Time series analysis and long short-term memory (LSTM) network prediction of BPV current density

• Published in: Energy & environmental science 2021-04, Vol.14 (4), p.248-2418
• Main Authors: Okedi, Tonny I, Fisher, Adrian C
• Format: Article
• Language: English
• Subjects: Biofilms; Current density; Decomposition; Electricity; Electron trajectories; First principles; Fluorescence; Long short-term memory; Microorganisms; Optimization; Photosynthesis; Time series
• ISSN: 1754-5692; 1754-5706
• DOI: 10.1039/d0ee02970j

Biophotovoltaics (BPVs) have increasingly gained interest due to their potential to generate low-carbon electricity and chemicals from just sunlight and water using photosynthetic microorganisms. A key hurdle in developing commercial biophotovoltaic devices is understanding the electron path from within the microorganism to the electrode. The complexities of competing cellular metabolic reactions and adaptive/temporal physiological changes make it difficult to develop first-principle models to aid in the study of the electron path. In this work, Seasonal and Trend Decomposition using locally estimated scatterplot smoothing or LOESS (STL) is applied to decompose the current density profile of electricity-generating BPV devices into their trend, seasonal and irregular components. A Long Short-Term Memory (LSTM) network is then used to predict the one-step-ahead current density using lagged values of the output and light status (on/off). The LSTM network fails to adequately predict the observed current density profile, but adequately predicts the light-controlled seasonal component with mean absolute errors of 0.007, 0.0014 and 0.0013 μA m −2 on the training, validation and test sets respectively. The improved performance in the latter is attributed to the removal of irregular patterns. An additional predictor, biofilm fluorescence yield, is proposed to improve predictions of both the observed current density and its seasonal component. This seminal work on the use of LSTM networks to predict the current density of biophotovoltaics opens doors for faster and more cost effective device optimisation, as well as the development of control software for these devices. LSTM networks are shown to predict the seasonal component of biophotovoltaic current density and photoresponse to high accuracy.