Based On Improved BP Neural Networks For Predicting Separation Percent In Electrodialysis Process

Electrodialysis(ED) is an electro-membrane process for separation of ions across chargedmembranes from one solution to another with the aid of an electrical potential difference usedas a driving force. The article is described that BP neural network and improved BPalgorithms apply to establishing laboratory-scale electrodialysis cell for modeling neuralnetworks, and predict separation percent(SP) of electrodialysis cell.BP neural network simulates the neural structure of human brain, applies to complexstructures of the networks which have a large number of processing units, and the processingunits can be connected to each other. The self-adaptive and self-learning abilities of BP neuralnetwork can deal with complex nonlinear system effectively, and achieve any nonlinearmappings from inputs to outputs. Using the feature of BP neural network, BP neural networkapplies to the electrodialysis process to predict separation percent in the article. However, BPneural network has its own weaknesses, improved BP algorithms are used to amend theweaknesses of BP neural network. In order to obtain the best prediction results, neuralnetworks train repeatedly, and compare experimental values of separation percent with theirpredictable values which are obtained by training networks. Finally, the best trainingparameters of the networks are obtained, namely, eight is the best number of the nodes of ahidden layer, the value of MSE is0.2504, the values of R of the training data and testingsamples are respectively0.9949and0.9921, the value of MSRE is0.0068, the training time is75; the best learning rate of adaptive learning rate method is0.05, the value of MSE is0.2464,the values of R of the training data and testing samples are respectively0.9967and0.9922,the value of MSRE is0.0057, the training time is32; the best momentum factor of additionalmomentum method is0.15, the value of MSE is0.2537, the values of R of the training dataand testing samples are respectively0.9958and0.9905, the value of MSRE is0.0062, thetraining time is18; the best increasing ratio of weights of flexible BP algorithm is1.2, thevalue of MSE is0.0942, the values of R of the training data and testing samples arerespectively0.9987and0.9974, the value of MSRE is0.0021, the training time is13. It isclear that when the nodes of the hidden layer are8, the best network structure is4:8:1.Comparing with the three algorithms of improved BP algorithms, adaptive learning ratemethod is more predictable accuracy than that of additional momentum method, andadditional momentum method shorten training time, but the predictable accuracy is not high.And flexible BP algorithm has the best fitting effect and the least time of training which canbe up to the target precision, so flexible BP algorithm is the best algorithm of improved BPalgorithms, adaptive learning rate method is the second one, additional momentum method isthe third one. The three algorithms of improved BP algorithms amend BP neural networkfrom different views, so the training results of improved BP algorithms are better than that ofBP neural network. And from the application view of neural networks, BP neural network andimproved BP algorithms can be applied to not only similar realm of membrane science, but also applied to other processes which have data with nonlinear relationships.