whole_SasanananSetta2009_thesis.pdf (17.19 MB)
Water treatment plant clarifier control : an artificial intelligence approach
thesisposted on 2023-05-27, 16:19 authored by Sasananan, S
The water treatment industry is currently facing growing pressure to provide efficient treatment at economical cost. The clarifier is one of the most important units in a water treatment plant. The impurity removal efficiency of the clarifier relies heavily on the skills of the operators and the fluctuation of raw water quality. The behaviour of the clarifier is complex and not able to be explained by simple mathematic equations. The goal of this thesis was to develop and evaluate intelligent control methods for clarifier control. The intelligent control system proposed conceptually controls the clarifier by mimicking human operator control. The control objectives here were not only to minimise the clarified water impurity but also minimise the operational cost. The intelligent control system developed in this thesis was an integrated system employing an Artificial Neural Network (ANN) and Genetic Algorithms (GAs), which were linked by Model Predictive Control (MPC) architecture. The intelligent control system was composed of three components: (i) an ANN clarifier process model, (ii) an Multiobjective genetic algorithm (MOGA) optimiser and (iii) a decision maker. The system was designed to alleviate the operational problems of clarifier control and help the operators to choose suitable control actions. The clarified water qualities and operational cost were mathematically taken into consideration for control action optimisation via the MOGA optimiser. Two case studies of solid contact clarifier operation were investigated in this thesis, at Bryn Estyn water treatment plant (BEWTP), Hobart, Australia and Bang Khen water treatment plant (BKWTP), Bangkok, Thailand. The ANN clarifier process models were developed using past operational data from both case studies. Past operational data were divided into training, validation and testing sets. To ensure the statistical equivalence among these three sets, clustering methods based on Self-Organising Map (SOM) networks were employed. The optimal temporal spans and architectures of the models were found by trial and error using the associated testing set. For the BEWTP case study, excellent predictive performance of clarified water colour (one time step ahead) is demonstrated with a coefficient of correlation (r\\(_2\\) ) of 0.88 and Mean Absolute Error (MAE) of 0.50 HU. There is also excellent prediction of clarified water turbidity, evidenced by coefficient of correlation (r\\(_2\\)) of 0.89 and Mean Absolute Error (MAE) of 0.11 NTU. For the BKWTP case study, the ANN model showed good performance prediction (one time step ahead) with a coefficient of correlation (r\\(_2\\)) of 0. 71 and Mean Absolute Error (MAE) of 0.65 NTU. These ANN models were reliable since their prediction errors were of the same magnitude as measurement errors. However, prediction further than one step ahead was not recommended since this resulted in prediction errors that were larger than input measurement errors. The intelligent control system was simulated using the testing set in order to assess its performance. Control actions were optimised to minimise clarified water impurities and operational cost by the MOGA optimiser and help the decision maker to choose the best set of control actions using the shortest normalised distance from the utopia point. However, defining of MOGA optimiser parameters was a prerequisite. Some parameters were chosen using guidelines from historical research, while the population size and number of generations were defined using a trial and error process. It was found that the optimal number of generations and population size was 200 and 80 respectively for the BEWTP case study and 400 generations and 100 populations for the BKWTP case study. According to the simulation results, all clarified water qualities complied with their operational targets and the operational cost was reduced by 8.3 percent and 3.4 percent for BEWTP and BKWTP respectively. A full-scale pilot plant test was conducted at BKWTP in Bangkok. Unfortunately, it was not possible to conduct a pilot test at BEWTP due to health and safety regulations. The full-scale pilot plant test was for verifying the performance of the proposed control system in real life situations and in comparison with the performance of human operators. One BKWTP clarifier was chosen to be controlled by the intelligent system and the other was selected for normal control by the human operators. For a period of about one month from August to September 2007, the full-scale pilot plant was run continuously with both intelligent and human clarifiers running simultaneously. In terms of operational cost, the intelligent control system ran the clarifier with a 2.4 percent saving over that of the human controlled clarifier. If the intelligent control system was installed for all clarifiers at BKWTP and they ran with full production capacity of 3.5 million cubic metres per day, the operational cost saving would have been about $AUS 153,000 per annum. The reduction in operator salary costs was another saving which was not included. In terms of water impurity removal, clarified water with mean turbidity of 6.4¬¨¬±1.5 NTU discharged from the intelligent clarifier which was found to comply with the operational target of seven NTU. However the human operator performed slightly better with mean clarified water turbidity of 6.3¬¨¬±1.5 NTU. The novel aspects of this work are in the establishment and implementation of intelligent control systems which are a combination of ANN modelling and a MOGA optimiser. Its performance is also assessed in real practice. This is the first time that an intelligent approach fully mimics how human operators control the solid contact clarifier. All of the control actions (i.e. both chemical and physical control actions) are used by the intelligent control system to minimise the clarified water turbidity and operational cost. The contributions based on the result of this research work include the following items. ‚Äö ANN process model method and the unique use of the SOM are successfully developed based on the particular operational data of each of these two case studies. ‚Äö To achieve the best prediction of the clarifier process model, there is a need for using both present and past (temporal) data. Although this was expected, the study shows exactly how long a period of data is needed to optimise the control actions. The optimal model architecture of each process model is obtained based on the operational data, and their time lags are 8 hours for BEWTP and 12 hours forBKWTP. ‚Äö The prediction of the clarifier process model deteriorates when the prediction is taken outside the training domain. One time step ahead (4 hour) prediction is reliable when compared with the measurement error. However, multiple steps (long range) prediction is not in favour of the ANN process model because errors accumulate over each time step. ‚Äö When the intelligent control system works outside its training domain, it has been shown that its performance is less satisfactory than when working inside its training domain since the performance of the intelligent control system depended upon the predictive performance of the clarifier process model. ‚Äö In this thesis, the \ill posed problem\" is avoided by the division of operational range according to raw water turbidity. Therefore the set of control actions were optimised in the specifically defined ranges of raw water turbidity. This method has been shown to be effective in real life application. In conclusion with the successful results of the pilot plant test both the operational targets of high water quality and cheaper operational cost are achieved using intelligent control. The intelligent control methods are proved able to work in real practice. This will overcome the limit of the manual skill of human operators and frees the clarifier operation from human error."
Rights statementNo access or viewing until 15 November 2011. Thesis (PhD)--University of Tasmania, 2009. Includes bibliographical references