Deep neural network-based optimization of thermoelectric atmospheric water generation under dynamic climatic conditions

Atmospheric Water Generation (AWG) with thermoelectric cooling (TEC) is a potential solution for decentralized freshwater provision, especially for dry climates. However, forecasting performance of such a system under different operating and ambient conditions is a difficult problem due to the nonlinear relationship involved. The present study aims to put forward a data-driven modeling strategy with a deep neural network (DNN) to predict water condensation rate ( m wc ) and effectiveness of the system ( Eff ). A dataset derived from validated thermodynamic simulations with a refined input space was used for building and training the model, involving inputs of ambient temperature, relative humidity, voltage, time of the day, and monthly encoding. With a highly accurate performance, a trained DNN has R 2 of 0.982 for m wc and 0.975 for Eff and respective Root Mean Squared Errors of 3.12 g/h and 0.064 g/Wh for m wc Eff , respectively. Model generalization has been confirmed through comparative validation across all months with no indication of over-fitting. Additionally, the research has also shown that voltage optimization using composite scores is viable, and that dynamic voltage management has the potential to increase effectiveness by up to 115 % whilst compared to using a constant 12 V input. These results support the implementation of deep models for intelligent control and optimization of AWG equipment to provide for responsive actions to dynamically varying ambient conditions.