As the national consumption level rises, the significance of the tourism sector within the national economy becomes more pronounced. Consequently, the share of tourism earnings in the Gross Domestic Product (GDP) continues to escalate. On the basis of the growth of people's economic level, the number of tourists in each scenic spot also increases, and tourist attractions and tourist cities also respond to the need to improve their own management quality and content to seek further development. The occurrence of tourism behavior also changes with the changes of weather, holidays, seasons and special circumstances. How to accurately predict the number of tourists is of great significance for both industry regulators and operators. Many scholars have also tried to use different models to predict tourism demand. AI methods have great advantages over traditional methods in adaptive learning ability and non-linear fitting ability, and have become a key research direction in academic circles in recent years. In this paper, we first improve the standard Drosophila algorithm by adaptively adjusting the fly population number and search step size, while optimizing the initial iteration position and improving the local search ability and search efficiency. Then the improved FOA algorithm is combined with the echo state network to build a two-stage combined prediction model named AFOA-ESN, obtain its key parameters through AFOA optimization ESN, and input the optimized parameters into ESN to form the final combined prediction model. Finally, the monthly data of a local sports ecological passengers was selected to test the prediction effect of AFOA-ESN. The results obtained after the trial show that the use of the AFOA-ESN model is more accurate than the accuracy of the results used by the autoregressive moving mean model, support vector machine model, BP neural network, standard ESN network, and other prediction models, while the convergence speed and prediction accuracy of AFOA-ESN outperformed standard ESN and FOA-ESN, demonstrating the effectiveness of model improvement.