The safety and economy of power grid operation are affected by load forecasting accuracy in the intelligent distribution network environment. It reduces the convergence speed and prediction accuracy of the algorithm, which randomly access the input neurons, neurons in hidden layer and output neurons between the weights and thresholds in BP algorithm. In order to obtain the optimal model of the network, this paper uses AFSA algorithm for the initial weights and threshold of BP algorithm for global optimization. The AFSA-BP short-term load forecasting model is established, based on the analysis of the power system load characteristics. In order to verify the accuracy of the algorithm, BP, LS-SVM, AFSA-BP algorithm is used to power load simulation, respectively. The RMSE value caculated by AFSA-BP, BP and LS-SVM algorithm are 0.0862, 0.2558 and 0.1522 respectively, which verifies that the AFSA-BP algorithm is suitable for short-term power load forecasting.