Metaheuristics‐optimized deep learning to predict generation of sustainable energy from rooftop plant microbial fuel cells

• Published in: International journal of energy research Vol. 46; no. 15; pp. 21001 - 21027
• Main Authors: Chou, Jui‐Sheng, Cheng, Tsung‐Chi, Liu, Chi‐Yun, Guan, Chung‐Yu, Yu, Chang‐Ping
• Format: Journal Article
• Language: English
• Published: Chichester, UK John Wiley & Sons, Inc 01.12.2022; Hindawi Limited
• Subjects: Accuracy; Algorithms; Animal learning; Aquatic plants; Artificial intelligence; Artificial neural networks; Biochemical fuel cells; building energy management; Capacity; Clean energy; Computation; computer vision‐based deep learning; Deep learning; Energy; Energy technology; Environmental factors; Freshwater plants; Fuel cells; Fuel technology; green energy conversion and forecasting; Heuristic methods; jellyfish search optimizer; Machine learning; Marine invertebrates; Mathematical models; Microorganisms; Model accuracy; Neural networks; Optimization; Optimization algorithms; Parameters; plant microbial fuel cell; Renewable energy; Solar energy; Solar energy conversion; Sustainability; Sustainable energy; Typha angustifolia; Urban environments; wireless sensor network application
• ISSN: 0363-907X; 1099-114X
• DOI: 10.1002/er.8538

Summary Plant microbial fuel cells (PMFCs) are an emergent green‐energy technology that continuously converts solar energy into electricity. Placing PMFCs on the roofs of urban buildings can help to create green urban environments even as they generate power. The power generation performance of PMFCs is affected by a range of environmental factors, so their power generation capacity is difficult to estimate. To develop an artificial intelligence model to forecast PMFC power generation accurately, relevant results obtained using shallow and deep learning techniques are compared for the first time. Once deep learning techniques had been identified as superior for this purpose, they were used with a bio‐inspired optimization algorithm to dynamically setting the model hyperparameters. The developed model can also be applied to estimate the power generation capacity of PMFC devices in the future. The model was trained using data collected from sensors in a site experiment that was carried out using PMFCs embedded with Chinese pennisetumin (Pennisetum alopecuroides), narrowleaf cattail (Typha angustifolia), dwarf rotala (Rotala rotundifolia), and no plant as a control group. The original data of device parameters, environmental parameters, and the measured power generation of PMFCs in numerical form were applied to train shallow learning and time‐series deep learning models. Meanwhile, the state‐of‐the‐art sliding window technique was used to establish a numerical matrix, which was converted into a 2D image‐like format to represent inputs for deep convolutional neural network (CNN) models. The accuracy in predicting the power generation capacity of PMFC devices showed that EfficientNet, an advanced type of CNN, was the best model among the shallow and deep learning techniques. These analytical results demonstrate the superior performance of deep CNNs in learning image features and their consequent suitability for constructing PMFC power generation forecasting models. To enhance the generalization performance of CNN, a newly developed bio‐inspired optimization algorithm, jellyfish search (JS), was incorporated into this model to determine the optimal hyperparameters, yielding the hybrid JSCNN model. This investigation revealed that the JS optimization algorithm can find better values of hyperparameters of the CNN and stabilize model accuracy. Notably, once the optimal hyperparameters have been obtained using JS, the computation time of the hybrid JSCNN model is shorter than that of the generic CNN model, supporting the need to determine the appropriate hyperparameter values in deep learning. This study is the first to use its particular setup and to offer its particular precautions; it therefore contributes to the body of domain knowledge and practicality of sustainable PMFCs. Placing PMFCs on the roofs of urban buildings can help to create green urban environments even as they generate power. Comprehensive AI models to forecast PMFC power generation using shallow and deep learning techniques are compared. Deep learning (DL) for computer vision demonstrates the superior performance in constructing PMFC power generation forecasting models. By incorporating jellyfish search optimizer to finetune the hyperparameters of the identified best DL model, the identified optimal model can help to increase the automation of dynamic power management systems.