CAISO Outage Reason Codes and Derate Depth

Authors: Brian Nguyen · Claire Kim · Created: 2026-08-25 · Tags: #Experiment #Outages #Weather

Notebook 11 models how many forced outages start on a county-day. This notebook asks what the reason codes say those outages are, and where the temperature response actually lives once the filings are read by reason.

from pathlib import Path
import sys
import warnings

import numpy as np
import pandas as pd
import polars as pl
from bokeh.io import output_notebook, show
from bokeh.layouts import row
from bokeh.models import ColumnDataSource, HoverTool, Range1d, Span
from bokeh.plotting import figure
from bokeh.resources import INLINE
from bokeh.transform import dodge
from scipy.stats import nbinom, poisson
from statsmodels.tools.sm_exceptions import (
    ConvergenceWarning,
    HessianInversionWarning,
)

PROJECT_ROOT = Path.cwd()
if PROJECT_ROOT.name == "notebooks":
    PROJECT_ROOT = PROJECT_ROOT.parent
sys.path.insert(0, str(PROJECT_ROOT))

SRC_DIR = PROJECT_ROOT / "src"
if str(SRC_DIR) not in sys.path:
    sys.path.insert(0, str(SRC_DIR))

from src.caiso_outage_nb_pipeline import (
    COUNTY_DAILY_FEATURES,
    build_caiso_county_nb_panel,
    build_caiso_market_zone_nb_panel,
    fit_negative_binomial_model,
    likelihood_ratio_block_tests,
    model_performance_by_group,
    overdispersion_summary,
    poisson_negative_binomial_comparison,
    resource_mapping_coverage_summary,
    rollup_county_predictions_to_market_zone,
    rolling_backtest_negative_binomial_model,
)
from src.caiso_outage_specification_search import (
    add_candidate_interaction_columns,
    add_extreme_weather_columns,
    candidate_specifications,
    coefficient_stability_across_expanding_fits,
    compare_specifications,
    extreme_specifications,
    fit_nb1_model,
    l1_regularization_path,
    purged_walk_forward_family_comparison,
    variance_family_comparison,
    walk_forward_specification_metrics,
)
from chart_theme import SERIES, corr_heatmap, html_table, style

DATA_DIR = PROJECT_ROOT / "data"
PROCESSED_DIR = DATA_DIR / "processed"
PROCESSED_DIR.mkdir(parents=True, exist_ok=True)

pd.options.display.max_columns = 80
pd.options.display.float_format = "{:.4f}".format

# The county fits sit on the alpha = 0 dispersion boundary, so statsmodels
# reports a failed convergence check and an uninvertible Hessian while still
# returning the Poisson-limit coefficients. The tables below carry the fitted
# alpha and the held-out scores, which is where a failed fit shows.
warnings.filterwarnings("ignore", category=ConvergenceWarning)
warnings.filterwarnings("ignore", category=HessianInversionWarning)

output_notebook(INLINE, hide_banner=True)

Setup: The Panel, the Filings, and the Block Tests

The same county-day panel as notebook 11, plus the matched filing-level events the reason codes are read from.

county_panel, mapped_events, resource_crosswalk, county_zone_weights = build_caiso_county_nb_panel(DATA_DIR)

county_feature_blocks = {
    "outage state": [
        "log1p_Y_lag1",
        "log1p_planned_start_count_lag1",
        "log1p_forced_start_mw_sum_lag1",
    ],
    "size": ["log_county_nameplate_capacity_mw", "log_county_plant_count"],
    "weather": ["cdd65", "hdd65", "temperature_f_std", "county_weather_basis_risk_flag"],
    "load": ["log_system_load_mean"],
    "calendar": ["weekend", "month_sin", "month_cos"],
    "zone effects": ["region_NP15", "region_SP15"],
}
county_block_tests = likelihood_ratio_block_tests(
    county_panel,
    feature_cols=COUNTY_DAILY_FEATURES,
    feature_blocks=county_feature_blocks,
    high_count_quantile=0.90,
    maxiter=300,
)

The target \(Y\) treats every forced start alike, but the report says why each generator was down (nature_of_work). Notebook 02 mapped that taxonomy at the system level; the county crosswalk built here lets us ask the sharper questions - which generators, in which counties, report temperature as the problem, and does a reason-aware target change the model?

