Power prediction for salinity-gradient osmotic energy conversion based on multiscale and multidimensional convolutional neural network

Abstract

Osmotic energy conversion (OEC) is a promising renewable energy utilization technology that directly convers salinity-gradient energy into electricity. However, most of current studies on the OEC power under different nanostructures and solution parameters were conducted experimentally or by simulation, which is costly and difficult to explore the optimal OEC device configuration. In this study, we propose a multiscale and multidimensional convolutional neural network-based power prediction model for salinity-gradient OEC. It can learn intrinsic characteristics embedded in multi-physical and nanopore geometric parameters that are closely related to the osmotic power generation, thus realizing accurate OEC power prediction. For model development and assessment, a numerical model of the salinity-gradient OEC device with conical nanopores was developed using COMSOL Multiphysics to generate training and test datasets. The test results show that the mean absolute percentage error between the predicted powers and real powers of the OEC device is only 0.309 % over 4077 typical operating conditions. Furthermore, the prediction performance of the proposed model outperforms other four comparative models employing widely-used deep learning algorithms, indicating its effectiveness and superiority in OEC power prediction. This study contributes to the optimal design and performance enhancement of OEC devices.