Improved surrogate modeling for multi-energy system design: Model architecture, sampling and scaling choices

Multi-energy systems (MES) are a key concept for developing more sustainable energy systems, but optimizing their design is computationally burdensome. This paper explores the development of machine-learning (ML) based surrogate models for the optimal design of MES. Surrogates are simple models, often ML-based, used to approximate detailed simulations, in this case MES design optimizations. These models provide instant responses, enabling fast comparisons and explorations of trade-offs between design variables. No related work proposes an ML procedure tailored to properties of the MES design application. Most related works use surrogates to predict system cost and other objectives. However, few works have used them to directly predict the optimal system design, and those that do show poor performance. This paper provides an extensive methodology tailored to properties of MES design problems to improve surrogate performance on small datasets. Four components were found to significantly improve surrogate performance: a careful and objective-oriented selection of samples, the use of upsampling to balance datasets, the use of non-linear rescaling methods, and a specific neural-network architecture called Mixture-of-Experts. These work together to turn the original design variable distribution (i.e., of the output) into a Gaussian-like data distribution, that can be more easily learned by the neural-network. The resulting surrogate model almost instantly predicts optimal energy system designs with high precision. This was tested across a wide variety of different climates, building types and decarbonization goals. Such surrogate models will make it much easier to explore different MES design options. • Energy system design shows irregularities and sudden changes in design variables. • Irregularities and splits reduce the learnability via neural-network. • A large exploration of modeling parameters is tested over a large pool of cases. • A simple mixture-of-experts architecture often addresses the issue. • Upsampling and a novel sampling strategy support the surrogate model’s training.