## Abstract

The amount of particles and organic matter in wash-waters and effluent from the processing of fruits and vegetables determines whether they need to be treated to fulfil regulatory standards for their intended use. This research proposes a novel technique in photovoltaic cell-based renewable energy in saline water analysis using the oxidation process and deep learning techniques. Here, the saline water oxidation is carried out based on photovoltaic cell-based renewable and saline water analysis is done using Markov fuzzy-based *Q*-radial function neural networks (MFQRFNN). The plan is entirely web-oriented to enable better control and effective monitoring of water consumption. This monitoring makes use of a communication system that collects data in the form of irregularly spaced time series. Experimental analysis has been carried out based on water salinity data in terms of accuracy, precision, recall, specificity, computational cost, and kappa coefficient.

## HIGHLIGHTS

This research proposes a novel technique in photovoltaic cell-based renewable energy in saline water analysis using the oxidation process and deep learning techniques.

Forecasting energy demand is an essential component of PV that aids in the planning of power generation as well as energy trading with a commercial grid.

Deep learning-based models hold great promise for forecasting consumer demands and RES energy generation.

## INTRODUCTION

Analysing water use is important because it helps building managers and operators adopt better usage planning strategies (Alshehri *et al.* 2021). Forecasting is crucial for effective consumption management and ongoing monitoring. Additionally, efficient detection and prevention of water leaks and waste in distribution networks and installations depend on accurate forecasting of use. There have been several approaches investigated to anticipate water demand and usage in close to real-time. Thereafter, more or less successful ways include statistical techniques, filtering and signal processing methods, fuzzy logic, intelligent strategies, and combinations of many models (Elsheikh *et al.* 2022). More recently, cutting-edge models like machine learning (ML) have demonstrated greater performance than traditional models. Deep neural networks (DNN) in particular have become effective forecasting techniques. Regardless of the approach, the forecasting performance's robustness is primarily influenced by context and environmental data, redundant measurement systems, and the short, medium, and long-term planning choices that need to be made. When no data are available to support the usage of a single water metre, projecting water demand remains a significant research challenge (Said *et al.* 2022). To forecast and prepare for the management of water quality, a variety of physical or mathematical models have been created, but they are complicated, time-consuming, and data-intensive. Users in poor nations that have limited data or background information find these models difficult to employ. On the basis of both linear and nonlinear correlations between input and output variables, statistical models of water quality have been constructed. Nevertheless, in situations with multivariate nonlinear connections, these models frequently fall short of accurately capturing the complexity of these events (Maddah *et al.* 2020). Numerous water quality indices have nonlinear, stochastic, and lagged correlations, making it difficult to develop a mathematical model to forecast events under these conditions using conventional methods (Parida *et al.* 2021). Recent years have seen research on artificial intelligence (AI) algorithms for water quality modelling. In order to create models that best capture these interactions, these algorithms investigate obscure and intricate correlations between input and output variables. The data required for AI models can often be obtained from remote sensing platforms, and they have a number of advantages over conventional physical and statistical models, including the following: they are less sensitive to missing data than conventional approaches, their structures are flexible, nonlinear, and robust, and they can handle massive amounts of data and data of various scales (Trappey *et al.* 2019).

Contribution of this research is as follows:

- 1.
To propose novel technique in photovoltaic cell-based renewable energy in saline water analysis using oxidation process and deep learning techniques.

- 2.
The saline water oxidation is carried out based on photovoltaic cell-based renewable and saline water analysis is carried out using Markov fuzzy-based

*Q*-radial function neural networks (MFQRFNN).

Organisation of this article is as follows: section 2 discusses existing techniques in renewable energy in saline water analysis, section 3 explains novel saline water analysis based on oxidation process and deep learning techniques, performance analysis with discussion is given in section 4 and the conclusion of this research with a future scope is given in section 5.

