Elsevier

Energy

Volume 161, 15 October 2018, Pages 361-369
Energy

Comparison of viscosity prediction capabilities of regression models and artificial neural networks

https://doi.org/10.1016/j.energy.2018.07.130Get rights and content

Highlights

  • Experimental optimization of transesterification reaction.

  • Production of waste cooking oil biodiesel having the lowest viscosity.

  • Effects of biodiesel fraction and temperature on biodiesel-diesel binary blends.

  • Viscosity prediction capabilities of regression models and ANN.

  • Rational model for viscosity changes vs. biodiesel fraction and temperature.

Abstract

Nowadays, biodiesel is seen as an alternative fuel to diesel fuel due to its many advantages such as higher density, cetane number and flash point. Although several methods are available for estimating fuel properties of biodiesel-diesel fuel blends, there is still the lack of works on the comparison of regression models and artificial neural networks (ANN) in predicting viscosities of the blends. Therefore, in this work, (1) optimum reaction parameters providing the lowest viscosity were determined for methanolysis of waste cooking oil, (2) waste cooking oil methyl ester was synthesized based on the determined optimum parameters, and it was mixed with diesel fuel on different volume ratios (3) viscosity measurements of the prepared blends were made at the temperature ranges between 273.15 K and 373.15 K, (4) changes in viscosity versus temperature and biodiesel fraction in blend were investigated and the rational model was proposed, finally (5) the predictive capability of rational model was compared to the three-term Vogel model, Bingham model and ANN by fitting to viscosity data measured by the authors and by Geacai et al. According to results, the measured values by the authors and Geacai et al. are the most accurately predicted by the rational model.

Introduction

In recent years, the development of renewable fuels is of increasing importance because of the problems associated with fossil fuels such as limitations in supply, intensified environmental degradation and their ever-increasing prices [[1], [2], [3]]. Among these fuels, biodiesel is considered to be a promising alternative for diesel engines, consisting a mixture of methyl or ethyl esters of long chain fatty acids by means of transesterification of natural triglycerides with a mono-hydroxyl alcohol (generally CH3OH and C2H5OH) in the presence of an eligible catalyst [[4], [5], [6]].

As a fuel, biodiesel has a number of benefits, compared to diesel fuel, such as biodegradable and non-toxic characteristics, domestic origin and higher density, flash point and cetane number [[7], [8], [9]]. Biodiesel can be easily mixed with diesel fuel at the different proportions to be used in diesel engines without large modifications [10,11]. It comprises of about 10–12% oxygen content (from fatty acids), resulting in a more advanced and faster overall combustion event, which decreases the emissions of about 14.2% HC and 26.8% PM, compared to diesel fuel [[12], [13], [14]]. Moreover, biodiesel possesses inherently greater lubricity than diesel fuel, resulting in longer engine component life. Adding biodiesel at even low levels (1–2%) improves the lubricity of ultra-low sulphur diesel fuel [15,16]. On the other hand, important shortcomings of biodiesel include lower volumetric heat content, poor oxidative and storage stability, inferior cold flow properties and higher viscosity, price and (generally) NOx emissions [[17], [18], [19]].

