Time-Series Decomposition Model combined with pre-trained LSTM: Real-time prediction of dissolved oxygen concentration suitable for automated aquaculture feeding system

Abstract

As part of automation in land-based aquaculture facilities, the widespread adoption of automated feed supply systems is actively taking place. The automated feed supply system must be capable of determining the appropriate feed quantity to supply at arbitrary feeding times based on the hunger level of cultured fish. However, the decline in dissolved oxygen concentration due to feed supply can pose a high mortality risk if not managed properly. Therefore, there is a need to ensure safety in real-time before the feed supply occurs. In this study, we proposed a method to predict dissolved oxygen concentration in real-time, even when feeding occurs at arbitrary times, and aimed to validate its performance. We combined a time-series data decomposition model with a pre-trained deep-learning model. We can obtain trend components for long-term changes in dissolved oxygen concentration through the time-series data decomposition method. Using two models, we predict seasonal components for short-term changes caused by feed supply at arbitrary times. Regular changes can be obtained through the time-series data decomposition model, while irregular changes can be obtained through the pre-trained deep learning model. Experimental results validate the performance of the proposed model by comparing the predicted results of regular and irregular feed supply with those of other deep learning models. Lastly, we discuss how the proposed model structure reduces complexity by simplifying the deep learning model's weight matrix elements, achieving good performance with fewer learning parameters.