1. Half of Matched Forced Starts Are Temperature Filings

matched_forced = mapped_events[
    mapped_events["county"].notna() & mapped_events["outage_type"].eq("FORCED")
]
reason_counts = (
    matched_forced.groupby("nature_of_work")["outage_mrid_str"]
    .nunique()
    .sort_values()
    .tail(10)
)

fig = figure(
    title="Matched Forced Outage Starts by Reason Code",
    y_range=reason_counts.index.tolist(),
    x_axis_label="Outage Starts",
    height=360,
    sizing_mode="stretch_width",
    toolbar_location="above",
)
fig.hbar(
    y=reason_counts.index.tolist(),
    right=reason_counts.tolist(),
    height=0.6,
    color=[
        SERIES[1] if reason == "AMBIENT_DUE_TO_TEMP" else SERIES[0]
        for reason in reason_counts.index
    ],
)
fig.add_tools(HoverTool(tooltips=[("reason", "@y"), ("starts", "@right{0,0}")]))
show(style(fig))

ambient_events = matched_forced[matched_forced["ambient_temperature_flag"]]
events_per_filer = ambient_events.groupby("resource_name")["outage_mrid_str"].nunique()
n_panel_days = county_panel["date"].nunique()
near_daily = events_per_filer[events_per_filer > 0.8 * n_panel_days]

ambient_audit = pd.DataFrame(
    [
        {
            "matched forced starts": int(matched_forced["outage_mrid_str"].nunique()),
            "ambient-temperature starts": int(ambient_events["outage_mrid_str"].nunique()),
            "ambient share": round(
                ambient_events["outage_mrid_str"].nunique()
                / matched_forced["outage_mrid_str"].nunique(),
                3,
            ),
            "ambient filers": int(events_per_filer.size),
            "near-daily filers (80%+ of days)": int(near_daily.size),
            "their share of ambient starts": round(
                float(near_daily.sum() / events_per_filer.sum()), 3
            ),
        }
    ]
)
html_table(pl.from_pandas(ambient_audit))

1 rows × 6 columns

matched forced starts	ambient-temperature starts	ambient share	ambient filers	near-daily filers (80%+ of days)	their share of ambient starts
152,080	77,021	0.506	189	20	0.437

AMBIENT_DUE_TO_TEMP is the largest forced-outage reason, about half of the matched forced starts in the panel. The audit row reframes it: 20 of the 189 filers file on more than 80% of the panel’s days - a five-year bar - and still carry 44% of the ambient starts, so for that roster an ambient filing is a routine daily derate, not an episodic breakdown. Notebook 02 section 4.3 maps the wider AMBIENT family.

2. Which Generators, in Which Counties

ambient_by_county = (
    ambient_events.groupby(["region", "county"])
    .agg(
        starts=("outage_mrid_str", "nunique"),
        mw_sum=("max_curtailment_mw", "sum"),
        generators=("resource_name", "nunique"),
    )
    .sort_values("starts")
    .tail(12)
    .reset_index()
)
county_labels = (ambient_by_county["county"] + " (" + ambient_by_county["region"] + ")").tolist()

fig = figure(
    title="Ambient-Temperature Outage Starts by County",
    y_range=county_labels,
    x_axis_label="Ambient-Temperature Starts",
    height=380,
    sizing_mode="stretch_width",
    toolbar_location="above",
)
county_source = ColumnDataSource(
    {
        "county": county_labels,
        "starts": ambient_by_county["starts"].tolist(),
        "mw_sum": ambient_by_county["mw_sum"].tolist(),
        "generators": ambient_by_county["generators"].tolist(),
    }
)
fig.hbar(y="county", right="starts", height=0.6, color=SERIES[0], source=county_source)
fig.add_tools(
    HoverTool(tooltips=[("county", "@county"), ("starts", "@starts{0,0}"),
              ("cumulative MW", "@mw_sum{0,0}"), ("generators", "@generators")])
)
show(style(fig))

county_weather_daily = county_panel[
    ["date", "region", "county", "cdd65", "temperature_f_mean"]
]
ambient_with_weather = ambient_events.merge(
    county_weather_daily, on=["date", "region", "county"], how="left"
)


def derate_temperature_correlation(derate_mw):
    """Correlate one filer's declared derate MW with its county temperature.

