Anomalous measurements bring huge challenges for existing learning-based distribution system optimization methods. Although several graphical learning-based methods have been developed to deal with abnormal measurements, their robustness is achieved by embedding the accurate topology structure of the distribution system, which is challenging to obtain in practice. To this end, this paper proposes a graphical learning-enabled deep reinforcement learning (DRL)-based voltage control method that is robust to anomalous measurements without reliance on accurate topology information. Instead of directly utilizing the inaccurate topology that is typically available in practice, an aggregated k-nearest neighbor-based extended graph is first developed to capture the potential node connection relationship according to available system measurements and the inaccurate given topology. Then, an adaptive multi-channel graphic learning (AMGL) model is designed to perform graph aggregation across both the extended feature space and the given topology space, mitigating the noise introduced by additional edges in the extended graph and contributing to multi-view representations for system observations. These representations are subsequently fed into the soft actor-critic algorithm, which informs real-time decisions to reduce the voltage deviation of the whole network. To mitigate the reliance on precise line parameters, a surrogate model is further developed on the basis of the AMGL model to yield reliable reward signals in the offline training process of the DRL agent. The extended graph-enabled AMGL allows the proposed method to maintain robust control performance against abnormal measurements using inaccurate topology information of the network. Comparative results with state-of-the-art graphical learning-based methods demonstrate the robustness of the proposed method in tackling both potential anomalous measurements and varying degrees of topology inaccuracies.