| China’s fast pace industrialization and growing population has led to several accidental surface water pollution events in the last decades. This has severely affected the safety of large populations downstream who are dependent of these waters for drinking. In other countries such as the USA, several accidental pollution events have forced them to develop early warning systems(EWS) for the protection of their drinking water sources. The government of China, in its 11 th Five Year plan, after the 2005 Songhua River incident, has pushed for similar actions. Despite recent government efforts, there are still many weaknesses and gaps in EWS in China such as the lack of pollution monitoring and advanced mathematical models to predict and forecast pollution events. The application of existing physical models for water quality prediction in China can be challenging due to information availability constraints. Data Driven Models(DDM) such as Artificial Neural Networks(ANN) have acquired recent attention as an alternative to physical models which require large amounts of data, do not take into account nonlinear hydrological properties, are computationally demanding and not always flexible.For a case study in a south industrial city in China, a DDM based on genetic algorithm(GA) and ANN, GA-ANN model, was tested to meet two objectives. The first objective was to increase the response time of the city’s early warning system to predict NH3-N, CODmn and TOC at station B using available measured variables at station A, 12 km upstream from station B. The second objective was to relate the measured physico-chemical variables at station A to corresponding toxicity data at Station A. At the same time, the capacity of the GA-ANN model to predict ahead of time was tested in both cases.The GA-ANN model using Nonlinear autoregressive with exogenous input(NARX) architecture was able to predict NH3-N, CODmn and TOC variables at station B using past measurements at station A. The model was also able to show which input variables were more sensitive for the prediction of the three aforementioned output variables. For NH3-N prediction at two hours ahead of time at station B the most sensitive input variables at station A were TOC, CODmn, TP, NH3-N and Turbidity with model performance giving a mean square error(MSE) of 0.0033, mean percent error(MPE) of 6% and regression(R) of 92%. For COD prediction at two hours ahead of time at station B the most sensitive input variables at station A were Turbidity and CODmn with model performance giving a MSE of 0.201, MPE of 5% and R of 0.87. For TOC prediction at two hours ahead of time at station B, the most sensitive input variables were Turbidity and CODmn with model performance giving a MSE of 0.101, MPE of 2% and R of 0.94. In addition, this study showed that the GA-ANN model can perform even better ahead of time and may have the capacity to predict for more than 8 hours ahead of time.In the case of toxicity prediction at Station A, the GA-ANN model seemed to perform well when using the NARX architecture with MSE of 4.3, MPE of 6% and R of 0.93. However, the model did not perform well as the time delay increased and it was not able to discern which physico-chemical input parameters were more sensitive. In this case, this was most probably due to insufficient available toxicity data to train the GA-ANN model.For future studies, the use of data-driven models, such as the GA-ANN model, can be very useful for water quality prediction in Chinese monitoring stations which already measure water quality data, but have not attempted any modelling in the past. This type of model is easy and fast to use. However, as observed in the case of toxicity prediction, a minimum of three years of past measurement records are needed to obtain a reliable model performance. |
PDF Full Text Request