Enhanced small-sample aviation accident prediction via an improved WCGAN incorporating neuralprophet and gradient penalty

Aviation accidents are characterized by their low frequency and high destructive power, making accurate prediction crucial for enhancing safety measures. Existing aviation safety prediction methods often struggle to extract meaningful trend information from limited accident data, which impedes effective risk management and decision-making. In response to these challenges, this paper proposes a novel small-sample aviation accident prediction method based on an improved Wasserstein distance conditional generative adversarial network (WCGAN). The generator integrates NeuralProphet for its hybrid architecture that combines Prophet's interpretability with neural networks' adaptability, effectively modeling sparse aviation accident data while mitigating overfitting common in time series prediction methods. Additionally, to address the mode collapse issue prevalent in traditional generative adversarial networks, the discriminator incorporates a gradient penalty (GP) mechanism, enhancing robustness and training stability. Using the publicly available Aviation Safety Reporting System (ASRS) dataset, the proposed method was rigorously evaluated against various existing aviation safety prediction techniques. The experimental results demonstrate remarkable improvement in accuracy, with the proposed method outperforming all baseline models by over 16%. This significant enhancement underscores the model’s capability to provide more accurate and actionable insights for safety decision-making.