Solcast day-ahead solar forecasting accuracy improves
Three years of countless sprints and algorithms. Tens of thousands in cloud server costs.
Three years without a result. Until today.
With a release earlier this week, the Solcast day-ahead irradiance and power forecasts are now improved by 15% to 35%, after three years of trying. The feeling is a mix of elation, and relief.
Day-ahead solar forecasting motivations
Customers had been telling us that our satellite-based actuals and our cloud-tracking short-term forecasts were amazing. But they said our day-ahead forecasts were "good but not great". It turns out there's a lot of places where day-ahead forecasting is still quite important, too - even as energy markets move to shorter range and more granular periods.
Solar forecasting NWP challenges and local clouds
The challenge was, we cover the earth with clouds, irradiance and solar power at 1-2km resolution, every 5 to 15 minutes. That's a lot of data, especially when you need to look back years, and you need to constantly update the algorithm so that it knows about yesterday. And at this resolution, cloud patterns are different everywhere - so the model features that work in one place may be useless in another.
This huge improvement is now live in the Solcast API, for all users. The data you get from the Solcast API will behave the same way, and you don't need to make any changes to your API integrations. The only difference is that the accuracy is now markedly better.
How much did the day ahead solar irradiance forecasting improve?
In our testing with thousands of places around the world, across different seasons the typical improvement you will see is in the range of 15 to 35% in RMSE (Root Mean Square Error) terms. This is quite a large improvement, which we are really proud of.
Here is a typical example of the "old" (previous Solcast algorithm - i.e. results you got from the Solcast API up until yesterday) and "new" (algorithm which is now live in the Solcast API). The new and old forecasts are issued day-ahead (about 24 hours ahead), and the "obs" are the actuals/truth from the Solcast satellite-based estimated actuals. The results are similar for any time horizon from +4 hours ahead to +72 hours ahead.