Improving a WRF-Based High-Impact Weather Forecast System for a Northern California Power Utility

Abstract

We describe enhancements to an operational forecast system based on the Weather Research and Forecasting (WRF) model for the prediction of high-impact weather events affecting power utilities, particularly conditions conducive to wildfires. The system was developed for Pacific Gas and Electric Corporation (PG&E) to forecast conditions in Northern and Central California for critical decision-making such as proactively de-energizing selected circuits within the power grid. WRF forecasts are routinely produced on a 2 km grid, and the results are used as input to wildfire fuel moisture, fire probability, wildfire spread, and outage probability models. This forecast system produces skillful real-time forecasts while achieving an optimal blend of model resolution and ensemble size appropriate for today’s computational resources afforded to utilities. Numerous experiments were performed with different model settings, grid spacing, and ensemble configuration to develop an operational forecast system optimized for skill and cost. Dry biases were reduced by leveraging a new irrigation scheme, while wind skill was improved through a novel approach involving the selection of Global Ensemble Forecast System (GEFS) members used to drive WRF. We hope that findings in this study can help other utilities (especially those with similar weather impacts) improve their own forecast system.

1. Introduction

High-impact weather events test the reliability and durability of power utilities and can be challenging to forecast and manage operationally. Examples of high-impact weather events include high winds from winter storms or offshore wind events, which can topple powerlines; hot and dry conditions, which can result in wildfires, especially when driven by high winds; and heat waves, which can result in high electric loads, leading to equipment stress and possible rotating power outages. We describe an operational forecast system for predicting weather conditions for Pacific Gas and Electric Company (PG&E), whose service territory covers most of Northern and Central California. PG&E is responsible for maintaining and operating gas and electric services to approximately 5.5 million electric customers and 4.5 million natural gas customers. The utility maintains and manages over 106,000 miles of overhead distribution lines and over 18,000 miles of overhead transmission lines [1].

Of critical interest during the past decade is the threat of wildfires. The devastating November 2018 Camp Fire, which began in the western foothills of the northern Sierra Nevada mountains, was the deadliest and most destructive wildfire in California history. The communities of Paradise and Concow were almost completely destroyed, 85 lives were lost, and overall damages were assessed at over USD 16 billion. The cause was traced to a hook on a PG&E transmission line that failed during a dry offshore wind event. A retrospective study found that utility equipment was responsible for up to 14 wildfires over a 9-year period [2] (p. 100).

The Camp Fire was driven by a strong downslope windstorm [3,4], and such events are common in both Northern and Southern California. These winds are commonly called Diablo Winds in Northern California and Santa Ana Winds in Southern California. They are a manifestation of offshore flow, in which high surface pressure over the Great Basin coupled with a surface trough off the California coast leads to strong downslope winds. Aloft, 500 hPa winds were from the north–northeast, nearly perpendicular to the northern Sierra Nevada crest, and a level of critical stability had ascended to around 600 hPa [3,4]. These conditions met the classic criteria for downslope windstorms, which include flow perpendicular to the mountain range and a stable layer near and above the mountain crest [5]. Similar conditions were associated with the Wine Country complex of fires, which occurred in the Coast Range during October 2017. That event was characterized by strong winds gusting to 30–48 m s⁻¹ near and downstream of the mountain crests [6].

An example of a winter storm with damaging winds occurred on 20–21 March 2023, associated with a rapidly strengthening mid-latitude “bomb” cyclone that came ashore in central California. The event was characterized by strong southerly winds, heavy precipitation, and heavy mountain snow. The National Weather Service issued numerous High Wind Warnings for the event along the California coast from San Francisco to San Diego. Winds gusted as high as 89 MPH (39.8 m s⁻¹) in Santa Clara County, and over 470,000 PG&E customers experienced outages [7]. Additional extreme winter storms occurred in January 2008 and February 2024, each disrupting service to over one million PG&E customers.

1.1. PG&E’s Mandate and Operations

In order to prevent the ignition of wildfires caused by PG&E’s electric assets during offshore wind events, PG&E implemented its Public Safety Power Shutoff (PSPS) program in 2018 [2]. The program involves proactively shutting off electricity in communities that are in high-threat areas and are predicted to experience extreme fire weather conditions. PSPS is a highly effective mitigation strategy for preventing utility-related wildfires and is layered atop other mitigation and prevention tools, such as undergrounding of power infrastructure, inspection programs, vegetation management, and system hardening.

In addition to PSPS, PG&E developed its Enhanced Powerline Safety Settings (EPSS) program in 2021. When enabled on distribution powerlines, EPSS reduces the number of fire ignitions caused by tree strikes, animal incidents, equipment failure, and other outage modes by as much as 68%. This is accomplished by adjusting sensitivity settings on equipment to deenergize powerlines within 0.1 s after a disturbance is registered. Although EPSS is an effective and relatively low-cost mitigation strategy, it can impact more customers per outage than PSPS as fuses are bypassed by design when it is enabled [2].

With these operational wildfire mitigation programs, there is a critical need for accurate high-resolution mesoscale forecasts days in advance to intelligently select powerlines for PSPS and EPSS. Such forecasts can reduce customer outages and the possibility of utility-caused wildfires, while increasing safety and reliability by enabling more targeted operational implementation of mitigation techniques.