## RELATED WORKS

A number of ML methods can be a potential solution to deal with multiple linear regression (MLR) methods’ limitations. Artificial neural networks (ANNs) are a promising alternative to traditional mathematical models for solving nonlinear issues in science and engineering because of their straightforward construction (Wang *et al.* 2022). According to work (Coşgun *et al.* 2021), ML algorithms and efficient use of field data are helpful in modelling water quality factors which may then be taken into consideration for treatment unit sizing. To further enhance the quality and sustainability of the wastewater treatment process, various optimisation techniques have been developed (Salem *et al.* 2022a). This method is incredibly efficient and simple to use for issues with highly linked input parameters. However, they are unable to produce an explicit equation between the input variables and the output variable(s) which is crucial for engineering problems with real-world applications (Jamei *et al.* 2021). The group method of data handling (GMDH) approach is a trustworthy and self-organising neural network sub-model among the numerous intelligence techniques. The detection of plant diseases, the modulus of a pressure metre, the temperature of the soil, the dispersion coefficients in water pipes, and the small strain shear modulus of grouted sands are just a few nonlinear problems where this method is successfully used (Hu & You 2022). Aslam *et al.* (2021) employed a method for evaluating air enteric viruses discharged from wastewater for irrigation and performed a quantitative microbial risk analysis (QMRA). They discovered that although it requires more computing time, the outcome may help to establish safe water recycling laws. In contrast to conventional methods, Sharma *et al.* (2022) proposed using ANN for disinfection, which may improve predictions of the inactivation of microorganisms over time as well as other physiochemical water parameters, such as temperature and pH (Potential of Hydrogen) value. Bu *et al.* (2021) develops the use of AI for a range of classifications as well as prediction difficulties. It also explains how hybrid AI is used to extract, classify, forecast, and model characteristics utilising a variety of algorithms and optimisation techniques (Tang *et al.* 2020). In a study (Sohani *et al.* 2022), ANNs were used to model the nonlinear operation of biotechnology. Along with these methods, a variety of ML methods have also been used to solve the issue, such as multi-layer perceptron (MLP) (Wang *et al.* 2021), support vector regression (SVR) (Nazari *et al.* 2022), and more recently random forest (RF) in Dave *et al.* (2020). In the scheme of groundwater prediction, the application of Extreme Gradient Boosting (XGB) is uncommon and is seen in only a few studies such as Pombo *et al.* (2022). To estimate ΔTWS, Salem *et al.* (2022b) carefully chose 11 hydrological factors, including total precipitation as well as snow cover. Regarding the idea of utilising prior pixels to anticipate present pixels, the literature does not go into enough detail. SVMs and RF were compared by Meng *et al.* (2021) for predicting current grid-based rainfall up to 3 h in advance, with input being previous radar-derived rainfall. According to Shalaby *et al.* (2022), entire water vapour photos were predicted using ANNs every 30 min using data from two prior photographs.

## SYSTEM MODEL

This section discusses novel techniques in photovoltaic cell-based renewable energy in saline water analysis using the oxidation process and deep learning techniques. The saline water oxidation is carried out using photovoltaic cell-based renewable and saline water analysis is carried out using Markov fuzzy based *Q*-radial function neural networks (MFQRFNN).

### Photovoltaic cell-based renewable energy in saline water analysis using oxidation process

*n*is the diode ideality constant;

*q*is the charge of electron;

*k*is the Boltzmann constant;

*T*is the temperature of junction; is the photo-current; is the reverse saturation currents of diode.

*c*and

_{i}*c*are the starting as well as ultimate concentrations of COD or TOC (mg/L), respectively. In order to maximise the factors affecting the removal of COD as well as TOC from saline solutions, RSM was utilised. pH, reaction time, salt content, and voltage at three coded levels were three independent variables that were looked at (Table 1). Preliminary screening studies as well as literature reviews provided the variable ranges. The experimental design used in this study was a two-level full factorial design with additional central as well as star points. According to Equation (4), total number of experiments (

*N*) can be computed:where

*N*

_{0}is the number of replications in the central point for evaluating net error (five replications),

*N*

_{a}is the number of two-level experiments, and

*N*

_{c}is number of star points (24). So, a total of 29 experiments were created. The software Design Expert 7 was used to do statistical design of trials as well as data analysis. The following equation can be used to represent the second-order method equation for the prediction of the ideal conditions: In the experiments, every variable was optimised. Analysis of Variance (ANOVA) was used to analyse data as well as identify interactions between independent variables of process as well as responses. The experiments were carried out at random to avoid systemic error. The performance of independent variables is determined by coefficients of second-order method, which interpret the degree of removal of researched parameters. Multiple regressions were used to assess the research data. Fischer test was used to control the statistical significance, while determination coefficients were used to control the quality of method fitting (

*F*-test).

