Spatio-temporal analysis and prediction for raw water quality of drinking water source by improved RNN algorithm

Abstract

The raw water quality of drinking water sources varies due to complex spatio-temporal patterns, which is crucial for rapid and accurate water quality prediction in water resource management. This paper proposes a spatio-temporal water quality prediction method that combines time series forecasting and spatial interpolation. The combination of deep learning methods and physical models aims to replace the semi-variogram fitting in Ordinary Kriging (OK) interpolation with a trained Recurrent Neural Network (RNN) for spatio-temporal water quality prediction. The results demonstrated that the combination of RNN and OK achieved the best predictive performance, with mean absolute errors for pH, dissolved oxygen, conductivity, turbidity, ammonia nitrogen, and chemical oxygen demand of 0.025, 0.025, 1.347, 2.029, 0.074, and 0.181, respectively. In temporal prediction, the LSTM model showed excellent performance, particularly when confronted with intricate time series data. Its mean absolute percentage error remained below 4.5 %. In terms of spatial prediction, the OK method displayed remarkable accuracy, with correlation coefficients (R²) ranging from 0.767 to 0.995. Compared to traditional water quality monitoring, the proposed method effectively captures the complex spatiotemporal dependencies of water quality changes, improving prediction accuracy and offering a novel approach for spatio-temporal water quality forecasting, which can be referenced for other water sources.