STOCHASTIC MODEL AND EXPERT SYSTEM FOR MUNICIPAL ACTIVATED SLUDGE PLANT CONTROL

Abstract

Activated sludge process has been widely accepted and practised as a standard treatment method for municipal wastewater. Albeit numerous research work put in since its inception for the improvement of process efficiency and minimisation of process inhibitions, much more work is still underway to achieve cost effective control automation especially with regard to dynamically varying input conditions. Recent time has witnessed a steep rise in population and industrialisation rendering wastewater treatment plant control extremely cost prohibitive. Fortunately over the past few decades widespread revolution in the fields of information technology and communication with the support of graphics, artificial intelligence, fuzzy logic programs promises a very important role in providing cost effective control system for wastewater treatment plants. This research work attempts to develop an expert system based control package that makes use of stochastic modelling of historical plant data to provide relative easiness in providing preventive operational controls in municipal wastewater treatment plants. Earlier works in activated sludge plant, were mainly concerned with development of steady state models for operation and control of the plants incorporating the basic bio-kinetic relationship between the different process variables. Notable contributions in this direction include J. Monod, Eckenfelder, McKinney, Lawrence and McCarty, Gaudy et al. [46]. These models were generally derived from the experience in the application of the process based on specific information collected from laboratory or pilot plant studies. An efficient time variant and adaptive control mode requires time variant dynamic models. A large number of dynamic models have been proposed in the literature. Such models are based on the mass balance differential equations which are commonly very complex, contain a large number of reaction constants and coefficients which are either unknown or known only approximately and require a large calibration effort as exemplified by the models proposed by Henz et al. [53]. However, The timeseries models developed from monitoring input-output data-series offer an efficient alternative, which does not require a detailed prior knowledge of the system [88], Introduction of the ARMA stochastic modelling methodology can be attributed to Box and Jenkins [27], Numerous applications of ARMA models can be cited in the area of hydrology [64] among others, while only a few have been reported in wastewater treatment plant modelling. Major works in this direction worth mentioning include Berthouex et al. [13], Capadoglio et al. [31], Novotony et al. [88], Capadoglio [32] etc., have reported development of wastewater treatment plant operational control using advanced time series models in combination with neural network modelling technology without a priori knowledge of mathematical inputoutput relationships. The use of artificial intelligence in diagnosing and controlling problems associated with wastewater treatment plant is a recent development in the wastewater engineering field. The expert system approach has the potential to resolve many problems frequently encountered by wastewater treatment operators. Some interesting and instructive efforts on wastewater treatment plant expert systems are Berthouex et al. [18, 21], Ladiges and Kayser,[66, 67], Tong et al. [121], Hirokama and Tsumura [55], Capadoglio [32], For most practising engineers, the planning, design and control ofa municipal wastewater treatment plant is a straight forward task involving average daily flows, hourly composite characteristics (or assumed values from regional figures in the absence of data) and design criteria available in various text books. Most of the present wastewater treatment plants have been designed and controlled using steady state models and parameters obtained from steady state treatability studies. Wastewater treatment plants, however, receive fluctuating characteristics and quantity of wastewater and accordingly it is not surprising that the effluent characteristics from the treatment plants vary substantially and frequently violate the effluent standards. Most of the plants are operated and controlled by intuition rather than by standardised procedures, which more often result in fluctuating efficiency of plants. In most cases the absence ofproper and timely monitoring in the plants also makes them less reliable and cost prohibitive. Another important characteristic of wastewater treatment plant operated in developing countries is that the operating personnel lack awareness of ill effects of malfunctioning of the plants. The plants also do not have n proper maintenance schedules and receive inadequate funds for proper maintenance of the system. Real time wastewater treatment plant data were collected from two different municipal wastewater treatment plants. The first plant caters to a large metropolis in India, with a designed capacity of 18 MLD wastewater. The second plant also had similar capacity but caters to a relatively smaller population. Either plants have been found to be presently overloaded than the designed capacity and are working supposedly on steady state flow conditions. The composite data pertaining to each day from hourly samples of influent parameters namely BOD, suspended solids and flow, operational data namely MLSS and effluent parameters namely BOD and suspended solids were collected from both the plants corresponding to period of one year. The data collected were considered as the base samples for analysis and model validation. As a prelude to development of expert system, the data collected from the wastewater treatment plants were subjected to detailed analysis in order to characterise the fluctuations in the plant parameters. Initially the series obtained for each parameter was tested for stationarity. Nonstationary series were made in to stationary series. The stationary series so obtained was then tested for any trend component and eliminated from the original data series. The trend free series was then tested for periodicity and the harmonic components were removed reducing the original series in to random variations. The residual data series was experimented with different forecasting models and most appropriate predictive model was selected for each of the parameters using minimal MSE as the evaluation criteria. The forecast model was subsequently tested for adequacy using the residual ACF and x2 test. The program was developed in C language with Unix environment for all the above computations and selection of appropriate forecasting model along with statistical evaluation. For the data collected from the first plant, initially ACF and PACF functions were plotted and found to posses significant correlation with respect to historical data. All the parameters are observed to have trend components and harmonic components. The residual series obtained after elimination of trend and harmonic components were 111 fitted with different forecasting models namely, AR and ARMA models of different orders. Based on the minimal MSE criteria the most appropriate model was selected. It was observed that for Influent flow ARMA(2, 0), SS ARMA(1, 0), BOD ARMA(1,1), effluent SS ARMA(2, 0), BOD ARMA(1,1), and MLSS ARMA(1, 0) give a reasonably accurate estimations. The individual coefficients of the model were also tested for significance using t-statistic. The adequacy of the model was further tested with residual ACF andx2 test Similar computations were also carried out for the data collected from the second plant and the results were observed to becomparable with those obtained for the first plant. The estimated values for each ofthe parameters were subsequently designed to form the input for the Expert System developed. The Expert System for the operational control of wastewater treatment plant was developed using commercially available expert system shell. The rule base was designed with domain knowledge maintained in IF - THEN format. In all 110 rules depicting major operational problems were incorporated with each rule having 4to 11 conditional statements. The knowledge base was developed from data accumulated through personal interviews of experts and also from operational manuals and literature. The knowledge base was categorised on the major problems pertaining to aeration tank, secondary settling tank and effluent quality. Two external programs were developed to estimate control variables such as F/M ratio and MCRT and were linked with the expert system developed. The developed Expert System was tested with actual data collected from the plant during the period of operational problems and outputs of the Expert System was found to be in close agreement with the actual control action undertaken in the plant. IV