GRIDCAST

Research overview

GridCast is a reproducible historical research platform for California electricity demand and NP15 day-ahead price forecasting. It presents observed CAISO data, direct-horizon forecasting experiments, evaluation artifacts, and a separate interactive dashboard.

The executed window covers 2024–2025. Results are historical research findings, not live operational claims or financial advice.

Problem definition

The task is to forecast hourly CAISO actual system demand and NP15 day-ahead total LMP at 1, 6, 24, 48, and 168 elapsed-hour horizons. The issue-time availability contract is explicit so target outcomes do not enter features before their assumed publication delay.

Data sources and preprocessing

Primary inputs are CAISO OASIS SLD_FCST/ACTUAL observations for CA ISO-TAC and PRC_LMP/DAM total LMP observations for TH_NP15_GEN-APND. The pipeline validates hour-long UTC intervals, rejects duplicates, preserves negative prices, and leaves one demand gap un-imputed.

Demand contains 17,543 observed hours; price contains 17,544, including 559 negative price observations. Raw request ZIPs and SHA-256 manifests preserve acquisition provenance.

Figure — Daily CAISO actual system demand and NP15 day-ahead total LMP from the verified historical export.

Feature engineering

Features use delayed historical lags at 0, 1, 2, 3, 6, 12, 24, 48, 72, and 168 hours, plus 24/168-hour rolling statistics, a weekly seasonal anchor, and calendar terms. Local time uses America/Los_Angeles while UTC remains canonical.

Historical weather reanalysis is limited to retrospective exploratory analysis and is excluded from model predictors.

Figure — Validation-only permutation importance for the 24-hour XGBoost diagnostic; this is not causal attribution.

Forecasting methodology

Baselines are persistence, seasonal naive, historical mean, and local weekday/hour mean. Fitted candidates are ridge regression, seasonal autoregression, XGBoost, LightGBM, and a compact CPU PyTorch GRU.

Quarterly expanding fits use training-only scaling. Candidate identity is selected from two validation quarters before held-out evaluation; every candidate remains in the result registry.

Model training and evaluation

Validation issues occupy January–June 2025. Held-out evaluation uses July–December 2025 with quarterly expanding refits, a 720-hour calibration block, and explicit label-availability purges.

The 24-hour validation-selected demand model is LightGBM; the corresponding price model is ridge with alpha 100. Each has 4,368 held-out eligible origins.

Figure — Held-out MAE across eligible origins for each candidate model and direct forecast horizon.

Results

For the verified 24-hour held-out experiment, demand LightGBM has MAE 1,191.45 MW, RMSE 1,638.88 MW, and empirical 90% interval coverage of 93.2%.

Price ridge-100 has MAE 6.03 USD/MWh, RMSE 7.79 USD/MWh, and empirical 90% interval coverage of 90.7%. These results should be read with the documented publication-vintage and single-hub limitations.

Figure — Observed values, direct 24-hour forecasts, and pre-fold 90% split-conformal intervals from the held-out replay.

Prediction uncertainty

Symmetric 90% split-conformal intervals use pre-fold calibration residuals and finite-sample rank correction. Serial dependence and distribution shift mean nominal coverage is evaluated empirically rather than guaranteed.

Figure — Held-out empirical interval coverage across forecasting horizons, reported against the documented nominal 90% target.

Model interpretation

Permutation importance uses validation data only, three repeats, a 12-hour sample stride, and MAE scoring. Correlated lags can share or mask importance; it is not causal attribution.

Limitations

CAISO extracts may include revisions unavailable at the historical issue time, so this is availability-aware research rather than a strict point-in-time vintage dataset. The price target is one NP15 hub; the work does not claim statewide retail-price forecasting.

Conclusions

GridCast documents a reproducible, availability-aware workflow for evaluating California demand and price forecasts. Results are bounded by the stated data-vintage, market-hub, and historical-window constraints.

Reproducibility

The separate repository preserves acquisition, preparation, training, report generation, model registry, tests, and a standalone dashboard. Reproduce from raw CAISO downloads for exact-input comparisons.