Salt water analysis . | Techniques . | Accuracy . | Precision . | Recall . | Specificity . | Computational cost . | Kappa coefficient . |
---|---|---|---|---|---|---|---|

S = 35–140 g/L | GMDH | 82 | 70 | 65 | 58 | 45 | 32 |

SVM | 84 | 72 | 68 | 62 | 48 | 33 | |

PRE_SWA_DL | 86 | 74 | 72 | 63 | 51 | 39 | |

T_{cond} = 20–30 °C | GMDH | 85 | 73 | 66 | 64 | 53 | 35 |

SVM | 88 | 76 | 69 | 66 | 55 | 38 | |

PRE_SWA_DL | 91 | 80 | 74 | 71 | 56 | 42 | |

T_{evap} = 60–80 °C | GMDH | 93 | 82 | 69 | 65 | 54 | 39 |

SVM | 94 | 84 | 72 | 69 | 56 | 43 | |

PRE_SWA_DL | 96 | 86 | 76 | 73 | 58 | 45 |

Salt water analysis . | Techniques . | Accuracy . | Precision . | Recall . | Specificity . | Computational cost . | Kappa coefficient . |
---|---|---|---|---|---|---|---|

S = 35–140 g/L | GMDH | 82 | 70 | 65 | 58 | 45 | 32 |

SVM | 84 | 72 | 68 | 62 | 48 | 33 | |

PRE_SWA_DL | 86 | 74 | 72 | 63 | 51 | 39 | |

T_{cond} = 20–30 °C | GMDH | 85 | 73 | 66 | 64 | 53 | 35 |

SVM | 88 | 76 | 69 | 66 | 55 | 38 | |

PRE_SWA_DL | 91 | 80 | 74 | 71 | 56 | 42 | |

T_{evap} = 60–80 °C | GMDH | 93 | 82 | 69 | 65 | 54 | 39 |

SVM | 94 | 84 | 72 | 69 | 56 | 43 | |

PRE_SWA_DL | 96 | 86 | 76 | 73 | 58 | 45 |

### Markov fuzzy-based *Q*-radial function neural networks (MFQRFNN) based saline water analysis

*X*is a realisation of the MRF represented by

_{t}*X*. The label field can be calculated for any feature extraction issue by maximising the posterior probability distribution given by the following equations:where the estimated labels are indicated by . Since is a constant prior probability, (6) simplifies to Equation (7):where the clique potential function of parameter vector is called

_{t}*θ*. The MAP estimate is shown here as x

_{t}. , the prior probability, and the likelihood function are the two components of Equation (7). It is possible to write the prior probability as the following equation:where is the energy function and

*z*is the partition function written as (a function of clique potentials). The clique potential function is represented by in the spatial domain. It can be defined by equations in terms of the bonding parameter of the MRF model Equations (9) and (10):where

*q*is a nearby site of

*s*and is the bonding parameter for the MRF model

*α*. It is understood that if

*X*is an MRF, it meets the spatial direction of the Markovianity property by the following equation:

_{t}*p*is the pixel at

_{ij}*I*in centre of a window of the dimensions (2

_{j}*Q*+ 1) (2

*Q*+ 1) in the input picture

*I*,

*M*. Type-2 membership function

*Q*. The mean of the k-middle is determined as Equation (26) for a

_{ij}*n*element set =

*x*

_{1},

*x*

_{2},

*x*

_{3},…,

*x*

_{n}retrieved from a window of size (2

*Q*+ 1) where, (2

*Q*+ 1) Equation (26):

*utilising the*

_{ij}*q*membership values for every window's dominant convolved feature. Matrix Δ

*can be written as Equation (31) in mathematics:*

_{ij}*v*and

*h*which represents the parameter :

*v*and

_{i}*h*from all of the datasets are obtained. First half of the formula is simple to calculate. Second half of the equation, which includes all 2|

_{j}*v*| + |

*h*| combinations of

*v*and

*h*, is equivalent to Equation (37) but it is challenging to solve because of the computational complexity:

_{2}O

_{2}(hydrogen peroxide) is yet only partially achieved. Following this revelation, the team concentrated on the WOR's production to better understand the mechanism. They found that when HCO

_{3}(bicarbonate) was present, substantial amounts of H

_{2}O

_{2}instead of O

_{2}(dioxygen) generated, surprisingly accomplishing simultaneous H

_{2}O

_{2}and H

_{2}(Hydrogen) production through PEC water splitting (Figure 4).

