ML Playground · Wildfire

Wildfire AI Series

A Quantitative Wildfire Risk Assessment pipeline on synthetic SP&L utility data. Five notebooks walk through the architecture behind ignition modeling, burn probability, conditional flame length, PSPS dollarization, vegetation contact, and Net Value Change. Every assumption is visible. Every output is inspectable.

Browse Notebooks ← ML Playground

A Working QWRA Pipeline

QWRA turns weather, fuel, and asset data into one number a utility planner can act on: annual expected dollar loss per asset. This series builds the same shape on open synthetic data, with the shortcuts and assumptions left where you can see them.

01 Ignition Spatial CV · per-cause hazards
→
04 Burn Probability Monte Carlo spread
→
04 Conditional FL 6-bin intensity
→
05 Net Value Change $ exposure per asset
→
02 / 03 Decisions PSPS & trim schedule
5Notebooks in the QWRA pipeline
8Wildfire datasets
4,900Fuel & terrain cells (70×70)
47,538Assets in the NVC layer
200Simulated fire seasons

The Five Notebooks

Each notebook stands alone and runs in Google Colab. They also compose. Notebook 04 produces the BP and cFL surface that notebook 05 monetizes. Notebook 01 produces the ignition surface that notebook 02 turns into PSPS decisions.

Notebook 01

Ignition Prediction

Intermediate ~60 min

Cell-hour ignition risk with spatial-block cross-validation, per-cause hazard decomposition (powerline, lightning, human, unknown), bootstrap CIs on Brier, and Expected Calibration Error. The spatial-CV gap is reported next to the time-only number so the autocorrelation tax is visible.

Spatial CV Per-cause Hazards Calibration
Open Notebook → Open in Colab
Notebook 02

PSPS Decision Support

Intermediate ~55 min

Multi-Attribute Value Function ranking. Expected loss avoided in dollars, minus customer-hour outage cost, minus medical-baseline penalty. The dollar framing CPUC reviewers want in WMP filings. Backtested against synthetic PSPS event history.

MAVF ($-denominated) Pareto Frontier Backtest
Open Notebook → Open in Colab
Notebook 03

Vegetation Contact

Intermediate ~55 min

30, 90, and 180-day contact probability per asset, with bootstrap CIs and per-horizon calibration. Cox proportional hazards where lifelines is installed; logistic-horizon fallback otherwise. Outputs a risk-prioritized trim schedule.

Survival / Logistic Calibration Risk Scheduling
Open Notebook → Open in Colab
Notebook 04

Monte Carlo Burn Probability

Advanced ~75 min

Simulate 200 fictional fire seasons over the 70×70 SP&L grid. A cellular-automaton spread model couples to wind alignment, slope direction, and fuel continuity. Outputs annual burn probability and a six-bin conditional flame-length distribution per cell. Runs in about a second.

Monte Carlo BP & cFL Cellular Automaton
Open Notebook → Open in Colab
Notebook 05

QWRA Net Value Change

Advanced ~75 min

End-to-end QWRA. BP × conditional flame length × response function × asset value, summed to annual expected dollar loss per asset, asset type, and feeder. This is the hardening-priority surface utility planners use to allocate covered-conductor, undergrounding, and pole-replacement budget.

QWRA Response Functions Hardening Priority
Open Notebook → Open in Colab

Wildfire notebook troubleshooting →

How This Maps to Production Practice

Utility wildfire-risk work usually has five layers. This series builds a working analogue for each on open synthetic data. The calibrated tools still matter. So does an inspectable reference.

Pipeline layer What production practice does Notebook analogue
Risk surface (QWRA) vendor Burn probability, conditional flame length, response functions, Net Value Change. The four-layer composition that powers most WMP risk filings. Notebooks 04 + 05
Operational spread forecasts vendor Fire-spread simulation coupled to terrain-modified winds and forecast weather. Runs on a rolling 1–72h cadence to drive PSPS decisions. Notebook 04 spread model (qualitative)
Ignition base rates consulting Per-cause ignition rates (wire-down, conductor-slap, vegetation contact, equipment) per circuit-mile-year, conditioned on weather and equipment age. Notebook 01 failure-mode decomposition
Risk frameworks (MAVF) utility Internal multi-attribute value frameworks that dollarize safety, reliability, and financial outcomes for PSPS and hardening calls. Notebook 02 MAVF + Notebook 05 NVC
Vegetation management utility LiDAR-derived strike-tree probability, species-specific growth models, fall-direction analysis, scheduled trim cycles. Notebook 03 horizon models (geometry-free)

Synthetic substrate disclaimer. SP&L is fictional Phoenix-offset utility data. Response functions, asset values, and ignition loss estimates are public-benchmark placeholders: ICE 2.0 customer interruption costs, CPUC-historical ignition loss ranges, generic transformer replacement values. Calibrate against utility-specific asset registers, claims history, and real fire-weather forcing before any operational use.

DNM Wildfire Expansion

The pipeline runs on a synthetic wildfire substrate added to the SP&L Dynamic Network Model. Phoenix-offset geography, fictional ecology of chaparral pockets, desert scrub, invasive grassland, and suburban WUI, layered over the existing distribution footprint.

fuel_models.parquet terrain.parquet fuel_moisture_daily.parquet wind_direction_hourly.parquet fire_weather_indices_hourly.parquet fire_history.parquet asset_vegetation_proximity.parquet psps_events.parquet

Run It Yourself

Every notebook ships with an Open-in-Colab badge. Clone the SP&L repo, point DNM_REPO_ROOT at the clone, and the eight wildfire datasets are wired into demo_data.load_demo_data. Swap in your own response functions and asset values when you are ready to move past the synthetic scaffold.

Get the SP&L Dataset Browse Notebooks