In the training process of neural network, the convergence speed of the algorithm is slow and easy to fall into the local extreme due to the random initialization of the network parameters. Therefore, an improved fruit fly optimization algorithm (IFOA) is proposed to optimize the initial connection weights and thresholds of neural network for global optimization. Firstly, the BPNN-DIOC model, that is add the connections from the input to the output based on the BP neural network is used to decrease the number of neurons required by the hidden layer, reduce the number of parameters adjusted in the training process of network and speed up the network training, to improve the prediction accuracy of power load and generalization ability of network. Then, combining IFOA and BPNN-DIOC, a load forecasting model based on IFOA optimized BPNN-DIOC is constructed. Finally, in order to verify the validity of this model, a simulation test was conducted using the data of AEMO in New South Wales on September 2015 as example, the average absolute error percentage of IFOA-BPNN-DIOC model is 0.6357% and the root mean square error is 0.0118, and the result is compared with the prediction results of other models in this paper. The results show that this model is a more effective method for short-term load forecasting.