The algebras known variously as left-symmetric algebras (or right-symmetric algebras), Vinberg algebras, or pre-Lie algebras are in general not power associative. These algebras occur in at least two apparently unrelated contexts. One is the study of flat affine structures, the other is the study of the algebraic structure of renormalization.

Let $[x, y, z] = (xy)z - x(yz)$ be the associator. An algebra is left-symmetric if
\begin{align}
(LSA)\qquad [x, y, z] = [y, x, z]
\end{align}
for all $x$, $y$, and $z$ (the opposite algebra of a left-symmetric algebra is right-symmetric, meaning it satisfies the identity $[x, y, z] = [x, z, y]$). 

These certainly do not have to be power associative. A simple example, taken from E. Kleinfeld's paper "Assosymmetric rings" is the following. The algebra is spanned by $x_{1}$, $x_{2}$, and $x_{3}$, with the non-zero products $x_{1}^{2} = x_{2}$ and $x_{2}x_{1} = x_{3}$. The only non-zero associator of the form $[x_{i}, x_{j}, x_{k}]$ is $[x_{1}, x_{1}, x_{1}] = x_{3}$. This shows that the algebra is left symmetric but not power associative. 

Left symmetric algebras were introduced by E.B. Vinberg (hence the terminology "Vinberg algebras") and used by him to give a sort of classification of homogeneous convex cones (the title of the English translation is "The Theory of convex homogeneous cones"). In a related context, a left invariant flat (torsion free) connection on a Lie group gives its Lie algebra a structure of a left symmetric algebra. (If $\nabla$ is an affine connection, then the product on vector fields defined by $xy = \nabla_{x}y$ satisfies $[x, y] = xy - yx$ if $\nabla$ is torsion-free, and satisfies (LSA) if $\nabla$ is moreover flat).  

Roughly contemporaneously, M. Gerstenhaber defined a product on the space of Hochschild cochains of an associative algebra and showed that makes the cochains into what he called a (graded) pre-Lie algebra. This alternative terminology reflects that for any left symmetric algebra the commutator $[x, y] = xy - yx$ satisfies the Jacobi identity, so is a Lie bracket (the condition (LSA) means that the left regular representation of the algebra is a Lie algebra homomorphism). In the context in which Gerstenhaber was working, the algebras are graded, meaning the rule of signs has to applied to all brackets, associators, etc.

There is a more sophisticated story involving operads and rooted trees and related to the work of Connes and Kreimer on the Hopf algebra of rooted trees, but I don't know the details well enough to recount it here. The key thing is that there is a product on rooted trees which makes them into a pre-Lie algebra. See the paper of Chapoton and Livernet ("Pre-Lie algebras and the rooted trees operad") and a survey by D. Manchon ("A short survey on pre-Lie algebras") for this point of view and references. For the more classical material related to cones and affine structures, in addition to Vinberg's article, some of the standard references are D. Segal's "Complete left-symmetric algebras" and J. Helmstetter's "Radical d'une algebra symetrique gauche". There are surveys by D. Burde and P. Cartier ("Vinberg algebras").