Accepted Manuscript

Productivity modelling of a developed inclined stepped solar still system based on

actual performance and using a cascaded forward neural network model

Mohammed Shadi S. Abujazar, Fatihah Suja, Ibrahim Anwar Ibrahim, A.E. Kabeel,
Suraya Sharil

PII: S0959-6526(17)32082-6

DOI: 10.1016/j.jclepro.2017.09.092

Reference: JCLP 10597

To appear in: Journal of Cleaner Production

Received Date: 01 August 2017

Revised Date: 03 September 2017

Accepted Date: 09 September 2017

Please cite this article as: Mohammed Shadi S. Abujazar, Fatihah Suja, Ibrahim Anwar Ibrahim, A.
E. Kabeel, Suraya Sharil, Productivity modelling of a developed inclined stepped solar still system
based on actual performance and using a cascaded forward neural network model, Journal of
Cleaner Production (2017), doi: 10.1016/j.jclepro.2017.09.092

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Highlights

 The performance of the stepped solar still using a cascaded forward neural network is
presented
 The effect of environmental parameters on the productivity of the system is studied
 In this study, three statistical error terms are used to evaluate the proposed model
 The RMSE, MAPE and MBE values of the proposed model are 22.48%, 18.51% and -
26.46%, respectively
 ACCEPTED MANUSCRIPT

1 Productivity modelling of a developed inclined stepped solar still system

2 based on actual performance and using a cascaded forward neural network
3 model

4 Mohammed Shadi S. Abujazar1, Fatihah Suja1, 2, Ibrahim Anwar Ibrahim3, A.E.

5 Kabeel4, Suraya Sharil5

6 1Department of Civil and Structural Engineering, Universiti Kebangsaan Malaysia, 43600

7 Bangi, Selangor, Malaysia
8 2Research Centre for Sustainable Technology (CESPRO), Universiti Kebangsaan Malaysia,

9 43600 Bangi, Selangor, Malaysia

10 3Department of Electrical, Electronic and Systems Engineering, Universiti Kebangsaan

11 Malaysia, 43600 Bangi, Selangor, Malaysia

12 4Mechanical Power Engineering Department, Faculty of Engineering, Tanta University,

13 Tanta, Egypt
14 5Department of Civil and Structural Engineering, Universiti Kebangsaan Malaysia, 43600

15 Bangi, Selangor, Malaysia

16
17 mohammedshadi@hotmail.com;
18 fati@ukm.edu.my;ibrahim.a.ibrahim@hotmail.com\ibrahim.a.ibrahim@ieee.org;kabeel6@ho
19 tmail.com; sharil_suraya@ukm.edu.my

20 Abstract

21 This paper presents a cascaded, forward neural network model for predicting the productivity
22 of a developed inclined stepped solar still system. The actual recorded data of the developed
23 inclined stepped solar still system is used to develop the proposed model. The results of the
24 predicted productivity are compared with that obtained from regression and linear models. In
25 this study, three statistical error terms are used to evaluate the proposed model: root mean
26 square error (RMSE), mean absolute percentage error (MAPE) and mean bias error (MBE).
27 The results show that the proposedcascaded forward neural network(CFNN) model more
28 accurately predicts the productivity of the system than the other modelsmentioned. The
29 RMSE, MAPE and MBE values of the proposed model are 22.48%, 18.51% and -26.46%,
30 respectively. Therefore, the CFNN model provides benefits for modelling the solar still.

31 Keywords: Solar still; solar desalination; productivity; ANN; prediction; modelling;

32 performance evaluation.

1
 ACCEPTED MANUSCRIPT

Nomenclature
Abbreviations
ANN Artificial neural network
CFNN Cascaded forward neural network
RMSE Root mean square error
MAPE Mean absolute percentage error
MBE Mean bias error
Symbols
𝑃 Productivity (mL/m2h)
𝐴𝑏 Area of basin liner (m2)
𝐴𝑔 Area of glass (m2)
𝐴𝑤 Area of water (m2)
As Area of Still (m2)
𝐶𝑝,𝑏 Specific heat of the basin (J/kgC)
𝐶𝑝,𝑔 Specific heat of the glass (J/kgC)
𝐶𝑝,𝑤 Specific heat of the water (J/kgC)
Ki Thermal conductivity
Li Thickness of the insulation
𝑚𝑏 Mass of basin (kg)
𝑚𝑒𝑤 Water productivity(kg/m2)
𝑚𝑔 Mass of glass (kg)
𝑚𝑤 Mass of water (kg)
hc,b ‒ w Convective heat transfer coefficient from basin to water (w/m2C)
hc,w ‒ g Convective heat transfer coefficient from water to cover (w/m2C)
ℎ𝑒,𝑤 ‒ 𝑔 Evaporative heat transfer coefficient from water to cover (w/m2C)
ℎ𝑟,𝑤 ‒ 𝑔 Radiative heat transfer coefficient from water to cover (w/m2C)
ℎ𝑟,𝑔 ‒ 𝑠𝑘𝑦 Radiative heat transfer coefficient from cover to sky (w/m2C)
ℎ𝑐,𝑔 ‒ 𝑠𝑘𝑦 Convective heat transfer coefficient from cover to sky (w/m2C)
ℎ𝑟,𝑤 ‒ 𝑠𝑘𝑦 Radiative heat transfer coefficient from water to sky (w/m2C)
ℎ𝑓𝑔 latent heat of vaporization (JKg-1)
I(t) Solar radiation on the horizontal surface (w/m2)
𝑃𝑔 Partial pressure at cover temperature (N/m2)
𝑃𝑤 Partial pressure at basin water temperature (N/m2)
𝑄𝑒,𝑤 ‒ 𝑔 Evaporative heat transfer from water to cover (w/m2)
Qc,b ‒ w Convective heat transfer from base to water by conduction (w/m2)
𝑄𝑐,𝑔 ‒ 𝑠𝑘𝑦 Convective heat loss from cover to sky (w/m2)
Qc,w ‒ g Convective heat transfer from water to cover (w/m2)
𝑄𝑟,𝑔 ‒ 𝑠𝑘𝑦 Radiative heat loss from cover to sky (w/m2)
𝑄𝑟,𝑤 ‒ 𝑔 Radiative heat transfer from water to cover (w/m2)
Qloss Heat loss from basin to ambient (w/m2)
𝑄𝑓𝑤 Heat taken by the replaced water (w/m2)
𝑇𝑎 Ambient temperature (oC)
𝑇𝑔 Glass temperature (oC)
𝑇𝑏 Basin Temperature (oC)
𝑇𝑠𝑘𝑦 Sky temperature (oC)
𝑇𝑤 water temperature (oC)

2
 ACCEPTED MANUSCRIPT

𝑉 Wind velocity in (m/s)

𝑈𝑏 Side heat loss coefficient from basin to ambient (w/m2C)
η Efficiency of solar still (%)
𝜀𝑒𝑓𝑓𝑒𝑐𝑡 Emissivity between water and cover
𝜀𝑔 Emissivity of glass
𝜀𝑤 Emissivity of water
𝛼𝑏 Absorptivity of basin material
𝛼𝑔 Absorptivity of glass cover material
𝛼𝑤 Absorptivity of water
𝜎 Stefan-Boltzmann constant (5.67x10-8 w/m2K4)
𝛽0 Intercept
𝛽1, 𝛽2, …, 𝛽𝑛 Regression coefficients
R² Deterministic coefficient
𝑀 The measured data
𝑋 The predicted data
33

34 1. Introduction

35 Clean, drinkable water is scarce. Approximately 800 million people lack access to
36 potable water; nearly 55% of global water consumption is anticipated by 2050 (Mashaly et
37 al., 2015). Approximately three fourths of the planet is covered by water (El-Samadony and
38 Kabeel, 2014), but only a very small amount of that water is drinkable;thus, potable water
39 plays a key role in the development of nations and countries (Sampathkumar et al., 2010). A
40 shortage of natural fresh water is a great challenge for many countries mainly due to
41 population explosion and enhanced living standards, along with industrial and agricultural
42 activities. Approximately 97% of the earth’s water is seawater, while another 2% is locked in
43 icecaps and glaciers. Less than 0.5% of the earth’s total fresh water resources come from
44 rivers and groundwater and are being depleted(Khawaji et al., 2008).

