ClimateSight: Climate Risk Engineering for Grid Infrastructure
The problem
Electric utilities own physical assets — poles, powerlines, transformers, substations — that fail or degrade faster under climate stress. The planner’s real question, how much will a warming climate cost this fleet, and where is the risk concentrated?, can’t be answered by a single model. It needs a trustworthy chain: from raw climate projections, through analysis-ready data, through physics-based impact models, to defensible cost distributions a decision-maker can act on.
At ICF, I designed and built the core of ClimateSight, a platform that delivers exactly that chain. The system ingests climate-model output, regrids and standardizes it to cloud-native datasets, runs per-asset impact models, and produces Monte Carlo cost distributions and loss-exceedance curves.
The climate data pipeline
The foundation is a 7-stage, separation-of-concerns pipeline that turns raw climate output into analysis-ready Zarr cubes on S3:
- Schema & time policy — CF-compliance and a UTC-only time standard that eliminates daylight-saving duplicate-timestamp ambiguities at the source.
- Lazy, S3-backed ingestion of CONUS404 and ERA5 via Intake/HyTest catalogs.
- Regridding from native projection to WGS84 via xESMF, with regrid-weight caching that delivers roughly 4× faster rebuilds.
- Derived-variable transforms, per-variable Zarr cube construction, and Dask cluster management on Coiled AWS profiles.
- Metadata consolidation that cut S3 GET requests from 100+ to 1–2, bringing dataset open times from ~30 seconds to under 1 second.
Across refactoring the experimental codebase into modular, tested, containerized packages and this performance engineering, I cut runtime and peak memory use by 40%+.
Asset-impact models
On top of the data layer sit physics-based impact models for four asset classes:
- Poles — lognormal wind fragility curves integrated against the wind distribution, with age-dependent strength decay.
- Powerlines — thermal ampacity derating coupling temperature, wind, and solar load (IEEE-738-aligned).
- Substations — flood vulnerability via Gumbel/GEV extreme-value fitting to return-period levels.
- Transformers — IEEE C57.91 / Arrhenius-based insulation-aging (hotspot temperature to aging-acceleration to lifespan loss). Transformer modeling was a collaborative effort with a colleague.
The consequence engine
A Monte Carlo cost simulation combines Bernoulli failure masking with cost draws (lognormal, CV-parameterized) to produce loss-exceedance curves — with Dask-safe monotonicity enforcement — plus Value-at-Risk, tail-VaR, and expected-annual-loss, reported with ensemble uncertainty bands across climate models and scenarios.
Hazard standardization & synchronized risk deltas
A distinguishing piece of the work was a configuration-driven pipeline that standardizes raw CoSMoS coastal-flood data — reprojecting from UTM to geographic coordinates into paired historic and future Zarr cubes — and a synchronized Monte Carlo sampler that applies identical random draws to both climate scenarios. This isolates the true climate signal from sampling noise, so a computed risk delta reflects climate change rather than randomness. It was a methodological requirement from the science lead, and it’s the kind of detail that separates a result you can defend from one you can’t. An automated paired-dataset validation framework checks dimensional and coordinate synchronization before any risk math runs.
The system’s formal climate and flood data contracts, and the overall architecture, were developed collaboratively with the ICF team.
Why it matters
Climate adaptation budgets are finite. By engineering the full path from raw projection to quantified, uncertainty-aware risk — and refusing to hide that uncertainty — this work helps utilities direct hardening investment where it protects the most people per dollar.