Do temperature features improve solar-output estimates?

Question and scope

The target is mean AC power across reporting inverters at a timestamp, estimated from weather observed at that same time. This is retrospective, contemporaneous estimation. It is not next-day forecasting, panel diagnosis or validation of Solar Revitalizer™ routing.

Method in brief

• Aggregate generation to mean inverter AC power per plant and timestamp; join weather one-to-one.
• Retain complete 22-inverter coverage, finite matched weather and irradiation above 0.01 in native source units. This sample restriction can introduce selection bias.
• Keep complete dates in chronological training, validation and final-test partitions. Fit each plant separately.
• Compare a linear-through-origin irradiation baseline, quadratic irradiation ridge regression, and ridge regression with temperature interactions. Choose regularization on validation MAE, then refit using training plus validation.
• Exclude contemporaneous DC power, yield counters and inverter identifiers from model features. Constrain predictions to nonnegative values.

What the comparison shows

Both flexible models reduced final-week error relative to the linear irradiation baseline. The temperature-interaction model gave little additional improvement for Plant 1 and had higher error than the quadratic model for Plant 2.

This is evidence for checking complexity against a useful baseline in this dataset and time window. It does not establish a causal temperature effect or broad transfer to other plants.

Inspect the evaluation code

The excerpt shows the chronological data split and training-only normalization used to calculate model error. Dataset credit: Ani Kannal, Solar Power Generation Data.

# Chronological evaluation and training-set error normalization.
# Input preparation and model fitting are described in the case methods.

train_dates, val_dates, test_dates = all_dates[:20], all_dates[20:27], all_dates[27:]
full = merged.inverter_count.eq(expected_inverters)
weather_valid = np.isfinite(merged[["IRRADIATION", "AMBIENT_TEMPERATURE", "MODULE_TEMPERATURE"]].to_numpy()).all(axis=1)
daylight = merged.IRRADIATION.gt(0.01)
clean = merged.loc[full & weather_valid & daylight].copy().sort_values("DATE_TIME")
clean["date"] = clean.DATE_TIME.dt.date
train = clean.loc[clean.date.isin(train_dates)].copy()
val = clean.loc[clean.date.isin(val_dates)].copy()
test = clean.loc[clean.date.isin(test_dates)].copy()
assert train.DATE_TIME.max() < val.DATE_TIME.min() < test.DATE_TIME.min()
assert not set(train.date) & set(test.date)
if min(len(train), len(val), len(test)) < 30:
    raise ValueError("Insufficient retained observations in a time partition")
xtrain, xval = features(train), features(val)
ytrain, yval = train.AC_POWER.to_numpy(), val.AC_POWER.to_numpy()
scale = float(np.quantile(ytrain, 0.95))

def metrics(y, pred, scale):
    return {"mae_reported_kw": float(np.mean(np.abs(y-pred))),
            "rmse_reported_kw": float(np.sqrt(np.mean((y-pred)**2))),
            "nmae_percent_train_p95": float(100*np.mean(np.abs(y-pred))/scale)}

Limits and next tests

• Two plants and one seven-day final evaluation window limit generalization.
• nMAE is normalized by each plant’s training AC-power 95th percentile, not rated capacity. Power values use the dataset’s reported units, which have not been independently verified.
• Complete-case and daylight filtering can change the population represented.
• Independent plants and periods, source metadata checks, distribution-shift tests and uncertainty calibration remain future work.
• Module-level disposition research requires separate condition, safety, performance and reference-outcome data.