Deep neural network analyses of water quality time series associated with water sensitive urban design (WSUD) features

Abstract

The abilities of Long short-term memory (LSTM), Gated recurrent units (GRU), Adaptive-network-based fuzzy inference system (ANFIS), Artificial neural networks (ANN), and Group method of data handling (GMDH) in predicting water quality time series associated with a cleansing biotope were evaluated. We examined continuous monitoring time series of chlorophyll-a, turbidity, and specific conductivity using YSI EXO datasondes at the Inlet and Outlet. Based on Root Mean Square Errors, Mean Absolute Percentage Errors, Mean Absolute Errors, Correlation Coefficients, and Theil’s U, the GRU generally was the most efficient model in predicting the Outlet water quality. AI models should find increasing implementation the area of ‘smart environment’. Ways forward for enhancing AI model performance were suggested to better consider data periodicity and explore a transfer function approach in which the water quality timeseries of one parameter is forecast based on an ensemble of other parameters.