45 Water is available in varying quantities in nature, and there is a shortage of fresh

46 water in some areas of many countries in the world(Hou and Zhang, 2008) as one third of the
47 earth's population lives in countries that suffer from the problem of the lack of distillate
48 water(Kabeel et al., 2017). As global drinkable water supplies become increasingly stressed
49 and populations are increasing, seawater desalination has become an increasingly sought-
50 after alternative for new water supplies in coastal areas. Accordingly, the utilization of solar
51 energy could be one promising renewable energy source for desalinating seawater in mostly

3
 ACCEPTED MANUSCRIPT

52 arid and semi-arid regions, where oil sources are both hard and costly to find(Mashaly et al.,
53 2015).

54 To solve this problem, various advanced desalination technologies using fossil fuels,
55 or electrical energy derived from fossil fuels, are used all over the world by either a thermo or
56 membrane process. These systems have the capacity to produce up to 10 tons of water/day.
57 The two most common methods of desalination (reverse osmosis and flash distillation) are
58 not capable of removing all the pure water from a saline or contaminated source and must
59 discharge up to 50% of their input as concentrate (Sethi et al., 2006). For bigger systems,
60 reverse osmosis and electro-dialysis are more economical, but for smaller units, simple solar
61 stills could be preferred because of their low costs. This has direct effects on global warming,
62 and it has high economical costs. Nevertheless, the huge shortage in the amount of fossil fuel
63 resources is attributed to the increasing tendency of replacing expensive energies with other
64 renewable energies. Solar stills could be one of the most effective and simple techniques
65 compared with other distillation processes. Although solar stills face the problem of low
66 productivity compared to their costs (El-Bahi and Inan, 1999), a low productivity refers to
67 three main parameters, such as environmental, design, and operation. Despite being a suitable
68 solution for the lack of a fresh water supply, low productivity impedes the application of a
69 solar still (Abujazar et al., 2016). However, for remote areas that lack of fresh water, where
70 land is available at a low cost and that have an abundance of solar radiation, the use of solar
71 energy is the cost-effective solution as an alternative source of energy. Since desalination is
72 an energy consuming process, when adding the fossil fuel problem, solar distillation becomes
73 more suitable and promising (El-Bahi and Inan, 1999). The use of solar stills as a cheap and
74 simple method for providing drinking water and their safe outcome on the environment are
75 the major attractions for research.

76 Therefore, the utilization of solar stills in arid and semi-arid regions at a small or large
77 scale could cover desalinated water needs by either seawater desalination or groundwater
78 purification(Kabeel and Almagar, 2013). To improve solar still productivity, many studies
79 have been conducted(El-Sebaii and El-Naggar, 2017; Feng and Tao, 2015; Kabeel et al.,
80 2017; Mahian et al., 2017). Most of these studies were carried out either by experimental
81 methods (Akash et al., 2000; AL-Karaghouli and Alnaser, 2004; Badran, 2007; Tabrizi and
82 Sharak, 2010) or based on modelling methods (Kumar and Kumar, 2014; Nijmeh et al., 2005;
83 Shukla and Sorayan, 2005; Velmurugan et al., 2009b; Zurigat and Abu-Arabi, 2004).

4
 ACCEPTED MANUSCRIPT

84 Meanwhile, modelling methods have been more wildly used than the experimental methods.
85 The modelling methods have advantages over experimental methods, such as no running
86 costs, time saving, simulation tests that cannot be carried out in laboratories, and better safety
87 (no hazards). Based on the literature (El-Samadony and Kabeel, 2014), modelling methods
88 include mathematical models, regression models, machine learning and artificial intelligence
89 (AI) models(El-Samadony and Kabeel, 2014; Feng and Tao, 2015). However, the accuracy of
90 the mathematical models is questionable, especially when they are dealing with highly
91 uncertain solar radiation (Eke and Demircan, 2013). On the other hand, regression models are
92 good in predicting the monthly averages of steeped solar still systems.

93 Due to the necessity for accurate and reliable modelling of the productivity of such
94 systems on an hourly basis, artificial neural network (ANN) models, as an alternative to the
95 classical models(Ata, 2015), have been successfully employed for this purpose due to their
96 capability of handling the high uncertainty of such data. ANNs incorporate meteorological
97 data to predict solar still performance; therefore, many research works are successfully
98 conducted using ANN models for modelling the solar still system. In (Santos et al., 2012), the
99 authors used an ANN model to predict the efficiency of a solar still with local meteorological
100 data, which included the daily ambient temperature, solar radiation, cloud cover, wind
101 velocity, and wind direction as inputs. Hamdan et al. (Hamdan et al., 2014) tested three ANN
102 topologies, namely, the Feed-forward neural network, Elman, and Nonlinear Autoregressive
103 Exogenous (NARX). This study aimed to study the performance of a triple solar still
104 operating under a Jordanian meteorological weather profile. The inputs of the mentioned
105 topologies were time, hourly variation of ambient temperature, glass cover temperature, water
106 in the upper basin temperature, water in middle basin temperature, water in lower basin
107 temperature, plate temperature, distillate volume and hourly solar radiation. In the results, the
108 Feed-forward neural network model had the best result in determining the required
109 performance. Moreover, the solar still systems need a comprehensive model to develop inputs
110 and outputs for the optimization of solar still performance by incorporating operational and
111 meteorological parameters, as this gap can be seen in the study of (Hamdan et al., 2014;
112 Santos et al., 2012). Based on (Hamdan et al., 2014; Santos et al., 2012), the main
113 meteorological and operational data that may affect the desalination process, in particular, the
114 processes of evaporation and condensation, were not included. Furthermore, the contribution
115 of each component in the modelling process was not specified.

5
 ACCEPTED MANUSCRIPT

116 Thus, the main objective of this paper is to propose and evaluate a prediction model to
117 predict the productivity of the solar still system. The novelty of this model is represented in
118 the ability of predicting the system productivity considering the uncertainty issue. The
119 cascaded forward neural network (CFNN) topology is utilized in this research work. Around
120 three months of experimental hourly data are used. To show the superiority of the proposed
121 model, the obtained results from the proposed model have been compared with those
122 obtained using mathematical linear and multiple-linear regression models.

123 2. Experimental Setup

124 In this research work, a copper inclined stepped solar still was designed and fabricated to
125 study the productivity of a copper stepped solar still. Figure 1shows the schematic diagram of
126 the stepped solar still. The developed stepped solar still, with internal dimensions of length is
127 equal to 1.8 m, width is equal to 1.2 m andheight is equal 0.20 m, which consists of 28 trays.
128 The experimental setup mainly consists of twoparts, namely, the seawater settling tank and
129 the stepped solar still. Here, the developed inclined stepped solar still was placed at an angle
130 of 30° with the horizontal, as presented inFigure 2.

131

132 Figure1. Schematic diagram of the stepped solar still.

6
 ACCEPTED MANUSCRIPT

133

134 Figure2. The developed stepped solar still with the main components.
135 In this system, the trays are made from copper sheet. Figure 3shows a cross-sectional
136 view of the copper trays dimensions inside the system,with internal dimensions of length is
137 equal to 1.2 m, width is equal to0.1039 m and height is equal to 0.03 m, which are placed
138 inside a box made from stainless steel and insulated by sawdust to a thickness of 6 cm on the
139 sides and bottom of the system. The developed stepped solar still is designed with a footprint
140 area reduction and effective evaporation area enlargement. Modifying the solar still design
141 from a flat basin into a step-wise basin with triangular cross section trays gave 55.6% greater
142 sun exposure area (effective evaporation area) and decreases the occupation area (footprint
143 area) by 13.6% when compared to a conventional basin type solar still. Seawater is supplied

7
 ACCEPTED MANUSCRIPT

144 to the inclined stepped solar still from a feeding tank, black in colour, under gravity force and
145 controlled with a water level sensor at a level of 3 cm for the trays. When solar radiation falls
146 on the condensing cover, water evaporates and condenses in the condensing glass cover. The
147 condensed water is collected at the bottom of the system and collected for further measuring
148 and laboratory tests.

