Abstract:

To accurately predict vehicle operation risks, this paper proposed an ISMA-BP model based on the ISMA and BP neural network. First, the slime mould population was initialized using Bernoulli chaotic mapping, and a multi-leader strategy was introduced to improve the algorithm's position update mechanism. Second, a T-distribution mutation strategy was adopted to enhance the algorithm's late-stage local exploitation capability, and an optimized dynamic weight coefficient strategy was proposed to dynamically adjust the search step size. ISMA was then used to optimize the weights and thresholds of BP neural network. Third, ablation experiments were designed, and six benchmark test functions were employed to comparatively analyze the algorithm's improvement process. Finally, actual floating car data from a city were utilized, and grey correlation-multiple regression was applied to predict vehicle operation risks. The results demonstrate that chaotic mapping and the multi-leader strategy can effectively improve ISMA's iteration efficiency and accuracy. Each incremental improvement reduces the error between the mean and optimal values by 5% to 35%, and ISMA's mean-optimal error is 5% to 30% lower than that of four other optimization algorithms. Compared with BP neural network models combined with five meta-heuristic algorithms, the proposed method reduces the mean absolute error (MAE) by 3.88%, 0.40%, 3.48%, 3.89% and 2.53%, the mean square error (MSE) by 0.34%, 0.06%, 0.21%, 0.32% and 0.17%, and the root mean square error (RMSE) by 4.76%, 0.69%, 3.53%, 4.52% and 3.09% respectively, indicating that the proposed method exhibits superior performance in vehicle operation risk prediction.

Key words: improved slime mould algorithm (ISMA), back propagation (BP) neural network, naturalistic driving, vehicle operation, risk prediction