1.2. PG&E’s Operational Mesoscale Modeling System (POMMS)

The operational wildfire mitigation programs are informed by high-resolution weather forecasts obtained via PG&E’s Operational Meteorological Modeling System (POMMS) along with fuel and vegetation characteristics driven by the weather forecasts. Based on the Weather Research and Forecasting (WRF) model [8,9], POMMS was initially fielded in 2014 (Table 1). Subsequent versions included finer resolution and 30-year WRF reanalyses to assist in past event analysis and for use in PG&E’s operational risk models.

Table 1. PG&E Operational Mesoscale Modeling System (POMMS) development.

POMMS Version	Year Implemented	WRF Version	Key Features
1	2014	3.5.1	Single 3 km grid using boundary conditions from a 12 km WRF run.
2	2018	4.0.2	Nested 3 km grid, MYNN surface layer scheme, RUC land surface model, 30-year reanalysis.
3	2020	4.1.2	Nested 2 km grid, Noah–MP land surface model, stochastically perturbed ensemble, 30-year reanalysis. See text and Table 2 for details.
4	2024	4.5.2	Nested 2 km grid, irrigation triggered by crop-growing season, GEFS-based ensemble, 30-year reanalysis.

Table 1. PG&E Operational Mesoscale Modeling System (POMMS) development.

POMMS Version	Year Implemented	WRF Version	Key Features
1	2014	3.5.1	Single 3 km grid using boundary conditions from a 12 km WRF run.
2	2018	4.0.2	Nested 3 km grid, MYNN surface layer scheme, RUC land surface model, 30-year reanalysis.
3	2020	4.1.2	Nested 2 km grid, Noah–MP land surface model, stochastically perturbed ensemble, 30-year reanalysis. See text and Table 2 for details.
4	2024	4.5.2	Nested 2 km grid, irrigation triggered by crop-growing season, GEFS-based ensemble, 30-year reanalysis.

Table 2. Summary of settings for current WRF version 4.1.2 configuration.

Parameter	Setting
Horizontal grid	18 km outer grid with nests of 6 and 2 km.
Vertical grid	51 levels with a 20 hPa top. Thickness of lowest level is 50 m.
Time step	Adaptive. Maximum time step is 144, 48, 16, and 5.33 s on the respective grids, where the fourth value refers to the 0.67 km on-demand nests. For 0.67 km grids spanning the Sierra Nevada, the maximum time step is reduced slightly for stability (120, 40, 13.33, 4.44 s).
Land use	MODIS 30 arcsec with lakes.Roughness length adjusted for two land use categories (see Table 3).
Radiation	Rapid Radiative Transfer Model for General Circulation (RRTMG) models for both long- and shortwave radiation.
Land surface model	Noah–MP. Using climatological albedo and Leaf Area Index (LAI) from GEOGRID files.
Surface layer scheme	MYNN surface layer.
PBL scheme	MYNN 3rd-order PBL scheme.
Microphysics	Thompson microphysics, which has ice, snow, and graupel processes suitable for high-resolution simulations, along with ice and rain number concentrations.The Kain–Fritsch cumulus parameterization is applied on the outer grid.
Diffusion and dispersion	Smagorinsky first-order closure is used for horizontal turbulent diffusion.Upper-level damping is applied.
Background models	GFS (0.25° grid) or ECMWF (0.125° grid). NOAA 1/12° SST analysis.
Data assimilation	3DVAR data assimilation is applied on the outer grid at T−3 h. Data assimilated include conventional surface and upper-air observations, aircraft data, and satellite-derived winds.

Table 2. Summary of settings for current WRF version 4.1.2 configuration.

Parameter	Setting
Horizontal grid	18 km outer grid with nests of 6 and 2 km.
Vertical grid	51 levels with a 20 hPa top. Thickness of lowest level is 50 m.
Time step	Adaptive. Maximum time step is 144, 48, 16, and 5.33 s on the respective grids, where the fourth value refers to the 0.67 km on-demand nests. For 0.67 km grids spanning the Sierra Nevada, the maximum time step is reduced slightly for stability (120, 40, 13.33, 4.44 s).
Land use	MODIS 30 arcsec with lakes.Roughness length adjusted for two land use categories (see Table 3).
Radiation	Rapid Radiative Transfer Model for General Circulation (RRTMG) models for both long- and shortwave radiation.
Land surface model	Noah–MP. Using climatological albedo and Leaf Area Index (LAI) from GEOGRID files.
Surface layer scheme	MYNN surface layer.
PBL scheme	MYNN 3rd-order PBL scheme.
Microphysics	Thompson microphysics, which has ice, snow, and graupel processes suitable for high-resolution simulations, along with ice and rain number concentrations.The Kain–Fritsch cumulus parameterization is applied on the outer grid.
Diffusion and dispersion	Smagorinsky first-order closure is used for horizontal turbulent diffusion.Upper-level damping is applied.
Background models	GFS (0.25° grid) or ECMWF (0.125° grid). NOAA 1/12° SST analysis.
Data assimilation	3DVAR data assimilation is applied on the outer grid at T−3 h. Data assimilated include conventional surface and upper-air observations, aircraft data, and satellite-derived winds.

Table 3. Roughness length adjustment in MPTABLE.TBL.

Category (LUINDEX)	Default Value (m)	Adjusted Value (m)
Evergreen broadleaf forest (2)	1.10	0.70
Urban (13)	1.00	0.60

Table 3. Roughness length adjustment in MPTABLE.TBL.

