Empowering energy sustainability: A cutting-edge ensemble network for precise power forecasting

Abstract

An appropriate forecasting model for electrical energy generation and consumption is considered a key demand of modern power management systems. Therefore, it has attracted considerable attention from researchers to develop efficient and precise energy forecasting methods. Energy forecasting for residential buildings and power generation sites is vital in energy future planning and convenient management. Mainstream energy forecasting methods including conventional Machine Learning (ML) and Deep Learning (DL) methods contain high levels of non-linearity between input and output which require more improvement with respect to robustness, forecasting performance, and generalization ability in terms of effective energy generation and consumption forecasting. To address these problems, this paper presents a hybrid forecasting model, integrating a modified Convolution Neural Network (CNN) and Gated Recurrent Units (GRU) followed by an attention mechanism. In this architecture, CNN is designed to extract spatial information in historical input data while GRU emphasizes temporal information sequentially. Additionally, the extracted features are then fed to the Soft Attention (SA) module, aiming to acquire the dominant patterns for final forecasting. To ensure a fair evaluation, this study conducted extensive experiments comparing the proposed CNNGRU-SA model with several competitive techniques over both power generation and consumption benchmarks, aiming to effectively determine and assess their performance. Overall, the experimental results justify that the proposed network achieved higher performance and outperformed across competitive networks over MSE, MAE, and RMSE. Hence, the proposed framework enhances the development of intelligent energy systems, supporting more effective and efficient power management within smart grid infrastructures.