In the background of adversarial engagement between fixed-wing unmanned aerial vehicles of both enemies and friends, reliable prediction of the future posture of the enemy UAV is of great significance. However, the overconfidence problem of most existing point estimation predictive models makes their predictive results lack credibility. To solve this problem, this paper proposes an R-Bayesian neural network (R-BNN) predictive model based on gradient balance parameters, which finely adjusts the gradient balance between the error cost loss function and complexity cost loss function in the BNN loss function by adjusting the numerical size of the parameter, so as to effectively balance the trade-off between model complexity and prediction accuracy, enabling the network to achieve better performance. This model places probability distributions on network weights to capture the uncertainty of predictive results, and further improves the reliability of predictions by integrating the various predictive results of the model through Monte Carlo sampling during the network prediction phase. In the verification stage, the paper compares long short-term memory neural network (LSTM), MultiLayer Perceptron (MLP) and R-BNN algorithms under the same conditions. The experimental results show that the R-BNN predictive model performs better in diving and climbing maneuvers with larger motions and has significantly higher prediction accuracy with limited data than the comparative algorithms, and can provide reliable support for military operations.