Datasetpublic

babycow/TD-HER-Dataset

TD-HER Dataset Pre-processed representations for reproducing all results in the paper "Target-Dependent Heterogeneous Expert Routing for Power System Frequency Extremum and Arrival-Time Prediction". Quick Start All experiment scripts default to data/ as the dataset root. No additional configuration is needed if the dataset is placed here. Dataset Structure data/ ├── ieee39/ # IEEE 39-bus New England system (10 generators) │ ├──… See the full description on the dataset page: https://huggingface.co/datasets/babycow/TD-HER-Dataset.

sourceHugging Faceupdated 5mo agoView on Hugging Face

1likes2downloads

Dataset Card

TD-HER Dataset

Pre-processed representations for reproducing all results in the paper "Target-Dependent Heterogeneous Expert Routing for Power System Frequency Extremum and Arrival-Time Prediction".

Quick Start

All experiment scripts default to data/ as the dataset root. No additional configuration is needed if the dataset is placed here.

Dataset Structure

data/
├── ieee39/                     # IEEE 39-bus New England system (10 generators)
│   ├── repA/                   # Representation A (tabular features)
│   │   ├── ms1/                #   ~10 ms early window (1 post-trigger step)
│   │   ├── ms5/                #   ~50 ms early window
│   │   ├── ms10/               #   ~100 ms early window (main setting)
│   │   ├── ms15/               #   ~150 ms early window
│   │   └── ms25/               #   ~250 ms early window
│   ├── repB/                   # Representation B (spatiotemporal tensor + static)
│   │   └── ms{1,5,10,15,25}/
│   ├── repC/                   # Representation C (graph: tensor + adjacency)
│   │   └── ms{1,5,10,15,25}/
│   ├── csv/                    # Raw ms10 CSV files (for robustness stress test)
│   └── adjacency/              # 10×10 Kron-reduced electrical-distance adjacency
│
└── ieee300/                    # IEEE 300-bus system (69 generators)
    ├── repA/ms10/              # Tabular features (same-distribution only)
    ├── repB/ms10/              # Spatiotemporal tensor + static
    ├── repC/ms10/              # Graph representation
    └── adjacency/              # 69×69 electrical-distance adjacency

Split Convention

Each ms{N}/ directory contains NumPy arrays following the naming convention:

File pattern	Description
`X_train.npy`, `y_train.npy`	Training set (L1 same-distribution)
`X_val.npy`, `y_val.npy`	Validation/calibration set (L1)
`X_test.npy`, `y_test.npy`	Held-out test set (L1)
`X_cross_cond_finetune.npy`	L2 cross-condition calibration set
`X_cross_cond_test.npy`	L2 cross-condition test set
`X_cross_cond_topo_finetune.npy`	L3 cross-topology calibration set
`X_cross_cond_topo_test.npy`	L3 cross-topology test set
`feature_names.json`	Ordered feature names for RepA
`meta.json`	Tensor dimensions and channel info for RepB

L2/L3 splits are only available at ms10 (the main early-window setting).

Prediction Targets

—y1 (y[:,0]): Signed COI frequency extremum (Hz)
—y2 (y[:,1]): Nonnegative time from disturbance trigger to frequency extremum (s)

Sample Counts

System	Split	Samples
IEEE 39-bus	Train	101,390
IEEE 39-bus	Validation/Calibration	12,673
IEEE 39-bus	Test	12,676
IEEE 39-bus	L2 cross-condition calibration	510
IEEE 39-bus	L2 cross-condition test	511
IEEE 39-bus	L3 cross-topology calibration	2,337
IEEE 39-bus	L3 cross-topology test	2,337
IEEE 300-bus	Train	7,391
IEEE 300-bus	Validation	923
IEEE 300-bus	Test	926

Channels

Channel	Source	Description
`FREQ`	PMU-observable	Bus frequency
`VOLT`	PMU-observable	Bus voltage magnitude
`ANGL`	PMU-observable	Bus voltage angle
`POWR`	PMU-computable	Generator-terminal active power
`SPD`	Generator-side state	Generator speed deviation (PMU-proxy compatible)

Paper Section Mapping

Data	Paper section	Experiment
ieee39/rep{A,B,C}/ms10	§IV.B–E	Main expert comparison, TD-HER routing, ablation, weights
ieee39/rep{A,B,C}/ms{1,5,15,25}	§IV.G, Fig. 10	Multi-window latency-accuracy study
ieee39/csv/*_ms10.csv	Supplementary §III	Robustness stress test (sensor noise, gaps)
ieee39/adjacency/	§IV (ST-GCN input)	Graph expert adjacency matrix
ieee300/rep{A,B,C}/ms10	§IV.F	300-bus scalability experiment
ieee300/adjacency/	§IV.F (ST-GCN input)	300-bus graph adjacency matrix

Data Generation

The raw time-domain simulation data were generated in PSS/E under randomized disturbances and operating conditions (see Section IV-A of the paper). The representations were constructed by the pipeline in data_proc/:

1.data_proc/extract_features.py — xlsx/CSV extraction from PSS/E output
2.data_proc/build_representations.py — RepA/RepB/RepC construction
3.data_proc/build_adjacency.py — Kron-reduced adjacency matrices

IEEE Test System RAW Topology Files

The graph-expert adjacency builder (data_proc/build_adjacency.py) consumes PSS/E .RAW topology files. These are standard public IEEE test cases and are not redistributed here.

File	Public source
`IEEE39.RAW`	IEEE PES Test Feeder Working Group / ICSEG
`IEEE300Bus_modified_noHVDC_v2.raw`	RLGC repository (`testData/IEEE300/`)

Place them in data/topology/ if you need to regenerate adjacency matrices from scratch (not required if using the pre-computed .npy files above).

Size

Component	Size
ieee39 (all windows)	15.5 GB
ieee300 (ms10 only)	2.5 GB
Total	18.0 GB

205 files total (184 .npy, 12 .json, 7 .csv, 2 other).

License

Released under the same license as the TD-HER repository.