The Great American Eclipse & the Monolithic Forecast
With respect to North American weather, the solar eclipse of August 21 (known as the “Great American Eclipse”) seems like ancient history. And while its impacts on sensible weather were short-lived and modest, the Great American Eclipse had a profound impact on variable energy resources, most acutely on solar power.
The eclipse was much anticipated nationally. Indeed, my birthplace of Hopkinsville, KY dubbed itself “EclipseVille” for the day, as it was not only in the path of total obscuration but also located at the point of maximum eclipse duration. The rarefied astronomical event drew some attention to an otherwise sleepy agricultural town in western Kentucky and to many other towns less well known.
Of course, the event was a great concern to utilities and power system operators across the country, particularly those in the western U.S., for which the partial-to-total obscuration would occur during the morning up-ramp in solar generation. On a clear and hot August day in California, a significant amount of morning energy demand can be displaced by generation from roof-top solar after sunrise, diminishing demand from utility-scale transmission-connected sources. But a principal worry for the California Independent System Operator (CAISO), responsible for managing the wholesale transmission of energy across the state, was the highly-amplified up-ramp that succeeds the eclipse as the moon moves away from the Sun. That is a ramp whose amplitude might be possible on a localized scale (due to a passing cloud) but generally is never seen simultaneously across hundreds of square miles and over 7000 MW of generation. Forecasting such an event is mission-critical, such that the CAISO can mitigate any supply reliability issues cost effectively.
The forecasting team at DNV GL is responsible for predicting the output of utility-scale and distributed-scale solar plants across the United States. We planned for the eclipse event five months in advance, warning forecasting clients of the eclipse effects on generation and its representation in the short-term forecast. The DNV GL Forecasting team uses multiple weather predictions from numerical weather prediction models from a variety of sources as seminal inputs to the 24/7 operational process of power prediction. We also run our own numerical weather prediction model, in multiple configurations, on an in-house Linux cluster based in DNV GL’s San Diego office. While this technology ensures high quality forecasts on any given day, days like August 21 are unique. Eclipses are not simulated by operational numerical weather prediction models, requiring the design and implementation of models that would make up for this deficiency. The result was an event forecasted so well that it improved over the CAISO’s own solar forecast for the day, as shown in the figure below for CAISO’s largest SP-15 Zone.
Figure 1: Day-Ahead Solar Power Forecast in MW for CAISO SP-15 Zone on August 21.
At the start of the eclipse, CAISO’s own forecast was off by nearly a gigawatt, and during the critical post-event hours, relative to the DNV GL forecast, CAISO’s forecast revealed an up-ramp twice as steep as reality. This event is a clear demonstration of DNV GL’s power forecasting accuracy, supported by a dedicated team of meteorologists and scientists supporting power producers, energy traders, and system operators, for which the costs of a bad forecast can total into the millions of dollars a year.
Such costs stand at odds with the common practice of relying on a single forecast provider for wind and solar generation.
Four days after the Great American Eclipse, Hurricane Harvey hit the Texas coast. Consider the consequences of entrusting your operation to a single forecast model’s prediction for this storm. While this is an extreme case, the analogy underscores that no forecast user should ever have too much trust in a single model nor in a single provider. But with few exceptions, utilities and ISOs are not seeking multiple opinions on their power forecasts, nor even making use of probabilistic forecasts, instead entrusting their operations to a single number. This practice ultimately exposes renewables to higher integration costs, revealing of an outdated modality for dealing with multiple sources of data. Some system operators like BPA are bucking this convention, requiring at least 2 forecast providers at all times and using relatively simple methods for blending them together based on recent performance.
Even with monolithic forecast guidance, alternative forecast solutions do not have to be invasive. Our internal numerical weather prediction now covers most of North America, allowing us to forecast at every wind and solar plant in the United States and Canada, and to provide both internally and externally on-demand, 7-day forecasts via ForecasterNow, with embedded situational awareness (see Figure 2).
This on-demand continental-scale delivery platform is the first of its kind in the wind and solar power forecasting market, with a goal to provide, in a non-intrusive way, that valuable second opinion. Hopefully, it meets a minimum goal of exposing users to easily-accessible alternative solutions and likely a better forecast.
Figure 2. Forecaster Now dashboard and map-based forecast page, showing real-time satellite-derived cloud during Hurricane Harvey.