This dataset addresses critical-load survival in distribution systems under sequential extreme events. Based on a 7-bus test system and the IEEE 123-bus distribution system, it provides multidimensional time-series scenarios including load demand, distributed generation outputs from photovoltaic and wind power, and line damage states. Combined with remote control switch actions, power-flow calculation, microgrid operational constraints, and reward evaluation, the dataset supports training and evaluation of topology reconfiguration policies using adversarial scenario generation network enhanced distributionally robust deep reinforcement learning. It can be used for distribution-system resilience defense, critical-load survival, extreme-event scenario generation, reinforcement-learning training, and comparative algorithm studies.
This dataset is completely synthetically generated through simulation and does not originate from specific literature, field measurements, or third-party downloads. System files include Case7_Data.xlsx and IEEE123_Data.xlsx with bus_data, line_data, load_data, sgen_data, slack_data, switch_data, and geo_data sheets. Scenario files include Case7_Scenario.xlsx and IEEE123_Scenario.xlsx with load, pv, wp, and disruption time-series sheets. The code loads load profiles, photovoltaic outputs, wind-power outputs, and line-damage states by scenario ID and time step, and simulates them in a Pandapower-based power-network environment.
The processing workflow builds a reinforcement-learning simulation environment using Python, PyTorch, Gymnasium, and Pandapower. After reading the system and scenario Excel files, each scenario is organized as a multidimensional time series containing load, photovoltaic output, wind-power output, and vulnerable-line status. Remote control switch actions update the distribution-system topology, power flow is calculated, and rewards are evaluated using critical-load connectivity to distributed generation, illegal switch operations, voltage violations, line losses, and power imbalance. DQN or PPO is used to train topology reconfiguration policies, while a TTN-enhanced AS-GAN generates adversarial scenarios that are iteratively added to the training set to improve robustness. Julia/JuMP/Gurobi code is used for deterministic and two-stage stochastic optimization comparisons.
The dataset has a clear tabular structure separating system data and scenario data. Scenarios are organized with a unified time-step structure, and each scenario covers a 24-hour horizon. Training, in-sample testing, and out-of-sample testing can be split through scenario indices in the code. Simulation quality is evaluated through power-flow convergence, switch-action feasibility, critical-load connectivity, reward values, survival rate, mean values, and standard deviations. The dataset is simulation-generated and does not represent real disaster records, real outage measurements, or real geographic observations; it should not be directly interpreted as the actual operating state of a specific regional grid.
This work is licensed under
CC BY 4.0 (Creative Commons Attribution 4.0 International License).
CSTR:11738.11.ncdc.db7489.2026
10.12072/ncdc.db7489.2026