Category (LUINDEX)	Default Value (m)	Adjusted Value (m)
Evergreen broadleaf forest (2)	1.10	0.70
Urban (13)	1.00	0.60

As described below, current operations of POMMS version 3 (v3) include both deterministic and stochastically perturbed WRF ensemble forecasts initialized with either Global Forecast System (GFS) or European Centre for Medium-Range Weather Forecasting (ECMWF) forecasts. We have found the WRF ensemble mean to have better skill than any single WRF deterministic model and to provide greater run-to-run consistency. We next describe how POMMS forecasts are utilized to make operational decisions and later describe enhancements to be made in the next release of POMMS, version 4 (v4).

1.3. Fire Potential Index and Fuel Moisture Calculations

The wildfire mitigation programs rely on daily forecasts of PG&E’s Fire Potential Index (FPI) model, a balanced random forest classification machine learning model that is currently driven by POMMS v3. The model combines gridded weather inputs (temperature, vapor pressure deficit, wind speed, friction velocity, and turbulent kinetic energy), dead fuel moisture (DFM), live fuel moisture (LFM) in shrubs, soil moisture, solar irradiance, topographic characteristics (such as ruggedness, a measure of terrain variability; and slope), type of combustible fuel (grass, shrub, timber, and low- and high-density urban), and vegetation characteristics to predict the probability of large and catastrophic fires [2] (pp. 767–773).

To obtain a gridded high-resolution DFM dataset, we force a process-based model developed by Nelson [10] with our historical reanalysis and operational forecasts. The Nelson model simulates the moisture content of 1 h, 10 h, 100 h, and 1000 h time-lagged fuel classes using several weather variables as input, including temperature, relative humidity, solar radiation, and rainfall. It models the transfer of heat and moisture within the fuel and the exchange of heat and moisture with the ambient environment. Comparison with field observations suggests that the model performs reasonably well [11].

For the LFM dataset, we adopt a machine learning-based framework where the daily LFM content for various vegetation species is modeled as a function of several weather predictors (including temperature, relative humidity, vapor pressure deficit, winds, rainfall, and solar radiation), with temporal scales ranging from short-term (1–7 days) to long-term (1–8 month) periods. A random forest regression model is used to relate the weather predictors to LFM content, while minimizing the root mean squared error (RMSE) of the moisture content against the historical LFM content observations from the U.S. Department of Agriculture (USDA) Field Sample Database. Based on the various verification metrics, this model compares favorably with similar models developed in other studies [12].

1.4. Motivation for Upgrading

We seek to enhance identification of weather and wildfire risks to minimize the impacts on PG&E’s customers and infrastructure through more targeted application of its PSPS and EPSS programs. Ongoing internal validation studies indicate the model has a low wind speed bias in higher terrain, a high wind speed bias in flatter and lower elevation terrain, and a warm and dry bias in summer.

Here we examine several techniques toward improving the model’s accuracy to enhance operational risk identification. These include upgrading to a newer version of WRF, introducing model irrigation, reducing the grid spacing, and implementing new forecast ensemble approaches.

2. Current Operational WRF Configuration

We previously undertook an extensive validation process in which we determined optimal settings for our current operational version of POMMS (v3). The process utilized surface wind, temperature, and humidity observations from selected Automated Surface Observing Stations (ASOS), Remote Automatic Weather Stations (RAWS), and the growing PG&E mesonet, which was inaugurated in 2018 and contained over 1550 stations as of August 2024. An emphasis was placed on predicting windstorms associated with wildfires.

The configuration is based on WRF version 4.1.2 (Table 2). A domain with 2 km horizontal grid spacing covering PG&E’s service territory is nested within grids of 6 and 18 km (Figure 1). The vertical grid has 51 levels and a 20 hPa top. Land use data are derived from a Moderate Resolution Imaging Spectroradiometer (MODIS) 30 s dataset. We selected the Noah–MP land surface model [13] and the Mellor–Yamada–Nakanishi–Niino (MYNN) surface layer and planetary boundary layer (PBL) schemes [14]. Based on initial verification results, we reduced the roughness lengths for a couple of land use categories, thereby improving a slight low wind speed bias in higher terrain (Table 3).

The forecasts are initialized using either GFS forecasts available on a 1/4° grid or ECMWF forecasts on a 1/8° grid, along with 1/12° NOAA sea surface temperature (SST) analyses. Cycling of soil temperature, soil moisture, and snow cover is applied, with values used from the previous run.

Observations are assimilated on the outermost grid, followed by a brief spin-up period in order for model fields to become internally consistent. Specifically, the outer grid is started 3 h before the nominal initialization time. 3DVAR data assimilation is applied on the outer grid using conventional surface and upper-air observations, as well as aircraft data, from the Meteorological Assimilation Data Ingest System (MADIS). The nested grids are spawned after the outer grid has run for 3 h.

2.1. Stochastically Perturbed Ensemble

In order to provide a measure of forecast uncertainty, several ensemble members are stochastically perturbed. Each member utilizes stochastically perturbed physics tendencies (SPPTs), the stochastic kinetic energy backscatter scheme (SKEBS), and the stochastically perturbed parameter scheme (SPP) [9]. Random perturbations are applied to all grids, and a different random seed is used for each member.

The mean value of all stochastically perturbed and deterministic runs is found to perform slightly better than any single deterministic run. It also provides better run-to-run consistency, which is important when scoping longer-range weather events.

2.2. Schedule

A GFS-initialized deterministic run is launched every day at 00:00, 06:00, 12:00, and 18:00 UTC, while an ECMWF-initialized deterministic run is launched every day at 00:00 and 12:00 UTC. This is supplemented by a 6-member stochastically perturbed ensemble that is run every day at 00:00 and 12:00 UTC. Half of the ensemble members are initialized with GFS, and the other half with ECMWF.