149

150 Figure 3. A cross-sectional view of the copper trays dimensions inside the system and the
151 vertical side of the steps.
152 2.1 Experimental procedure

153 In this research, the copper inclined stepped solar still distillation system was
154 designed, installed and tested outdoors at the Department of Civil and Structural Engineering,
155 Faculty of Engineering and Built Environment, UKM, Bandar Baru Bangi, Selangor,
156 Malaysia (latitude 2.939671°N and longitude 101.78784°E). The fabrication was done by
157 (GMI) Technical and Consultancy Services SDN. BHD. The experiments were carried out for
158 12 hours a day from 8:00 am to 7:00 pm in 12 runs (5 days for each run) for three months
159 during the period of the 27th of September 2016 to the 23rd of December 2016. Here, this case
160 study is conductedunder a tropical climate were the climate profileisalmost stable during the
161 year, therefore, a low deviation of the main inputs (solar radiation, ambient temperature and
162 humidity) is expected for the other months (Khatib et al., 2012a). Moreover, different
163 duration of desalination studies has been conducted by researchers. In (Hanson et al., 2004),
164 14 days duration of laboratory and field study dataset are used for testing the solar still in
165 Southern New Mexico, USA, by New Mexico State University. In addition, (Ismail, 2009)
166 used six days duration of outdoor experiment of transportable hemispherical solar still in
167 Dhahran climatic conditionsin Iran. While, (Tabrizi and Sharak, 2010)carried out an
168 experimental study of an integrated basin solar still with a sandy heat reservoir were
169 conducted for 3days under Iranian climatic conditions. Based on the tropical climate

8
 ACCEPTED MANUSCRIPT

170 characteristic and the literature, using three months of the hourly dataset may not effect on
171 the objective of this research work.

172 Different parameters are measured during the experiments. These parameters included
173 solar radiation, glass inner and outer cover temperatures, vapour temperature, basin (tray)
174 water temperature, and basin temperature. The measurements are taken on an hourly basis. In
175 addition, the collected freshwater productivity is also measured on an hourly basis. Moreover,
176 the other meteorological parameters, such as ambient temperature wind velocity, humidity
177 and cloud cover were collected from the “AccuWeather” meteorological website
178 (AccuWeather, 2016). All experimental measurements aimed to evaluate the performance of
179 the stepped solar stills under the climatic conditions of Bangi, Selangor, Malaysia.

180 Seawater samples were collected from Sepang Gold coast beach using a plastic
181 aspirator bottle, 25 Litres, as shown inFigure 4, and transferred to the feeding tank at the
182 study site. Seawater in the still is heated by solar radiation, the formed water vapour is
183 condensed at the inner glass surface, and the water droplets glide along the glass. Then, the
184 condensed water is collected in a stainless-steel flask for volume recording. The depths of the
185 seawater inside the trays are maintained at 3 cm using an electric valve connected to a
186 Finetek FD3002D mini single float switch.

187

188 Figure4. Seawater collection samples from Sepang Gold coast beach using plastic aspirator
189 bottle, 25 Liters.
190

9
 ACCEPTED MANUSCRIPT

191 2.2 Measuring system

192 A hole in the side of the inclined stepped solar still is made to allow for the
193 installation of thermocouples for measuring the variations of the glass outer and inner,
194 vapour, water and basin temperatures using a Thermocouple Maltec-T (Type –K) sensor, the
195 sensors installed via one hole placed on the mid of the side wall of the still. Then, the sensors
196 are distributed into different places to measure different points of temperature as shown in
197 Figure 5. In addition, the solar radiation is measured by a Tenmars TM-750 sensor, and the
198 data are recorded using a paperless data logger (OHKURA) with 6 channels. The water flow
199 is due to gravity force and controlled using an electric valve connected to a water Magnetic
200 float level sensor. A summary of the used measurement and control instruments is shown in
201 Table 1, while Figure 5illustrates the different points of the solar still device where
202 temperatures are measured.

203 Table 1. Measuring Instruments with accuracy, ranges and errors.

Sr. instrument Accuracy Range % error

No.
1 Thermocouple Maltec-T (Type -K) ±0.1°C (-100 to150) °C 5%
2 Finetek FD3002D Mini Single float ± 0.1% (25°C) 5% FS 0.01% F.S /°C
switch
3 Data logger OHKURA Paperless 0.1% +1 digit -10.00 ～ +10.00 0.01%
chart paper 6 channel.
4 Tenmars TM-750 ±0.1°% 2000W/m2 0.38W/m2/°C
204

10
 ACCEPTED MANUSCRIPT

205

206 Figure 5. Position of different points for measuring temperature variations.

207 3. Modelling of a stepped solar still

208 Modelling is one of the most important elements of thermal system design for
209 studying system behaviour and optimizing its components for better performance.
210 Predominantly, the solar still system is analysed using a set of equations that represent the
211 physical behaviour of the system. This may be a simple way for modelling the system, which
212 may affect the accuracy of the results as well. In this study, three models, namely, the
213 mathematical linear model, regression model and artificial intelligence (AI) models, are used
214 for predicting the productivity of the steeped solar still.

215 3.1. Mathematical linear model

216 Productivity of a solar still can be predicted by mathematical modelling equations

217 related to energy and mass balance equations, as solar radiation is the main source of energy,
218 which depends on the time during the day. The basic energy and mass balance equations must
219 be time dependent because solar intensity differs from one time to another (AYBAR, 2007;
220 Panchal et al., 2011b). Solar still performance can be calculated based on productivity,
221 efficiency, and internal heat and mass transfer coefficients. The performance of a solar still is
222 directly proportional to the internal heat transfer coefficient and distillate output, and the

11
 ACCEPTED MANUSCRIPT

223 convective heat transfer, radiative heat transfer and evaporative heat transfer coefficients
224 (Panchal et al., 2011b).

225 3.1.1 Convection heat transfer

226 According to a few studies (Panchal et al., 2011a, 2011b), the heat transfer inside the
227 solar still occurred by free convection, and the convective heat transfer coefficient is
228 influenced by the buoyancy force due to density differences in the heated air.

229 The rate of convective heat transfer between the basin and water surface (Qc,b ‒ w) is
230 taken as the following equation (Velmurugan et al., 2008):

231 Qc,b ‒ w = hc,b ‒ w𝐴𝑏(𝑇𝑏 ‒ 𝑇𝑤) (1)

232 where hc,b ‒ wis the convective heat transfer coefficient from basin to water (W/m2C), 𝐴𝑏 is
233 the area of the basin liner (m2), 𝑇𝑏 is the basin temperature (oC), and 𝑇𝑤 is the water
234 temperature (oC).
235 Meanwhile, the convective heat transfer between the basin and water is 135 W/m2C
236 (Zurigat and Abu-Arabi, 2004), where the heat loss from basin to ambient (Qloss) can be
237 calculated by equation (2).

238 Qloss = Ub(Ab + As)(𝑇𝑏 ‒ 𝑇𝑎) (2)

239 Where the side heat loss coefficient from basin to ambient can be calculated from
240 equation (3) (Kabeel et al., 2012).

1 𝐿𝑖
241 𝑈𝑏 = 𝐾𝑖 (3)

242 The rate of convective heat transfer between a water surfaces to the condensing glass cover is
243 calculated by equation (4)(Velmurugan et al., 2008):

244 Qc,w ‒ g = hc,w ‒ g𝐴𝑤(𝑇𝑤 ‒ 𝑇𝑔) (4)

245 where Qc,w ‒ grepresents the convective heat transfer from water to cover (W/m2), hc,w ‒ g is
246 the convective heat transfer coefficient from water to cover (W/m2C), 𝐴𝑤 is the area of water
247 (m2), and 𝑇𝑤 is water temperature (oC).
248 The convective heat transfer coefficient (hcw) can be calculated by equation (5):

12
 ACCEPTED MANUSCRIPT

(𝑃𝑤 ‒ 𝑃𝑔)(𝑇𝑤 + 273)

249 ℎ𝑐,𝑤 ‒ 𝑔 = 0.88 [(𝑇𝑤 ‒ 𝑇𝑔) + ]1/3 (5)
(268.9𝑥103 ‒ 𝑃𝑤)

250 where 𝑇𝑔 is the glass temperature (oC), 𝑃𝑔 is the partial pressure at cover temperature (N/m2),
251 and 𝑃𝑤 is the partial pressure at basin water temperature (N/m2).
5144
252 𝑃𝑤 = exp (25.317–(273 + 𝑇 )) (6)
𝑤

5144
253 𝑃𝑔 = exp (25.317–(273 + 𝑇 )) (7)
𝑔

254 3.1.2 Radiative Heat Transfer

255 According to (Panchal et al., 2011a, 2011b), the radiative heat transfer occurs through
256 solar energy. Thus, the radiative heat transfer from a water surface to the condensing cover (
257 𝑄𝑟,𝑤 ‒ 𝑔) is calculated by

