Twenty-three notebook-style guides for the ML techniques that matter in utility operations. Each guide is a working model with real inputs, real outputs, and a plain-language methodology. All run against the SP&L synthetic distribution dataset, so you can clone, train, and validate without touching CEII data or signing an NDA.
Eight starting guides. Each builds a working model end-to-end against SP&L data. Pick the problem closest to what you actually need to solve.
Random Forest classifier on SP&L fault records. Step-by-step from feature engineering through validation and deployment.
Gradient Boosting regression on five years of weather-correlated load history. Day-ahead and hourly forecasts at the feeder level.
Power-flow simulation with OpenDSS. Compute hosting capacity per feeder, identify circuits with headroom and circuits at the edge.
XGBoost failure-probability model for transformers using age, loading, fault history, and weather exposure. Output is a prioritized inspection list.
Graph algorithms for fault isolation and service restoration. Compare automated FLISR sequences against the actual outage history.
Conservation voltage reduction with rule-based control and Q-learning. Quantify kW savings from a CVR strategy on each feeder.
Monte Carlo across DER adoption trajectories, EV penetration, and storage. Identify which feeders need infrastructure investment first.
Isolation Forest and autoencoders on AMI data. Detect equipment degradation, theft, and measurement errors.
Eight deeper-cut guides building on the beginner set. PyTorch, SHAP explainability, time-aware validation, reinforcement learning, stochastic optimization.
Multi-class outage cause prediction with XGBoost. SHAP explainability and time-aware validation.
PyTorch LSTM that predicts 24 hours ahead using weather and temporal features.
LightGBM surrogate models trained on computed hosting capacity values. Screen circuits in milliseconds.
Time-to-failure modeling with the lifelines library. Hazard curves and predictive maintenance budgets.
Reinforcement learning for restoration optimization and microgrid islanding on the SP&L distribution network.
Deep Q-Network for multi-device Volt-VAR control with PyTorch.
Cost-benefit analysis under uncertainty. Choose upgrades that perform across multiple DER scenarios.
Variational Autoencoders for probabilistic anomaly scoring. Real-time streaming sliding-window detection.
Four guides on the plant side. Boiler feed pumps, feedwater systems, rotating-equipment diagnostics, and the path to a digital twin.
Isolation Forest plus rolling statistics on boiler feed pump vibration data. Detect bearing wear and seal failures.
Linear regression and residual analysis on BFP performance vs unit load. Identify pump drift.
Multi-class classification with SHAP and Random Forest. Distinguish bearing wear, misalignment, cavitation.
Physics-informed ML using OEM pump curves and a composite health index. The first step toward a true plant digital twin.
Three guides for distribution power flow analysis using OpenDSS directly. Voltage profiles, hosting capacity, loss analysis.
Distribution power flow on the SP&L synthetic utility model. Voltage profiles, line loading, system losses.
Iterative power flow with voltage constraints. Solar PV hosting capacity on the SP&L distribution model.
Distribution system loss analysis from OpenDSS power flow. Top loss contributors, seasonal patterns, reduction strategies.
Common gotchas (data path issues, OpenDSS install problems, version mismatches) are documented on the troubleshooting page. If you hit something not covered, send it.