To address the problem of data dispersion and complex dynamic changes in safety risk prediction for small airlines, a hybrid model of temporal convolutional network bidirectional long short term memory (TCN-BiLSTM) based on the modified black kite algorithm (MBKA)was proposed. Firstly, a dynamic weighting civil aviation safety risk indicator system was constructed, which solved the problems of redundant indicators and high computational complexity in traditional methods through multi-dimensional quantitative mapping of event levels, types, causes, and stages. Secondly, the proposed timing adaptive attack strategy, linear lock optimization strategy, and variable scaling factor differential evolution mechanism significantly enhance the global search capability and convergence efficiency of MBKA, enabling it to efficiently optimize the hyperparameters of TCN-BiLSTM. The experimental results show that compared with traditional TCN-BiLSTM, PSO-TCN-BiLSTM model based on particle swarm optimization (PSO) and BKA-TCN- BiLSTM models, the MBKA-TCN-BiLSTM model reduces mean absolute error (MAE) by 18.44%, 5.81%, and 3.95%, and root mean square error (RMSE) by 17.21%, 7.77%, and 8.72%, respectively. The coefficient of determination (R²) increases by 4.59%, 1.72%, and 1.90%, respectively. The results not only provides lightweight and high-precision risk prediction tools for small airlines, but also provides a universal framework for modeling and optimizing complex time-series data, which can be extended to dynamic risk prediction in fields such as transportation and energy.