258 𝑄𝑟,𝑤 ‒ 𝑔 = ℎ𝑟,𝑤 ‒ 𝑔𝐴𝑤(𝑇𝑤 ‒ 𝑇𝑔) (8)

259 where 𝑄𝑟,𝑤 ‒ 𝑔is the radiative heat transfer from the water to the cover (W/m2) and ℎ𝑟,𝑤 ‒ 𝑔 is
260 the radiative heat transfer coefficient from water to cover (W/m2C).
261 Moreover, the radiative heat transfer coefficient (ℎ𝑟,𝑤 ‒ 𝑔) can be presented as:

262 ℎ𝑟,𝑤 ‒ 𝑔 = 𝜀𝑒𝑓𝑓𝑒𝑐𝑡 𝜎[(𝑇𝑤 + 273)2 + (𝑇𝑔 + 273)2](𝑇𝑤 + 𝑇𝑔 + 546) (9)

263 where 𝜀𝑒𝑓𝑓𝑒𝑐𝑡 is the emission between water and the cover and 𝜎 is the Stefan-Boltzmann
264 constant (5.67x10-8 W/m2K4).
265 Additionally, 𝜀𝑒𝑓𝑓𝑒𝑐𝑡 can be formulated as:

1 1
266 𝜀𝑒𝑓𝑓𝑒𝑐𝑡 = (𝜀 + 𝜀𝑤 ‒ 1) ‒ 1 (10)
𝑔

267 where 𝜀𝑔 is the emission glass (0.88) and 𝜀𝑤 is the emission water (0.96).

268 3.1.3 Evaporative Heat Transfer

269 Based on (Panchal et al., 2011a, 2011b), the incident solar energy inside a solar still
270 evaporates water into steam. Therefore, the evaporative heat transfer (𝑄 𝑒,𝑤 ‒ 𝑔) can be
271 calculated as(Velmurugan et al., 2008):

272 𝑄 𝑒,𝑤 ‒ 𝑔 = ℎ𝑒,𝑤 ‒ 𝑔𝐴𝑤(𝑇𝑤 ‒ 𝑇𝑔) (11)

13
 ACCEPTED MANUSCRIPT

273 where 𝑄 𝑒,𝑤 ‒ 𝑔is the evaporative heat transfer from the water to the cover (W/m2) and ℎ𝑒,𝑤 ‒ 𝑔
274 is the evaporative heat transfer coefficient from water to cover (W/m2C). The ℎ𝑒,𝑤 ‒ 𝑔can be
275 calculated by(Velmurugan et al., 2009a):
(𝑃𝑤 ‒ 𝑃𝑔)
276 ℎ𝑒,𝑤 ‒ 𝑔 = 16.27 × 10 ‒ 3 × ℎ𝑐,𝑤 ‒ 𝑔 × (𝑇 (12)
𝑤 ‒ 𝑇𝑔)

277 It is also assumed that, the makeup water is at atmospherictemperature and takes heat
278 from basin. The heat taken by thereplaced water is estimated from (Kalidasa Murugavel et
279 al., 2010),

280 𝑄𝑓𝑤 = 𝑚𝑒,𝑤𝐶𝑝,𝑤(𝑇𝑎 ‒ 𝑇𝑤) (13)

281 The radiative heat transfer between the condensing cover (glass) and the sky is given
282 by(Kabeel et al., 2012; Velmurugan et al., 2009a)

283 𝑄𝑟,𝑔 ‒ 𝑠𝑘𝑦 = ℎ𝑟,𝑔 ‒ 𝑠𝑘𝑦𝐴𝑔(𝑇𝑔 ‒ 𝑇𝑠𝑘𝑦) (14)

284 where 𝑄𝑟,𝑔 ‒ 𝑠𝑘𝑦is the convective heat loss from cover to sky (W/m2), ℎ𝑟,𝑔 ‒ 𝑠𝑘𝑦is the radiative
285 heat transfer coefficient from glass to sky (W/m2C), and 𝑇𝑠𝑘𝑦 is the sky temperature (oC).

286 The radiative heat transfer coefficient ℎ𝑟,𝑔 ‒ 𝑠𝑘𝑦 between the condensing cover (glass)
287 and the sky is given by (Velmurugan et al., 2009a),

𝜀𝑔 𝜎[(𝑇𝑔 + 273)4 ‒ (𝑇𝑠𝑘𝑦 + 273)4]

288 ℎ 𝑟,𝑔 ‒ 𝑠𝑘𝑦 = (𝑇𝑔 ‒ 𝑇𝑠𝑘𝑦) (15) where

289 the effective sky temperature is taken from(Zurigat and Abu-Arabi, 2004),

290 𝑇𝑠𝑘𝑦 = 𝑇𝑎 ‒ 6 (16)

291 The convective heat transfer between the condensing cover (glass) and the sky is
292 given by (Zurigat and Abu-Arabi, 2004),

293 𝑄𝑐,𝑔 ‒ 𝑠𝑘𝑦 = ℎ𝑐,𝑔 ‒ 𝑠𝑘𝑦𝐴𝑔(𝑇𝑔 ‒ 𝑇𝑠𝑘𝑦) (17)

294 where ℎ𝑐,𝑔 ‒ 𝑠𝑘𝑦 is the convective heat transfer coefficient from cover to sky (W/m2C).
295 ℎ𝑐,𝑔 ‒ 𝑠𝑘𝑦is given by(Kabeel et al., 2012),

296 ℎ𝑐,𝑔 ‒ 𝑠𝑘𝑦 = 2.8 + 3.0 𝑉 (18)

14
 ACCEPTED MANUSCRIPT

297 where 𝑉 represents the wind velocity (m/s).

298 In general, the solar still has several components, and its performance can be
299 predicted by energy balance equations of these components. Separate energy balances can be
300 described by the behaviour of the glass cover, the basin water and the basin as shown
301 inFigure 6.

302

303 Figure6. Energy flow in stepped solar still.

304 The heat transfer mechanism begins when solar energy is incident on the basin water.
305 Table 2shows the constants and various design parameters for the theoretical model. Here,
306 the heat balance equation of the basin water can be described by the behaviour of the system
307 with the following assumptions (Panchal et al., 2011a, 2011b):

308 1. There is no vapour leakage in the solar still.

309 2. It is an air tight basin; hence, there is no heat loss.
310 3. The heat capacities of the cover, absorbing material and insulation are negligible.
311 4. There is no temperature gradient across the basin water and glass cover of the solar still.
312 5. The water level inside the basin is maintained at a constant level.
313 6. Only film type condensation occurs in place of drop type condensation.
314

15
 ACCEPTED MANUSCRIPT

315 Based on the above equations, the energy balance models for the glass cover, basin water and
316 basin can be formulated as follows:
317  Energy balance for the glass cover(Kabeel et al., 2012; Velmurugan et al., 2009b):
𝑑𝑇𝑔
318 𝑄𝑟,𝑔 ‒ 𝑠𝑘𝑦 + 𝑄𝑐,𝑔 ‒ 𝑠𝑘𝑦 + (𝑚𝑔 ∗ 𝐶𝑝,𝑔 ( )) = 𝛼 𝐴
𝑑𝑡 𝑔 𝑔 𝐼(𝑡) + (𝑄𝑐,𝑤 ‒ 𝑔 + 𝑄𝑟,𝑤 ‒ 𝑔 + 𝑄𝑒,𝑤 ‒ 𝑔) (19)

319  Energy balance for the basin water(Kabeel et al., 2012; Velmurugan et al., 2009b):
𝑑𝑇𝑤
320 𝛼𝑤𝐴𝑤𝐼(𝑡) + 𝑄𝑐,𝑏 ‒ 𝑤 = (𝑚𝑤 ∗ 𝐶𝑝,𝑤)𝑤 ( )+ 𝑄
𝑑𝑡 𝑐,𝑤 ‒ 𝑔 + 𝑄𝑟,𝑤 ‒ 𝑔 + 𝑄𝑒,𝑤 ‒ 𝑔 + 𝑄𝑓𝑤(20)

321  Energy balance for the basin(Kabeel et al., 2012; Velmurugan et al., 2009b):
𝑑𝑇𝑏
322 𝛼𝑏𝐴𝑏 𝐼(𝑡) = 𝑚𝑏 ∗ 𝐶𝑝,𝑏 ( ) + (𝑄
𝑑𝑡 𝑐,𝑏 ‒ 𝑤 + 𝑄𝑙𝑜𝑠) (21)

