Hi,
I have been trying to devise a solar power generation forecast using weather forecast information. I have power generation data for a particular PV system for several time periods during the past year. So far I have used a neural network, but it isn't working out as nicely as I would like it to. The relevant parameters appear to be cloudiness, humidity, precipitation, time of the day, solar elevation and direct irradiance of the PV system assuming clear sky conditions.
The problem is that the neural network is not learning some of the things that I know for sure. There is no power output at night. I have mitigated this problem by reducing the analysis to the hours between sunrise and sunset. Power output during daytime hours is always > 0, even assuming the worst possible weather conditions. I have mitigated this problem by calculating the minimum and maximum power output at each hour and refining the final result by confining it between these two boundaries. It seems like a dirty fix though and it isn't working very well because different seasons will have different maximum/minimum values. I can calculate direct and diffuse irradiance quite accurately for clear sky conditions. Not in absolute terms, but at least in relative terms. It would be nice if I could tell the system these and ask it to find out how these are modified by the given weather forecast data. Unfortunately, I don't have irradiance data to work with.
I know the tendencies of the influences of weather data. More clouds > less direct irradiance, more humidity > more scattering and thus less direct irradiance – but possibly more diffuse irradiance (or less). Higher precipitation > darker sky > less direct or diffuse irradiance. I just don't know the exact relationships.
Is there a way to tell a system what I already know rather than just throwing data at it and letting it figure everything out by itself?
Regards, Malte
Best Answer