In general you use smoothing functions to remove the noise, not the trend (which in general is what we want to identify or study). I think the best is by trial and error. Smooth with increasing smoothing parameters and check the data by eye, once you don't have that many short spikes but a quite "smooth" curve, than it's done. Be careful to don't smooth too much, otherwise you will remove interesting signals. important issue is also what method to use for smoothing, moving window? LO(W)ESS?

In my data there are  trend and seasonality and i want to remove trend only from the daily data of water consumption data.

maybe you could use fast fourier transform or wavelets to get the trends at different time-scales. Or you smooth on a scale of 2-3 days.