Establishment of a water nitrite nitrogen concentration prediction modelbased on stacked autoencoder-BP neural network

Nitrite nitrogen is toxic to the aquatic animals. Monitoring the concentration of nitrite nitrogen is very critical for the culture of aquatic animals. Due to the high cost of the current commercial electrode sensor which is used to measure the concentration of nitrite nitrogen in water, this kind of sensor is very difficult to be popularized on a large scale. Therefore, it is an urgent need to develop another novel method to predict the concentration of nitrite nitrogen in water. In this paper, taking the advantage of the established online water monitoring system in our laboratory, the temperature, pH value, dissolved oxygen and oxidation-reduction potential were recorded from the water in tanks. Meanwhile, the actual concentration of nitrate nitrogen in water was measured using alpha-naphthalene colorimetric method. The data after pretreatment were used as the original data to be used for SAE neural network training. Thereafter, unsupervised greed training method was applied. The learnt characteristics were used for the supervision and training of BP neural network. The model was optimized using the back propagation (BP) algorithm. The prediction model R² of the nitrite nitrogen after training was 0.95, and root mean square error of the prediction (RMSEP) was 0.099 71, indicating that the model could accurately predict the concentrations of nitrate nitrogen in water. The established model will pave a new way for developing online system for monitoring the water nitrate nitrogen concentration in the future.