Energy evaluation optimization in photovoltaic-load coupled systems using a transformer-based and Bayesian-optimized probabilistic neural network

Abstract

In response to the escalating complexity of energy Internet architectures, this paper proposes a novel two-stage energy evaluation framework integrating a Transformer-based Bayesian-optimized Probabilistic Neural Network (TBP model) with Hadoop-powered distributed processing to address critical challenges in microgrid interaction and real-time energy optimization under dynamic storage-distribution conditions. The framework combines transformer’s parallel temporal encoding, Bayesian hyperparameter optimization, and probabilistic neural network’s (PNN) probabilistic outputs to achieve superior prediction accuracy and robustness, demonstrating three key innovations: (1) a cloud-Internet of things (IoT) integrated distributed cluster system enabling scalable renewable energy coordination, (2) a density-curve–filtered photovoltaic anomaly detection mechanism with intelligent rule-based control, and (3) a robust hybrid objective function accommodating both real-time and non–real-time monitoring constraints. Experimental results demonstrate that the proposed method significantly reduces photovoltaic curtailment and frequency deviations compared to baseline models, thereby improving energy utilization efficiency and operational stability. This work contributes a scalable, intelligent solution for energy-interactive microgrids and offers practical applicability to future smart grid systems with high penetration of renewables.