ST-SafeMAPPO: code, processed NREL ARPA-E PERFORM / OPSD test systems and simulation configurations for minute-level power system production simulation with deep reinforcement learning and safety constraints

Abstract

Reference implementation, processed test systems and simulation configurations accompanying the manuscript Minute-Level Power System Production Simulation Algorithm Based on Deep Reinforcement Learning and Safety Constraints (Zheng Li, Juan Su, Yuhao Fu, Yue Xia). The archive contains: (1) the Python source code of ST-SafeMAPPO – the physics-informed spatio-temporal graph attention network (PI-STGAT) with the DC-PTDF electrical coupling distance, the L1 exact-penalty differentiable safety layer (DSL; OSQP forward pass, KKT implicit-differentiation backward pass, warm start, dual-variable feedback), and the CTDE multi-agent PPO training framework with the auxiliary projection loss; (2) the five baselines used in the paper (MLP-MAPPO-DSL, IPPO-DSL, MAPPO-Penalty, MACPO and the rolling Relaxed-MPC-QP benchmark); (3) the processed NREL ARPA-E PERFORM (NYISO, 11 load zones, 5-min actuals, 2018-2019) and Open Power System Data (Germany, 15-min, 2019) profiles together with the test systems built on the IEEE 118-bus and IEEE 30-bus networks (device tables, PTDF, electrical distance, control regions, typical days A/B/C, 30-day long-run window); (4) all YAML configuration files (hyper-parameters of Table 1, system parameters, scenario definitions) and the scripts for downloading and preprocessing the original data, training, and the evaluation protocols of Figs. 4-13 / Table 2; (5) unit tests. Original data: NREL ARPA-E PERFORM datasets (https://data.openei.org/submissions/5772, CC BY 3.0 US) and Open Power System Data, Time series package version 2020-10-06 (https://doi.org/10.25832/time_series/2020-10-06). Code: MIT licence. Processed tables: CC BY 4.0.