## PERFORMANCE ANALYSIS

The experiment was carried out using a computer with the following technological specs: Intel Core i5 7200U, 8 GB of RAM, a 1 TB hard drive, and NVIDIA GTX 760MX graphics. Additionally, Python 3.5 environments were used to mimic implementation of the suggested approach. We conducted a statistical study by evaluating predicted performance to establish the findings of the suggested strategy.

### Dataset description

The feed flow rate (*F* = 400–600 L/h), permeate flux (Pflux (L/h · m^{2})), condenser inlet temperature (*T*_{cond} = 20–30 °C), evaporator inlet temperature (*T*_{evap} = 60–80 °C), and feed salt concentration (*S* = 35–140 g/L) made up the datasets. The principal output was permeate flux. To train NN, data have also been separated into three categories: training, validation, and testing. The training division is typically utilised to get model specifications. While testing division confirms its performance, the validation division checks precision of continuous training to prevent overfitting.

Table 1 gives a comparative analysis between the proposed and existing technique based on salt water analysis. Here the salt water analysis has been carried out based on salinity and temperature analysis of water like *S* = 35–140 g/L, *T*_{cond} = 20–30 °C, *T*_{evap} = 60–80 °C. Parameters analysed are in terms of accuracy, precision, recall, specificity, Computational cost, and Kappa coefficient.

*S*= 35–140 g/L based on water salinity. Here the proposed technique attained accuracy of 86%, precision of 74%, recall of 72%, specificity of 63%, Computational cost of 51%, and Kappa coefficient of 39%, the existing GMDH (Group Method of Data Handling) attained accuracy of 82%, precision of 70%, recall of 65%, specificity of 58%, computational cost of 45%, Kappa coefficient of 32%, and the support vector machine (SVM) attained accuracy of 84%, precision of 72%, recall of 68%, specificity of 62%, computational cost of 48%, and Kappa coefficient of 33%.

*T*

_{cond}= 20–30 °C based on water salinity. Proposed technique attained accuracy of 91%, precision of 80%, recall of 74%, specificity of 71%, computational cost of 56%, and Kappa coefficient of 42%. The existing GMDH attained accuracy of 85%, precision of 73%, recall of 66%, specificity of 64%, computational cost of 53%, and Kappa coefficient of 35%. SVM attained accuracy of 88%, precision of 76%, recall of 69%, specificity of 66%, computational cost of 55%, and Kappa coefficient of 38%.

*T*

_{evap}= 60–80 °C based on water salinity. Proposed technique attained accuracy of 96%, precision of 86%, recall of 76%, specificity of 73%, computational cost of 58%, and Kappa coefficient of 45%. The existing GMDH attained accuracy of 93%, precision of 82%, recall of 69%, specificity of 65%, computational cost of 54%, and Kappa coefficient of 39%, and SVM attained accuracy of 94%, precision of 84%, recall of 72%, specificity of 69%, computational cost of 56%, and Kappa coefficient of 43%. Hence, deep learning-based models hold great promise for forecasting consumer demands and RES energy generation. However, the efficiency of photovoltaic renewable energy is low compared to other sources of energy. Energy from the sun is intermittent and unpredictable and can only be harnessed in the presence of sunlight.

## DISCUSSION

Seawater is too salty and has a high concentration of salt ions to be used for drinking or household purposes. Seawater treatment facilities are less effective and dependable. Deep learning systems have the potential to improve efficiency and accuracy of salt particle analysis in saltwater, which will benefit water treatment plant performance. Before implementing the system, the contribution aids manufacturers and research in evaluating the desalination system's productivity in a beta environment. Forecasting energy demand is an essential component of PV that aids in planning of power generation as well as energy trading with commercial grid. Deep learning-based models hold great promise for forecasting consumer demands and RES energy generation.

## CONCLUSION

This research proposed a novel method in photovoltaic cell in saline water analysis using oxidation process and deep learning techniques based on *Q*-radial function neural networks. Forecasts of renewable availability points are produced using physics-informed ML methods, whose inaccuracy is measured by fitting residuals of the predictor to various statistical distributions using an unsupervised approach. The proposed technique attained accuracy of 96%, precision of 86%, recall of 76%, specificity of 73%, computational cost of 58%, and Kappa coefficient of 45%. In order to more thoroughly analyse the issue of water treatment and create future water treatment systems that are incredibly accurate, we intend to implement our proposed study using hybrid DL methods.

## DATA AVAILABILITY STATEMENT

All relevant data are included in the paper or its Supplementary Information.

## CONFLICT OF INTEREST

The authors declare there is no conflict.