Accurate Power Load Prediction Via a Hybrid Network Based on Automatic Feature Association

Abstract

Amidst the rising need for sustainable energy solutions, precise forecasting of power load has emerged as a pivotal aspect of efficient energy management. However, achieving accurate load forecasts remains challenging, requiring a delicate balance between precision and computational efficiency. This study presents a novel approach to general power load forecasting, leveraging a hybrid network based on automatic feature association to enhance the ability of power load generation and predictions. The proposed network investigates the effects of various neural network interactions on real-world power load data within the network layer, continually adjusting parameters to assess their impact. Convolution based operation as well as an integration of attention and bidirectional long short-term memory (BiLSTM) network are employed to learn comprehensive features from power load data, serving as inputs for the network layers. By integrating the two learning-based structure, the hybrid network is trained hierarchically, utilizing multi-model fusion, interaction, and infiltration strategies. Experimental findings demonstrate that the proposed approach yields competitive performance compared to traditional single-model methodologies. This highlights the effectiveness of utilizing interconnected neural networks to improve the precision of load prediction, thereby promoting more efficient energy management strategies.