<p>Beyond Fifth Generation (B5G) and Sixth Generation (6G) networks require ultra-reliable, low-latency performance and extreme scalability for emerging applications like extended reality and Internet of Everything (IoE). With augmented traffic intensity and mobility of users, pressure to provide resources adaptively at distributed edge nodes grows exponentially. Edge resource scaling in B5G/6G settings remains a daunting task due to fluid workloads, probabilistic traffic patterns, and haphazard edge resources. Conventional heuristics and traditional static models are incapable of meeting real-time and context-aware expectations for optimal edge behavior. This has created an urgent need for intelligent, adaptive, and scalable resource prediction mechanisms. The aim of this research, Edge Resource Scaling in B5G/6G Networks using Diffractive Deep Sparse Graph Attention Neural Network with Orchard Algorithm (2D-SGrAN-2N + OrA), is to develop a deep learning-based workflow starting with system-wide traffic data collection from Base Stations (BSs). Such data are preprocessed using Fuzzy Min–Max Rough Sets (Fuz-MMRS) to address uncertainty and missing data. High-level temporal and spatial features are extracted using the Path Enhanced Transformer (PET), followed by edge resource prediction using the Diffractive Deep Sparse Graph Attention Neural Network (2D-SGrAN-2N), a combination of Diffractive Deep Neural Networks (2D-2N) for multi-layer traffic decomposition and topology-aware Sparse Graph Attention Networks (SGrAN) topology-aware prediction. The Orchard Algorithm (OrA) optimizes model parameters for real-time deployment. The model attains 99.27% prediction accuracy, 99.61% SLA compliance, 99.83% latency reduction, and 99.12% resource utilization efficiency. This research establishes that 2D-SGrAN-2N + OrA provides a robust and scalable solution for real-time edge resource scaling in B5G/6G infrastructures.</p>