Predicting solar output

The Case:

According to a recent analysis by Bloomberg 50% of the world’s energy will come from solar and wind by 2050. The output of these energy sources is mostly dependent on weather conditions.

This means that solar panels and wind turbines cannot be turned on or off on demand and any reduction in their output needs to be compensated by other assets such as gas power plants. Short term fluctuations in solar output can thus cause outages and destabilize the power grid.

The Challenge:

The world's weather system is one of the most complex systems known to man. Minor changes in the composition of the atmosphere, direction of currents or even the conditions on the surface of the Sun can have an impact on the weather here on Earth. This makes short term weather forecasting extremely difficult.

To make matters worse climate change makes our old weather models completely obsolete.

The Solution:

To increase the reliability of solar panels and to reduce the aforementioned risk we’ve developed a novel algorithm to predict PV panel output. The algorithm uses a novel neural network architecture that takes in both history production data as well as weather data to make accurate predictions of the output for the next 24 hours.

The Outcome:

We've developed this algorithm using data from Ampiron and TenneT from the last 4 years of their operations in Germany. The error rate that our model managed to achieve on this dataset is 3.22%. The error rate of a local DSO’s forecasts was reported at 5%.