Deep Network Hyperparameter Adaptive Research Based On Hybrid Strategy Whale Optimization

In industrial systems,fault diagnosis and prediction of equipment is of great importance.In equipment operation,predicting the occurrence of faults in advance by predicting equipment working conditions and operating and maintaining them in time is a reliable means to ensure the safety of industrial systems.This means is difficult to achieve better results by human observation and prediction.Using deep neural networks,a good prediction system can be constructed,and then reliable prediction results can be obtained.However,the performance of neural networks is limited by the selection of neural network hyperparameters,and it is quite important to achieve hyperparameter adaptivity because of the difficulty of manual selection and low reliability.In this paper,we study the hyperparameter adaption problem of the prediction model based on long and short-term memory artificial neural network,transform it into a mathematical optimization problem,and use the proposed hybrid strategy whale optimization algorithm to seek the optimal solution to construct the HSWO-LSTM prediction model,reduce the influence of human selection of parameters on the model,and improve the accuracy and efficiency of the model.In this paper,firstly,we need to transform the superparametric adaptive problem based on LSTM prediction model into a mathematical optimization problem,and establish suitable objective functions and constraints.Secondly,the HSWO algorithm is proposed for the shortcomings of The Whale Optimization algorithm such as unbalanced population initialization,slow convergence rate at the later stage and trapped local optimum cannot be easily escaped.This algorithm uses chaotic mapping to deal with the imbalance of population initialization,and introduces nonlinear convergence factors and adaptive weights in the search process to balance the global retrieval and local exploitation levels of HSWO.In addition,some adaptations are introduced in the bracketing process to solve the problem of getting into a local optimum that cannot be easily jumped out.Then,the global convergence of the HSWO algorithm is proved from theory.Further,the HSWO algorithm is demonstrated in experiments with single-peak and multi-peak functions to have significant advantages over the WO algorithm and the PSO algorithm in terms of the effect of finding the best and speed.Finally,based on these studies,the HSWO-LSTM prediction model is further proposed for the prediction of thermal hydraulic systems in nuclear power plants.The hyperparameters in the LSTM network are adaptively selected by the HSWO algorithm,and the best prediction model is obtained by training and testing the model,after which the real data are used as the network input and the corresponding five-point prediction values are output.By comparing and analyzing the results of prediction experiments and comparison experiments,the HSWO-LSTM prediction model proposed in this paper outperforms the LSTM prediction model with manually selected hyperparameters.