This paper studies the back-propagation neural network (BPNN), the particle swarm optimization back-propagation neural network (PSO-BPNN), the firefly-optimized back-propagation neural network (FA-BPNN), and the Fibonacci-optimized back-propagation neural network (IM-FSM-BPNN) for the tracking of the maximum power point of photovoltaic modules under local shadows, and the above-mentioned algorithms fly photovoltaic power generation tracking in solar drones. The results show that: First of all, under partial shading, the power prediction accuracy of IM-FSM-BPNN is the lowest, the tracking time is the longest, and the robustness is poor, the reason is the large number of control parameters and the dependence on the initial values of the parameters. The power pre-diction accuracy of FA-BPNN is the highest and the robustness is better, because it can effectively avoid the problem of gradient disappearance during the training process. Secondly, In the increase of sample data and the application of solar drones, it is found that the prediction effect of the FA-BPNN is good and the limitations of the IM-FSM-BPNN . Finally, the influence of parameter changes on the prediction results is discussed. IM-FSM-BPNN, PSO-BPNN and FA-BPNN are more suitable for multi-sample data prediction than BPNN, and IM-FSM-BPNN is more suitable for smaller learning rates than the other three algorithms. The average tracking time and power average prediction accuracy of the four algorithms fluctuate with the number of hidden layer nodes.