Integrating quantum key distribution (QKD) with existing optical networks is highly desired to reduce the deployment costs and achieve efficient resource utilization, and some point-to-point transmitting experiments have verified its feasibility. Nevertheless, there are still many problems in the realistic scenario where QKD coexists with dynamic data traffics. On the one hand, the conventional static channel allocation schemes cannot guarantee the quality of quantum channels in the presence of the time-varying noises. On the other hand, considering the complex noise generation caused by dynamic classical data traffics with variable characters, it is challenging to achieve online high-performance quantum channel assignments. To address these problems, we propose a machine learning based noise-suppressing channel allocation (ML-NSCA) scheme. In this scheme, the LightGBM based ML framework is trained to predict the optimal channel allocations with lowest noise impacts, according to which, the quantum channels are periodically reallocated to guarantee high secure key rate. To improve the accuracy and scalability of the ML framework, we also optimize the method of feature extraction during the training process. The performance evaluation results indicate that the proposed scheme can effectively resist the dynamic noise impacts in the realistic optical networks and obtain higher secure key rate with less operation complexity than the previous schemes.