Operation state gated topology aware multimodal fusion for simulation based active distribution network risk assessment

Active distribution planning requires day-ahead identification of buses exposed to voltage and loading stress. Separating injection forecasting from feeder simulation leaves the contribution of multimodal fusion to node-risk estimation unresolved. We develop OSTMF-Net, an operation-state-gated topology-aware multimodal fusion network that combines temporal and graph experts through convex, residual and direct heads selected on a chronological validation holdout. Public load and NASA POWER weather variables are mapped to the IEEE 123-node OpenDSS feeder. Quasi-static time-series simulations generate deterministic 24-h absolute-risk and positive-increment targets under base, high-photovoltaic, heavy-load and combined-stress conditions. Evaluation measures agreement with OpenDSS-derived risk surrogates within this reproducible simulation benchmark. Among matched fusion methods across eight correlated scenario-target cases, OSTMF-Net obtains the lowest root mean square error (RMSE) in six and remains within 1% of the lowest value in seven. Mean case-wise RMSE reductions are 9.77%, 17.37% and 3.72% relative to a validation-selected expert, uniform averaging and linear stacking, respectively; gains over stacking concentrate in absolute-risk cases. A matched state-adaptive control yields mean RMSE 0.03460 versus 0.03080 for OSTMF-Net. Forecast perturbations, sensor-input corruption and label sensitivity assess robustness to input and target changes. On a disjoint 2013 chronology on the same feeder, OSTMF-Net records RMSE 0.03067, high-risk F1 0.795 and Top-10 agreement 1.000. The results connect complementary temporal and topology information with state-dependent correction, supporting auditable day-ahead bus-risk screening on the evaluated OpenDSS benchmark.