Author：Chi-Yuan Lin(1), Chao-Neng Wang(2)
(1)Planning and Information Division, Fisheries Research Institute,
(2)Department of Bioinformatics and Medical Engineering, Asia Universit
There are several key water quality parameters that can each directly affect aquatic animals’ health. However, in the complex dynamic aquaculture outdoor environment, water quality parameters also influence each other. Thus, maintaining balance in water quality parameters is fundamental for the health and growth of culture organisms. Nevertheless, auxiliary use of the latest information technology, like automated electron-control aeration and feeding systems with IoT sensors and AI calculations, is essential for modern fish farms if they are to achieve high yields and stay competitive. In order to further reduce prediction costs and improve prediction efficiency, this paper proposed a multi-scale forecasting method based on AI machine learning models. The most important dissolved oxygen key parameter was chosen to decompose and reconstruct by analyzing their correlation. The machine learning models was developed by DecisionTreeRegressor, XGBoostRegressor, RNN, LSTM, and GRU methods and dissolved oxygen datasets were actually IoT-sensed from aquaculture pools in Gangshan Dist., Kaohsiung City, Taiwan.
In this study, we select a total of 343 samples to training and testing including twelve environmental water-quality and weather factors. All of the eleven factors are closely related to DO. For example, the higher the water temperature, the lower the concentration of DO. The higher the salinity, the lower the concentration of DO, etc. The water quality data are split into two parts: the first 75% of water quality data are used for modeling training and the last 25% of data as testing data to analyze the prediction performance of our model. The DO series is shown in Figure 1.
The prediction resulted of five model are show in Table 1. The relative MSE, RMSE and MAR differences between five model. XGBoostRegressor model are 0.0111, 0.1058 and 0.0774; Decision-TreeRegressor model are 0.220, 0.1483 and 0.0944; RNN model are 0.0125, 0.1122 and 0.0852; LSTM model are 0.0113, 0.1066 and 0.819; GRU model are 0.0116, 0.1081 and 0.795 in the text period, respectively. Figure 2 shows the forecasting results of the first set of data. The fitting accuracy of XGBoost-Regressor model is higher than the other four models.
The model-learning results showed that XGBoostRegressor model had the best prediction performance based on their MAE, MSE, and RMSE. Moreover, through verified the validity of 5 learning models, XGBoostRegressor model also had a significantly more reliable performance and a higher prediction precision than the other models. This study primary demonstrates that our machine learning model is proven to be an effective approach to predict aquaculture dissolved oxygen.
(extract and reprint from proceeding of AFAF 2022 conference)