Kinematic viscosity is one of the key fuel characteristics. A fuel with high viscosity such as biodiesel tends to form larger droplet size, which causes poor fuel atomization and incomplete combustion, increase in engine deposits and greater fuel pump energy demand [20,21]. On the other hand, a fuel of very low viscosity can result in leakage or increased wear due to insufficient lubrication for the fuel-injection pumps [22]. Although simple equipment is generally used for measurement of biodiesel viscosity at a certain temperature, it takes time and is generally not practical especially when the measurements are performed for different biodiesel-diesel fuel blends at various temperatures [23]. Therefore, researchers often use regression models or artificial neural networks (ANNs) to obtain viscosity data for their thermodynamic or dimensional engine modelling studies. A number of models were presented for prediction of viscosity of biodiesel-diesel fuel binary blends in the existing literature such as: Kanaveli et al. [24] examined twelve mixing models (such as Bingham, Cragoe, Latour, Shu models etc.) developed to estimate viscosities of biodiesel-diesel fuel blends. Two hundred and thirty-one fuel blends including three diesel fuels and seven biodiesels (eleven blending ratio) were prepared for testing the prediction capabilities of the models. Some of the mixing models were modified by using alternative forms or new constants to improve their prediction capabilities. The most accurate estimation was obtained by the modified Shu and Barrufet & Setiadarma's mixing models. In the study performed by Lapuerta et al. [25], kinematic viscosities were measured at 40°C for ethanol and n-butanol blends with diesel and biodiesel. Stabilizing additives were not used for any blends. The best known equations (linear, Arrhenius [26], Kendall-Monroe, Bingham, Grunberg-Nissan and McAllister) were tested for the prediction of the viscosities of the blends in the whole range of alcohol contents (2.5%, 5%, 10%, 20%, 40% and 75% on volume basis). Grunberg-Nissan model provided more accurate correlation for butanol-biodiesel, ethanol-biodiesel and butanol-diesel blends while McAllister model with three interaction parameters achieved better fit for ethanol-diesel blends. In the study carried out by Geacai et al. [27], viscosity data for binary blends of biodiesel with benzene, toluene and diesel fuel measured in the temperature range of 293.15–K 323.15 K were reported. The experimental data were used to evaluate the accuracy of different models (Grunberg-Nissan, McAllister's two-parameter, Wilke and Andrade) for viscosity prediction. Grunberg-Nissan, including molecular interaction parameter (Gij), and McAllister models gave better results than simplified Grunberg-Nissan, not including Gij, and Wilke models for all binary blends systems. Gülüm and Bilgin [21] produced hazelnut oil biodiesel by means of transesterification reaction, and measured viscosities and densities of the hazelnut oil biodiesel-diesel fuel blends (B5, B10, B15, B20, B50 and B75) at 10, 20, 30 and 40°C. The exponential, rational and two-term power models were newly proposed for density-temperature, viscosity-blending ratio and viscosity-temperature variations. In addition, the models were compared the well-known models (linear, quadratic, Arrhenius type, one-term power and Liew models) using various viscosity data measured by different authors. According to results, the models proposed by Gülüm and Bilgin were found to the best predictors for density and viscosity predictions, compared to the well-known models. Aminian and ZareNezhad [28] proposed neural network-based models (i.e. support vector machine, an adaptive neuro-fuzzy inference system, and feed-forward neural network model trained by genetic algorithm, simulated annealing and Levenberg-Marquardt) to predict the viscosities of various biodiesels. The models based on neural network were compared to well-known models (Yuan, revised Yuan, Krisnangkura and Ceriani). The authors pointed out that the proposed models optimized by genetic algorithm and simulated annealing have more accurate for predicting the viscosity. Although a number of studies dealing with the development of ANN and regression models to predict viscosity of blends including vegetable oil, biodiesel, alcohol and diesel fuel have been also presented, there are still lacks detailed studies on comparing the predictive capabilities of ANN and new regression equations for the viscosity estimation of different diesel-biodiesel blends at several temperatures. Thus, in the present study, (1) waste cooking oil biodiesel (B100) was produced, and mixed with the commercial diesel fuel (DF) at different volume percentages. (2) Kinematic viscosities of the biodiesel-diesel binary blends were measured at various temperatures. (3) Changes in viscosity with biodiesel fraction in blend and temperature were assessed and the rational model, previously proposed by the authors in Ref. [21], was used to estimate their viscosity. (4) Finally, the predictive ability of the rational model was compared to ANN and Bingham (as a function of biodiesel fraction in blend) and three-term Vogel models (as a function of blend temperature), commonly used in the literature for estimating viscosities, for the viscosity data measured by the authors and by Geacai et al.

Section snippets

Biodiesel production

Methanol (99.80% purity), potassium hydroxide (pure grade), anhydrous sodium sulphate and waste cooking oil were used to synthesize biodiesel (B100) by means of transesterification. The important physicochemical properties of used waste cooking oil are listed in Table 1. As shown in this table, since free fatty acid and water contents of the oil meet DIN 51605 limiting value, only alkali-catalyzed transesterification was performed. In biodiesel production, 0.75%, 1.00%, 1.25% of catalyst

Effect of temperature

Viscosity is a quantitative measure of a fluid's resistance to flow [36] or it may considered as the integral of the interaction forces of molecules [37]. Temperature has a strong effect on viscosity compared to pressure [36]. Little increases occur in viscosities for gases and the most liquids with increasing pressure [36]. On the other hand, viscosity of liquids decreases with increasing temperature, whereas for gases vice versa [38]. This difference in the effect of temperature on viscosity

Conclusions

This work deals with (1) production of waste cooking oil methyl ester having the lowest viscosity, (2) measurement of kinematic viscosities of the waste cooking oil methyl ester-diesel fuel blends at different temperatures according to related international standards, and finally (3) development and comparison of regression models (rational, three-term Vogel and Bingham) and artificial neural network (ANN) in estimating viscosities of blends measured by the authors and Geacai et al. [27]. The

References (42)

Cited by (0)

View full text