During critical weather events, PG&E has the ability to launch a fine-scale 0.67 km selectable nest on demand within the 2 km domain for tactical decision-making. These runs are launched using a GUI (Figure 2). To maintain computational stability in grids that span the Sierra Nevada, the maximum adaptive time step is reduced slightly (Table 2).

2.3. Historical Reanalysis

A 30-year WRF reanalysis was constructed using Climate Forecast System Reanalysis (CFSR) and CFSv2 data [15]. Each reanalysis year was initialized on July 29 at 00:00 UTC. This allowed for a 3-day spin-up period, which was discarded. The reanalysis year thus begins on August 1 of each year. Spectral nudging was applied on the two outermost grids at the mid- and upper levels. The reanalysis is updated annually.

The reanalysis is utilized by PG&E meteorologists and data scientists to train fire probability and outage probability models. The latest version of FPI was trained using a novel fire occurrence dataset [16] and hourly weather and fuel moisture reanalysis data. PG&E’s outage model also leverages hourly reanalysis data and the location and time of over 500,000 power outages, as well as other model features, to train a multi-classification machine learning model. The models are then deployed operationally to match the schedule and resolution of POMMS and are the foundation of PG&E’s operational wildfire mitigation decisions.

The size, scope, and frequency of Diablo wind events can be estimated using the reanalysis. For example, PG&E constructed a climatology of Diablo wind occurrence over PG&E’s service territory for the 31-year period of 1990–2020 (Figure 3). The criteria were as follows:

Wind direction between 350° and 112.5° (north to east–southeast)

Wind speed ≥ 8.9 m s⁻¹ (20 mph)

Relative humidity ≤ 25%

The above criteria were met for a minimum of 3 consecutive hours over an extent of at least 900 km².

Each event must be spaced 6 h apart from another event.

In this climatology, a strong peak of more than 12 days per year is noted over the High Sierra. Elsewhere, the Coast Range north of San Francisco Bay and the northern Sierra Nevada experience about five to six Diablo wind events per year. These locations coincide with areas where PSPS operational mitigations are most likely to be invoked.

3. Methodology

3.1. WRF Experiments

To improve the forecast system, we selected a recent version of WRF (version 4.5.2, released in December 2023) as a candidate for inclusion in POMMS v4. We investigated and explored options that have been added or updated since v3. Examples include the irrigation scheme in Noah–MP and the use of a gravity wave drag scheme. Experiments on finer grids provided guidance on the trade-off between the accuracy afforded by higher resolution versus higher computing costs. Finally, we were interested in evaluating the use of intelligently sub-selected Global Ensemble Forecast System (GEFS) [17] members to drive an improved WRF ensemble. These experiments are summarized in

Table 4. WRF experiments.

Experiment ID	WRF Version	Grid Spacing (km)	Irrigation	Gravity Wave Drag	Description
C41	4.1.2	2.0	No	No	POMMS v3 Control
C45	4.5.2	2.0	No	No	POMMS v4 Control
IRR	4.5.2	2.0	Yes	No	Triggered irrigation
F45	4.5.2	2.0	No	Option 1 (outermost two grids)	Final POMMS v4 configuration
F1P5	4.5.2	1.5	No	Option 1 (outermost two grids)	Final POMMS v4 configuration (1.5 km grid)
F1P0	4.5.2	1.0	No	Option 1 (outermost two grids)	Final POMMS v4 configuration (1.0 km grid)
FGM	4.5.2	2.0	No	Option 1 (outermost two grids)	Final POMMS v4 configuration (GEFS initialization, mean value)
FG25, FG75	4.5.2	2.0	No	Option 1 (outermost two grids)	Final POMMS v4 configuration (GEFS initialization, 25th and 75th percentiles)

.

3.1.1. Initialization

Rather than using global reanalysis data to initialize our experiments, we chose operational global forecast model runs that would have been available at the time of initialization, with the intent of providing a more realistic characterization of errors at typical forecast lead times. For most experiments, we used GFS forecasts available on a 0.25° global grid with 3 h output intervals. Each GFS run was initialized 6 h prior to the event, and a 3 h spin-up period was applied on the outermost WRF grid. For example, the 24 October 2019 00:00 UTC WRF run utilized the 23 October 18:00 UTC GFS forecast, and the outer grid was initialized on 23 October at 21:00 UTC.

Ensemble experiments were initialized using GEFS forecasts available on a 0.5° global grid with 6 h output intervals, and a 6 h spin-up was applied on the outermost WRF grid. (As an exception, case 1 utilized a 1° GEFS grid.)

3.1.2. Control Experiments

The initial set of experiments compare the POMMS v3 configuration (WRF version 4.1.2; C41) with a similar configuration based on WRF version 4.5.2 (C45). As with the current operational version, we reduced the roughness length for two vegetation categories (Table 3). We also made two minor configuration changes. First, C45 uses the default MYNN PBL scheme closure level of 2.6, whereas C41 uses the third-order closure. Second, C45 uses the SST fields from the initial model (GFS) rather than from a NOAA analysis.

3.1.3. Irrigation and Gravity Wave Drag

Much of California is used for agriculture, with large concentrations of irrigated crops in the Central Valley, Salinas Valley, and various wine-growing regions. Irrigation adds moisture to the soil and lower atmosphere and is potentially important for reducing model dry and warm biases. One research group using an earlier version of WRF (3.8.1) determined that the addition of irrigation helped to alleviate a warm, dry bias in the crop-growing season in the central United States [18].