323 A constant water level is controlled to be 3 cm inside the trays using a water level
324 control sensor. In addition, the absorptivity of the still (𝛼𝑏), the glass (𝛼𝑔) and the water (𝛼𝑤)
325 are taken as 0.95, 0.05 and 0.0475, respectively, while the specific heat of the seawater (𝐶𝑝,𝑤)
326 is calculated. Based on equations (19-21), the basin water temperature (𝑇𝑤), the glass cover
327 temperature (𝑇𝑔) and the basin temperature (𝑇𝑏) formulas can be presented as follows:

328 𝑇𝑤 = 𝑇𝑤 + 𝑑𝑇𝑤 (22)

329 𝑇𝑔 = 𝑇𝑔 + 𝑑𝑇𝑔
330 (23)

331 𝑇𝑏 = 𝑇𝑏 + 𝑑𝑇𝑏
332 (24)

333 As a result, the hourly productivity of a steeped solar still system can be estimated
334 by(Phadatare and Verma, 2007),
𝑞𝑒,𝑤 ‒ 𝑔 ℎ𝑒,𝑤 ‒ 𝑔(𝑇𝑤 ‒ 𝑇𝑔)
335 𝑚𝑒𝑤 = ℎ𝑓𝑔 𝑋3600 = ℎ𝑓𝑔 𝑋3600 (25)

336 Finally, Table 2 shows the constants and various design parameters which are used in
337 developing the mathematical linear model.

338 Table 2. Constants and design parameters used in the mathematical linear model

Parameter Value Ref.

16
 ACCEPTED MANUSCRIPT

αb, Absorptivity of basin material 0.95 (Shukla, 2011; Zurigat and

Abu-Arabi, 2004)
αg, Absorptivity of glass cover material 0.05 (Shukla, 2011)
αw, Absorptivity of water 0.05 (Shukla, 2011; Zurigat and
Abu-Arabi, 2004)
ɛg, Emissivity of glass 0.88 (Zurigat and Abu-Arabi,
2004)
ɛw, Emissivity of water 0.96 (Zurigat and Abu-Arabi,
2004)
hc,b-w, convective heat transfer coefficient 135, (Zurigat and Abu-Arabi,
from basin to water W/m C2 2004)
Ϭ, Stefan-Boltzmann constant (5.67x10-8 5.67x10-8, (Shukla, 2011)
w/m2K4) W/m2K4
mw, Mass of water (kg) 52.36, kg Current study
dt, Time intervals (s) 60 s Current study
Ab, Area of basin liner (m2) 2.16, m 2 Current study
Cg, Specific heat of the glass (J/kgC) 800, J/kg (Zurigat and Abu-Arabi,
C 2004)
Cp,w, Specific heat of the water (J/kgC) 4178, (Zurigat and Abu-Arabi,
J/KgC 2004)
Cb, Specific heat of the basin (J/kgC) 385, (Specific heat of materials,
J/KgC 217AD)
339
340 3.2. Regression model

341 The regression model is one of the most famous models that can be used for analysing
342 multi-factor data. A regression model is a statistical model that is used to predict the response
343 of a multi-inputs model by expressing the relationship between the input variables of interest
344 (dependent variable and/or independent variables) and the output of the system. The linear
345 model is the simplest regression model, which expresses the relationship between a single
346 explanatory variable (x) and the response (y) for the model in a straight line by the general
347 formula as (Ameen et al., 2015)

348 𝑦 = 𝛽0 + 𝛽1𝑥 (26)

349 where 𝛽0 is the intercept and 𝛽1 is the slope of the line.

350 On the other hand, the multiple regression model is more appropriate to express the
351 type of systems that consider a level of fluctuation in the inputs. Meanwhile, this model
352 considers the influence of multiple explanatory variables on a response variable. The basic
353 formula for a multi-regression linear model can be represented by (Ameen et al., 2015)

354 𝑦 = 𝛽0 + 𝛽1𝑥1 + 𝛽2𝑥2 + … + 𝛽𝑛𝑥𝑛 (27)

17
 ACCEPTED MANUSCRIPT

355 where 𝛽0 is the intercept and (𝛽1, 𝛽2, …, 𝛽𝑛) are the regression coefficients.

356 In this paper, the multiple linear regression model is utilized to model the productivity
357 of the solar still system using actual hourly based data. The model inputs are solar radiation (
358 𝑆), ambient temperature (𝑇𝑎), month number (𝑀), day number (𝐷), number of hours per day (
359 𝐻), wind speed (𝑊), humidity (𝐻𝑚), cloud cover (𝐶𝐶), vapour temperature (𝑇𝑣), water
360 temperature (𝑇𝑤), basin temperature (𝑇𝑏) and the difference between the inner and outer
361 surface of glass temperature (𝛥𝑇), while the output is the productivity (𝑃), as formulated in
362 Eq. (28). Here, the deterministic coefficient (R²) of the proposed model is 0.905.

363 𝑃 =‒ 290.517 + 0.016 ∗ (𝑆) ‒ 0.156 ∗ (𝑇𝑎) ‒ 10.167 ∗ (𝑀) + 0.588 ∗ (𝐷) ‒ 4.775 ∗ (𝐻)
364 + 1.867 ∗ (𝑊) + 0.555 ∗ (𝐻𝑚) ‒ 0.412 ∗ (𝐶𝐶) ‒ 8.129 ∗ (𝑇𝑣) ‒ 25.475 ∗ (𝑇𝑤)
365 + 41.552 ∗ (𝑇𝑏) + 14.927 ∗ (𝛥𝑇)
366 (28)

367 3.3. The proposed Cascaded forward neural network model

368 The Artificial Neural Network (ANN) is an artificial intelligence (AI) technique, and
369 is known as the information processing paradigm, which is inspired by the human biological
370 nervous systems to process information(Artrith and Urban, 2016; Chiteka and Enweremadu,
371 2016). The ANN can identify patterns and learn from their interactions with the environment
372 (Jafar et al., 2010). The construction of an ANN contains three main parts; the input layer, the
373 hidden layer(s) and the output layer, in which these layers contain parallel units called
374 neurons(Dharma et al., 2017). These neurons are connected by huge weight links, which
375 allow the transformation of the information between the layers. The ANN model basically
376 depends on two main stages to predict the response of the various systems: the training stage
377 and the testing stage. In the training stage, the neurons receive the inputs over their incoming
378 connections and then combine these inputs by a nonlinear operation with the targeted output
379 to find the best values of the weight links. Thus, the relationship between the input and the
380 output variables is recorded to predict the new data. Meanwhile, in the testing stage, a part of
381 the input data is used to test the performance of the system and the predicted data is
382 compared with the actual data. The main advantage of the ANN model is that it is able to
383 solve complex problems that are not easy to solve using conventional models and is able to
384 solve the problems that do not have an algorithmic solution or that the algorithmic solution is
385 too complex to be defined(Shabanpour et al., 2017).

18
 ACCEPTED MANUSCRIPT

386 Different ANN topologies are used for modelling in the literature(Kashyap et al.,
387 2015; Khatib et al., 2012b; Qazi et al., 2015). These topologies can be categorized as feed-
388 forward back propagation neural networks (FFNNs), probabilistic neural networks, general
389 regression neural networks (GRNNs), radial basis function networks (RBFs) and cascaded
390 forward neural network (CFNNs).The FFNN topology is the first and simplest topology of an
391 ANN, which is used as the back-propagation (BP) algorithm to update the weight links
392 during the learning stage. The RBF and GRNN topologies are mainly used when the size of
393 the input data is relatively small. Although the CFNN topology is used, the BP algorithm is
394 the FFNN topology, but each neuron in a CFNN is dependent on all the previous layer
395 neurons. In addition, the CFNN can handle a wide range if the input data has a large memory
396 capacity.

397 The construction of the CFNN topology starts with the input and output neurons. The
398 output neurons previously existed in the neural network; then, new neurons are fed into the
399 network. By adding new neurons into the network, the network itself tries to maximize the
400 correlation magnitude between the inputs and outputs by comparing the network’s residual
401 with and the new calculated error. This process continues until gaining the lower error value
402 in the network. This is the reason for labelling this network as a cascade. In general, CFNN
403 contains three main layers: the input, hidden and output layers. The variables that are the
404 network inputs are set into the input neurons in the input layer, and these variables are then
405 distributed on the hidden layer neurons. In the hidden layer, these variables are multiplied by
406 the bias (1.0) and the weight (calculated into the construction stage to decrease the prediction
407 error) and then added to the sum going into the neuron. Meanwhile, the resulting value from
408 this process will pass through a transfer function to represent the output value (Khatib et al.,
409 2012b). The schematic diagram of the CFNN topology is illustrated in Figure 7.

