Predictive Modeling of Large-Scale Commercial Water Desalination Plants: Data-Based Neural Network and Model-Based Process Simulation

Abstract

This paper presents a methodology and practical guidelines for developing predictive models for large-scale commercial water desalination plants by (1) a data-based approach using neural networks based on the backpropagation algorithm and (2) a model-based approach using process simulation with advanced software tools ASPEN PLUS and SPEEDUP and compares the relative merits of the two approaches. This study utilizes actual operating data from two of the largest multistage flash (MSF) and reverse osmosis (RO) desalination plants in the world. Our resulting neural network and process simulation models are capable of accurately predicting the actual operating data from commercial MSF desalination plants, but the accuracy of a neural network model depends on both the proper selection of input variables and the broad range of data with which the network is trained. A neural network model can handle noisy data more effectively than statistical regression and performs better in predicting the performance variables of both MSF and RO desalination plants. Our neural network model compares favorably with recent neural network models developed by others in accurately predicting actual operating data from commercial MSF desalination plants. When compared to a data-based neural network, a properly validated model-based process simulation (as in the case of MSF desalination plants) can more effectively quantify the effects of varying operating variables on the desalination performance variables. When it is difficult to develop a model-based process simulation (as in the case of RO desalination plants), we can use a data-based neural network to accurately predict the desalination performance variables.