Regret-Calibrated Margin Ensemble Learning for Voltage-Stability Constrained Rolling-Horizon Dispatch in Multi-Infeed Receiving-End Grids
Description
This dataset contains the research materials associated with the article “Regret-Calibrated Margin Ensemble Learning for Voltage-Stability Constrained Rolling-Horizon Dispatch in Multi-Infeed Receiving-End Grids.” It includes the case-study code, input and processed operating data, configuration files, continuation-power-flow-based static voltage stability margin labels, trained model artifacts, and selected numerical results. The case study is based on a modified IEEE 39-bus receiving-end system with three DC infeeds and evaluates 48 candidate dispatch actions through a 24-hour rolling-horizon simulation. The package also provides scripts for margin-label generation, operating-cost calculation, margin-requirement calibration, surrogate-model training, dispatch replay, and performance evaluation, together with the outputs used for benchmark comparisons, calibration analysis, and collapse-risk assessment. These materials support the methodology and case studies reported in the paper and provide the underlying data for its principal figures, tables, and quantitative findings.
Files
Steps to reproduce
The research data were generated through a fully computational workflow; no laboratory instruments or chemical reagents were used. A MATPOWER IEEE 39-bus test system was adapted to represent a receiving-end grid with three DC infeeds, renewable generation, and reactive-power support devices. Hourly load, wind, and photovoltaic profiles were derived from the RTS-GMLC dataset, while thermal-generation parameters and operating costs were based on PGLib-UC data. For each hourly operating condition, 48 candidate dispatch actions were constructed and screened using AC power flow. Continuation power flow was then performed along predefined stress directions to calculate the static voltage-stability margin of each feasible candidate. The resulting context–action records were divided chronologically into training data from 1–21 January 2020, validation data from 22–26 January, and a held-out test day on 27 January. Heterogeneous regression models were trained in Python using NumPy, pandas, SciPy, scikit-learn, and LightGBM, and their predictions were combined and calibrated on the validation set according to dispatch regret. The trained models were finally evaluated through a 24-hour rolling-horizon dispatch simulation with a six-hour look-ahead horizon, in which candidate actions were ranked using predicted margins and subsequently assessed against the stored CPF results. The supplied scripts, YAML configuration files, processed data, trained artifacts, and output tables document this workflow and allow other researchers to reconstruct the data-generation, model-training, and evaluation procedures.
Institutions
- Shanghai Jiao Tong UniversityShanghai, Shanghai