Abstract: The increasing integration of distributed energy resources and frequent topology changes in active distribution networks significantly increase the complexity of state estimation. To address this issue, a heterogeneous data fusion-based representation learning method is proposed. A conditional variational autoencoder is employed to map node power injections and network topology into a low-dimensional latent space, effectively capturing the intrinsic dependencies between continuous operating states and discrete topology. Trained on historical operational data, dedicated prediction networks are developed for voltage, phase angle, active and reactive power, and network topology. A compact discrete encoding of switch combinations based on fundamental-cycle decomposition is further integrated with topology-constraint regularization and an adaptive loss function to reduce the modeling complexity of heterogeneous variables and improve training efficiency. In the IEEE 33-bus and 136-bus test systems, the probabilities of radial topologies existing under a Hamming distance ≤ 1 reach 96.8% and 98.2%, respectively, enabling efficient compression of the discrete topology space. Leveraging CVAE to fuse heterogeneous information including continuous power variables and discrete topologies, the prediction of 6000 samples only takes 0.378 s. The results demonstrate that the proposed method can achieve fast and high-precision real-time estimation of multiple state variables under dynamic topologies with favorable scalability.