Jules Thibault

Dynamic modelling of the activated sludge process: Improving prediction using neural networks

A procedure has been developed to improve the accuracy of an existing mechanistic model of the activated sludge process, previously described by Lessard and Beck [Wat. Res. 27, 963–978 (1993)]. As a first step, optimization of the numerous model parameters has been investigated using the downhill simplex method in order to minimize the sum of the squares of the errors between predicted and experimental values of appropriate variables. Optimization of various sets of parameters has shown that the accuracy of the mechanistic model, especially on the prediction of the dissolved oxygen (DO) in the mixed liquor, can be easily improved by adjusting only the values of the overall oxygen transfer coefficients, KL a. Then, in a second step, neural network models have been used successfully to predict the remaining errors of the optimized mechanistic model. The coupling of the mechanistic model with neural network models resulted in a hybrid model yielding accurate simulations of the five key variables of the activate sludge process.

A neural network methodology for heat transfer data analysis

Abstract Neural networks have been until very recently a topic of academic research. Recent developments of powerful learning algorithms and the increasing number of applications in a great number of disciplines suggest that neural networks can provide useful tools for modelling and correlating practical heat transfer problems. This paper presents an introduction to computing with neural networks. To evaluate the potential of neural networks for correlating heat transfer data, three different examples are solved, using a three-layer feedforward neural network. Two different learning algorithms, including the traditional backpropagation algorithm, are used to teach the neural network. It is shown that neural networks can be used to adequately correlate heat transfer data.

Process modeling with neural networks using small experimental datasets

Abstract This paper reports some work done to improve the modeling of complex processes when only small experimental data sets are available. Various solution strategies based on feed-forward and radial basis function (RBF) neural networks have been tested for three problems including two wood pulp applications. Experimental data sets obtained from D-optimal design and from a random selection throughout the experimental space were compared for their ability to lead to the proper model. In addition, the influence of activation functions, the number of levels in stacked neural networks and the composition of the training data sets have been studied. The study shows that designed training data sets are more desirable than random experimental sets, due to their higher orthogonality. The use of neural network is a powerful tool for modeling complex processes even when only a small set of data is available for training. However, special care must be exercised to insure that good predictive models are obtained.

Multi-objective optimization for chemical processes and controller design: Approximating and classifying the Pareto domain

In recent years, the development of multi-objective optimization techniques for simultaneously optimizing multiple and conflicting objectives have received wide attention in the literature. In this paper, three algorithms for generating the Pareto domain were studied for their efficiency to generate a well-defined Pareto domain as a first step in the development of a multi-objective optimization strategy. Twelve standard test cases, which have been used frequently in the literature, were considered along with two engineering problems, namely, the determination of optimum operating conditions for the production of gluconic acid and the determination of optimal tuning parameters for a PI controller. These multi-objective optimization problems were selected in order to evaluate the robustness and versatility of the algorithms studied. The results of three of these test cases and both engineering problems are presented. The results identify a robust optimization strategy that generates a Pareto domain using a dual population evolutionary algorithm and classifies it using net flow, a technique that incorporates the knowledge of an expert into the optimization routine.

Modelling of coagulant dosage in a water treatment plant

Abstract The coagulation-flocculation is a major step in the drinkable water treatment process allowing the removal of colloidal particles. The water treatment facilities of the City of Sainte-Foy have been well instrumented and process variables such as temperature, pH, turbidity, conductivity of raw and treated water along with actual coagulant dosage available have been measured and stored each 5 min for several years. Using such a data bank, the objective of this paper is to report on the development of a neural network predictor of coagulant dosage in order to facilitate process operation. Feedforward neural models have been built using a quasi-Newton method along with the early stopping approach to avoid overfitting. Annual and seasonal models have been built and their performances are discussed.

High-titer adenovirus vector production in 293S cell perfusion culture.

Human 293S cells culture for recombinant adenovirus production is traditionally carried out in batch at a maximum of 6 × 105 cells/mL. A previous report demonstrated that fed‐batch, applied to the adenovirus/293S cells system, improves the volumetric production of viral proteins by increasing the cell density at which cells can be infected, up to 2 × 106 cells/mL, without reducing the per‐cell yield of product. To increase this cell density limit, the adenovirus production was performed in a perfusion system where the cells were separated by means of a tangential flow filtration device. 293S cell growth to 14 × 106 cells/mL was achieved in 10 days, at a medium renewal rate of 1 volume of medium per reactor volume and day (VVD). For adenovirus production, three 293S cell cultures were perfused at 1 VVD in parallel and infected at an average density of 8 × 106 cells/mL. One of the cultures was set at 37 °C and the two others at 35 °C. After a rapid initial cell loss, the average cell density stabilized at 5.75 × 106 cells/mL, 12 h postinfection, which was 8 times higher than the cell density in the batch control. This allowed the production of 3.2 × 109 infectious viral particles/mL (IVP/mL) at 37 °C and 7.8 × 109 IVP/mL at 35 °C, this last result being 5.5 times higher than the control. To our knowledge, this nonconcentrated titer is the highest value that has ever been published for adenovirus vector production. These observations lead to the conclusion that perfusion is an efficient tool to maintain, at high cell density, a specific production rate level sufficient to increase significantly the adenovirus volumetric production. Furthermore, it shows that perfusion at 35 °C can improve viral titer by 2.4‐fold compared to 37 °C, in accordance with a previous study on adenovirus batch production.

