Investigation on the Effects of Inherent Factor and Time- Interval Decision in Artificial Neural Networks for Determining the Optimal Coagulant Dosage in Drinking Water Treatment

Artificial neutral network (ANN) approach has been used in the drinking water treatment process for determining optimal coagulant dosage in the process. In such approach, a significant amount of input variables is typically required. To remedy the constraint, this study investigates the effects of input variables on the output performance of ANN models, using the real time coagulant dosage data collected in a drinking water treatment plant in Taipei, Taiwan. In this paper, ANN models are developed and applied to predict the optimal coagulant dosage, considering the effects of different model input parameters. Past coagulant dosages, defined as the inherent factor, along with the time interval between two dosages, were considered in investigating the effects and evaluated by the Pearson correlation coefficient. The root mean square error criterion was used in evaluating the performance of ANN models, and the Levenberg-Marquardt method in combination with weight decay regularization was applied to avoid over-fitting. From the results of this study, it is found that the ANN models with the inherent factor provide the optimal coagulant dosage prediction, while the increasing use of the inherent factor does not significantly improve the predictability of ANN models. The use of the inherent factor is also found useful when influent water does not provide sufficient information on water quality. *Corresponding author Email: [email protected] INTRODUCTION Coagulation is an important stage in drinking water treatment processes for the maintenance of acceptable treated water quality and economic plant operation. In practice, the required coagulant dosage in water treatment is usually evaluated by jar tests and the operator’s experience. However, the jar test is time-consuming and less adaptive to change in raw water quality in real time. Artificial neural network (ANN) is an approach to the coagulation process with several advantages, such as dosage prediction, human mistake elimination, time saving, and chemical expenses reduction. In previous studies, ANN has been used to develop models simulating the alum dosing process. Gagnon et al. [1] developed an ANN model to predict the optimal alum dosage for the Ste-Foy water treatment plant in Quebec, Canada. Baxter et al. [2] employed the ANN models to predict the treated water turbidity and color, at the Rossdale water treatment plant in Edmonton, Alberta, Canada. Joo et al. [3] developed a similar model for Chungju water treatment plant in Korea. Van Leeuwen et al. [4] developed the ANN model for the prediction of optimal alum dosage based on jar tests conducted on surface waters collected in Southern Australia. Maier et al. [5] adopted the database of van Leeuwen et al. [4] to predict optimal alum dosage and treated water quality variables. From all of the above studies, the model inputs consists of raw water quality variables, whereas the model output is the optimal alum dosage needed to achieve the desired treated water quality. Zhang and Stanley [6] included the turbidity of the treated water as an input in addition to a number of raw water quality variables in their ANN model to predict the optimal alum dosage at the Rossdale water treatment plant. Yu et al. [7] used the same input as Zhang and Stanley [6] in their 194 J. Environ. Eng. Manage., 19(4), 193-201 (2009) model for the prediction of optimal alum dosage at a water treatment plant in Edmonton AB, Canada. Lamrini et al. [8] have adopted the Levenberg-Marquardt method in ANN to predict the coagulant dosage for the raw water with the turbidity higher than 1000 NTU (Nephelometric turbidity unit). Influent water does not provide information on water quality, the predicting model can not be used. Based on these concepts, a project was initiated to study the potential capacity of ANN process control in a drinking water treatment. Li et al. [9] investigated the global robust exponential stability of interval neural networks with timevarying delays and found the delay-dependent and delay-independent criteria, which are derived via a linear matrix inequalities approach. Lu and He [10] reported a time delay parameter into the ANN model to characterize the signal transmission delays, and prove the condition given to ensure that the existence and uniqueness of equilibrium also guarantee the global exponential stability of the ANN by construction of a suitable Lyapunov functional. He et al. [11] investigated the stability problem for ANN with time varying interval delay and reported less conservative stability criteria by relationship between the time-varying delay and its lower and upper bounds, when calculating the upper bound of the derivative of Lyapunov functional. Zhang et al. [12] reported the robust asymptotical stability issue of linear interval time-variant/timeinvariant systems with uncertain delays, which combine Lyapunov-Krasovskii functions, parameterized first-order model transformation and the transformation of interval matrices into norm-bounded uncertainty. Corzo and Solomatine [13] reported that the modular approach to dividing the data and establishing hydrological forecasting Multi-Layer Perceptron (MLP) ANN models and optimizing the overall model structure assures accurate representation of the subprocesses constituting a complicated natural phenomenon. The present study focuses on the improvement of the ANN models and it describes the nonlinear relationships between optimal coagulant dosage and raw water characteristics such as turbidity, streaming current, conductivity, and pH. The coagulation process involves many complex physical and chemical mechanisms, which are difficult to model using the traditional methods. Cybenko [14] reported the neural approach is well suited for this type of problem and it offers the benefit of very short computational times and it is likely to depict some nonlinear relationships between system inputs and outputs. The goal of this study is to find out effects of different time intervals; effects of the inputs with and without the inherent factor; different inherent-predicting models on ANN estimation of the optimal coagulant dosage. Inherent factors indicate that past event will affect present event, for example, coagulant dosage at time t, D(t), will be affected by D(t-1). METHODOLOGY 1. Data Collection and ANN Analysis One of the earliest adaptive techniques in engineering and computing science is ANN. McCulloch and Pitts [15] developed the first model of a neuron as a computational device. Rosenblatt [16] implemented the first learning neural network. In the present research, ANN models are developed to assist water treatment plant operations in order to determine the real time coagulant dosage for drinking water treatment plant (2,700,000 m d) in Taipei, Taiwan. Poly aluminum chloride (PACl) was used as coagulant. The treatment processes include the coagulation, flocculation, sedimentation, filtration, and disinfection. In order to obtain the input and output data required to develop and validate ANN models, water samples are collected from drinking water treatment in Taipei, Taiwan. Desired water quality variables are determined, including the turbidity, streaming current, conductivity and pH values. The quality of treated water should follows the drinking water quality standards (DWQS) such as, the maximum of turbidity is 2 NTU; the range of the pH values is between 6.0 and 8.5 [17]. Collected data were screened by the limit values in DWQS. To improve the precision of ANN models, input variables are selected by the Pearson correlation coefficient. The Pearson correlation coefficient for each input and output dataset is calculated and shown in Table 1. The root-mean-square normalized error (RMSE) criterion is used as a performance index to compare the prediction capability of ANN models trained by each dataset. The RMSE criterion is known to be descriptive when the prediction capability among predictors is compared [18].