Software evolves continuously. As a consequence, software systems tend to become increasingly complex and, as such, more difficult to change. A software system's complexity is for a large part determined by its structure, or architecture. In this thesis we investigate how to reduce the risks and costs associated with the evolution of software architectures. Automation and abstraction are two basic software engineering techniques to deal with complexity. In this thesis we investigate the applicability of model-driven engineering, a new software development approach based on abstraction and automation, to support the evolution of software architectures. The main research question we address in this thesis is: "How can evolution of software architectures be supported?". Three subquestions related to industrial integration, software product lines, and automation further clarify the scope of our work. We first conducted a survey among several software development organisation to inventory the state-of-the-practice in software engineering technologies. Some trends we observed from this inventory include: the informal use modelling in industry, the use of product-line approaches, and the importance of the evolutionary aspect of software. Next, we investigated how to support four tasks related to software architecture evolution: evaluation, conformance checking, migration, and documentation. We aim to automate this support where possible. To this end, we employ model-driven software development technologies. For each of the software evolution tasks, we present a case study that investigates how that task can be supported. The informal use of modelling in industry calls for a normalisation step to enable the integration of evolution support in practice. Several chapters address the impact of the use of product-line approaches on the evolution support. Although the increased scope make such support more difficult to develop, the return on investment for the model-driven support is much improved. The model-driven evolution support follows a similar three-step pattern. First, a set of source models is preprocessed into a form suitable for the application of model transformations. Then, model transformations are applied that do the actual work, such as conformance checking or a migration. Finally, the resulting models are postprocessed in a resulting into a desired target form.