Prediction of temperature elevation for seawater in multi-stage flash desalination plants using radial basis function neural network

https://doi.org/10.1016/j.cej.2010.05.060Get rights and content

Abstract

In this paper, a radial basis function (RBF) neural network model was developed for estimating temperature elevation (TE) in multi-stage flash (MSF) desalination processes. The constructed artificial neural network (ANN) model use as input variables the boiling point temperature (BPT) and salinity. The developed RBF neural network was found to be precise in predicting TE from the input variables. The performance of the ANN model was analyzed by mean squared error (MSE). The developed RBF neural network was found to be highly precise in predicting TE for the new input data, which are kept unaware of the trained network showing its applicability to estimate the TE for seawater in MSF desalination plants better than the empirical correlations, thermodynamic models and MLP neural network.

Introduction

Desalination is the natural continuous process, which is essential for the water recycle. Desalination methods are classified into two major processes: thermal and non-thermal. Thermal distillation involves phase changes and it includes multi-stage flash (MSF), vapor-compression (VC), and multi-effect (ME). Non-thermal processes do not involve phase changes and includes reverse osmosis (RO), electro-dialysis (ED) and ion exchange (IE) [1]. RO plants can be considered as ideal processes for the seawater desalination from several viewpoints.

An ideal system requires least operating resources that are recoverable from its product if desired. Both RO and MSF plants are non-linear processes, which should operate with performance optimization under specific constraints. Although the MSF process as well as the ME process consumes a larger amount of energy than the RO process, about 18 kWh/m3 for MSF, 15 kWh/m3 for ME, and 5 kWh/m3 for RO, the reliable performance of the thermal desalination processes MSF and ME made highly competitive against the RO process.

At present, MSF units with large production capacity have the largest sector in the desalination industries.

MSF plants are used for the production of potable water and process water from seawater and brackish water. Saline water is steam heated and then led into a series of stages where reduced pressure leads to immediate boiling (flash) without the need to supply additional heat [2]. A schematic diagram of a MSF desalination process is shown in Fig. 1.

In addition, MSF desalination plants especially large ones are often paired with power plants in a cogeneration configuration. Waste heat from the power plant is used to heat the seawater, providing cooling for the power plant at the same time.

This reduces the energy needed from one-half to two-thirds, which drastically alters the economics of the plant, since energy is by far the largest operating cost of MSF plants [3], [4].

Modeling of MSF plants are well established in Refs. [5], [6], [7], [8], [9], [10], [11], [12]. The steady and unsteady state models [8], [9], [10], [11], [12] can be used for evaluating the design characteristics of the process and study the transient behaviors, respectively. In MSF plants, the incoming seawater passes through the heating stages and is heated further in the heat recovery sections of each subsequent stage.

After passing through the last heat recovery section, and before entering the first stage where flash boiling occurs, the feed water is further heated in the brine heater using externally supplied steam. This raises the feed water to its highest temperature (boiling point temperature or top brine temperature), after which it is passed through the various stages where flashing takes place.

The seawater BPT is usually calculated by summing up the BPT of pure water at a given pressure and the TE due to salinity. It is increases the danger of corrosion and scaling in the plant. Thus, a proper knowledge of TE can lead to the better optimization and control of the system and prevent the errors in calculating the design of process equipments.

Several investigations are performed to model the TE predictions of various source of seawater based on the empirical correlations [9], [14], [15]. El-Dessouky and Ettouney [13] developed an empirical correlation for calculating TE as a function of BPT and the salinity in weight percent of seawater. In addition, a detailed model incorporating neural networks for physical properties estimation describes the MSF desalination process. In previous work [16], for estimating TE in MSF plants, for each source of experimental data an MLP neural network model was constructed.

The choice of the input variables is the key to insure complete description of the systems, whereas the quality and the number of the training observations (experimental data) have a critical impact on the reliability and the performance of the neural network. The most important variables that affect the TE estimations are the BPT and salinity of seawater [16]. Therefore, experimental data are required for accurate estimating of TE for given BPT (degree of Celsius) and salinity (weight percent).

Experimental data are usually not available in the wide range of operating conditions and therefore should be predicted using accurate models. Using experimental data, the TE for seawater was evaluated from such models.

ANN is another type of modeling procedure. ANNs have superiority as compared with other conventional modeling techniques. The advantage of ANN is that it does not need any knowledge about the process. ANN, however, is capable of modeling highly complex and non-linear systems with large numbers of inputs and outputs. ANNs have been widely used in many fields such as process modeling, control, optimization and prediction [17], [18], [19].

In this paper, an RBF neural network model as an adequate powerful tool was developed for predicting TE over a wide range of operating conditions, which are based on the available experimental data.

Section snippets

RBF neural network background

The objective of this work is to explore the use of a RBF neural network for the prediction of TE in MSF desalination plants. A RBF consists of an input layer, hidden layer and output layer with the activation function of the hidden units being radial basis functions (Fig. 2).

Normally, an RBF consists of one hidden layer, and a linear output layer. One of the most common kinds of radial basis function is the Gaussian bell-shaped distribution. The response of the hidden layer unit is dependent

Simulation results

It should be mentioned that the empirical correlations such as the El-Dessouky and Ettouney correlation [13], could not provide an adequate predictions of temperature elevation for the wide range of operating conditions. However, the proposed NRBF neural network model as an adequate powerful tool can be used to predict the TE in the range of available experimental data as well as for some seawater compositions, which are not represented in the training data set. Table 3 represents the

Conclusions

In order to proper operation of MSF desalination plants, an RBF neural network has been used to predict the temperature elevation during the MSF process. In an MSF process, top brine temperature is one of the important parameters, which could be obtained at a proper TE estimation. Furthermore, the danger of corrosion and the energy consumption reduced and the design of process equipments (e.g., heat transfer area, the size of flash chamber) will calculate in such a way to minimize the total

References (25)

  • W.T. Hanbury et al.

    Desalination Technology

    (1993)
  • J.H.S.C. Low Tay et al.

    Vacuum desalination for water purification using waste heat

    Desalination

    (1996)
  • Cited by (27)

    • A review on state-of-the-art applications of data-driven methods in desalination systems

      2022, Desalination
      Citation Excerpt :

      The most important highlights drawn from Table 3 are presented below. It can be seen from Table 3 that most of the studies used ANNs to predict the performance of various desalination systems [37,63–78]. This lies in the fact that the ANN model has the privilege of performance anticipation of non-linear systems with high generalization capability and accuracy.

    View all citing articles on Scopus
    View full text