19
 ACCEPTED MANUSCRIPT

410

411 Figure7. The schematic diagram of the CFNN topology.

412 The values of the weights and biases are estimated in each neuron. This estimation is
413 conducted by a supervised training process with the learning data. During the training
414 process, the ANNs minimize the error rate between the outputs and targets by adjusting the
415 values of weights and biases. Here, the back-propagation (BP) learning algorithm is used for
416 the learning CFNN. The basic training process can be presented by

𝑇𝑟𝑎𝑖𝑛𝑖𝑛𝑔
417 (𝐴𝑁𝑁𝑢(𝐿,𝑁),𝐼,𝑇) → (𝐴𝑁𝑁𝑡) (29)

418 where 𝐴𝑁𝑁𝑢 is the untrained ANN, 𝐿 is the hidden layers, 𝑁 represents the neurons
419 per layer, 𝐼 is the training inputs data, 𝑇 is the training targets and 𝐴𝑁𝑁𝑡 represents the
420 trained ANN. In this work, we employed a TRAINLM function as a training function.

421 The numbers of hidden neurons mainly play a role in network performance. However,
422 there is no direct formula in the literature to find the optimal number of neurons. Most of the
423 studies in this field (Blum. A, 1992; Boger and Guterman, 1997; Caudill, M., & Butler, 1994;

20
 ACCEPTED MANUSCRIPT

424 Linoff and Berry, 1997; Swingler, 1996) calculate the number of neurons using the trial and
425 error method. In case the numbers of hidden neurons are lower than the optimal numbers,
426 high training and generalization errors may occur. While if the number of hidden neurons are
427 higher than the optimal numbers, then over-fitting and high variance may occur. Thus, the
428 optimal number of hidden neurons must be calculated to gain the best performance of the
429 network (Geman et al., 1992).

430 Based on this, the number of neurons in the proposed CFNN based model is
431 optimized using the iterative method based on the minimum value of root mean squared error
432 (RMSE). Therefore, the flowchart of the optimization and the prediction processes using the
433 proposed CFNN based method is presented in Figure 8.

21
 ACCEPTED MANUSCRIPT

434

435 Figure 8. The flowchart of the CFNN based method prediction procedure.

436 The stages of the proposed CFNN based method for predicting the productivity can be
437 described as follows:

438 1. As a first step, set the input samples and variables into the CFNN algorithm. In this
439 paper, the inputs are solar radiation, ambient temperature, month number, day
440 number, number of hours per day, wind speed, humidity, cloud cover, vapour
441 temperature, water temperature, basin temperature and the difference between the
442 inner and outer surface of glass temperature. The output is the productivity.

22
 ACCEPTED MANUSCRIPT

443 2. Set the initial number of neurons to be equal to 1 neuron; then, start the training
444 process. This process is repeated by increasing the number of neurons by 1 neuron
445 until reaching 100 neurons to find the best number of neurons in the hidden layer.
446 3. Calculate the RMSE for each trial and save these trials in a matrix. In this paper, the
447 design space is equal to 100 trials.
448 4. Optimize the number of neurons based on the minimum value of the RMSE.
449 5. Set the best number of neurons to train the CFNN algorithm in order to prepare it for
450 the testing stage to predict the productivity. Here, 70% of the dataset is used for
451 training, while 30% of the dataset is used fortesting and validation.
452
453 3.4. Evaluation Criteria

454 Different evaluation criteria can be used to evaluate the modelling techniques. In this
455 study, three statistical error terms are used. First, mean absolute percentage error (MAPE),
456 which expresses the accuracy of the models. Second, mean bias error (MBE), which
457 represents the average deviation between the actual and predicted data in the proposed
458 models. Finally, root mean squared error (RMSE), which stands for the efficiency indicator
459 for the models (Ibrahim and Khatib, 2017). Meanwhile, the MAPE can usually be expressed
460 as a percentage by

|
𝑀 𝑖 ‒ 𝑋𝑖
461
𝑛 1
𝑀𝐴𝑃𝐸 = 𝑛∑𝑖 = 1 𝑀𝑖 | (30)

462 where 𝑀 is the measured data and 𝑋 is the predicted data.

463 In addition, the value of the MBE expresses the behaviour of the predicted data. In a
464 case where the value of MBE is positive, it indicates an overestimation in the predicted
465 values and vice versa. However, the value of MBE can be calculated as

1
𝑛
466 𝑀𝐵𝐸 = 𝑛∑𝑖 = 1(𝑋𝑖 ‒ 𝑀𝑖) (31)

467 Finally, the RMSE, as mentioned above, represents the efficiency of the model based
468 on the variation between the predicted and actual data as well. Based on that, a large positive
469 RMSE means that there is a big deviation between the predicted and actual data and vice
470 versa. The RMSE can be represented as

1 𝑛
471 𝑅𝑀𝑆𝐸 = ∑
𝑛 𝑖 = 1(𝑋𝑖 ‒ 𝑀𝑖) (32)

23
 ACCEPTED MANUSCRIPT

472 In this research work, the value of RMSE and MBE are in ml/m²h. To express these
473 values as percentages, RMSE and MBE values are divided by the average of the hourly
474 productivity, which are approximately 182.4 mL/m²h.

475 4. Case study based on a Malaysian weather profile

476 To show the effectiveness of the proposed model, a case study based on a Malaysian
477 weather profile has been utilized in this research work. Hourly meteorological data for
478 approximately three months at the Faculty of Engineering and Built Environment, UKM,
479 Selangor, Malaysia (101.7713° E, 2.921065° N) recorded by the Department of Civil and
480 Structural Engineering is used to develop the proposed model. This dataset contains solar
481 radiation, ambient temperature, wind speed, humidity, cloud cover, vapour temperature,
482 water temperature, basin temperature, the difference between the inner and outer surface of
483 glass temperature and the productivity of the solar still. The period for the utilized data is set
484 from the 27th of Sep. 2016, to the 23rd of Dec. 2016. Figure 9shows samples of the training
485 independent input variables data for the proposed model. Figure 9.a shows a sample of the
486 hourly solar radiation data, and the average solar radiation is 501.39 W/m2. Figure 9.b shows
487 a sample of the hourly ambient temperature, and the average ambient temperature is 29.42
488 °C. Moreover, Figure 9.c, Figure 9.d and Figure 9.e show wind speed in km/h, humidity as
489 percentages and cloud cover as percentages, respectively.

24
 ACCEPTED MANUSCRIPT

490

491 Figure9. Samples of the training independent input variables data (solar radiation, ambient
492 temperature, wind speed, humidity and cloud cover) for the proposed CFNN model
493 In addition, Figure 10 shows samples of the training dependent input variables data
494 for the proposed model. Here, Figure 10show the water temperature, vapour temperature,
495 basin temperature and the difference between the inner and outer surface of glass
496 temperature, respectively.

25
 ACCEPTED MANUSCRIPT

497

498 Figure10. Samples of the training dependent input variables data (water temperature, vapor
499 temperature, basin temperature and the difference between the inner and outer surface of
500 glass temperature) for the proposed CFNN model.
501

502 The aforementioned hourly training data mentioned in Figure 9and Figure 10are
503 utilized to develop the proposed CFNN model. To optimize the number of neurons in the
504 proposed model, the numerical method based on RMSE is implemented, as illustrated in
505 Figure 8.

506 Based on that, the optimization curve of the number of neurons is shown in Figure 11,
507 while a sample of the RMSE at the 100 neuronsprobability is mentioned in Table 3.

26
 ACCEPTED MANUSCRIPT

508

509 Figure11. The optimization curve for the number of neurons in the CFNN model based on the
510 RMSE values.
511 Table 3. The RMSE values for 100 probabilities of hidden neurons in CFFN model. The bold
512 value identifies the best number of neurons.