    A unit that filed once, or that files the same MW every time, has no
    correlation to report; returning NaN keeps those groups out of the
    ranking instead of dividing by a zero standard deviation.
    """
    paired = pd.concat(
        [derate_mw, ambient_with_weather.loc[derate_mw.index, "temperature_f_mean"]],
        axis=1,
    ).dropna()
    if len(paired) < 2 or paired.nunique().min() < 2:
        return np.nan
    return paired.iloc[:, 0].corr(paired.iloc[:, 1])


filer_depth = (
    ambient_with_weather.groupby(["resource_name", "region", "county"])
    .agg(
        starts=("outage_mrid_str", "nunique"),
        max_mw=("max_curtailment_mw", "max"),
        corr_mw_temperature=("max_curtailment_mw", derate_temperature_correlation),
    )
    .reset_index()
)
filer_depth = (
    filer_depth[filer_depth["starts"] >= 200]
    .sort_values("corr_mw_temperature", ascending=False)
    .round({"max_mw": 1, "corr_mw_temperature": 3})
)
html_table(
    pl.from_pandas(
        filer_depth.rename(
            columns={
                "resource_name": "generator",
                "starts": "starts",
                "max_mw": "max derate (MW)",
                "corr_mw_temperature": "corr of derate MW with county temperature",
            }
        )
    ),
    max_rows=40,
)

68 rows × 6 columns (first 40 of 68)

generator	region	county	starts	max derate (MW)	corr of derate MW with county temperature
GWF HENRIETTA PEAKER PLANT UNIT 2	ZP26	Kings	536	10.30	0.936
GWF HENRIETTA PEAKER PLANT UNIT 1	ZP26	Kings	536	10.00	0.934
Mammoth G2	ZP26	Mono	1,093	10.50	0.668
McGrath Beach Peaker	SP15	Ventura	1,504	48.60	0.566
Pio Pico Unit 1	SP15	San Diego	1,846	111.30	0.528
LIVE OAK LIMITED	ZP26	Kern	1,325	49.70	0.490
GATEWAY GENERATING STATION	NP15	Contra Costa	1,686	585.00	0.470
Pio Pico Unit 3	SP15	San Diego	1,856	112.00	0.439
MCKITTRICK LIMITED	ZP26	Kern	1,165	47.50	0.431
Hanford Peaker Plant	ZP26	Kings	548	98.50	0.421
Midway Peaking Aggregate	ZP26	Fresno	898	119.90	0.371
Pio Pico Unit 2	SP15	San Diego	1,849	112.70	0.356
Mira Loma Peaker	SP15	San Bernardino	1,389	47.20	0.355
MIDWAY SUNSET COGENERATION PLANT	ZP26	Kern	1,658	248.00	0.341
Panoche Peaker	ZP26	Fresno	743	50.00	0.324
Tracy Combined Cycle Power Plant	NP15	San Joaquin	550	336.00	0.261
Sunrise Power Project AGGREGATE II	ZP26	Kern	1,561	586.00	0.229
Bear Mountain Limited	ZP26	Kern	1,352	49.20	0.213
CHALK CLIFF LIMITED	ZP26	Kern	1,328	48.70	0.211
Mountainview Gen Sta. Unit 3	SP15	San Bernardino	1,225	555.00	0.211
Marsh Landing 3	NP15	Contra Costa	1,801	209.00	0.209
OTAY MESA ENERGY CENTER	SP15	San Diego	1,654	603.70	0.198
Marsh Landing 1	NP15	Contra Costa	1,737	204.20	0.192
Mariposa Energy	NP15	Alameda	496	198.00	0.191
Mountainview Gen Sta. Unit 4	SP15	San Bernardino	1,417	555.00	0.188
DOUBLE "C" LIMITED	ZP26	Kern	1,306	52.20	0.186
Marsh Landing 2	NP15	Contra Costa	1,719	202.70	0.186
Stanton 2	SP15	Orange	737	49.60	0.160
Marsh Landing 4	NP15	Contra Costa	1,726	204.30	0.152
Sentinel Unit 3	SP15	Riverside	590	105.70	0.146
Los Medanos Energy Center AGGREGATE	NP15	Contra Costa	1,606	389.00	0.143
HIGH SIERRA LIMITED	ZP26	Kern	1,312	52.40	0.120
La Paloma Generating Plant Unit #1	ZP26	Kern	1,645	267.00	0.116
Russell City Energy Center	NP15	Alameda	1,365	615.20	0.111
PANOCHE ENERGY CENTER (Aggregated)	ZP26	Fresno	1,069	417.00	0.098
HIGH DESERT POWER PROJECT AGGREGATE	SP15	San Bernardino	728	850.00	0.080
BADGER CREEK LIMITED	ZP26	Kern	1,272	48.10	0.074
LA PALOMA GENERATING PLANT, UNIT #4	ZP26	Kern	1,640	267.00	0.073
KERN FRONT LIMITED	ZP26	Kern	1,306	52.40	0.065
King City Energy Center, Unit 1	ZP26	Monterey	735	44.60	0.064

The roster is Kern’s Central Valley fleet, San Diego’s peakers and combined-cycle plants, Contra Costa’s CCGTs, Alameda’s Russell City, San Bernardino’s Mira Loma and Mountainview, plus Mammoth’s geothermal in Mono County - gas turbines and combined-cycle plants almost throughout.

