The forecast works based on the clear-sky solar irradiance onto your PV plane. The orientation has to be as accurate as you can make it, for both inclination and azimuth, and the total solar power peak figure has to correct too.
Each plane has to be set up with its own figures, combining into one ‘virtual’ surface is always going to produce inaccurate values. I have two planes (east and west) and the total generation curve produced cannot be replicated using a single plane ‘equivalent’.
HA allows you to set up more than one instance of forecast.solar, thus ideally one for each plane. The energy dashboard very nicely adds up all the integrations and shows the total (dotted) line.
forecast.solar allows only so many free API calls, but normally HA will only call each integration once per hour, and the last time I looked the ‘free’ account allows for 12 per hour.
So, set up two integrations (call them east and west or whatever they are) one with declination 42, power set to max watts peak for the 9 panels, and the other with declination 35 degrees and max watts peak for 5 panels. If you have a system of 14 panels, and it has a peak power of say 4.2 kW, one side will be 9/14th of 4200 peak watts, and the other 5/14th total power (assuming your panels are identical).
As for azimuth, this is the direction the panels face, with 90deg being due east, 180 due south, and 270 due west. A quick and dirty way to check this out is to use something like www.suncalc.org to zoom in on your roof, play with the time bar (top of map), and when the sun lines up with facing your panels note the azimuth. Alternatively line the sun up with the ridge along the panel edge and add/subtract 90 degrees.
Now you should have an ‘accurate’ forecast. After that, the forecast is designed to deal with the weather. On a nice sunny day, you should see a nice full curve, and this should match quite well to your actuals. Expect differences due to shading, early/late damping, and the issue of the 1-hour shift that is the subject of much debate above. You may, in time, be able to adjust the morning/evening to better fit forecast-actual by adding the damping factor. I use 0.4 myself to compensate for shading from houses etc on the horizon so as to flatten the morning/evening forecast.
The total inverter power is just there to stop the system forecasting say 5kW in high summer when your inverter can only actually produce 3kW, but with split planes like this as I have, my east/west arrangement means that my inverter limit is never reached.
And then there is cloud and rain. Yes, it is a forecast, so it is going to be out a bit. Also I find that the forecast does not match too well for planes that are not facing south - the production curve becomes quite lopsided. Having two equally sized planes I have managed to balance out the damping so that the error on one plane is typically cancelled out by the error in the other. In your case, the 9/14 and 5/14 split means that the overall combined error may be more likely to show at one end of the day.
However, having put a lot of time and effort into this, I now collect my own forecast using the API calls directly and processing the result myself. Today my own graph shows (for forecast and for actual power)
And here is the HA energy dashboard, for effectively the same data.
The integration calls forecast.solar once per hour, and the energy graph (dotted line) is updated either hourly or three hourly. At such times you may see the line go up or down dramatically - this is a side effect of various things being updated and then pulled together for the graph. It usually settles out again although I do not believe that the historical part of the curve is accurate.
There are differences, for various reasons, but over time the forecast is mostly good to very good.