Hybrid deep learning architecture for temperature gradient control of a solid oxide electrolysis cell under fluctuating wind power

• A novel hybrid deep learning architecture (HDLA) framework • Combination of data-driven machine learning with multi-physics simulations • Reducing SOEC temperature gradients from ±20 to ±5°C under dynamic wind power • 99 % energy utilization efficiency and stable operation even at fluctuating renewable energy input • HDLA enhances grid-SOEC synergy, enabling carbon-neutral chemical production The co-electrolysis of CO₂ and H₂O through solid oxide electrolysis cells (SOECs), powered by renewable energy sources, offers a promising pathway to achieving carbon neutrality in the chemical industry. However, the inherent intermittency of renewable energy generation, such as wind power, leads to unstable power input for electrolysis. This variability induces significant thermal stress in SOECs, potentially causing cracks or even system failure. To address this challenge, a hybrid deep learning architecture (HDLA) was developed to control the temperature gradient of SOECs. The architecture combines a convolutional neural network (CNN) and a long short-term memory (LSTM) model for wind power prediction, a multi-physics model for temperature gradient simulation, and a linear neural network regression model to simulate the temperature distribution in SOECs. Training and verification are conducted using 16 datasets from an industrial wind farm. The results demonstrate that the application of HDLA successfully reduce the temperature gradient of SOECs from ±20°C to ±5°C. Additionally, the potential wind power utilization achieved near-complete wind power utilization, increasing from 18 % to 99 %. This real-time control strategy, which optimizes flow regulation, effectively mitigates thermal stress, thereby extending the lifespan of SOECs and ensuring continuous carbon reduction, efficient conversion, and utilization.