Real-time coordination of distributed energy resources (DERs) is crucial for regulating the voltage profile in distribution grids. By capitalizing on a scalable neural network (NN) architecture, machine learning tools can attain decentralized DER decisions by minimizing the average loss of prediction. This paper aims to improve these learning-enabled approaches by accounting for the potential risks associated with reactive power prediction and voltage deviation. Specifically, we advocate to measure such risks using the conditional value-at-risk (CVaR) loss based on the worst-case samples only, which could lead to the learning efficiency issue. To tackle this issue, we propose to accelerate the training process under the CVaR loss objective by selecting the mini-batches that are more likely to contain the worst-case samples of interest. Numerical tests using real-world data on the IEEE 123-bus test case have demonstrated the computation and safety improvements of the proposed risk-aware learning algorithm for decentralized DER decision making in distribution systems.