Recent work in the evolutionary computation field suggests that the implementation of the principles of modularity (functional localization of functions), repetition (multiple use of the same sub-structure) and hierarchy (recursive composition of sub-structures) could improve the evolvability of complex systems. The generation of neural networks through evolutionary algorithms should in particular benefit from an adapted use of these notions. We have consequently developed modular encoding for neural networks based on attribute grammars (MENNAG), a new encoding designed to generate the structure of neural networks and parameters with evolutionary algorithms, while explicitly enabling these three above-mentioned principles. We expressed this encoding in the formalism of attribute grammars in order to facilitate understanding and future modifications. It has been tested on two preliminary benchmark problems: cart-pole control and robotic arm control, the latter being specifically designed to evaluate the repetition capabilities of an encoding. We compared MENNAG to a direct encoding, ModNet, NEAT, a multi-layer perceptron with a fixed structure and to reference controllers. Results show that MENNAG performs better than comparable encodings on both problems, suggesting a promising potential for future applications.