WRF now has a dynamic irrigation scheme within Noah–MP to estimate irrigation water requirements and apply water in the irrigated croplands (Figure 4). The scheme irrigates based on sprinkler, drip, and surface-flooding methods. We tested the option in which irrigation is triggered based on crop season planting and harvesting dates. According to a standard WRF Noah–MP table, these dates are April 22 to October 28. This experiment, denoted as IRR, is a variant of C45.

We tested a basic gravity wave drag parameterization that includes the subgrid topography effects of gravity wave drag and low-level flow blocking. In our experiments, we elected to apply the gravity wave drag parameterization to the two outermost grids only. These tests also included the triggered irrigation of the IRR experiment and are referred to as F45. This experiment served as the baseline for subsequent experiments.

3.1.4. Higher-Resolution Horizontal Grid

Decreasing horizontal grid spacing allows for better resolution of terrain and land surface features, but at significant added computational cost. We compared results with grid spacings of 1.5 km and 1 km (experiments F1P5 and F1P0, respectively) against our baseline 2 km grid (

Table 5. Horizontal grid spacing used in experiments. The top row is the baseline 2 km grid, while the bottom two rows are for higher-resolution experiments. The number of grid points is given in parentheses.

Experiment	Outer Grid	Middle Grid	Inner Grid
All except F1P5, F1P0	18 km (270 × 270)	6 km (316 × 316)	2 km (397 × 481)
F1P5	22.5 km (150 × 157)	7.5 km (208 × 229)	1.5 km (531 × 641)
F1P0	15 km (206 × 217)	5 km (256 × 289)	1 km (796 × 961)

).

3.1.5. Initialization with GEFS Forecasts

The GEFS is a stochastically perturbed 30-member ensemble based on the GFS. The ensemble members utilize the GFS analysis perturbed by a 6 h ensemble Kalman filter (EnKF) forecast ensemble. Model uncertainty is introduced using the stochastically perturbed physics tendencies (SPPTs) and stochastic kinetic energy backscatter (SKEB) schemes [17].

The computational cost would be prohibitive if we were to initialize a high-resolution WRF forecast corresponding to each member. We therefore tested a novel forecast strategy using 9 representative GEFS members that were dynamically selected to maintain the large-scale flow diversity of the entire GEFS ensemble. The intended outcome is a WRF ensemble that is more accurate than a single WRF forecast at a higher resolution, yet also that provides meaningful information about forecast uncertainty.

This selection strategy involves analyzing the GEFS forecast 500 hPa geopotential height field (Z500) over an area roughly the extent of the 6 km nest, averaged over the third day of the forecast. It consists of two steps that are intended to sample the mean and diversity of the ensemble [19]. The first step involves a self-organizing map (SOM) [20,21], which is an artificial neural network used as a clustering method to group together events with similar structure. The SOM analysis is used to classify the GEFS ensemble into five nodes. We identified the five GEFS members closest to the centroid of each SOM node, thereby sampling the mean flow field. In the second step, we performed a principal component analysis (PCA) [22], a technique for reducing the dimensionality of a field while increasing interpretability. We identified the four GEFS members with the largest magnitude of the two leading principal components, thereby sampling the diversity of the flow field. Overall, we selected the 9 GEFS members that captured both the mean and outlier behavior of the large-scale flow in the full ensemble.

To compare with other experiments, we analyzed the mean, 25th percentile, and 75th percentile of the various validation measures from the resulting WRF ensemble. In this manner, the GEFS-based forecasts were calibrated at each observation location.

3.2. Validation

The surface observing networks used in the validation are shown in Figure 5, with station counts shown in

Table 6. Station counts for observing networks used in the validation.

Network	Number Used
ASOS	44
RAWS	46
PG&E mesonet	1152

. The observing networks are:

ASOS. Winds are measured at the WMO standard of 10 m above ground level (AGL) and air temperature and dew point temperature at 2 m AGL.

RAWS. Winds are measured at 6.1 m (20 ft) AGL, while air temperature and relative humidity are measured at between 1.2 and 2.5 m AGL.

PG&E mesonet. The height of the instrument package varies depending on the available mounting points at each location, ranging from 2.1 to 14.3 m AGL, with a mean of 7.4 m AGL.

Note that the large majority of RAWS and PG&E mesonet stations are located within wildfire-prone regions.

We eliminated any stations previously determined through statistical analysis to be of questionable quality. We then adjusted the WRF wind speed at 10 m to the anemometer or instrument height under the assumption that wind speed follows a logarithmic profile [23]. Further, we performed a gust factor analysis for each station, where the gust factor is defined as hourly maximum wind gust divided by hourly maximum wind speed [24]. Specifically, we examined a 60-day moving average spanning all of the cases (or the period of record, if shorter) of the gust factor for each station. Stations that failed to show a relatively constant gust factor or a stable seasonal pattern were omitted from the validation.

We validated the following surface variables:

Wind speed at 10 m. Separate categories were validated depending on the threshold: all wind speeds, and wind speeds (whether observed or simulated) exceeding a threshold of 5 or 10 m s⁻¹.

Air temperature at 2 m.

Vapor pressure deficit (VPD) at 2 m. Because the moisture content of saturated air is exponentially related to temperature, validation metrics based on relative humidity and dew point temperature can be misleading. We therefore chose VPD as our humidity variable. A measure of dryness, it is simply the difference between the saturation vapor pressure (a function of temperature) and actual vapor pressure (a function of specific humidity). VPD is increasingly used as a component in fire weather indices such as the Hot–Dry–Windy Index [25]. Because of its dependence on temperature, VPD errors are strongly influenced by temperature errors.

