A time series model adapted to multiple environments for recirculating aquaculture systems

Environmental time series modeling of recirculating aquaculture systems provides the basis for the design of intelligent and foreseeable agricultural facilities. The modeling accuracy of environmental factors plays an important role, which could help grasp the environmental situation and change trend of the recirculating aquaculture system, assist in early warning when the environment factor level exceeds the normal data range, and combine with the control method to improve the accuracy of environmental control. The traditional time series model is difficult to predict complex situations, which is mainly due to the effective integration of multi-dimensional data. Our goal is to make improvements to the traditional time series model. The proposed multiple graph fusion network (GraphTS) fuses multi-sensor Spatio-temporal information using a multi-graph fusion method based on Gated Recurrent Unit (GRU) and graph attention neural network. We collected two recircu-lating aquaculture datasets with various features and applications to test GraphTS's performance. Comparing the average metrics of predictor outcomes of proposed GraphTS with the standard model LSTM, the average margin of error (AME) is reduced by 37% and 13%, and the Pearson correlation Coefficient (PCC) is improved to 97% and 96% for two datasets, respectively. The best results are also achieved on the discrete traffic prediction dataset. It shows the adaptability and multi-dimensional information gathering ability of GraphTS.

Automated summary

Environmental time series modeling is crucial for the design of intelligent and predictable agricultural facilities. The accuracy of modeling environmental factors is important for grasping the situation and trend of recirculating aquaculture systems, providing early warnings for abnormal environmental conditions, and improving the accuracy of environmental control. The proposed GraphTS, a multi-graph fusion network based on GRU and graph attention neural network, shows improvement in prediction performance compared to the traditional LSTM model, with reduced average margin of error and improved Pearson correlation coefficient.