关键词: 溶解氧, 长短时记忆网络, 损失函数, 平均绝对百分比误差

Abstract:

Dissolved oxygen (DO) is vital to aquaculture industry and affects the yield of aquaculture. Low DO in water can lead to death of crabs, therefore, it is important to measure DO accurately. However, the DO sensors are usually expensive and often lost function due to corrosion in water environmental and adsorption of different materials on their surface, which result in the inaccuracy in measured DO values. It is thus important to develop effective methods to estimate DO concentrations by using other environmental variables, which may reduce farmers' cost because DO sensors are not used. In this research, the collected environmental data, including temperature, pH, ammonia nitrogen, turbidity, were used to estimate DO concentrations in crab ponds. The data were preprocessed to eliminate missing values and outlier. Correlation analysis was applied to determine the relationship between environmental variables (temperature, pH, ammonia nitrogen, turbidity) and DO to show the rationale of using these four variables to forecast DO concentration. Principal component analysis was used to reduce the dimension of environmental data to reduce computation cost. For DO concentration estimation, it is more important to make the estimation of DO concentration at low values more accurate because DO concentration at low values is dangerous to crabs. This implies that estimation of DO concentrations at low or high values should be treated differently and applied different rates. Based on the Long Short-Term Memory (LSTM), a low DO concentration estimation model of Low Dissolved Oxygen Long Short-Term Memory(LDO-LSTM), which can improve the estimation accuracy of low DO values was proposed by optimizing the loss function of LSTM back propagation. The loss function of LDO-LSTM was based on the Mean Absolute Percentage Error (MAPE). According to the trend of DO, the true DO and the estimated DO values were applied weight functions. The Root Mean Square Error (RMSE) and the MAPE were used to evaluate the performance of LDO-LSTM and LSTM in DO estimation. Experimental results show that the value of RMSE and MAPE were stable at about 0.1 for LSTM and LDO-LSTM in forecasting DO when dissolved oxygen was higher than 6mg/L and the value of RMSE and MAPE of LDO-LSTM were lower than LSTM by 0.25 and 0.139. The results prove that the proposed method can not only provide desirable estimation accuracy for DO concentrations at high values but also make the estimated DO concentrations at low values more accurate. This research is expected very useful in reducing aquaculture costs and improving accuracy in forecasting DO especially at low values.

Key words: dissolved oxygen, long short-term memory, loss function, mean absolute percentage error