In recent years the energy-efficiency of software has become a key focus for both researchers and software developers, aiming to reduce greenhouse-gas emissions and operational costs. Despite this growing awareness, developers still lack effective strategies to improve the energy-efficiency of their programs beyond the well-established approaches that optimize for runtime performance. In this paper we present a dynamic adaptation algorithm that uses energy consumption feedback to optimize the energy-efficiency of data-parallel applications, by steering the level of parallelism during runtime through external control. This approach is especially suited to functional languages, whose side-effect-free nature and strong semantic guarantees allow for easier code generation and straightforward scalability of the parallelism of programs. Through a series of experiments we evaluate the effectiveness of our approach. We measure how well the adaptation algorithm adapts to runtime changes, and we evaluate its effectiveness compared to a hypothesized oracle that knows the optimal level of parallelism, as well as a runtime-optimising-based approach. We show that in a fixed-workload scenario we approach the theoretical best energy-efficiency, and that in changing workload scenarios the adaptation algorithm converges towards an optimal level of parallelism that minimizes energy consumption.