Elsevier

Applied Soft Computing

Volume 55, June 2017, Pages 402-412
Applied Soft Computing

Full length article
A hybrid computational intelligence framework in modelling of coal-oil agglomeration phenomenon

https://doi.org/10.1016/j.asoc.2017.01.054Get rights and content

Highlights

  • Optimizing coal-oil agglomeration process by genetic programming-support vector regression.

  • Five inputs are oil dosage, agitation speed, agglomeration time, temperature, and pH.

  • Formulated organic matter recovery (OMR) models fit the experimental data.

  • 2-D and 3-D parametric analysis of the formulated models monitors OMR of the process.

  • Controlling pH is vital for optimizing the OMR process.

Abstract

The phenomenon of Coal–Oil agglomeration for recovering the coal fines by agitating the coal-water slurries in oil is often practised by coal industry to ensure a safe and healthy environment. Experimental procedure for implementing this phenomenon is complex which involves three main mechanisms: crushing, ultimate and proximate analysis. Past studies have often focused on studying this phenomenon by the application of statistical modelling based on response surface designs. The response surface designs hold an assumption of pre-definition of the model structure, which may introduce uncertainty in the predictive ability of the model. Alternatively, the computational intelligence approach of Genetic programming (GP) that evolves the explicit models automatically can be used. However, the effective functioning of GP is often affected by its nature of producing the models of complex size. Therefore, this work develops a hybrid computational intelligence approach namely, Support vector regression-GP (SVR-GP) to study the coal-oil agglomeration phenomenon. Experimental studies based on five inputs, namely, oil dosage, agitation speed, agglomeration time, temperature, and pH are used to measure the organic matter recovery (OMR (%)) from the coal water slurries. A hybrid computational intelligence approach of SVR-GP is proposed in formulating the relationship between OMR (%) and the five inputs. The performance comparison and validation of the SVR-GP model is done based on the coefficient of determination, root mean square error and mean absolute percentage error. 2-D and 3-D surface analysis based on parametric and sensitivity approach is then conducted on the proposed model to find the relevant relationships between OMR (%) and inputs. The findings suggest that the pH of coal slurry has a significant effect on the OMR (%) and hence is important for reducing coal waste generation and promoting a cleaner environment.

Introduction

The world has witnessed the drastic climate change over the years with deteriorating environmental conditions. This significant change is attributed to the excessive energy consumption in various industries. With shortage of natural resources, energy conservation and environmental protection are main focus of governments across the globe and is seen as a problem of national interest [1]. A main contribution to energy consumption comes from the burning of coal, which is reliable, affordable and safer to be transported in comparison with other fossil fuels [2], [3].

Consequences of excessive usage of coal in industry sectors are severe, resulting in momentous generation of coal waste which is seen as fly ash or coal-water slurry thereby creating a polluted environment. Long term consequences endanger the longer survival of human and plant life around the mining fields. Thus, the reduction in generation of coal waste is essential which can be achieved by reducing the coal-water slurries through extraction of coal fines [3], [4].

From this perspective, coal-oil agglomeration as an increasingly popular phenomenon is used for recovery of coal fines and also affects beneficiation of coal. The mechanism of the phenomenon is depicted in Fig. 1. The agglomeration mechanism is based on the fact that as oil is dispersed in the water containing solid particles, the immiscible liquid will wet some solid particles. On agitating the mixture, the liquid ones will get adsorbed on solid surface resulting in an increase of the particle size [4]. In the case of coal-oil agglomeration, the coal particles containing minerals increase in size without any change in minerals. As the size of coal particles decreases, the coal-oil agglomerates become stronger. In brief, this phenomenon can easily process low ranks and oxidized coal and enlarges the size of coal fines thus making it easier to recover and also increases its calorific value. In this phenomenon, the retained mineral matter gets separated out from the coal matrix. However, the process of implementation of coal-oil agglomeration is complex as it involves five main inputs, namely oil dosage, agitation speed, agglomeration time, temperature, and pH which affect the organic matter recovery (OMR) from coal water slurries [3], [5], [6], [7], [8], [9], [10]. The objective should be to maximise the OMR which is only possible by suitable adjustment of the inputs. To achieve this objective, there is a strong desire to construct an explicit and meaningful model.

