DCPP Intake Temperature Forecast

8-day forecast (today plus 7 days ahead) of DCPP cooling water intake daily maximum temperature, blending three independent ocean/weather forecast models (CMEMS, Open-Meteo, and NOAA HYCOM) with NDBC buoy 46215's own water temperature via quantile regression.

Forecast Methodology

What the model does

The forecast above blends three independent forecast-capable sources - the Copernicus Marine (CMEMS) global ocean analysis-and-forecast model, the Open-Meteo Marine weather-model-based sea surface temperature forecast, and a NOAA/NCEP HYCOM ESPC-D-V02 Forecast Model Run - each combined with NDBC buoy 46215's own water temperature (extrapolated forward via its own trend + seasonal-harmonic + autoregressive model) as a second predictor, via quantile regression. The three sources' individual predictions are then averaged (equally weighted) into a single consolidated "Blend" forecast, which is what feeds this page's chart. Unlike the legacy model, there is no fixed, one-time-fit set of coefficients: the regression refits against a rolling 120-day training window every time the forecast runs, so it adapts as conditions change through the year.

Why multiple ocean/weather models instead of the buoy alone

The previous version of this page forecast the intake's daily maximum from NDBC 46215's own daily minimum reading alone. That buoy-only approach was independently backtested against this multi-source approach and found to perform substantially worse, so it has been replaced. The buoy's own reading is still used - as a second predictor alongside each ocean/weather model - since it retains real forecast-relevant information, but it is no longer the sole input.

Why three separate sources, blended

CMEMS and Open-Meteo are independent global ocean/weather analysis-and-forecast systems, each already extending several days into the future on their own. The NOAA HYCOM source is split into two roles: a historical archive (spanning back to the start of the current year) supplies the training data for its regression fit, while a separate, frozen "Forecast Model Run" (one fixed model cycle, refreshed daily, covering 8 days forward) supplies its forward-looking predictor value - since a single forecast cycle on its own has essentially no history to train against, it borrows the coefficients already fit from the HYCOM historical archive rather than fitting its own. Blending all three sources' predictions, rather than relying on any single one, is more robust to any one source being temporarily degraded or unavailable on a given day (the blend falls back gracefully to whichever sources do have data).

Why quantile regression

As with the legacy model, an ordinary (mean) regression of intake daily maximum against each source's predictor would systematically underpredict on the days that matter most - the actual hot days - since a day selected for an unusually high outcome is, almost by definition, above its own conditional average. Quantile regression targets "how hot could it plausibly get" at a chosen percentile rather than "what is typical."

Data and fitting method

Each source's regression is fit independently (iteratively reweighted least squares, the same quantile-regression method as the legacy model) against a rolling 120-day training window of paired source-predictor / DCPP-intake-maximum observations, refit on every run rather than fit once and held fixed.

Forecast skill

[Placeholder - see TODO comment in this page's source. Insert validated bias/RMSE/skill metrics for this model here once available, in the same style as the retired table this replaced, rather than leaving this unquantified indefinitely.]