Machine learning enabled thermoelectric generator design and optimization

Zhu, Yuxiao (2024) Machine learning enabled thermoelectric generator design and optimization. University of Southampton, Doctoral Thesis, 154pp.

Record type: Thesis (Doctoral)

Abstract

With increasing global recognition of environmental protection and carbon emission reduction, renewable energy development is drawing heightened attention. At the same time, energy conversion efficiency needs to be improved, requiring techniques such as energy harvesting. In this context, thermoelectric generators (TEGs), which can recycle thermal energy, have garnered significant interest from researchers.

Traditional modelling of thermoelectric generators (TEGs) typically involves two approaches: theoretical models and mathematical models. The theoretical, 1-D and 2-D mathematical models often suffer from reduced accuracy due to the omission of specific parameters. On the other hand, the 3-D mathematical model, commonly known as 3-D finite element analysis (FEA), is hindered by slow computational speeds. To address these limitations, this thesis proposes a novel approach using artificial neural networks (ANNs) to model TEGs, achieving accuracy and computational efficiency.

This thesis first demonstrates the application of ANNs in constructing a forward model for a thermoelectric generator. This method attains computation speeds thousands of times faster than 3D FEA while preserving 98% accuracy compared to the results from 3D FEA. Furthermore, when integrated with optimization algorithms, this model can effectively optimize the structure of the thermoelectric generator, demonstrating a significant advancement in modelling and design efficiency.

Later in this work, an ANN has been applied to build accurate and fast forward modelling of the segmented thermoelectric generator (STEG). More importantly, an iterative method is adopted in the ANN training process to improve accuracy without increasing the dataset size. This approach strengthens the proportion of the high-power performance in the STEG training dataset. Without increasing the size of the training dataset, the accuracy was increased from 92% to 98%. Coupling with a genetic algorithm, the trained artificial neural networks can optimise design within 10 seconds for each operating condition. It is over 4,000 times faster than the optimization performed by the conventional FEA model. Such an accurate and fast modeller also allows the mapping of the STEG power against different parameters.

Then, the hybrid TEG system is analyzed. Radiative cooling (RC) can provide a continuous temperature difference, which a TEG can convert into electrical power. This novel combination of radiative cooling with TEG expands the category of sustainable energy sources for energy harvesting. Using 3D FEA, this system provides a systematic analysis of the concept of RC-TEG by investigating the impact of radiative cooler properties, TEG parameters, and environmental conditions to provide a complete picture of the performance of RC-TEG devices. The capability of RC-TEG to provide continuous power supply is simulated using real-time environmental data from both Singapore and London on two different days of the year, demonstrating continuous power supply sufficient for a wide range of IoT devices in all four scenarios.

Finally, this thesis introduces an ANN-based model designed to predict the performance of hybrid PV-TEG systems. Utilizing a cyclic approach, the ANN model incorporates various factors, including PV coating, morphology, TEG geometry parameters, temperature-dependent material properties, and environmental conditions like solar irradiance and convection. The model's integrated nature allows independent use of PV and TEG components, enhancing its adaptability and generalizability. Remarkably, compared to FEA simulations, the ANN model demonstrates over 98% accuracy and a significant boost in computational efficiency, with a 6,000-fold increase in simulation speed. This efficiency enables extensive parameter sweeps, offering insightful analysis into the influence of various factors on the PV-TEG system's performance.

Overall, the ANN model's rapid processing capabilities are particularly beneficial for large-scale simulations and practical applications in renewable energy technology.