Integration of Machine Learning and Metaheuristic Optimization for Smart Distribution System Planning

Abstract

The growing penetration of photovoltaic, wind, and storage systems has intensified the need for optimization frameworks that are both computationally efficient and uncertainty-resilient. This study proposes a Machine Learning–Enhanced Sea Lion Optimization framework for optimal siting and sizing of distributed energy resources in radial distribution networks. The framework uniquely integrates deterministic optimization, surrogate-assisted learning, and probabilistic scenario analysis. Voltage profiles are converted into spectrograms and scalograms, from which Local Binary Pattern, Gabor, and wavelet-energy descriptors are extracted to represent time–frequency characteristics. A Gradient Boosting surrogate model trained on these features accurately predicts active power loss, voltage deviation, and stability indices, significantly reducing the computational burden of repetitive load-flow evaluations. The modified Sea Lion Optimization algorithm, enhanced with chaotic search and adaptive weighting, drives the optimization process, while Monte Carlo sampling incorporates load and renewable uncertainties. Applied to the IEEE 33-bus system, the proposed optimization algorithm achieves 96.4 % reduction in power loss, improves minimum voltage to 0.996 p.u., and yields annual cost savings of $80,750 with superior convergence speed. Statistical validation using Friedman, Nemenyi, ANOVA tests confirm its robustness and scalability,
establishing proposed method as an effective hybrid framework for smart distribution network planning under uncertainty.