No. of Neurons RMSE (ml/m²h)

1 45.47309
2 49.5691
: :
26 48.6133
27 41.0109
28 57.4272
: :
99 103.9798
100 87.3183
513

514 From Figure 11and Table 3, it is clear that the best number of neurons in the proposed
515 CFNN model is 27 hidden neurons. After the number of neurons has been optimized and set
516 for predicting the productivity, the training process was conducted using 70% of the data set,
517 while 30% of the data were used to test and validate the proposed model. A scatter plot of the
518 dataset in training, validation and testing process using the proposed model is illustrated in

27
 ACCEPTED MANUSCRIPT

519 Figure 12. However, Figure 13 shows an hourly sample of the predicted productivity based
520 on the CFNN model.

521

522 Figure12. Scatter plot of the dataset in training, validation and testing process using CFNN
523 topology.

28
 ACCEPTED MANUSCRIPT

524

525 Figure13. Hourly sample of the predicted the solar still productivity by CFNN model.
526 From Figure 12., it is clear that the values of R in the training, validation, and testing
527 processes are 0.9964, 0.9809 and 0.9815, respectively, while the total value of R in this
528 topology is 0.9916. To show the superiority of the proposed CFNN model, the same case
529 study dataset is used to predict the productivity by multiple-linear regression and linear
530 models; then, the data obtained by these models are compared with those obtained by the
531 proposed model. Figure 14 shows the predicted productivity of the system using the CFNN,
532 regression and linear models. Moreover, Table 4 summarizes the statistical error values of the
533 predicted data by the CFNN, regression and linear models.

29
 ACCEPTED MANUSCRIPT

534

535 Figure14. Hourly sample of the predicted the solar still productivity by CFNN, regression and
536 linear models.
537 Table 4. The statistical error values of CFNN, regression and linear models in the prediction
538 of the productivity for a solar still.

Model RMSE (mL/m²h) RMSE (%) MAPE (%) MBE (mL/m²h) MBE (%)
CFNN 41.01 22.48 18.51 -26.46 -14.51
Regression 50.21 27.53 23.21 -43.24 -23.71
Linear 80.36 44.06 29.43 -72.34 -39.67
539

540 From Figure 14and Table 4, it is clear that the CFNN model is more accurate in
541 predicting the productivity of the solar still than the regression and linear models. The
542 statistical error values show that the RMSE for the CFNN model (22.48%) is lower by
543 18.43% and 48.98% than those for the RMSE obtained by regression (27.53%) and linear
544 (44.06%) models, respectively. The MAPE for the CFNN model (18.51%) is lower by
545 20.25% and 37.10% than those for the MAPE obtained by the regression (23.21%) and linear
546 (29.43%) models, respectively. Moreover, the MBE value for the CFNN model (-14.51%) is
547 lower by 30.80% and 63.42% than those for the MBE obtained by the regression (-23.71%)
548 and linear (-39.67) models, respectively. Here, the negative values for the MBE mean that the
549 predicted results are under-forecasted. These results prove the superiority of the CFNN model
550 in predicting the productivity of the stepped solar still compared with the regression and
551 linear models.

30
 ACCEPTED MANUSCRIPT

552 5. Conclusion

553 A CFNN model for predicting the productivity of a developed inclined stepped solar
554 still system is presented in this paper. Hourly experimental data for three months from a
555 developed inclined stepped solar still system are used in developing the proposed model. The
556 model’s’ inputs are solar radiation, ambient temperature, month number, day number,
557 number of hours per day, wind speed, humidity, cloud cover, vapour temperature, water
558 temperature, basin temperature and the difference between the inner and outer surface of
559 glass temperatures. The model’s output is the productivity. Three statistical error terms,
560 namely, RMSE, MAPE, and MBE, are employed to evaluate the accuracy of the proposed
561 model. Based on the results, the CFNN model shows higher accuracy in modelling the
562 stepped solar still system and exceeds the regression and linear models. The RMSE, MAPE,
563 and MBE values for the proposed CFNN model are 22.48%, 18.51%, and −14.51%,
564 respectively. Thus, the CFNN model is recommended for modelling and characterizing the
565 productivity of the inclined stepped solar still.Moreover, this study proves that modelling
566 based on the ANN technique may save time and reduce the financial expenses and efforts as
567 the ANN is an effective tool for modelling the solar still.

568 6. Acknowledgement

569 The authors wish to extend sincere thanks to the many individuals and organisations for their
570 kind support in the study, in particular, Universiti Kebangsaan Malaysia through its GUP-
571 2016-020 funding.

572

31
 ACCEPTED MANUSCRIPT

573 References:

574 Abujazar, M.S.S., Fatihah, S., Rakmi, A.R., Shahrom, M.Z., 2016. The effects of design
575 parameters on productivity performance of a solar still for seawater desalination: A
576 review. Desalination 385, 178–193. doi:10.1016/j.desal.2016.02.025

577 AccuWeather, 2016. Malaysia Weather [WWW Document]. URL

578 https://www.accuweather.com/en/my/malaysia-weather (accessed 12.23.16).

579 Akash, B.A., Mohsen, M.S., Nayfeh, W., 2000. Experimental study of the basin type solar
580 still under local climate conditions. Energy Convers. Manag. 41, 883–890.

581 AL-Karaghouli, A.A., Alnaser, W.E., 2004. Experimental comparative study of the
582 performances of single and double basin Solar-stills. Appl. Energy 77, 317–325.
583 doi:10.1016/S0306-2619(03)00124-7

584 Ameen, A.M., Pasupuleti, J., Khatib, T., Elmenreich, W., Kazem, H.A., 2015. Modeling and
585 Characterization of a Photovoltaic Array Based on Actual Performance Using Cascade-
586 Forward Back Propagation Artificial Neural Network. J. Sol. Energy Eng. 137, 41010.

587 Artrith, N., Urban, A., 2016. An implementation of artificial neural-network potentials for
588 atomistic materials simulations: Performance for TiO2. Comput. Mater. Sci. 114, 135–
589 150.

590 Ata, R., 2015. Artificial neural networks applications in wind energy systems: a review.
591 Renew. Sustain. Energy Rev. 49, 534–562. doi:10.1016/j.rser.2015.04.166

592 AYBAR, H.S., 2007. A review of desalination by solar still, in: Solar Desalination for the
593 21st Century. pp. 207–214. doi:10.1007/978-1-4020-5508-9_15

594 Badran, O.O., 2007. Experimental study of the enhancement parameters on a single slope
595 solar still productivity. Desalination 209, 136–143.

596 Blum. A, 1992. Neural Networks in C++: An Object-Oriented Framework for Building
597 Connectionist Systems, Volume 1, 1st ed. John Wiley & Sons.

598 Boger, Z., Guterman, H., 1997. Knowledge extraction from artificial neural network models.
599 IEEE Syst. Man, Cybern. 1997 1–12.

600 Caudill, M., & Butler, C., 1994. Understanding Neural Networks: Computer Explorations: A

32
 ACCEPTED MANUSCRIPT

601 Workbook in Two Volumes With Software for the Macintosh and PC Compatibles. Mit
602 Press, Cambridge, MA.

603 Chiteka, K., Enweremadu, C.C., 2016. Prediction of global horizontal solar irradiance in
604 Zimbabwe using artificial neural networks. J. Clean. Prod. 135, 701–711.
605 doi:10.1016/j.jclepro.2016.06.128

606 Dharma, S., Haji Hassan, M., Chyuan Ong, H., Hanra Sebayang, A., Susan Silitonga, A.,
607 Kusumo, F., Milano, J., 2017. Experimental study and prediction of the performance and
608 exhaust emissions of mixed biodiesel blends in diesel engine using Jatropha curcas-
609 Ceiba pentandra artificial neural networks Experimental study and prediction of the
610 performance and exhaust emissio. J. Clean. Prod. doi:10.1016/j.jclepro.2017.06.065

611 Eke, R., Demircan, H., 2013. Performance analysis of a multi crystalline Si photovoltaic
612 module under Mugla climatic conditions in Turkey. Energy Convers. Manag. 65, 580–
613 586.

614 El-Bahi, A., Inan, D., 1999. Analysis of a parallel double glass solar still with separate
615 condenser. Renew. Energy 17, 509–521.

616 El-Samadony, Y. a. F., Kabeel, A.E., 2014. Theoretical estimation of the optimum glass
617 cover water film cooling parameters combinations of a stepped solar still. Energy 68,
618 744–750.

