<p>Accurate prediction of PM_2.5 concentrations is of great significance for societal productivity, public safety, and human health. However, the dynamic and variable temporal distribution of PM_2.5 concentration will decrease the generalization ability of prediction models in future prediction. Beyond this challenge, the noisy data is usually intertwined with the training data, which is likely to mislead the model to extract robust features from observed data. To address these two problems, we propose an adaptive hybrid prediction model based on Temporal Convolutional Network (TCN). This model aims to extract robust and fine-grained features from dynamic temporal distribution while distinguishing between normal and noisy data. Specifically, we first preprocess the time-series data by split it into long-term and short-term features using two sliding time windows. During the training stage, we employ TCN with a Maximum Mean Discrepancy (MMD) to characterize their temporal distribution dependencies and correlations, while an additional MMD penalty term is incorporated to enhance model capability of distinguishing between normal and noisy data. Finally, we construct a predictor to output prediction results based on the short-term features. Extensive experiments indicate that our proposed model outperforms other baseline models across different datasets.</p>