Reassessment of the estimation of dissolved oxygen concentration profile and KLa in solid-state fermentation

Abstract Oxygen mass transfer in aerobic microbial growth systems is often a limiting factor for optimal growth and productivity. Oxygen mass transfer has been widely studied in submerged fermentations but has attracted as yet little attention for solid state fermentations. The parallel to submerged fermentation has led to the incorrect interpretation and use of the overall oxygen mass transfer coefficient ( K L a ) to assess the ability of a particular fermentation system to supply the oxygen to microorganisms. The use of K L a , as traditionally defined, should be used with caution in solid substrate fermentation systems because there is no convection on the liquid side of the medium, and oxygen is consumed in the biofilm. Hence, K L a must be redefined for solid state fermentation. In this paper, the use of oxygen mass transfer coefficients in solid state fermentations is clarified. Published literature data were analysed with a simple pseudo-steady-state model and used to discuss the influence of the biofilm thickness, the dissolved oxygen diffusion coefficient, the convective gas mass transfer coefficient, and the gas flow rate on the oxygen mass transfer coefficient in solid state fermentations.

SOLIDS TRANSPORTATION MODEL OF AN INDUSTRIAL ROTARY DRYER

ABSTRACT A complete simulation model has been developed for an industrial rotary dryer to account for the heat and mass exchange between the solids and the gas. This simulator is mainly composed of three models: solids transportation model, furnace model, and gas model. The solids transportation model is the modified Cholette-Cloutier model It consists of a series of interactive reservoirs which are subdivided into an active and a dead compartments to account for the characteristic extended tail of the residence time distribution (RTD) curves observed in industrial dryers. To expand the validity of the model, experiments have been performed in an industrial rotary dryer to obtain RTD curves under different mineral concentrate and gas flow rates. This paper describes these experiments and presents the variation of the average residence time and model parameters as function of solids and gas flow rates.

Adsorbent screening for biobutanol separation by adsorption: kinetics, isotherms and competitive effect of other compounds

Butanol, considered as one of the best renewable alternatives for gasoline, has attracted significant attention in recent years. However, biobutanol production via fermentation is plagued by the low final product concentration due to product inhibition. It is possible to enhance productivity by selectively removing biobutanol from the fermentation broth. Adsorption is one of the most promising and energy-efficient techniques for butanol separation and recovery. In the present study, different adsorbents were tested by performing kinetic and equilibrium experiments to find the best adsorbent for butanol separation. Activated carbon (AC) F-400 showed the fastest adsorption rate and the highest adsorption capacity amongst ACs and zeolites tested. AC F-400 also showed the highest affinity toward butanol and to a lesser extent for butyric acid whereas its adsorption capacity for the other main components present in acetone–butanol–ethanol fermentation broths was very low. In addition, the butanol adsorption capacity was not affected by the presence of ethanol, glucose and xylose while the presence of acetone led to a slight decrease in adsorption capacity at low butanol concentrations. On the other hand, the presence of acids (acetic acid and butyric acid) showed a significant effect on the butanol adsorption capacity over a wide range of butanol concentration and this effect was more pronounced for butyric acid.

pH prediction and final fermentation time determination in lactic acid batch fermentations

Abstract For the control of lactic acid batch fermentations, the prediction of the pH over a long time horizon and the determination of the final fermentation time are very useful information. Lactic acid batch fermentations, with uncontrolled pH, are modelled with a feedforward neural network to generate the model (pH versus time) of a reference fermentation of Streptococcus thermophilus strain on skim milk. In essence, the feedforward neural network is used as a general nonlinear model to store the information of a series of well-behaved fermentations. This neural network stores the information of previous fermentations and defines a reference fermentation. This reference fermentation is used to perform a comparison with the actual fermentation for the on-line prediction of future pH values and of the final fermentation time. This time, which occurs at a predetermined pH, is predicted with an accuracy of less than 20% at the onset of fermentation and with a much better accuracy as the fermentation proceeds. The future pH values are predicted with a mean error of 0.05 pH for a 3 hour prediction horizon. The prediction is obtained with four geometrical methods by sliding the curve of the reference fermentation along the curve of the actual fermentation. The procedure for sliding the reference curve depends on the geometrical method used. The distinct advantage of neural networks over other class of models is the plasticity of its structure which allows to easily capture the shape of the fermentation curve.