Accurate Residential Net Load Forecasting With Hybrid Statistical and Machine Learning Techniques Enhanced by Advanced Decomposition Method

Abstract

The increased use of renewable energy (RE) in residential settings is driving the demand for smart energy management systems (EMS) to optimally coordinate the purchasing and selling of power to the grid. Given the uncertain and volatile nature of RE and electricity consumers’ changing load behavior, the EMSs require accurate net load forecasting, which has become a tough challenge. Unlike prior approaches that might employ a single, fixed Long Short-Term Memory (LSTM) architecture for all data components regardless of their underlying complexities, the model proposed in this paper introduces a specifically designed advanced hybrid machine-learning algorithm for highly accurate net load forecasting in residential areas equipped with solar panels. This involves applying more complex architectures to high-frequency data components and simpler architectures for lower-frequency components. This tailored approach is crucial for optimizing network complexity and computational burden while maintaining or enhancing forecasting performance, leading to comparable accuracy with significantly shorter training times compared to fixed architectures. This paper provides brief descriptions of powerful forecasting tools to accurately predict different variables in power systems, and presents a case study of these models in different combinations. This paper further compares them across three different time resolutions of data, and discusses their characteristics and different evaluation metrics. Compared to other methods that have similar computational efficiency, the proposed model demonstrates superior accuracy as the time resolution of the input data increases in order to accommodate more intricate details in the RE generation and consumers’ load behavior.