Neural Network-Driven Optimization: Enhancing Speed in Energy Systems
Abstract
This work explores the use of neural network surrogate models to approximate outputs such as installed renewable capacities, storage sizing and system costs, thereby reducing simulation runtimes while maintaining predictive accuracy. Training datasets are generated to ensure broad coverage of input scenarios, including economic variables like electricity prices and capital expenditures, as well as climatic factors influencing renewable generation. The surrogate models are integrated with outputs from the oemof framework, which represents the regional energy system through graph-based optimisation. Performance assessment employs metrics such as root mean square error, mean absolute percentage error, and coefficient of determination, with attention to prediction accuracy across typical and extreme operational conditions. A data-driven surrogate based on a feedforward multilayer perceptron was developed to emulate the optimization results of the energy system model. The surrogate provides accurate capacity predictions with an order-of-magnitude lower computational cost, enabling extensive scenario and sensitivity analyses.
Keywords
decarbonisation, energy system, energy system modelling, open source, optimization, neural network, multi-layer perceptrons, convolutional neural networks, recurrent neural networks