The case of algebraic curves was already implicit in the Riemann-Roch theorem. Not explicit, because for a curve C the coherent groups Hi vanish for i > 1; but H1 does enter implicitly. In fact the basic relation of the theorem involved L(D) and L(K-D), where D is a divisor and K a divisor of the canonical class. After Serre we recognise L(K-D) as the dimension of H1(D), where now D means the line bundle determined by the divisor D. That is, Serre duality in this case relates groups H0(D) and H1(KD*), and we are reading off dimensions (notation: K is the canonical line bundle, D* is the dual line bundle, and juxtaposition is tensor product of line bundles).
In this formulation the theorem can be rearranged to read as a calculation of the Euler characteristic h0(D) - h1(D), in terms of the genus of the curve and the degree of D. It is this expression that can be generalised to higher dimensions.
Serre duality of curves is therefore something very classical; but it has interesting light to cast. For example, in Riemann surface theory, the deformation theory of complex structures is studied classically by means of quadratic differentials (namely sections of L(K2). The theory of Kodaira and Spencer identifies deformations via H1(T), where T is the tangent bundle K*. The duality shows why these approaches coincide.
The origin of the theory lay in Serre's earlier work on several complex variables. In the generalisation of Alexandre Grothendieck, Serre duality becomes a part of coherent duality in a much broader setting. While the role of K above in Serre duality is the determinant line bundle of the cotangent bundle, in generality K cannot be a single sheaf in the absence of some hypothesis of non-singularity on V. The formulation in full generality uses a derived category and Ext functors, to allow for the fact that K is now represented by a chain complex of sheaves. The statement of the theorem is recognisably Serre's, however.