Optimization of the Layout of Distributed Photovoltaic Access Points Using Sparse Autoencoders and Graph Embedding

Efficient planning of distributed photovoltaic (PV) access is increasingly important for intelligent power systems and electromagnetic energy infrastructure, yet existing approaches often suffer from insufficient extraction of high-dimensional electrical features and inadequate representation of network topology. To address these challenges, this paper proposes a hybrid optimization framework that combines Sparse Autoencoders (SAEs) with graph embedding for coordinated PV layout design. A three-layer SAE regularized by Kullback–Leibler divergence learns sparse feature representations emphasizing voltage deviation, line loading, and other electrically sensitive variables associated with PV integration. Simultaneously, the distribution network is modeled as a weighted graph in which Node2Vec generates topology-aware node embeddings that preserve electrical proximity. A bi-level optimization architecture is adopted, where the upper layer employs NSGA-II to maximize photovoltaic penetration and the lower layer performs OpenDSS-based power flow analysis with an electrical similarity mechanism for adaptive capacity adjustment under voltage constraints. Experimental results demonstrate that the proposed SAE effectively extracts key electrical characteristics while accurately preserving both global and local topological structures. The resulting layout strategy significantly improves technical performance and economic efficiency by increasing PV utilization and reducing operational costs. Beyond photovoltaic planning, the proposed feature–topology collaborative optimization paradigm provides a practical computational framework for electromagnetic energy distribution networks and intelligent grid infrastructure requiring adaptive spatial configuration and reliable power delivery.