WTConv-TimesNet for Road Icing State Classification with IWOA-Based Hyperparameter Optimization

Road icing is a complex and highly dynamic phenomenon that poses a serious risk to winter road safety. However, nonlinear evolution, multiscale temporal dependence, and rapid transitions between adjacent stages still make it difficult to accurately identify icing states from multivariate environmental time-series data. In this work, a road icing state classification model based on TimesNet is developed. We integrate wavelet transform convolution (WTConv) into the original TimesNet architecture to strengthen multiscale time-frequency feature extraction, enabling more effective capture of high-frequency dynamics and abrupt local variations. To address the hyperparameter sensitivity commonly observed in icing scenarios, an Improved Whale Optimization Algorithm (IWOA) is employed and uses Pearson correlation analysis to select informative and physically meaningful features from multi-source monitoring data. Experiments on a real-road dataset show that the proposed IWOA-TimesNet-WTConv model improves overall accuracy from 92.72% to 98.83% compared with the baseline TimesNet model. In addition, feature selection yields a further 1.04 percentage-point gain in overall accuracy and increases the Macro F1-score from 0.9691 to 0.9809, indicating reduced redundancy and more stable discrimination under transitional icing conditions. Overall, the proposed method provides a practical and effective data-driven solution for intelligent road icing monitoring and early warning in complex winter road environments.