The statistical measures assessed are root mean square error (RMSE), bias (mean error), and correlation. RMSE was chosen over mean absolute error (MAE) because it gives greater weight to more extreme errors.

3.3. Case Descriptions

We selected six recent cases to evaluate across WRF experiments, all of which were associated with high-impact weather events within the PG&E service territory. One case involved an early spring storm with strong winds and widespread rain and snow. The remainder involved a combination of hot, dry, and windy conditions associated with wildfires. For each case, a series of 120 h WRF forecast experiments was run, with validation focused on a 48 or 72 h period encompassing the critical period of the event (

Table 7. List of case studies. Each WRF run is 120 h in duration. Validation periods are 48 or 72 h.

Case	Description	Classification	WRF Start Time (UTC)	Validation Start (UTC)	Validation Period (h)
1	Intense downslope windstorm	High winds; hot and dry	2019-10-24 00:00	2019-10-25 00:00	72
2	Late-season strong offshore wind event	High winds	2021-01-17 00:00	2021-01-18 00:00	48
3	Dixie Fire in northern Sierra Nevada	Hot and dry	2021-07-11 12:00	2021-07-12 12:00	48
4	Heat wave and Mosquito Fire (northern Sierra Nevada)	Hot and dry	2022-09-04 00:00	2022-09-05 00:00	72
5	Northwesterly wind event in Central and Southern California	High winds	2023-02-19 12:00	2023-02-20 12:00	48
6	Late winter storm	High winds; heavy precipitation	2023-03-19 12:00	2023-03-20 12:00	48

).

3.3.1. Case 1: Intense Downslope Windstorm, October 2019

A strong offshore northerly wind event occurred on 26–28 October 2019 as a quick-moving, positively tilted shortwave trough embedded in a highly amplified ridge pattern over the eastern North Pacific Ocean that dropped down from southwestern Canada and into the Great Basin (Figure 6). It became even more amplified, with its axis becoming more positively tilted, as it dug into the Desert Southwest. A maximum wind gust of 102 MPH (45.6 m s⁻¹) was recorded at the Pine Flat Road PG&E weather station (PG305, elevation 3308 ft [1008 m]) in Sonoma County on 27 October at 08:00 PDT, and numerous stations across northern California recorded gusts of 30 m s⁻¹ or more during the event. The Redding–Sacramento (KRDD-KSAC) pressure gradient, an important metric for forecasting northerly winds, peaked at 10.6 hPa, exceeding the historical 99.9 percentile for this metric. The wildfire threat led PG&E to de-energize approximately 935,000 customers, making it by far the largest PSPS event in the company’s history. PG&E’s inspectors found more than 550 instances of utility equipment damage and related hazards in the de-energized areas, any of which could have potentially sparked a fire had powerlines been energized.

3.3.2. Case 2: Offshore Wind Event, January 2021

A late-season strong offshore wind event occurred on 18–19 January 2021 during an unusually dry and abnormally warm winter. A small shortwave trough moving through a highly amplified eastern North Pacific ridge then moved offshore of California and helped to pull another positively tilted trough in the Western U.S. westward toward Central California (Figure 7). It ultimately became a strong cut-off low that moved off the Southern California coast. At the same time, a strong ridge of high pressure built into the Pacific Northwest. Owing to the abnormally dry fuels and lack of recent precipitation, PG&E issued a PSPS affecting 5200 customers in the southern Sierra Nevada foothills. Wind gusts exceeding 35 m s⁻¹ were recorded in that area. PG&E’s inspectors identified 423 incidents of utility equipment damage and related hazards resulting from high winds experienced in the de-energized areas.

3.3.3. Case 3: Dixie Fire in Northern Sierra Nevada, July 2022

The Dixie Fire started on 13 July 2021 in the Feather River Canyon of the northern Sierra Nevada mountains, burning over 960,000 acres before being fully extinguished in October. Caused by a tree falling onto a PG&E distribution line, it was the largest wildfire of the 2021 California wildfire season and the largest single wildfire in the state’s history [26]. It was associated with very hot temperatures and a significant annual precipitation deficit (Figure 8).

3.3.4. Case 4: Heat Wave and Wildfire Conditions, September 2022

A significant heat wave occurred over California on 4–9 September 2022 as a strong upper-level high-pressure area built westward from the Four Corners region (Figure 9). For example, San Jose International Airport (KSJC) reached 106 °F (41.1 °C) on the 6th, breaking the maximum temperature record for September. Also on that day, the all-time peak electrical load of 52,061 MW was recorded by the California Independent System Operator (CAISO). Several large wildfires ignited across the state during the first nine days of the month. The Mosquito Fire, California’s largest wildfire of the year, ignited on the 6th and spread through the Sierra Nevada foothills of Placer and El Dorado counties west of Lake Tahoe. It eventually burned more than 75,000 acres [26].

3.3.5. Case 5: Northwesterly Wind Event, February 2023

On 20–21 February 2023, a shortwave trough originating in the Gulf of Alaska strengthened and amplified into a longwave trough as it moved south off the Pacific Northwest coast, bringing high winds and cold temperatures to California. An associated strong jet stream developed over Oregon and Northern California (Figure 10). Wind gusts exceeded 26 m s⁻¹ in the San Francisco Bay Area and 35 m s⁻¹ in the southern Sierra Nevada foothills. Nearly 400,000 PG&E customers experienced a service disruption during the event.

