**7. ANN as a prediction method for the performance of desalination systems**

Artificial intelligence (AI) is quickly progressing every day. Computers can easily perform the difficult tasks for the humans to do. AI is first contrived in 1950. AI passed with steps and stages; some failed and others succeeded. Deep learning with powerful computers is the main keys of the success of AI. With time, AI can solve numerous industrial problems including the image processing, recognition of speech, language processing, optimization, prediction, robotic automation, categorization, self-driving cars …etc. AI is of numerous branches, and we will focus, in this chapter, on the neural networks (NNs). NNs help the computers to teach themselves from observing the data to predict the system performances. Preparing data for neural network processing is typically the most difficult and time-consuming task you will encounter when working with neural networks. In addition to the enormous volume of data that could easily reach millions and even billions of records, the main difficulty is in preparing the data in the correct format for the task in question.

#### **7.1 Application of ANN on desalination systems**

Contemporary experience indicates that artificial neural networks (ANN's) may be specifically appropriate to offer tools to assist desalination plant operators in day-to-day operations. Prediction of the productivity of the desalination units is an essential consequence to evaluate its potentiality to provide potable water with involving in conducting more experiments. AI-based methods are stated as robust tools to obtain the correlation between the process parameters and responses [123–127]. This chapter suggests that ANNs are particularly appropriate as the basis for the development of tools to aid in the various phases of operating a desalination plant. In solar energy applications, different kinds of ANNs predicted and optimized the performance of collectors, cells, distillers, etc. [128]. The most application of ANNs is solving the issues of designing and procedures in electric power systems. Due to the similarities between power plants and desalination plants, the methods of ANNs can be used to predict the performance of desalination plants [129]. The merits of applying ANN included that it handles large amounts of datasets, can discover complex nonlinear relationships among all parameters (dependent and independent), and can find all interactions among all tested parameters.

Zarei and Behyad [130] introduced an ANN to test the effective parameters of the greenhouse on water productivity such as width, length, height of evaporator, and roof transparency. The method obtained an acceptable agreement with the experimental results. Tayyebi and Alishiri [131] used a nonlinear inverse model control strategy based on neural network for predicting the performance of MSF desalination plant. The proposed NNs consisted of three layers identified from input–output data and trained with a descent gradient algorithm. The set point tracking performance of the proposed method was studied when the disturbance is present in MSF unit. Three controllers were provided for checking the saline temperature, last stage level, and salinity. The results obtained that a neural network inverse model control strategy is robust and highly promising to be implemented in such nonlinear systems. An ANN model based on field data was built to investigate the vacuum membrane distillation (VMD) performance at various factors [132]. The introduced model could accurately predict the unseen data of VMD. The correlation coefficient of the overall agreement between ANN predictions and experimental data was found to be more than 0.994. Aish et al. [133] utilized ANN to forecast reverse osmosis performance in Gaza Strip through predicting the next week values of total dissolved solids (TDS) and permeate flow rate of the product water. Multilayer perceptron (MLP) and radial basis function (RBF) neural networks were trained and developed with reference to feedwater parameters including the pressure, pH, and conductivity to predict permeate next week values of flow rate. MLP and RBF neural networks were used for predicting the next week TDS concentrations. Both networks are trained and developed with reference to product water quality variables including the water temperature, pH, conductivity, and pressure. The prediction results showed that both types of neural networks are highly satisfactory for predicting TDS level in the product water quality and satisfactory for predicting permeate flow rate.

Santos et al. [134] simulated the distiller yield based on local weather data such as the solar irradiance, air temperature, glass temperature, and air speed and direction by ANNs. Essa et al. [51] introduced an enhanced ANNs incorporated to Harris Hawks optimizer to anticipate the distiller yield. The optimizer was utilized to get the optimal structure and network factors. Nevertheless, traditional artificial neural networks suffer from trapping in local optima problem during learning process. To overcome this problem, RVFL is proposed as it determines the output weights without involving in time-consuming learning process. Besides, RVFL has a direct connection among the input and output nodes, which is of a substantial impact on the network performance. This aids in avoiding the overfitting issue occurring in other ANNs. Using these merits, the conventional RVFL networks has some restrictions like that there are various conformations can be utilized, and

*Thermal Desalination Systems: From Traditionality to Modernity and Development DOI: http://dx.doi.org/10.5772/intechopen.101128*

this can affect the quality of water productivity final prediction. Then, Ensemble Random Vector Functional Link Networks (EnsRVFL) [135], which relies on RVFL as a basic modeling, were utilized to eliminate these restrictions of RVFL. A random vector functional link (RVFL) network integrated with artificial ecosystem-based optimization (AEO) algorithm was introduced to predict the seawater greenhouse (SWGH) performance [52]. Power consumption and yield of SWGH were predicted *via* RVFL-AEO modeling. Besides, the statistical analyses with various statistical tools were conducted to find out the neural network efficacy. The statistical tools revealed a complete match between the field and modeling data. The RVFL-AEO performance was compared with that of conventional RVFL. RVFL-AEO obtained an improved performance as compared by RVFL, which indicates the role of AEO in obtaining the optimal RVFL factors that improved the model accuracy.
