Disclosed in the invention is a three-stage-diagnosis-based large-scale photovoltaic array fault diagnosis and positioning method. The method is characterized in that faults of the photovoltaic array are classified into four types including an assembly short-circuit type, an assembly open-circuit type, a shadow type, and a multi-fault type under the circumstances that the photovoltaic array is in a formation mode of series connection and parallel connection. During a first-stage fault diagnosis execution process, whether a fault occurs at a branch circuit where a photovoltaic grid-connected inverter as well as the type of the fault is determined by using a power incremental ratio method; during a second-stage fault diagnosis execution process, whether faults occur at all photovoltaic assembly strings of the branch circuit of the photovoltaic grid-connected inverter is determined by using a current and voltage similarity method; and during a third-stage fault diagnosis execution process, positioning of an internal fault assembly of the fault photovoltaic assembly string is carried out by using a trained three-layer BP fault diagnosis neural network. According to the invention, on-line real-time detection of the operation state of the large-scale photovoltaic array is carried out based on the three-stage diagnoses, thereby realizing accurate and rapid fault type identification and fault positioning.