Active distribution networks with distributed energy resources, energy storage systems, and remotely operated switches exhibit coupled continuous–discrete behavior: nodal injections vary continuously, whereas feeder reconfiguration changes the network topology. This paper develops a topology-dependent neural surrogate for fast alternating-current power flow evaluation in reconfigurable distribution systems. The proposed model incorporates switch statuses together with load-scaling factors and active and reactive power setpoints of distributed generation and energy storage systems, while graph-theoretic connectivity and radiality tests are used during scenario generation to retain only admissible feeder configurations. A fixed-dimensional target representation is adopted to predict voltage magnitudes at all buses, active power flows on fixed physical branches, and total feeder losses under changing radial topologies. The surrogate is trained using Newton–Raphson solutions generated in Pandapower for 100,000 feasible operating scenarios of a modified 33-bus benchmark active distribution feeder. Test results show close agreement with the Newton–Raphson reference solutions, with voltage, flow, and loss coefficients of determination of 0.99971, 0.99946, and 0.99912, respectively. The mean absolute error of total feeder loss is 0.0079 MW. Batched central processing unit inference requires 0.54 ms per scenario, corresponding to an approximately 44-fold throughput speedup relative to the Newton–Raphson solver. The results indicate that the proposed surrogate is suitable for rapid screening and many-query reconfiguration or optimization studies, provided that final operational decisions are validated using a full alternating-current power flow solution.