The invention discloses a probabilistic power flow depth learning method based on a physical model. The method comprises steps that 1), power system data is acquired; 2), a loss function for the probabilistic power flow analysis depth neural network is established, and a coding parameter theta of the depth neural network is updated; 3), the depth neural network is utilized to perform depth learning of the probabilistic power flow of a power system; and 4), a probabilistic power flow depth learning calculation model is established. Through combining the data-driven technology with a physical mechanism of a power domain, and problems of huge calculation cost and calculation accuracy balancing difficulty in solving the probabilistic power flow are solved.