As one of the main renewable energy sources in electric power systems， the photovoltaic （PV） power generation takes on a transition development trend from centralized-based to combination （centralized and distributed） grid connection. The demand for low-voltage distributed photovoltaic to participate in grid peak shaving is becoming increasingly urgent. To solve the difficulty of precise real-time awareness for the low-voltage distributed-PV active power in automatic generation control （AGC）， a data-driven real-time modeling scheme is proposed by using the directly-collecting data from the benchmark PV station integrated with the meteorological data. The AGC real-time modeling of distributed-PV clusters is built with the uses of density-based spatial clustering algorithm with noise （DBSCAN） and eXtreme gradient boosting（XGBoost）machine learning. Based on the feasible region projection theory， the adjustable power boundaries of distributed-PV clustering aggregation are described by adopting Minkowski sum approach and vertex search optimization method. The distributed-PV precision modeling method proposed in this paper enhances the observation frequency at the dispatch and control master station without additional collection devices， achieves minute-level active power of distributed-PV， and provides a feasible region of distributed-PV that meets power grid operation constraints. Furthermore， a multi-dimensional weighting PV peak shaving strategy is also studied by taking the security （indexed by regulation rate） and the fairness （indexed by abandoned power rate， load factor and installed capacity factor） into consideration. The effectiveness and usefulness of the present models are verified properly form a practical case combining with the actual power grid dispatch and control master station at a certain provincial level. On the premise of power grid security， the model balances the maximum utilization of renewable energy and regulation fairness.