Physics based models are difficult to be formulated because this phenomenon is complex involving multiple inputs and hence difficult to be understood. An alternative route would be formulating models based only on the given data [11], [12], [13], [14], [15], [16]. Statistical methods such as design of experiment methods (Taguchi design and response surface methodology (RSM)) have been applied [6], [7], [8], [9], [10], [17], [18], [19], [20], [21]. However, these methods rely on some assumptions (pre-definition of structure of the model and the non-correlated residuals) and build on the entire data base without consideration of testing of method on the test data samples [22]. This means, the statistical models can do well on the training data but the performance is highly uncertain on the data sample beyond the given sample range.

Alternatively, computational intelligence (CI) approaches like genetic programming (GP), artificial neural network, and support vector regression (SVR) can be applied to model the process parameters [23]. These methods do not require statistical assumptions and complement the limitations of statistical methods such as RSM and ANOVA. Among these methods, GP has the ability for evolving the explicit models based only on the given data [23]. The functioning of SVR [24] is based on structural risk minimization (SRM) principle, which is responsible for imparting good generalization ability. However, the problem with SVR is that it cannot produce the functional expression of the model [25]. On the other side, the problem with GP is that it is prone to the over-fitting problem and therefore needs extensive control on settings of its parameters [26]. The size of the GP model can be controlled by imposing constraints on its parameters such as the depth and number of the nodes. However, it then does not assure a free and self-adapting ability of the GP model to fit and generalize the given data [27]. The past studies [22], [23], [24], [28] on hybrid methods of ANN performed by authors and related experts in this field have also argued on the same. An extensive review [29] performed on System Identification of the processes has also reveal that the hybridization of the GP approach with other potential CI methods could improve the computational accuracy of the GP model [30]. This has motivated authors to work on developing a holistic hybrid approach that can be used to model the relationships between OMR and the five inputs accurately.

From this literature, the following research questions require thorough investigation.

  • (1)

    What is the relationship between maximum organic matter recovery and the five inputs (oil dosage, agitation speed, agglomeration time, temperature, and pH) of the coal-oil agglomeration phenomenon? What methodology should be adopted to model such a complex relationship?

  • (2)

    How does OMR change with respect to each of the inputs of the coal-oil agglomeration phenomenon?

  • (3)

    What should be the appropriate value of the five inputs (oil dosage, agitation speed, agglomeration time, temperature, and pH) of the coal-oil agglomeration phenomenon that maximises the OMR from coal?

In this context, this work proposes a hybrid computational intelligence approach, namely, SVR-GP that combines the advantages and features of both methods. A step by step procedure for modelling OMR as a function of five inputs is shown in Fig. 1. Firstly, the experiments are performed to measure OMR (%) from the given coal samples. The data obtained from the experiments is then normalized and the two CI methods (SVR-GP and GP) are applied to extract the relationships between OMR (%) and five inputs. The experimental studies on coal-oil agglomeration phenomenon for measuring OMR based on the five inputs are discussed Section 2. Section 3 introduces the proposed hybrid SVR-GP CI approach and its parameter settings. Section 4 presents the performance analysis of the proposed model based on cross-validation, hypothesis tests and three error metrics. Section 5 illustrates the 2-D and 3-D surface analysis of the SVR-GP model. In Section 6, the conclusions arising from the study are discussed.

Section snippets

Experimental database of coal-oil agglomeration

The experimental details (set up, coal processing, etc) of the coal-oil agglomeration phenomenon are referred from the study conducted by Kumar et al. [3]. The coal used in this study was obtained from southern eastern coal fields ltd of Chhattisgarh (India). This type of coal is extensively used in thermal power plant located in Korba. With huge demand of electricity, the usage of coal becomes excessive thus resulting in momentous generation of coal-water slurries and fly ash. This excessive

Numerical modelling by hybrid heuristic optimization approach

To understand the notion of hybrid SVR-GP approach, each framework is firstly discussed. The framework of SVR is developed on the principle of statistical learning theory [31]. It is also referred to as the extension of support vector machine algorithm, which is used to solve classification problems. The ability of SVR in modelling complex symbolic regression problems has been found [31]. The main advantage of SVR over the traditional soft computing methods is that it is not based on any

Performance evaluation of the hybrid OMR models

The performance analysis of the two OMR (%) models formulated from the two methods (GP and SVR-GP) is evaluated against the actual data (Table 1) based on the following metrics:CoeffecientofDetermination(R2)=(i=1n(AiAi¯)(MiMi¯)i=1n(AiAi¯)2i=1n(MiMi¯)2)2Rootmeansquareerror(RMSE)=i=1N|MiAi|2NMeanabsolutepercentageerror(MAPE)(%)=1ni|AiMiAi|×100

