Abstract: The uncertainty of high proportion of distributed generation brings great challenges to the stable operation of the islanded distribution networks. In this paper, for the short-term prediction of source and load based on the traditional distribution model has the shortcomings of sharp peaks and heavy tails, the distribution model of source and load prediction error under multi-scenarios and different time scales is constructed by combining the bidirectional long and short-term memory (BiLSTM) neural network with the nonparametric kernel density method (KDE). On this basis, the system operates and regulates in multi-temporal operation and control process, taking into account the short-term meteorological fluctuations in the uncertainty. Then, the distflow model is relaxed by using the mixed integer second-order cone programming and the probabilistic Distflow of the system is obtained from the stochastic response surface (SRSM). Based on the stochastic response surface to improve the Sobol' method, the global sensitivity analysis model of the isolated island distribution network operation risk taking into account the source and load uncertainty is established. For this reason, a risk real-time risk assessment and regulation strategy based on BiLSTM-SRSM method is proposed. Finally, the feasibility of the proposed method is verified using a radial distribution system with the IEEE33 nodes.