads have made models more accurate. On the other hand, these methods need a lot of sensitive information about how much people use something gathered centrally to train a forecasting model. This isn't good for privacy and scalability. Moreover, models may become less accurate over time due to changing conditions. In this work, we propose a framework for energy consumption forecasting that exploits adaptive learning, federated learning, and edge computing concepts. A central server aggregates numerous long short-term memory (LSTM) models that users at various locations trained with their energy consumption data to create a generalized model that uses adaptive learning to detect data drifts and enhance forecasting at the edge layer. Our findings show that adaptive federated learning performs better than centralized learning while preserving privacy, reducing communication overhead, lowering the forecast error rate by 8%, and decreasing the training time by approximately 80%.
科研通智能强力驱动
Strongly Powered by AbleSci AI