To address the limitations of traditional models in mining multi-dimensional temporal features and handling class imbalance, A multi-task driving risk prediction model based on Class Balance loss (CB)- Asymmetric Loss (ASL)-TimesNet and is proposed. A sliding time window was employed to extract multi-dimensional temporal features, improving the objectivity and granularity of risk level labeling. Based on these features, four types of labels were constructed: forward collision risk, rear collision risk, lateral collision risk, and steering intention. In terms of model design, the TimesNet architecture was incorporated to effectively capture periodic variations and dynamic evolution in multi-dimensional temporal features, thereby enhancing the modeling capability of temporal information in complex driving scenarios. Meanwhile, a hybrid loss function integrating CB and ASL was devised to improve prediction performance under class-imbalanced conditions. Experimental results demonstrate that the proposed CB-ASL-TimesNet model achieves an average accuracy of 0.908 6 in driving risk and steering intention prediction tasks. Compared with the traditional machine learning model CatBoost, the proposed approach yields a 16% improvement, and it outperforms the mainstream time-series model Gate Recurrent Unit(GRU) by 5.8%, verifying the significant effectiveness of the proposed model in enhancing prediction performance.