The composition of advanced composite materials aims at leveraging the different advantageous properties of their constituents. However, most of the constituents exhibit inelastic constitutive behavior of some kind. In addition to the heterogeneous mesostructure and interactions between the constituents, this inelastic behavior leads to huge computational effort. This is especially noticeable in numerical investigations of the material within the framework of multiscale modeling. Those are required for obtaining an accurate constitutive behavior on the macroscopic scale for structural simulations. By employing Machine Learning methods for computational homogenization, it is possible to reduce the computational effort required to evaluate the anisotropic, inelastic mesostructural behavior while retaining high accuracy in the representation. Generally, one unit cell with representative characteristics of the composite material is chosen as to model the heterogeneous mesostructure, the so-called Representative Volume Element (RVE). This presents a simplification of the actual composition. To resolve this within the Machine Learning (ML) based homogenization approach at hand, multiple differently constituted Statistical Volume Elements (SVE) are utilized. Thereby, it is possible to account for the naturally occurring fluctuations in the composition of the mesostructure whilst the computational effort for evaluating the constitutive model is not significantly increased compared to ML based homogenization of an RVE. Additionally, the inelastic behavior introduces history dependency to the material behavior. By considering time series data which include previous loading states as history variables, the formulation at hand is also enabled to capture this material property for the SVE. The applicability of this approach is demonstrated for a numerically generated set of heterogeneous mesostructures with different compositions and elasto-plastic constituent materials. Thereafter, the results are compared with references from direct numerical modeling using Finite Element Method.