To address key challenges in large‑scale residential distributed photovoltaic (DPV) integration into distribution networks for peak shaving, where traditional distributed model predictive control (DMPC) suffers from convergence difficulties, high computational complexity, and low communication coordination efficiency under high‑dimensional coupled constraints, this paper proposes an adaptive projection‑based DMPC (AP‑DMPC) method. First, a two‑stage solution mechanism of "unconstrained optimization + projection" is constructed, decomposing the original constrained distributed optimization problem into two independent steps: unconstrained collaborative optimization and feasible‑region geometric projection. This effectively decouples the strongly coupled constraints among subsystems, significantly improving per‑iteration computational efficiency and enhancing convergence robustness in ultra‑large‑scale scenarios. Then, an adaptive scaling factor based on the unconstrained solution magnitude is introduced into the iterative process, dynamically adjusting the step size according to the real‑time solving status of each subsystem, thereby significantly accelerating convergence toward the global optimum. Finally, combined with a streaming computing architecture, a real‑time collaborative optimization framework tailored to the radial topology of distribution networks is designed. Through theoretical analysis, the optimal backbone node configuration is determined to minimize aggregated communication delay and support high‑concurrency, low‑latency online dispatching of massive DPV resources, effectively overcoming communication and computation bottlenecks in large‑scale distributed resource coordination. Experimental results demonstrate that, under real‑time peak‑shaving scenarios requiring millisecond‑level response, the proposed method can stably maintain the peak‑shaving error within 0.1% while fully satisfying real‑time operational requirements.