Weather risks, investment alignment and power system resilience

Code accompanying Divergent weather risks reveal misaligned investment and reshape power system resilience priorities.

The study distinguishes weather-related increases in net load (demand minus modeled wind and solar generation) from customer outages. It compares these risks with generation additions, distribution investment and federal mitigation awards, then examines future priorities across 32 climate, demand and policy combinations. Historical net load covers 1980–2019 and 18 US planning subregions; the outage estimation panel covers 2,268 counties. Future comparisons cover 2030–2050. The downscaling network supplies wind inputs for generation modeling. Historical outage estimation uses separate weather and outage records.

Release status

This is the current code-only release candidate, not a self-contained, one-command reproduction package. The bounded Fig. 6 implementation is included. Older analyses and figure versions remain for provenance and must not be mistaken for the current workflow.

Before public release, resolve these limitations:

• Data, trained weights and fitted artifacts are not bundled. Public availability of a source does not establish that all required inputs can be rebuilt by the included scripts.
• Many scripts retain /data, /datasets/goes_repro or author-workspace paths. Environment variables are supported only by the workflows listed below, not by every stage.
• MANIFEST.csv records historical origins, not a complete inventory of later additions or a complete execution order.
• Final assembly of Figs. 1–5 and several SI plotting programs are outside this directory. Native plotting output is not necessarily the submitted panel selection or layout.
• Requirements retain archived version pins. Installation and end-to-end execution have not been revalidated in a clean environment for this revision.

Do not execute every numbered file or every plotting program with a wildcard. Some are alternative analyses, diagnostic branches or superseded calculations.

Environments and paths

Software stack	Requirements file	Main use
Analysis	requirements.txt	Arrays, analysis and plotting
Demand	requirements-load.txt	TELL, calibration and statistical dependencies
Siting	requirements-siting.txt	CERF and geospatial processing
GPU	requirements-gpu.txt	PyTorch downscaling training and inference

Review script imports before installation: TELL, CERF and some acquisition clients are installed separately, and some analysis scripts require packages listed in another stack. The requirements files describe the archived setup, not a verified universal installation recipe. Do not combine all four stacks without resolving version constraints.

Example, within an already selected compatible environment:

python -m pip install -r requirements.txt

Full reconstruction needs substantial weather storage and GPU resources for downscaling. Earlier workspace estimates were approximately 18 TB and an L40S-class GPU; these are planning estimates, not benchmarked requirements for this release. Plotting can use existing analysis artifacts without retraining, but the artifacts and definition stamps must be provided. No synthetic-data demo is included.

Variable	Implemented workflow	Purpose
GOES_PLOT_DATA_ROOT	Shared native plotting for Figs. 1–5	Analysis-artifact mirror
GOES_PLOT_OUTPUT_DIR	Most current native Figs. 1–5 plots	Output directory
GOES_DATA_ROOT	Bounded Fig. 6	Original data archive
GOES_FIG6_WORK	Bounded Fig. 6	Baseline, fits, projections and diagnostics
GOES_FIGURE_DIR	Fig. 6 export	Output directory
GOES_ARIAL_FONT	Fig. 6 export	Optional Arial TTF path

Some map/grid paths remain hard-coded even with these variables set. Inspect headers before running. Install Arial for the intended figure typography; do not redistribute proprietary fonts without permission.

Analysis stages

This table describes scientific roles, not a complete execution graph.

Directory	Role
01_acquire_data	Weather, outage, demand, fleet, investment and demographic acquisition/preprocessing
02_downscale_wind	Wind downscaling: cache, normalization, training and inference
03_wind_solar_generation	Generation, temporal interpolation and loss treatment
04_demand_model	County/authority demand, calibration and future demand
05_capacity_siting	Future capacity and policy-dependent siting
06_netload_panel	Historical/future net load aggregation
07_hazard_calendar	Hazard definitions, event calendars and regression controls
08_adequacy_analysis	Net load comparisons, regressions and robustness; includes older branches
09_outage_attribution	Historical attribution, population comparisons and investment alignment
10_future_adequacy	Future peaks, duration, hazard frequency and variance decomposition
11_damage_projection	Bounded cyclone outage projection and legacy diagnostics
12_figures	Native panels, current exports and older layouts

Keep shared modules with their consumers. In particular, hazard-definition stamps, pre-event screens, path definitions and model utilities are not optional plotting files. Circulation indices remain regression controls even though the submitted Fig. 1 no longer displays the earlier circulation panels.

