Distributed photovoltaic system has wide range of points, coupled with a severe lack of available data, making it difficult to detect equipment failures in a timely manner, and is prone to long-term operation with faults, reducing the life cycle power generation. An anomaly detection method of distributed photovoltaic system based on measurement data is proposed, taking advantage of the characteristic that abnormal faults in photovoltaic system will ultimately affect power generation output. Firstly, the characteristics of solar irradiance on sunny days are analyzed, and a clear sky day screening mechanism is proposed, conducting correlation analysis on different power stations, and obtaining photovoltaic power stations with high output correlation as horizontal references.Then the output curves of the tested power station on different clear days for longitudinal comparison are selected to eliminate various interference factors in abnormal detection. The output data excluding the above interference is input into the quantile regression temporal convolutional network model for training to obtain the fitting range of photovoltaic normal output, and then abnormal output of distributed photovoltaic power station is detected according to the normal output range. The simulation analysis using actual photovoltaic system data shows that the proposed method can accurately identify distributed photovoltaic systems with faults and anomalies, promoting the refined operation and maintenance of distributed photovoltaics.