Hamiltonian DNN-Optimized Adaptive Droop Control for Balancing Battery State of Charge in DC Microgrids by the RBMO

This paper proposes a novel approach called The Hamiltonian Deep Neural Networks -Red-Billed Blue Magpie Optimizer (HDNN-RBMO) in, which balances the State of Charge (SoC) in batteries within DC microgrids using adaptive droop control (ADC). Initially, data are collected from the Battery SoC Dataset and passed through a pre-processing stage. In this phase, Multivariate Fast Iterative Filtering (MFIF) is applied to handle missing values in the input data. The processed data are then fed into an HDNN, which is utilized to predict energy flows and battery SoC, enabling balanced operation within the microgrid via adaptive droop control. Traditionally, HDNNs do not incorporate advanced optimization strategies to fine-tune parameters for accurate prediction. To address this limitation and assess the weight characteristics of the HDNN, the RBMO is introduced, thereby enhancing its capability to forecast energy demands accurately. The proposed HDNN-RBMO model is carried out in Python and computed utilizing performance metrics like F1-score, calculation time, precision, recall, specificity, accuracy, and error rate. This study’s novelty lies in integrating a Hamiltonian-based deep learning framework with a nature-inspired optimizer RBMO for adaptive droop control. Unlike traditional models, this approach allows for dynamic parameter tuning and preserves system dynamics, ensuring robust performance under various load conditions, heterogeneous battery capacities, and real-time SoC balancing challenges. The results are contrasted against existing state-of-the-art methods, including Particle Swarm Optimization (PSO), Wild Horse Optimizer (WHO), and Atom Search Optimization (ASO), showcasing the improved method’s superior presentation. The intended approach introduces a novel combination of rule-guided parameter control, adaptive optimization, and multi-metric evaluation, offering improved performance across both binary and multi-class classification problems. The proposed method combines randomized parameter initialization under hyperparameter constraints with a dynamic parameter adjustment mechanism. Unlike conventional approaches, RBMO incorporates a multi-metric fitness function including accuracy, F1 score, and specificity that allows it to adapt effectively to both binary and multi-class classification tasks. This integration results in improved convergence behavior and robust performance across diverse datasets. Compared to traditional optimization methods such as GA and PSO, RBMO offers faster convergence, better handling of class imbalance, and superior accuracy through its dynamic and multi-metric optimization strategy.