Artificial Neural Network Backpropagation for Real-Time Coagulant Dose Prediction in Drinking Water Treatment

A drinking water treatment plant is crucial for fulfilling the increasing water demand. As an integrated series, the coagulation unit is the most basic unit for removing particulates and reducing turbidity. However, the time gap to determine an effective coagulant dose was almost 6 h, which cannot accommodate the fluctuations in water inlet quality. A noteworthy method to reduce time is to use artificial neural network backpropagation (ANN-BP) to predict an optimum dose. The dataset consisted of five parameters (pH, temperature, conductivity, color, and turbidity) for a month of primary data sampling and historical jar test data from 2018 to 2022. The F-test result, F-value (6038,779) > F-table (2.21923), showed that one or more parameters had a statistically significant influence on the coagulant dose. Subsequently, a t-test excluded pH and temperature, with p-values lower than 0.05. Empirical models were developed through trial-and-error variations of the input layers (three and five parameters), hidden layers (2-10 nodes), and an output. The models with the lowest MSE and highest R2 were [5-6-1] (R2 =- 0.96051; MSE = 0.00179) and [3-4-1] (R2 = 0.97755; MSE =0.00102). In conclusion, [3-4-1] is recommended because it has the lowest MSE and the highest R2.