619 El-Sebaii, A.A., El-Naggar, M., 2017. Year round performance and cost analysis of a finned
620 single basin solar still, Applied Thermal Engineering. Elsevier Ltd.
621 doi:10.1016/j.applthermaleng.2016.08.215

622 Feng, M., Tao, Y., 2015. A Mathematical Model for the Performance of a Horizontal
623 Convective Solar Still. ASME Summer Heat Transf. Conf. 1–9.

624 Geman, S., Bienenstock, E., Doursat, R., 1992. Neural Networks and the Bias/Variance
625 Dilemma. Neural Comput. doi:10.1162/neco.1992.4.1.1

626 Hamdan, M. a., Khalil, R. a. H., Abdelhafez, E. a. M., 2014. Comparison of Neural Network
627 Models in the Estimation of the Performance of Solar Still Under Jordanian Climate. J.
628 Clean Energy Technol. 1, 238–242.

629 Hanson, A., Zachritz, W., Stevens, K., Mimbela, L., Polka, R., Cisneros, L., 2004. Distillate

33
 ACCEPTED MANUSCRIPT

630 water quality of a single-basin solar still: Laboratory and field studies. Sol. Energy 76,
631 635–645. doi:10.1016/j.solener.2003.11.010

632 Hou, S., Zhang, H., 2008. A hybrid solar desalination process of the multi-effect
633 humidification dehumidification and basin-type unit. Desalination 220, 552–557.

634 Ibrahim, I.A., Khatib, T., 2017. A novel hybrid model for hourly global solar radiation
635 prediction using random forests technique and firefly algorithm. Energy Convers.
636 Manag. 138, 413–425.

637 Ismail, B.I., 2009. Design and performance of a transportable hemispherical solar still.
638 Renew. Energy 34, 145–150.

639 Jafar, R., Shahrour, I., Juran, I., 2010. Application of Artificial Neural Networks (ANN) to
640 model the failure of urban water mains. Math. Comput. Model. 51, 1170–1180.

641 Kabeel, A.E., Almagar, A.M., 2013. Seawater Greenhouse in Desalination and Economics,
642 in: IWTC17 Istanbul. pp. 5–7.

643 Kabeel, A.E., Teamah, M.A., Abdelgaied, M., Aziz, G.B.A., 2017. Modified pyramid solar
644 still with v-corrugated absorber plate and PCM as a thermal storage medium. J. Clean.
645 Prod. doi:10.1016/j.jclepro.2017.05.195

646 Kabeel, a. E., Khalil, A., Omara, Z.M., Younes, M.M., 2012. Theoretical and experimental
647 parametric study of modified stepped solar still. Desalination 289, 12–20.

648 Kalidasa Murugavel, K., Sivakumar, S., Ahamed, J.R., Chockalingam, K.K.S.K., Srithar, K.,
649 2010. Single basin double slope solar still with minimum basin depth and energy storing
650 materials. Appl. Energy 87, 514–523.

651 Kashyap, Y., Bansal, A., Sao, A.K., 2015. Solar radiation forecasting with multiple
652 parameters neural networks. Renew. Sustain. Energy Rev. 49, 825–835.
653 doi:10.1016/j.rser.2015.04.077

654 Khatib, T., Mohamed, A., Sopian, K., Mahmoud, M., 2012a. Solar energy prediction for
655 Malaysia using artificial neural networks. Int. J. Photoenergy 2012.
656 doi:10.1155/2012/419504

657 Khatib, T., Mohamed, A., Sopian, K., Mahmoud, M., 2012b. Assessment of artificial neural

34
 ACCEPTED MANUSCRIPT

658 networks for hourly solar radiation prediction. Int. J. Photoenergy 2012.

659 Khawaji, A.D., Kutubkhanah, I.K., Wie, J.M., 2008. Advances in seawater desalination
660 technologies. Desalination 221, 47–69.

661 Kumar, D., Kumar, P., 2014. Mathematical Modeling of Conventional Solar Still Coupled
662 With Solar Air Heater. IJISET 1, 379–385.

663 Linoff, G., Berry, M.J.A., 1997. Data mining techniques: For marketing, sales, and customer
664 support, John Willey and Sons, New York. John Wiley & Sons, Inc., New York.

665 Mahian, O., Kianifar, A., Heris, S.Z., Wen, D., Sahin, A.Z., Wongwises, S., 2017. Nanofluids
666 effects on the evaporation rate in a solar still equipped with a heat exchanger. Nano
667 Energy 36, 134–155. doi:10.1016/j.nanoen.2017.04.025

668 Mashaly, A.F., Alazba, A.A., Al-Awaadh, A.M., Mattar, M.A., 2015. Predictive model for
669 assessing and optimizing solar still performance using artificial neural network under
670 hyper arid environment. Sol. Energy 118, 41–58.

671 Nijmeh, S., Odeh, S., Akash, B., 2005. Experimental and theoretical study of a single-basin
672 solar sill in Jordan. Int. Commun. Heat Mass Transf. 32, 565–572.

673 Panchal, H.N., Doshi, M., Chavda, P., Goswami, R., 2011a. Effect of Cow dung cakes inside
674 basin on heat transfer coefficients and productivity of single basin single slope solar still.
675 Int. J. Appl. Eng. Res. Dindigul 1, 675–690.

676 Panchal, H.N., Shah, P.K., Of, O., 2011b. Modelling and verification of single slope solar
677 still using ANSYS-CFX. Int. J. energy Environ. 2, 985–998.

678 Phadatare, M.K., Verma, S.K., 2007. Influence of water depth on internal heat and mass
679 transfer in a plastic solar still. Desalination 217, 267–275.

680 Qazi, A., Fayaz, H., Wadi, A., Raj, R.G., Rahim, N.A., Khan, W.A., 2015. The artificial
681 neural network for solar radiation prediction and designing solar systems: A systematic
682 literature review. J. Clean. Prod. 104, 1–12. doi:10.1016/j.jclepro.2015.04.041

683 Sampathkumar, K., Arjunan, T. V., Pitchandi, P., Senthilkumar, P., 2010. Active solar
684 distillation-A detailed review. Renew. Sustain. Energy Rev. 14, 1503–1526.

685 Santos, N.I., Said, A.M., James, D.E., Venkatesh, N.H., 2012. Modeling solar still production

35
 ACCEPTED MANUSCRIPT

686 using local weather data and artificial neural networks. Renew. Energy 40, 71–79.

687 Sethi, S., Walker, S., Drewes, J., Xu, P., 2006. Existing & Emerging Concentrate
688 Minimization & Disposal Practices for Membrane Systems. Florida Water Resour. J. 38,
689 40–45.

690 Shabanpour, H., Yousefi, S., Saen, R.F., 2017. Forecasting efficiency of green suppliers by
691 dynamic data envelopment analysis and artificial neural networks. J. Clean. Prod. 142,
692 1098–1107. doi:10.1016/j.jclepro.2016.08.147

693 Shukla, S.K., 2011. Water Distillation Using Solar Energy System with Lauric Acid as
694 Storage Medium. Int. J. Energy Eng. 1, 1–8.

695 Shukla, S.K., Sorayan, V.P.S., 2005. Thermal modeling of solar stills: An experimental
696 validation. Renew. Energy 30, 683–699.

697 Specific heat of materials, 217AD. engineeringtoolbox [WWW Document].

698 Swingler, K., 1996. Applying Neural Networks: A Practical Guide. Morgan Kaufmann, New
699 York.

700 Tabrizi, F.F., Sharak, A.Z., 2010. Experimental study of an integrated basin solar still with a
701 sandy heat reservoir. Desalination 253, 195–199.

702 Velmurugan, V., Naveen Kumar, K.J., Noorul Haq, T., Srithar, K., 2009a. Performance
703 analysis in stepped solar still for effluent desalination. Energy 34, 1179–1186.

704 Velmurugan, V., Pandiarajan, S., Guruparan, P., Subramanian, L.H., Prabaharan, C.D.,
705 Srithar, K., 2009b. Integrated performance of stepped and single basin solar stills with
706 mini solar pond. Desalination 249, 902–909.

707 Velmurugan, V., Senthil Kumaran, S., Niranjan Prabhu, V., Srithar, K., 2008. Productivity
708 enhancement of stepped solar still - Performance analysis. Therm. Sci. 12, 153–163.

709 Zurigat, Y.H., Abu-Arabi, M.K., 2004. Modelling and performance analysis of a regenerative
710 solar desalination unit. Appl. Therm. Eng. 24, 1061–1072.

711

36