The problem of estimating a probability distribution from a closely timed sample? I have a dataset, which is the hourly mean wind speed of each day(24 points each day), for 20 years. And I'm planning for using this dataset to estimate the probability distribution of hourly mean wind speed
The problem is that, clearly, the wind speed between consecutive hours are correlated, the speed at 5:00 wouldn't be too far from at 6:00.
So, will there be problem using 24 points each day (insteand of 2 points each day) to estimate the probability density distribution? 

My thoughts:
Why I'm asking this? If I use the result of rolling a dice 500 times, the results between each rolling are independent. And that is totally ok to use it to model the probability of dice-rolling. 
But the hourly wind speed data are not the same, the data are correlated. And I'm thinking will this difference affect my modelling result (e.g. PDF)?
 A: A parametric (or empirical) probability distribution is not necessarily the best way to model what is essentially a random process.  The broader question you should be asking is, "what am I trying to do?  What is the goal?"
For example, if you are interested in estimation, you may be saying, "I want to know the average historical wind speed."  In this case, you don't need to fit a distribution to your data.  If you say, "I want to know the true expected value of the wind speed with 95% confidence," then that would still not require much in the way of model fitting:  you would presumably make some assumptions about the standard error and maybe get a conservative interval estimate from that.
But if your interest is in prediction, then none of the above applies:  instead, you will use a time series to model the process, taking into account seasonality and autocorrelation.
The point I am making here is that before we apply the tools of statistical inference to data, we have to state with precision and clarity the research goal.  Don't just start modeling something without having a properly specified purpose in mind.
A: The autocorrelation in wind speed and direction does not affect your ability to take the mean (through there are separate issues of whether you take direction into account: if the wind blows at $10$ km/h from the west and then at $10$ km/h from the east, is the average $0$ km/h or $10$ km/h?)
Where the autocorrelation and lack of independence do have an effect is on the uncertainty associated with the sample mean: the standard error of the sample mean will be larger with autocorrelated data, so you may need a longer sample to narrow the error.  
That should not be a particular worry for you, given the $20$ years of your sample is and the comparatively short-term effects of wind autocorrelation: it might have been different if for example your sample had been taken every second for two days (a similarly sized data set, but not giving so much useful information).  
A: Why are the hourly wind observation correlated? Did you try to use some tests for correlation?
Nevertheless, you could take the first differences in order to "remove" correlation and then estimate the PDF for the differences. 
A: After playing a while with my data, I find the plot would be informative in answering this question
The 1st is 2 points/day (0, 1200), the 2nd is 24 points/day (for each hour), time range is 1973 ~ 2014


The 3rd is 2 points/day (0, 1200), the 4th is 24 points/day (for each hour), time range is 1995 ~ 2014


So, from the plot, I'd say the time correlation is negligible.
