You simply have to try. You could use standard linear filters (see the help and documentation for filtfilt or conv), you might use median-type-filters (see: medfilt1, medfilt2), you might get something out of Lee's sigma-filters (see wiener2), you might want a filter with some combined characteristics of these 3 (look for bilateral filters or susan-filters on the file exchange) you might want to use loess or lowess (see smooth), you might get good results with a spline-approximation to your data (see csaps or spap2). There are busloads of additional filters you might use with different characteristics.
But since you ask this question it is clear that you need to learn to understand the characteristics of the variations you want to keep in your data (daily/diurnal variations to be kept, or do you want to remove them too?) and the disturbances you want to suppress. I guess (hard stress on that this is guessing now) that your data have noise that has annoyingly long tails in their distribution - meaning that you have large noise-spikes that are rather intermittent. If that's the case you need a filter that is robust to such spikes. When I try to understand the noise-distribution of data I'm looking at I typically do something like this:
d_med = medfilt1(d,7); % or 5 or 9...
subplot(2,1,1)
hist(d-d_med,-100:100); % or whatever range is suitable for your data
subplot(2,1,2)
hist((d-d_med)./d_med,-10:0.1:10); % or whatever range is suitable for your data
That way I get some sense of how much the data scatters around the local medians and has been a rather useful tool to get a feel for the noise-distribution when the ideal noise-free variations are not that sharp.
HTH
