Main content area

Eutrophication modelling of Amirkabir Reservoir (Iran) using an artificial neural network approach

Homayoun Aria, Shiva, Asadollahfardi, Gholamreza, Heidarzadeh, Nima
Lakes & reservoirs 2019 v.24 no.1 pp. 48-58
algal blooms, ammonia, ammonium, data collection, electrical conductivity, environmental factors, eutrophication, lakes, monitoring, neural networks, nitrates, nitrites, nitrogen dioxide, oxygen, phosphates, prediction, temperature, turbidity, water quality, Iran
The complicated non‐linear relationships between water quality and environmental parameters involved in predicting algal blooms necessitate a new approach, using data‐driven modelling. Accordingly, a multilayer perceptron (MLP) and time delay neural network (TDNN) were used to predict the eutrophication status of two monitoring stations in the Amirkabir Reservoir in Iran. Six scenarios for each monitoring station were performed to select a significant, independent input using 12 years of monthly data. The final inputs were temperature, turbidity, phosphate (PO₄), nitrate (NO₃), nitrite (NO₂), ammonium (NH₃), dissolved oxygen (DO) and electrical conductivity (EC). Applying an MLP neural network to the upstream monitoring station with 21–38 neurons in the first and second hidden layers, the minimum mean squared errors (MSE) in training, validating and testing were 0.083, 0.81 and 1.95 cells/100 ml, respectively. Further, when the TDNN network was used with the same neuron numbers in the hidden layer for the similar monitoring station, the minimum MSE values for model training, validating and testing were 0.06, 0.72 and 1.76 cells/100 ml, respectively. For the Beylaghan monitoring station, using the MLP neural network with 29–23 neurons in the first and second hidden layer, the minimum MSE values gained in training, validating and testing were 0.181, 0.58 and 0.95 cells/100 ml, respectively. Using the TDNN network with the same neurons in the hidden layers of the MLP neural network for the station, the minimum MSE values for training, validating and testing were 0.152, 0.43 and 0.84 cells/100mL, respectively. Thus, TDNN exhibited a high accuracy and workability, compared to the MLP. Sensitivity analysis of the Amirkabir Reservoir dataset indicated increasing the value of nitrate is the first factor, followed by turbidity and NH₃, having the greatest impacts on eutrophication prediction.