A deep learning algorithm for multi-source data fusion to predict water quality of urban sewer networks

Point source pollution in urban drainage networks, which is difficult to monitor and control, has been regarded as an intractable problem. To solve the problem, key water quality indicators must be tracked in the evaluation and prediction of sewer water quality. However, some of these important chemical indicators (e.g. biological oxygen demand (BOD₅), chemical oxygen demand (COD), ammonia nitrogen (NH₄⁺-N), total nitrogen (TN), and total phosphorus (TP)) require a great deal of time and effort to measure, which will adversely affect the prediction in a sewage network. Existing statistical methods and machine learning algorithms cannot effectively solve the detection time problem or provide limited accuracy. Moreover, the lack of various factors taken into account in these methods results in unsatisfactory predictive performance. Few studies consider the impact of urban multi-source data on water quality prediction of sewer networks while developing statistical methods or machine learning algorithms. To address this problem, we propose a deep learning approach based on multi-source data fusion. This approach takes into account the following indicators to comprehensively analyze and predict drainage water quality: environmental indicators (such as area and diameter); social indicators (such as population); water quantity indicators (such as drinking water supply, sewage flow, water velocity, and liquid level); and easily monitored water quality criteria indicators (such as pH, temperature, and conductivity). To test the effectiveness of this method, we conducted a case study in a city in southern China. By comparing this new method with the linear method (multiple linear regression, MLR) and traditional learning algorithm (multilayer perception, MLP), it is found that the deep learning algorithm—which includes recurrent neural network (RNN), long-short term memory (LSTM), and gated recurrent unit (GRU)—has good predictive performance, in which GRU shows superior ability in predicting the chemical index of water quality and the learning curve is faster. The results showed that the GRU achieved 0.82%–5.07% higher R² than RNN and LSTM, 9.13%–15.03% higher R² than traditional machine learning algorithms, and 37.26%–43.38% higher R² than linear methods.