3.3.6. Case 6: Late Winter Storm, March 2023

As previously mentioned, on 20–21 March 2023, a rapidly strengthening mid-latitude “bomb” cyclone associated with a cold, deep, upper-level low originating from the Gulf of Alaska dropped south off California (Figure 11). The surface low had a minimum pressure of 986 hPa at 18:00 UTC just before making landfall. Radar and satellite imagery near the time of landfall revealed two individual mesoscale low-pressure centers rotating around the main low, creating a Fujiwhara effect. The event produced strong southerly winds, heavy precipitation, and heavy mountain snow, causing a service disruption to more than 500,000 PG&E customers.

4. Results

A summary of the validation results by experiment, for all cases and observing networks combined, is presented in Figure 12 and listed in

Table 8. Validation of wind speeds by threshold and experiment for all cases and observing networks combined. Threshold values are for either the simulated or the observed speed exceeding the threshold.

	Wind Speed (All Speeds; m s−1)				Wind Speed (Speeds ≥ 5 m s−1)			Wind Speed (Speeds ≥ 10 m s−1)
Experiment	N	RMSE	Bias	Correlation	N	RMSE	Bias	N	RMSE	Bias
C41	324,769	2.04	0.39	0.48	63,663	3.96	−0.37	8406	6.81	−3.40
C45	324,769	1.99	0.32	0.55	65,346	3.91	−0.05	9118	6.44	−2.27
IRR	324,769	1.98	0.30	0.56	64,781	3.92	−0.08	9226	6.41	−2.18
F45	324,769	1.96	0.28	0.57	64,669	3.94	−0.05	9517	6.34	−1.91
F1P5	324,769	1.92	0.27	0.60	64,110	3.83	−0.05	9306	6.18	−1.95
F1P0	324,769	1.91	0.29	0.61	65,071	3.80	0.10	9709	6.02	−1.53
FGM	324,769	1.80	0.24	0.65	60,424	3.57	−0.54	7744	6.20	−3.87
FG25	324,769	1.81	−0.13	0.62	54,469	3.87	−1.46	7280	6.82	−5.12
FG75	324,769	1.92	0.60	0.65	69,885	3.58	0.47	9225	5.83	−1.82

and

Table 9. Validation of temperature and vapor pressure deficit by experiment for all cases and observing networks combined.

	Temperature (°C)				Vapor Pressure Deficit (hPa)
Experiment	N	RMSE	Bias	Correlation	N	RMSE	Bias	Correlation
C41	328,497	3.40	0.40	0.88	328,490	6.59	2.98	0.82
C45	328,497	3.50	1.24	0.89	328,490	7.38	4.46	0.84
IRR	328,497	3.43	1.12	0.89	328,490	6.98	4.01	0.84
F45	328,497	3.34	1.11	0.89	328,490	6.83	3.90	0.85
F1P5	328,497	3.28	1.04	0.89	328,490	6.72	3.71	0.85
F1P0	328,497	3.13	0.98	0.90	328,490	6.44	3.54	0.86
FGM	328,497	3.23	1.07	0.90	328,490	6.43	3.51	0.86
FG25	328,497	3.15	0.60	0.90	328,490	5.95	2.54	0.86
FG75	328,497	3.42	1.54	0.89	328,490	7.10	4.44	0.85

. Validation data counts by variable and event for each experiment are shown in

Table 10. Validation data counts by variable and event for each experiment.

Case	Temperature	Vapor Pressure Deficit	All Wind Speeds	Winds ≥ 5 m s−1	Winds ≥ 10 m s−1
1	42,762	42,758	41,869	7860	1386
2	38,248	38,248	37,975	19,405	4002
3	51,990	51,989	51,381	5896	408
4	88,078	88,077	81,071	4920	58
5	53,280	53,279	52,725	15,939	2197
6	54,139	54,139	53,748	10,649	1466
All	328,497	328,490	324,769	64,669	9517

. Complete validation results, disaggregated by case and observing network, are provided in Supplementary Table S1.

4.1. Control Experiments

Wind speeds in the version 4.5.2 control configuration (C45) were improved compared with the version 4.1.2 configuration (C41) for all wind thresholds and for all statistical measures (RMSE, bias, and correlation; Figure 12a–c; Table 8). In both experiments, it is evident that WRF had a high wind speed bias overall but a low bias at higher wind speeds. For instance, in C45, the overall wind speed bias was +0.32 m s⁻¹, while for speeds ≥ 10 m s⁻¹ the bias was −2.27 m s⁻¹. At the highest threshold, C45 reduced the slow bias by 33% and reduced RMSE by 5% compared to C41.

Despite these improvements in wind speeds, the previous version (C41) still performed better overall for temperature and VPD (Figure 12d,e; Table 9). Although both experiments exhibited warm and dry biases overall, the biases were greater in C45 than in C41. This was particularly true for warm season events (Table S1). We note that VPD error is highly correlated with temperature error, because saturation vapor pressure is an exponentially increasing function of temperature.

4.2. Irrigation and Gravity Wave Drag

Irrigation triggered based on crop season (experiment IRR) reduced the warm, dry bias during the warm season events compared to C45. The most significant impact was during the July 2021 case, in which the temperature RMSE and bias were reduced by 10% and 19%, respectively, with concomitant reductions in VPD error (Table S1).