Table 2 shows the values of error metrics (R2, RMSE and MAPE) of the GP model and the SVR-GP model on the five data sets obtained

2-D and 3-D analysis and the relationships between OMR and inputs from the SVR-GP model

This section performs the 2-D and 3-D surface analysis on the obtained SVR-GP model. The mathematical formulae and the notations used for the analysis are discussed [33], [34], [35], [36]. The OMR (%) values are computed from the SVR-GP model using the values of these given inputs. Fig. 6 shows the results of the 2-D analysis which illustrates the main effect of each of the input on the OMR (%). It clearly shows that the OMR (%) increases linearly up to oil dosage of 35% and after this, it

Conclusions

This study highlights the need for optimizing the OMR of coal-oil agglomeration phenomenon for promoting a cleaner environment. The novelty of the present work lies in the proposition of a hybrid SVR-GP model for the optimization of coal-oil agglomeration phenomenon. Experimental studies conducted measures the OMR (%) based on the oil dosage, pH, temperature, agitation speed and agglomeration time. Performance comparison analysis of the GP model and the SVR-GP model on the measured data

Acknowledgments

The study was supported by projects ref. M4061473 and M060030008 in Nanyang Technological University, Singapore.

References (41)

Cited by (13)

  • Prediction of electron beam weld quality from weld bead surface using clustering and support vector regression

    2023, Expert Systems with Applications
    Citation Excerpt :

    The superior performance of SVM is due to the enhanced generalization capability, ability to avoid local minima and sparse representation of the solutions. It is to be noted that neural network (NN) works based on the Empirical Risk Minimization (ERM) principle, whereas SVM uses the principle of Structural Risk Minimization (SRM) (Garg et al., 2017). The above-mentioned optimization problem could be solved by introducing a dual set of variables and constructing a Lagrange function.

  • Hierarchical polynomial-based fuzzy neural networks driven with the aid of hybrid network architecture and ranking-based neuron selection strategies

    2021, Applied Soft Computing
    Citation Excerpt :

    A practical and feasible method is to use experimental data to model the system and determine the effective input–output relationship [1,2]. As a kind of effective modeling methodology, hybrid architecture using computational intelligence (CI) [3–5] such as fuzzy set, neural networks, and granular computation has gotten more and more attention recently. And it has a far-ranging of applications in the fields of environment, energy, medical imaging, etc. [6–8].

  • Application of hybrid support vector regression artificial bee colony for prediction of MMP in CO<inf>2</inf>-EOR process

    2020, Petroleum
    Citation Excerpt :

    Other correlations such as Alston et al. [15], Sebastian et al. [16], Eakin and Mitch [17], Lee [18], Dong [19], and Yuan et al. [20] have focused directly on specific oil reservoirs, and this cannot satisfy the level of comprehensiveness and generalization required by various other oil reservoirs where the characteristics are different. A good number of studies have been carried out on the application of various artificial intelligence (AI) techniques such as artificial neural network (ANN), gene expression programming (GEP) and support vector regression (SVR) in many fields of science and engineering [21–30]. In petroleum and reservoir engineering, these techniques are becoming more attractive in resolving many problems of practical importance such as well-test analysis [31], PVT parameters estimation [32–34], history matching [35], WAG optimization [36], bottom hole pressure estimation [37], reservoir characterization [38–40] and others.

  • An integrated computational intelligence technique based operating parameters optimization scheme for quality improvement oriented process-manufacturing system

    2020, Computers and Industrial Engineering
    Citation Excerpt :

    Although the CI techniques have the potential to provide more accurate and flexible surrogate model for complex industrial process, the inherent demerit (i.e., black box) and increasing complexity (e.g. deep learning model) increase the difficulties on interpretation of underlying industrial process and seeking a solution for optimization formulation. Fortunately, GP, as one of CI approaches, has the ability for evolving the explicit models based only on the given data (Garg et al., 2017). To the best of our knowledge, there exists limited research on GP method for operating parameters optimization, especially for process manufacturing system (Bagheri, Nazari, & Sanjayan, 2019; Ghimire, Deo, Downs, & Raj, 2019).

  • Fine coal recovery from washery tailings in Turkey by oil agglomeration

    2018, Powder Technology
    Citation Excerpt :

    Moreover, reduction of droplet size of oil because of enhanced emulsification improved the recovery. Similar findings were reported in the same range of agitation rates [33,40,53]. Removal of ash and sulphur were also improved generally with higher agitation rates despite a slight decrease in ash rejection after 1400 rpm.

View all citing articles on Scopus
View full text