Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Modelling of a water treatment plant operation

Gyparakis Stylianos

Full record

---

Summary

The monitoring of the main operational variables and water quality characteristics of a Water Treatment Plan (WTP) is a critical issue for all WTP operators, for the production of human consumption (drinking) water. WTP operators often seek a quick, reliable and easy-to- use method for predicting the dosages of the water treatment chemicals used, which is their main daily concern.This PhD thesis focuses on the modeling of a WTP operation, using ANN and MLR analysis models. Also, examines the water quality characteristics and the main WTP operational variables correlation, their interaction and mainly focuses on the prediction of the necessary water treatment chemicals dosages in a WTP. The innovation of this study lies in the fact that it takes into account the extended experience of a WTP operator.In this Thesis, the studied case study comes from the Aposelemis WTP, in which a prediction of the output variables of the ANN and MLR models is made, regarding the required dosages of the water treatment chemicals, based on the observed water quality and other operational variables. The estimated main operational variables of the WTP include: the dosages of residual ozone (O3), anionic polyelectrolyte (ANPE), polyaluminum chloride sulfate (PACl) and chlorine gas (Cl2(g)). Daily water sample analysis results and recordings from the WTP SCADA, covering a period of 38 months (1,188 values for each of the 14 measurable variables), were used as input parameters for the ANN modelling. Specifically, the input parameters of the ANN model include: the raw water flow (Q), the raw water turbidity (T1), the treated water turbidity (T2), the residual free chlorine of the treated water (Cl2), the residual aluminum concentration of the treated water (Al), the water turbidity at the inlet of the filtration beds (T3), the daily water height difference in the reservoir of the Aposelemis dam (ΔH), the raw water pH value (pH1), the treated water pH value (pH2) and the daily electricity consumption (El) at the Aposelemis WTP. The output parameters of the ANN include: the concentration of residual ozone (O3) after the ozonation process, the dosage of anionic polyelectrolyte (ANPE), the dosage of polyaluminum chloride sulfate (PACl) and the supply of chlorine gas (Cl2(g)).A total of 304 different ANN models were constructed and based on the best value of the test performance index (tperf) of them, the scenario with 100 neural networks, 100 nodes, 42 hidden nodes, 10 inputs and 4 outputs was finally selected. This ANN model achieved very good simulation results, which suggests that ANNs are potentially useful tools for predicting the main WTP operational variables.Also, four (4) different scenarios were examined using Multiple Linear Regression Analysis (MLR) with dependent variables: the residual ozone (O3), anionic polyelectrolyte dosage (ANPE), poly-Aluminum chloride sulfate (PACl) dosage and chlorine gas flow (Cl2(g)), as well as were used ten (10) independent operational and water quality variables.According to the results of R2 and R, the ANN model had a better performance compared to the MLR analysis model for predicting the dosages of the used water treatment chemicals. Based on the criterion R²> 0.5, the ANN performance was satisfactory in predicting the dosages of the three water treatment chemicals: ANPE (R2= 0.772), PACl (R2= 0.742) and Cl2(g) (R2= 0.838, +23% compared to the corresponding value of the MLR model and R= 0.95, +11% compared to the corresponding value of the MLR model). Accordingly, the prediction of the MLR model was evaluated as satisfactory for predicting the dosage of Cl2(g) only (R2= 0.681, R= 0.82500).According to the RMSE results, the MLR model performed better for three (RMSEANPE= 0.05 mg/L, RMSEPACl= 0.08 mg/L and RMSECl2(g)= 0.10 kg/h) of the four dependant variables (drinking water added chemicals), than the ANN model, which performed better for only one water treatment chemical (RMSEO3= 0.02 mg/L).In general, if someone wants to use the scenarios and prediction models (ANN or MLR) to predict Cl2(g) dosages, then it is preferable to use the one with the smallest RMSE. If one is interested in having a future prediction of water treatment chemical dosages with a prediction model with the best fit, then it is preferable to use the model with the largest R2 value. Also, the ozone dosage variable (O3) presented low R2 values, in all cases, probably due to the large variation of its values.This study further reinforces the point of view that ANNs are useful decision support tools for a WTP operator, which can simulate with great accuracy and adequacy the decisions regarding the dosages of the water treatment chemicals used, which is the main and daily concern of the operator of such a facility.It is recommended future research to be conducted to further increase knowledge on the prediction of water treatment chemicals, using models such as ANNs, as accurate prediction models, and MLR analysis models as flexible, fast and reliable prediction models. In particular, further research could be conducted on the prediction of chemicals used in a WTP, using ANNs with a smaller number of variables to ensure greater flexibility, without, however, substantially reducing the reliability of the prediction model.This will enable to establish ANNs as forecasting models in the water sector and in the daily operation of the WTPs. In addition, future research could include investigating the use of other comparison criteria, such as MAE (Mean Absolute Error), MAPE (Mean Absolute Percentage Error) or the NSE (Nash-Sutcliffe) performance index. Furthermore, it is suggested to conduct sensitivity and uncertainty analyses on the most influential variables, which could further improve the modeling process. Finally, since the main limitation of the current study is that the models have been trained with data from a single WTP, it is suggested as future work to include data from more corresponding WTPs, in order to increase the robustness of the models and their universal applicability.

Available Files

Services