The introduction of gravity wave drag (experiment F45) brought about further incremental improvements. Compared with C45, the new baseline configuration with gravity wave drag on the two outermost domains and triggered irrigation (F45) improved upon almost all validation measures. For instance, the bias of wind speeds ≥ 10 m s⁻¹ improved by 16%.

4.3. Higher-Resolution Horizontal Grid

Not surprisingly, runs with horizontal grid spacings of 1.5 and 1 km (experiments F1P5 and F1P0) outperformed their baseline 2 km counterpart (F45). For instance, comparing the 1 and 2 km results, the slow bias for winds ≥ 10 m s⁻¹ was reduced by 20%, the warm bias was reduced by 12%, and the dry bias was reduced by 9%. Concomitant improvements in RMSE and correlation were also noted. We report on the computational aspects of these higher-resolution runs in Appendix A.

4.4. Ensemble Forecasts Initialized with GEFS

The mean (FGM) and percentile values (FG25, FG75) of the nine-member ensemble initialized with GEFS generally provided greater accuracy compared with all other experiments on a 2 km grid. They also compared favorably with finer-grid experiment (F1P5, F1P0) metrics.

For all wind speeds, FGM and FG25 had the lowest RMSE and the smallest bias. FG25 actually had a small negative bias. These results indicate that the positive wind speed bias in the deterministic runs was reduced at least in some of the GEFS members. For winds ≥ 10 m s⁻¹, FG75 reduced the RMSE by 14% and reduced the slow wind bias by almost 50% compared to the current operational baseline (C41). Furthermore, FG75 had the lowest RMSE and a bias that was smaller than that of all experiments except the 1 km grid (F1P0).

For temperature, the 25th percentile (FG25) provided the highest accuracy of any of the WRF version 4.5.2 experiments. FG25 had the lowest RMSE in all experiments and the lowest bias in almost all experiments except C41. Similarly, for VPD, the 25th percentile (FG25) had the lowest RMSE and bias in all configurations.

5. Discussion and Conclusions

An operational WRF-based system for predicting high-impact weather was built in 2014 for PG&E and has been improved upon since then. We have constructed a testing and validation environment for exploring new features and techniques with the latest available version of WRF. To our knowledge, this is the first study documenting an original weather forecast system in use by power utilities.

Compared with our current system based on WRF version 4.1.2, a configuration based on version 4.5.2 resulted in significantly reduced wind speed error. This was especially true for the highest speeds, which are associated with high-impact weather such as wildfires.

Several studies have noted systematic warm biases in weather and climate models during the summertime over the Great Plains (see [18] and references therein). Much of this can be attributed to model soil moisture conditions that are too dry. This inhibits evapotranspiration, preventing a repartitioning of net radiation into latent heating. One result is that fewer shallow clouds are available to cool the surface [27]. We noted that such biases were also present in WRF and that they actually increased somewhat in the baseline version 4.5.2 configuration compared with the 4.1.2 configuration. This motivated additional experiments involving the use of WRF’s irrigation scheme.

A study using an earlier version of WRF coupled with a provisional irrigation scheme was able to reduce the warm, dry bias over the Central United States during summertime [18]. We therefore tested the irrigation scheme in WRF version 4.5.2, finding that the addition of irrigation triggered by the crop-growing season does indeed reduce the summertime warm and dry bias. We recognize, however, that the irrigation scheme in WRF has certain limitations. For instance, irrigation in WRF is managed through static tables for various land use types, but land use patterns may change over time. Further, the application of model irrigation may vary significantly from real-world values. In particular, actual irrigation might not be applied during droughts when there is a shortage of surface water, or in periods of abundant precipitation. Modeled irrigation could be a valuable tool, but we have limited control over seasonality and areal coverage. For these reasons, the application of modeled irrigation should be closely monitored, particularly over long simulations such as year-long reanalyses.

Experiments using higher-resolution grids were consistently more accurate than their 2 km counterparts, though at a considerably higher computational cost (Appendix A). This prompted additional experiments involving the use of a WRF ensemble to optimize across forecast skill and cost.

We introduced a novel approach in which a WRF ensemble is initialized using a dynamically selected subset of GEFS ensemble members. These members are selected based on the GEFS mean large-scale flow and a representative sample of its diversity. The intended outcome is a WRF ensemble that, when calibrated by observations, is more accurate than a single WRF forecast at a higher resolution, yet also provides meaningful information about forecast uncertainty. We find that judiciously chosen members of a WRF ensemble initialized in this manner with GEFS can be more accurate than higher-resolution deterministic runs. For instance, the 75th percentile of the GEFS-based ensemble initialized results improved wind forecasts greater than 10 m s⁻¹ and outperformed the higher-resolution experiments. Additionally, we found that using the 25th percentile for temperature and VPD provided the greatest reduction in the summertime warm and dry bias. These findings can be recalibrated once the ensemble forecast system has been in routine operation for a period of time. A further advantage of the ensemble approach is that it reduces run-to-run variations that can impair operational use of forecast results. This is especially important when scoping longer-range (up to 5 days) events.

Finally, two points are worth noting. First, our ensemble selection methodology can be applied to other global and regional ensemble forecast systems, such as the ECMWF Ensemble Prediction System [28]. Second, the WRF forecasts presented here are limited by the grid scale representations of terrain, land use, and vegetation. This is particularly true regarding the necessarily smoothed terrain field required by WRF. In recent years, machine learning models, especially deep learning models, have been used to improve weather forecasts produced by mesoscale models such as WRF by employing very high-resolution terrain data [29]. This approach holds promise for cost-efficiently improving wind forecasts in complex terrain and is worthy of future investigation.