It is more difficult to determine experimentally the structure of membrane protein than that of soluble protein. In order to develop a reliable model for predicting protein structure, it is necessary to perform model optimisation on the largest (representative) set of membrane proteins of known structures with mutual similarities below 30%. Existing algorithms for selection of representative sets of membrane proteins of alpha-type do not use information about the complexity of structure, although it is expected that the models will be more reliable in prediction if they are developed on a more complex protein structures. Consequently, the concept of a random model based on two secondary structures was introduced and noticed that the formula for estimation of its accuracy is connected with the complexity of structure. Then, the concepts of the binomial and segmental random model were introduced, as well as formulae for the number of possible realizations of protein model structure, showing the analogy with entropy, were developed. The segmental model is best suited to the membrane protein structure in which several adjacent amino acids form segments having regular secondary structure of alpha type. The number of realizations of model structure of segmental random model is related to the complexity of structure showing a significant correlation with the number of transmembrane segments. It is involved in developed algorithms, and the best one is based on original analysis of the number of common neighbours between proteins in the initial set. Applications of these algorithms to databases of membrane proteins of known structures produce larger representative sets of structures which are significantly more complex than those published in the literature.