As a first step, the settings have to be absolutely correct. My solar panels face east-north-east and west-south-west, so the HA integration azimuth figures used are 75 and 255.
The elevation should be measured with a clinometer or by depth-height to calculate an angle. Zero is flat, my UK house roof is 35 and quite shallow. Free standing ideal is I think closer to 50, so 25 maybe incorrect (and the flatter they are the closer the two array solar peaks will be).
The basic solar pv model is accurate for South facing arrays, but becomes increasingly unreliable away from due South. North of East and West seems to be particularly difficult since the sun faces the panels near dawn/dusk, when atmosphere and horizon shading are most prevalent. Solar noon gives the best sun and estimates, but the sun is glancing incident on the array.
The weather is unpredictable. Cloud will reduce solar pv to about 15% of clear sky.
My forecast for the past hour was 600 Wh, I got 200. On a clear, sunny day, outside of December to January, I get a good match.
Personally, I don’t see that the 'next hour forecast is accurate for my practical use.
In time, based on either the solar forecast, the’ clear sky’ estimates, and/or my own ‘best day’ history, used together with my local weather forecast, I can see an opportunity to develop a better-fit next-hour estimate by down-rating based on cloud cover forecast.
My cloud cover is 100%. All day. I expect about 1.8 kwh today, not the 5 kwh forecast.
To try and answer your other question, only forecasting will give you an idea of what the solar pv will be in the next hour.
However, the statistics platform can offer averages and other figures, sampled over a given time period. Using a linear average sampled over 1hour, periods of passing cloud for 5-10 minutes can be “averaged out”. This is very much back-casting, and assumes that the next hour will be about the same as the average of the last hour.
I am most certainly not a statistics wizard, so I continue to explore and experiment from time to time to see if I can get a better model. A simple linear average would be an easy place to start.