Application of artificial neural networks and genetic algorithms for crude fractional distillation process modeling

📝 Abstract

This work presents the application of the artificial neural networks, trained and structurally optimized by genetic algorithms, for modeling of crude distillation process at PKN ORLEN S.A. refinery. Models for the main fractionator distillation column products were developed using historical data. Quality of the fractions were predicted based on several chosen process variables. The performance of the model was validated using test data. Neural networks used in companion with genetic algorithms proved that they can accurately predict fractions quality shifts, reproducing the results of the standard laboratory analysis. Simple knowledge extraction method from neural network model built was also performed. Genetic algorithms can be successfully utilized in efficient training of large neural networks and finding their optimal structures.

💡 Analysis

This work presents the application of the artificial neural networks, trained and structurally optimized by genetic algorithms, for modeling of crude distillation process at PKN ORLEN S.A. refinery. Models for the main fractionator distillation column products were developed using historical data. Quality of the fractions were predicted based on several chosen process variables. The performance of the model was validated using test data. Neural networks used in companion with genetic algorithms proved that they can accurately predict fractions quality shifts, reproducing the results of the standard laboratory analysis. Simple knowledge extraction method from neural network model built was also performed. Genetic algorithms can be successfully utilized in efficient training of large neural networks and finding their optimal structures.

📄 Content

 Abstract—This work presents the application of the artificial neural networks, trained and structurally optimized by genetic algorithms, for modeling of crude distillation process at PKN ORLEN S.A. refinery. Models for the main fractionator distillation column products were developed using historical data. Quality of the fractions were predicted based on several chosen process variables. The performance of the model was validated using test data. Neural networks used in companion with genetic algorithms proved that they can accurately predict fractions quality shifts, reproducing the results of the standard laboratory analysis. Simple knowledge extraction method from neural network model built was also performed. Genetic algorithms can be successfully utilized in efficient training of large neural networks and finding their optimal structures. I. INTRODUCTION rtificial neural networks (ANN) as well as genetic algorithms (GA) are popular machine learning technologies. They were both designed in analogy to structures and processes occurring in nature. Due to their desired properties, they are widely applied to solve various problems of modeling and optimization [1-5]. Although those techniques are utilized mostly separately, together they can extend the range of their possible applications. They can be applied to problems, where it is difficult to recover a clear analytical solution [6,7]. The example can be crude fractional distillation, which is a highly nonlinear process [8]. It also includes many random disturbances caused by many unpredictable factors like mechanical or measurement problems [9]. Because of the above obstacles, data from such process can be perfectly suited to verify the behaviour and properties of artificial neural networks trained and optimized by genetic algorithms. A The aim of this work is to apply artificial neural networks, trained and structurally optimized by the genetic algorithm, to model laboratory quality measurements of main crude oil fractional distillation products. A. Artificial neural networks Artificial neural network is a structure built by many interconnected basic elements called artificial neurons. It resembles the natural tissues in brain, which consists from many nervous cells. Observation of the natural neurons behaviour resulted in uncovering its basic operation 
principles and interesting properties. The first mathematical description of artificial neuron was done by McCulloch and Pitts in 1943. Rosenblatt in 1957 refined his idea creating modern artificial neuron called simple perceptron [10,11]. Perceptron artificial neuron can be perceived as a transducer of many input signals giving one output. This led to deriving mathematical description of the neuron, which is presented on Fig. 1 and characterized by equation (1). y = F(∑ i=0 n (ui wi )+w0) (1) where: y – value of the output signal, F – activation function, n – number of input signals, u – value of input signal number i, w – weight value of the connection number i. Weights located on input connections are coefficients, which are set during learning process. They can resemble storage of gained knowledge. Every input value is multiplied by weight coefficient and added at the end. This sum is then an argument for activation function of choice. Activation function determines the properties of the artificial neuron. Typically the unipolar sigmoid activation function, defined by equation (2), is commonly used [10, 11]. Chart showing its curve was presented on Fig. 2. Results of this function are included in range from 0 to 1. β parameter in this equation is responsible for the sigmoid function steepness. Learning process is basically equivalent to shape Application of artificial neural networks and genetic algorithms for crude fractional distillation process modeling Łukasz Pater PKN ORLEN S.A. Płock, Poland, Faculty of Mathematics and Computer Science, Nicolaus Copernicus University, Toruń, Poland Email: pater.lukasz@gmail.com Fig 1. Artificial neuron – simple perceptron [11]. modification of the activation function. The result of this function is the output value, which can be final or become input for another artificial neuron [10,11]. F(x)= 1 1+e−β x (2) where: F – type of activation function, x – sum of weights multiplied by input values, β – sigmoid function steepness parameter.
In order to extend the single perceptron classification capabilities of nonlinear relationships, many artificial neurons are grouped into many layers, creating multi layer perceptrons (MLP). Layers between the input and output layer are called hidden layers. This class of artificial neural network is widely used for a great majority of problems, because it can model highly nonlinear data relationships. Structure of artificial neural network also has to be chosen carefull

This content is AI-processed based on ArXiv data.