Transition towards carbon-neutral power systems has necessitated optimization of power dispatch in active distribution networks (ADNs) to facilitate integration of distributed renewable generation. Due to unavailability of network topology and line impedance in many distribution networks, physical model-based methods may not be applicable to their operations. To tackle this challenge, some studies have proposed constraint learning, which replicates physical models by training a neural network to evaluate feasibility of a decision (i.e., whether a decision satisfies all critical constraints or not). To ensure accuracy of this trained neural network, training set should contain sufficient feasible and infeasible samples. However, since ADNs are mostly operated in a normal status, only very few historical samples are infeasible. Thus, the historical dataset is highly imbalanced, which poses a significant obstacle to neural network training. To address this issue, we propose an enhanced constraint learning method. First, it leverages constraint learning to train a neural network as surrogate of ADN's model. Then, it introduces Synthetic Minority Oversampling Technique to generate infeasible samples to mitigate imbalance of historical dataset. By incorporating historical and synthetic samples into the training set, we can significantly improve accuracy of neural network. Furthermore, we establish a trust region to constrain and thereafter enhance reliability of the solution. Simulations confirm the benefits of the proposed method in achieving desirable optimality and feasibility while maintaining low computational complexity.