Figure mapping

This follows the manuscript assembly workflow, not historical filename numbering.

Submitted figure	Program(s) in 12_figures/	Final selection
Fig. 1	02_fig1_full_oc.py	Hazard maps, demand/generation estimates and lag panel
Fig. 2	03c_fig2_merged.py	Demand–outage comparison, outage associations and county burden; also exports SI intensity/timing panels
Fig. 3	05c_si_adequacy_alignment.py	Despite its si filename, supplies the main generation-investment figure
Fig. 4	05_fig4_v2.py + 05b_fig4_alignment_panels.py	Population, undergrounding and federal-funding panels
Fig. 5	06_fig5_adequacy_oc.py	Five selected panels, including hazard frequency and variance decomposition
Fig. 6	07_fig6_damage.py	Reads completed bounded-model results; does not fit the model

The author workspace assembles Figs. 1–5 using submission_Revised/scripts_codex/assemble_native_figures_2026_10_01.py and SVG helpers in close_remaining_figures_2026_09_28.py. These files are outside this release directory. They select existing SVG axes and change layout/labels, not scientific data. Their system dependencies include lxml, PyGObject/librsvg and Cairo. A standalone release still needs these files packaged with portable paths.

03_fig2_v2_oc.py and 04_fig3_v8.py represent earlier figure organizations, not the submitted Figs. 2 and 3. 06c_fig5_hazard_frequency.py is a standalone export of a panel already supplied by the current Fig. 5 program. Do not use these as replacement entry points.

The current SI also contains downscaling, demand, generation and siting diagnostics. Not all their final plotting programs are present here; this directory alone does not yet provide complete SI reproduction.

Current Fig. 6

Follow the explicit fitting, projection, uncertainty and verification order in 11_damage_projection/README.md. A separately fitted fractional-logit auxiliary model bounds daily predictions between zero and 24 outage hours per customer. Historical PPML attribution remains a separate analysis.

After all model artifacts and diagnostics are complete:

export GOES_DATA_ROOT=/path/to/data
export GOES_FIG6_WORK=/path/to/completed/fig6_work
export GOES_FIGURE_DIR=/path/to/figure_output
python 12_figures/07_fig6_damage.py

The hottest realization gives approximately +9.56% in modeled excess tropical cyclone outage hours, with a combined 95% interval of −1.23% to +20.17%. This is a relative change, not percentage points of observed outage share. The interval includes zero. Storm occurrence is replayed and network vulnerability is fixed, so this is not an unconditional forecast of future outages.

Do not use 06_r5_dose2.py or 07_dose6.py for the submitted result. Their former 1.74/1.37-percentage-point results are superseded. Keep 09_daily_projection.py when packaging the bounded model: although its standalone unbounded projection is not the submitted method, project_bounded_outage_2026_10_01.py imports its design utilities.

Documented data sources

Data are not redistributed here. Check access conditions and third-party licenses at download time; this table is a source inventory, not a claim of verified endpoint availability.

Domain	Source
Historical/future weather	Thermodynamic Global Warming
Downscaling reference	CONUS404; acquisition endpoints recorded in stage 01
Reanalysis indices	ERA5 vapor transport and geopotential height
Tropical cyclones	HURDAT2
Outages	EAGLE-I
Demand	EIA-930, county-hourly demand, FERC Form 714
Fleet	EIA-860/EIA-923; GODEEEP for comparison
Capacity/siting	GCAM-USA, CERF, GRIDCERF, Cambium
Investment/mitigation	EIA-861, FERC Form 1 and FEMA awards
Population	American Community Survey five-year estimates, 2022

Trained checkpoints, model metadata and fitted artifacts need documented distribution locations before a complete release. Do not assume that every inference script loads the same checkpoint solely because an earlier README named G_deploy.pth.

License and contact

Code is released under the MIT License. Third-party data and software retain their own licenses. Scientific questions should go to the corresponding author listed in the manuscript; code issues can be reported through the repository issue tracker.