Spatial weights would be relevant in non-homogeneous settings in which one expects the behaviour at different regions of space to be different. For instance, if there is an obstacle or a boundary, a weight that depends on the distance to the boundary would be natural in order to capture boundary effects. If the initial data is originally assumed to be concentrated at the origin, then weights involving the distance $|x|$ to the origin are also natural. Similarly, weights involving time $t$ are sometimes natural in evolution equations, particularly if one is trying to describe decay or blowup in time.

More generally, if there is a natural singular set in physical space or frequency space, then it is natural to weight one's spaces around that set. The $X^{s.b}$ spaces mentioned in Willie's answer are a good example of this in the frequency domain (and Sobolev spaces themselves reflect the privileged nature of the frequency origin for many PDE, as the zero set for the symbol of the underlying linear operator (e.g. the Laplacian)).

If one needs to prevent the solution from concentrating all its mass or energy into a ball, then Morrey or Campanato spaces are occasionally useful.

As for the frequency-based refinements to Sobolev spaces (e.g. Besov and Triebel-Lizorkin, but also Hardy spaces, BMO, BV, etc.), these are "within logarithms" of Sobolev spaces, in the sense that if the ratio between the finest and coarsest spatial scale of interest (or equivalently, the ratio between the highest and lowest frequency scale of interest) is comparable to $N$, then the ratio between a Besov or Triebel-Lizorkin norm and its Sobolev counterpart (as plotted for instance on this type diagram: http://terrytao.wordpress.com/2010/03/11/a-type-diagram-for-function-spaces/ ) is at most a power of $\log N$. Because of this, Sobolev spaces generally suffice for all "subcritical" or "non-endpoint" situations in which one does not have to contend with a logarithmic pileup of contributions from each scale. If one is working in a critical setting (which is more or less the same thing as a scale-invariant or a dimensionless setting), these refinements can often be necessary to stop the logarithmic divergences caused by such things as the failure of the endpoint Sobolev inequality, e.g. $H^{n/2}({\bf R}^n) \not \subset L^\infty({\bf R}^n)$. (In this particular case, one can sometimes replace the Sobolev space $H^{n/2}$ with the smaller Besov space $B^{n/2}_1$ to recover the endpoint embedding, although there is no free lunch here and this will likely make some other estimate in one's analysis harder to prove.)

In general, unless one is perturbing off of an existing method, one does not proceed by randomly picking function spaces and hoping that one's argument closes. Often the function spaces one ends up using are dictated by trying to directly estimate solutions (or approximations to solutions). For instance, if one is trying to establish a local well-posedness result for a semilinear evolution equation in some standard space, e.g. $H^s({\bf R}^n)$, one can try to expand the $H^s$ norm of that solution using a basic formula such as the Duhamel formula or the energy inequality. In trying to estimate the terms arising from that formula by harmonic analysis methods (e.g. Holder inequality, Sobolev embedding, etc.), one is naturally led to the need to control the solution in other norms as well. If all goes well, all the norms on the right-hand side can be controlled by what already has on the left-hand side plus the initial data, and then one has a good chance of closing an argument; if not, one often has to tweak the argument by either strengthening or weakening the norms one is trying to control, as dictated by what the harmonic analysis is telling you. The final norms one uses to close the argument often arise from a lengthy iteration of this procedure (which unfortunately is often hidden from view in the published version of the paper, which usually focuses on the final choice of spaces that worked, rather than the initial guesses which didn't quite work but needed to be perturbed into the final choice).

Ultimately, in PDE one is usually more interested in the *functions* themselves, rather than the *function spaces* (though there are exceptions, e.g. if one is taking a dynamical systems perspective, or is relying on a fixed-point theorem exploiting the global topology of the function space). The reason that function spaces appear so prominently in PDE arguments is that functions have an infinite number of degrees of freedom, and the basic physical features of such functions (e.g. amplitude, frequency, location) are not easy to define directly in a precise and rigorous fashion. Function space norms serve as mathematically rigorous proxies for these physical statistics, but in the end they are only formal tools (with the exception of some physically natural norms or norm-like quantities, such as the mass or energy) and one should really be thinking about the physical features of the solution to the PDE directly. I discuss this point at http://terrytao.wordpress.com/2010/04/02/amplitude-frequency-dynamics-for-semilinear-dispersive-equations/ in the setting of semilinear dispersive equations (but there are similar perspectives for other PDE also).