Deep learning of temporal renewable energy patterns for optimizing Power-to-X processes

Abstract

The inherently intermittent nature of renewable energy and its geographic variations make the optimisation of Power-to-X processes dependent on the local temporal characteristics of solar and wind energy generation profiles. By using green ammonia production as a case study, this paper demonstrates that deep convolutional and recurrent neural networks can successfully learn key characteristics of renewable power profiles (capacity production, peak supply and intermittency) and accordingly determine the capacity of the solar panels/wind turbines, electrolysers and hydrogen storage to minimise the cost of green ammonia production. By learning implicit cost relationships, deep neural networks can also predict the economically optimal process design with an error of 5-15%. As a result, neural networks can be implemented for rapid screening of locations for initial economic feasibility of Power-to-X processes, providing the foundations for more challenging optimizations such as those considering decades of weather-based renewable power profiles and complex energy system models.