3. The Heat Signal Lives in the Depth, Not the Count

County-day CDD65 barely correlates with the count of ambient filings - the near-daily filers file regardless. What temperature moves is how many MW each filing declares lost. Pooled first, over county-days with at least one ambient filing:

ambient_county_daily = (
    ambient_with_weather.groupby(["date", "region", "county"])
    .agg(
        ambient_starts=("outage_mrid_str", "nunique"),
        ambient_mw=("max_curtailment_mw", "sum"),
        cdd65=("cdd65", "first"),
    )
    .reset_index()
)
pooled_depth = pd.DataFrame(
    [
        {
            "county-day quantity": "ambient filing count",
            "corr with CDD65": round(
                ambient_county_daily["cdd65"].corr(ambient_county_daily["ambient_starts"]), 3
            ),
        },
        {
            "county-day quantity": "summed ambient derate MW",
            "corr with CDD65": round(
                ambient_county_daily["cdd65"].corr(ambient_county_daily["ambient_mw"]), 3
            ),
        },
    ]
)
html_table(pl.from_pandas(pooled_depth))

2 rows × 2 columns

county-day quantity	corr with CDD65
ambient filing count	0.123
summed ambient derate MW	0.152

Then the same relationship one generator at a time, for filers with at least 200 ambient starts:

roster = filer_depth.sort_values("corr_mw_temperature")
roster_labels = (roster["resource_name"] + " (" + roster["county"] + ")").tolist()
roster_source = ColumnDataSource(
    {
        "generator": roster_labels,
        "correlation": roster["corr_mw_temperature"].tolist(),
        "starts": roster["starts"].tolist(),
        "color": [
            SERIES[0] if value > 0 else SERIES[7]
            for value in roster["corr_mw_temperature"]
        ],
    }
)
fig = figure(
    title="Derate Depth vs County Temperature by Generator",
    y_range=roster_labels,
    x_axis_label="Correlation of Filed Derate MW With County Temperature",
    height=640,
    sizing_mode="stretch_width",
    toolbar_location="above",
)
fig.hbar(y="generator", right="correlation", height=0.7, color="color",
         source=roster_source)
fig.add_layout(Span(location=0, dimension="height", line_color="black", line_width=1))
fig.add_tools(
    HoverTool(tooltips=[("generator", "@generator"),
              ("correlation", "@correlation{0.000}"), ("starts", "@starts{0,0}")])
)
show(style(fig))

The two strongest bars, point by point - each dot one of that unit’s ambient events:

exemplars = filer_depth[filer_depth["starts"] >= 300].head(2)

fig = figure(
    title="Derate Depth Against County Temperature, Two Exemplar Units",
    x_axis_label="County Daily Mean Temperature (F)",
    y_axis_label="Filed Derate (MW)",
    height=380,
    sizing_mode="stretch_width",
    toolbar_location="above",
)
renderers = []
for color, unit in zip(SERIES, exemplars["resource_name"]):
    unit_events = ambient_with_weather[
        ambient_with_weather["resource_name"] == unit
    ]
    renderers.append(
        fig.scatter(
            unit_events["temperature_f_mean"].to_numpy(),
            unit_events["max_curtailment_mw"].to_numpy(),
            color=color, size=6, alpha=0.5, legend_label=unit, name=unit,
        )
    )
fig.add_tools(
    HoverTool(renderers=renderers, tooltips=[("unit", "$name"),
              ("temperature (F)", "@x{0.0}"), ("derate MW", "@y{0.0}")])
)
fig.legend.location = "top_left"
show(style(fig))

• Per unit, depth tracks temperature closely. The twin Henrietta peaker units in Kings County track county temperature at 0.936 and 0.934, Mammoth’s geothermal unit at 0.67, and a band of Valley and San Diego peakers between 0.4 and 0.6. The big CCGTs are weaker over five years than over one (Gateway 0.47, Marsh Landing 0.15 to 0.21): their filings mix seasonal maintenance into the temperature signal.

