Obviously I'm well aware of how an energy grid works - I've been following this subject for years. You can even follow that here:
(Your URL produces an error for me - probably for others as well.)
In California, we have tracking, at 5 minute intervals, for the relative makeup of the majority of the state's grid:
Armed with information like that, and a bit of machine learning, it wouldn't be too hard to write code that looks at current usage and recent history, and a model that predicts what's likely to happen in the next 12 hours, say, and you'll be right most of the time. Particularly if you lean heavily on what you've seen in the last hour or two. That would make for a "charge when there's plenty of renewable energy sources online" algorithm that would be right most of the time.
And using an algorithm like that, across many millions of phones (keep in mind that they're not all going on one monolithic answer/schedule, each could be checking on their current local situation), could have a noticeable positive impact.
You go on to do a lot of handwaving about bad forecasting models. Yes, there are bad forecasting models out there, that are used to do stupid things - most of those are at a systemic level, affecting what kind of infrastructure states and/or municipalities choose to provide and how they chose to run that infrastructure (surely you know about the situation in Texas). I'm talking about "what does the current situation look like right this minute, and based on past experience, what is it likely to look like in the next few hours". That's a
much smaller target to hit.
Apple doesn't have to predict whether clouds are going to impact solar production over the next week. They only have to look at what is
actually going onto the grid right now, and make some educated guesses about the next few hours.