An embodiment of the invention of a device and a method for photovoltaic short-term load prediction of a power distribution network and aims to solve technical problems of high calculation quantity and error accumulation effects in single prediction of a large quantity of load nodes of an existing power distribution network and low prediction precision in superposition prediction of all loads in a whole area due to neglect of characteristic differences of different loads. The method for photovoltaic short-term load prediction of the power distribution network includes: S1, preprocessing missing data and bad data of net load power data formed by distributed photovoltaic and grid load; S2, dividing the net load power data meeting preset conditions into five cluster loads according to a fuzzy C clustering method; S3, respectively subjecting each cluster load to load prediction according to a similar day and support vector machine based short-term load prediction method, and subjecting load prediction results of each cluster load to superposition.