• Pooled, it washes out. County-day CDD65 reaches only 0.12 against ambient filing counts and 0.15 against summed derate MW.

• A few units deepen in cold instead. San Diego’s MMC Escondido aggregate (-0.43) and El Cajon (-0.27) file their largest derates on cold days; Alameda’s two GT units flatten to about -0.05 over five years. The mechanism is not settled here.

4. A Reason-Aware Count Target Does Not Rescue Weather

County-day correlations of the weather features against every outcome measure, with no-filing days counted as zero ambient MW:

ambient_mw_daily = (
    ambient_events.groupby(["date", "region", "county"])["max_curtailment_mw"]
    .sum()
    .rename("ambient_mw_sum")
    .reset_index()
)
correlation_frame = county_panel.merge(
    ambient_mw_daily, on=["date", "region", "county"], how="left"
)
correlation_frame["ambient_mw_sum"] = correlation_frame["ambient_mw_sum"].fillna(0.0)

CORRELATION_VARIABLES = {
    "CDD65": "cdd65",
    "HDD65": "hdd65",
    "temperature dispersion": "temperature_f_std",
    "all forced starts": "Y",
    "ambient starts": "ambient_temperature_start_count",
    "plant-trouble starts": "plant_trouble_start_count",
    "forced MW": "forced_start_mw_sum",
    "ambient MW": "ambient_mw_sum",
}
correlation_matrix = correlation_frame[list(CORRELATION_VARIABLES.values())].corr()
correlation_square = pl.DataFrame(
    {"variable": list(CORRELATION_VARIABLES)}
    | {
        label: correlation_matrix[code].tolist()
        for label, code in CORRELATION_VARIABLES.items()
    }
)
corr_heatmap(correlation_square, "variable",
             "County-Day Correlations: Weather Against the Outcome Family")

CDD65 tops out at 0.08 against every outcome, HDD65 is mildly negative against all five, and only temperature dispersion registers (0.10 against all forced starts, 0.14-0.21 against the ambient measures). CDD65 against HDD65 at -0.55 is definitional; ambient starts against all forced starts at 0.89 follows from ambient filings being half of \(Y\). The block tests make the same point on the model’s own likelihood:

ambient_block_tests = likelihood_ratio_block_tests(
    county_panel,
    feature_cols=COUNTY_DAILY_FEATURES,
    feature_blocks=county_feature_blocks,
    target_col="ambient_temperature_start_count",
    high_count_quantile=0.90,
    maxiter=300,
)
trouble_block_tests = likelihood_ratio_block_tests(
    county_panel,
    feature_cols=COUNTY_DAILY_FEATURES,
    feature_blocks=county_feature_blocks,
    target_col="plant_trouble_start_count",
    high_count_quantile=0.90,
    maxiter=300,
)
block_lr_by_target = (
    county_block_tests[["block", "lr_chi2"]]
    .rename(columns={"lr_chi2": "all forced starts (Y)"})
    .merge(
        ambient_block_tests[["block", "lr_chi2"]].rename(
            columns={"lr_chi2": "ambient-temperature starts"}
        ),
        on="block",
    )
    .merge(
        trouble_block_tests[["block", "lr_chi2"]].rename(
            columns={"lr_chi2": "plant-trouble starts"}
        ),
        on="block",
    )
    .round(1)
)
html_table(pl.from_pandas(block_lr_by_target))

6 rows × 4 columns

block	all forced starts (Y)	ambient-temperature starts	plant-trouble starts
outage state	31,615.90	21,725.40	2,989.80
size	5,570.60	2,502.30	2,770.20
zone effects	1,665.40	1,301.00	204.90
calendar	168.40	176.30	182.10
weather	47.40	845.10	80.80
load	18.70	115.00	59.50

Block importance (lr_chi2) for three count targets, same features and split throughout. The ambient-temperature count is persistence again: its weather block rises to 845 (from 47 on total \(Y\)) while persistence rises to 21,725, because the near-daily filers file tomorrow whenever they filed today. Plant trouble is a fleet story - size and persistence roughly tied, weather real but small.

Findings

• Half of matched forced starts are routine ambient-temperature capacity filings rather than equipment failures.

• The heat signal lives in filed derate MW per unit, not in county-day counts: the temperature response is a depth phenomenon.

• Re-cutting the count target by reason does not rescue weather - persistence still dominates every reason-specific target.

• The next weather-outage model is unit-level derate depth: filed MW as a function of temperature per generator - the physical supply channel behind weather-price coincidence.