grid-stability-mlp-full

Grid Stability MLP (Full) — Research Artifact

A scikit-learn Pipeline (StandardScaler + MLPClassifier, hidden layers (64, 32)) trained to classify power-grid stability (stable / unstable) on the UCI Electrical Grid Stability Simulated Data set.

This is the model our paper "Beyond Accuracy: An Edge-Deployability Trade-off Analysis of Machine Learning Models for Power Grid Stability Classification" identifies as the strongest practical candidate among seven models compared on accuracy, inference latency, and on-disk size: it is statistically tied on accuracy with the top classical method (gradient boosting via XGBoost), while being faster and about 6x smaller on disk. See the paper for the full trade-off analysis and the decision rule behind this pick.

Intended use

This is a research and reproducibility artifact, released so others can inspect, retrain, and independently validate it — not a certified or production-ready component.

• Appropriate: further research on edge-deployable grid-stability classification; retraining or fine-tuning on real (non-simulated) substation/microgrid telemetry; benchmarking against other models or real edge hardware.
• Not appropriate as-is: direct deployment in any protection, trip, or safety-critical decision path. The model was trained and evaluated only on simulated data (see Limitations), and — as the paper argues at length — even a fast, accurate classifier of this kind is best used as an early-warning / monitoring signal upstream of, and independent from, the certified, deterministic protection relays that actually govern breaker-trip decisions (IEC 61850-5 / IEEE 1646). It has not been certified for, and should not be wired into, any such trip circuit.

Training data

UCI Electrical Grid Stability Simulated Data (Arzamasov, 2018, CC BY 4.0) — a simulation of a 4-node star-topology Decentral Smart Grid Control (DSGC) scheme, not live SCADA/PMU telemetry. See metadata.json for the exact feature list, hyperparameters, and held-out test metrics.

Usage

import joblib

pipeline = joblib.load("grid_stability_mlp_full.joblib")
prediction = pipeline.predict(X)  # X: DataFrame with the 12 input columns listed in metadata.json

Limitations

• Trained and evaluated on simulated data from a single, small (4-node) grid topology — generalization to other topologies or to real grid telemetry is untested.
• No adversarial or noisy-channel robustness evaluation was performed.
• Reported latency/footprint figures in the paper were measured on a development workstation, not certified edge hardware.

License

Code and model weights: MIT (matching the parent repository). The training dataset is licensed separately under CC BY 4.0 by its original author (Arzamasov, 2018) — see above.

Citation

If you use this model, please cite the paper (full citation forthcoming in the public repository) and the original dataset (Arzamasov, 2018, UCI Machine Learning Repository, doi: 10.24432/C5PG66).