Research And Application Of Ammonia Nitrogen Prediction Model Based On Random Weight Neural Network

In recent years,due to the excessive discharge of ammonia nitrogen in industrial sewage and urban sewage,some rivers and lakes have been eutrophicated,and people’s living and ecological environment have been seriously affected.The on-line prediction of ammonia nitrogen concentration is an important link in sewage treatment and an important guarantee to prevent water pollution.However,the wastewater treatment process is a complex system characterized by multivariable coupling,large lag,large time variation,etc.,so it is difficult for traditional chemical detection methods and mechanism model detection methods to realize online measurement of effluent ammonia nitrogen,which cannot provide a basis for decision-making in time.Random weight Neural Network（RWNN）,which has the ability to approximate arbitrary nonlinear systems,has been widely used in chaotic time series prediction.Therefore,this paper intends to establish an RWNN based ammonia nitrogen prediction model for the purpose of realizing accurate online ammonia nitrogen measurement,and carry out research on the structure design,learning algorithm and practical application of the effluent ammonia nitrogen prediction model.Firstly,the prediction method of ammonia nitrogen based on extreme learning machine（ELM）was designed.In the ELM,the prediction performance of the ammonia nitrogen prediction model will be affected if the network structure is too large or too small.To solve this problem,a pruning algorithm for smoothing L0 regularization term is proposed.In the process of network model training,the output weights with low contribution are pruned to reduce the redundancy of network structure and improve the performance of network generalization.The experimental results show that the proposed method can effectively sparse the network structure and improve the prediction accuracy and generalization performance of the ammonia nitrogen prediction model.Secondly,the prediction method of ammonia nitrogen based on Echo State Network（ESN）is designed.While ESN avoids the hassle of gradient explosions and vanishing,but its reservoir structure is often too large,which tends to produce ill-posed solutions and affect the prediction effect.To solve this problem,a robust echo state network model（RESN-SOL）based on sparse online learning algorithm is proposed.This model not only introduces the ε-insensitive loss function to improve the robustness of the network,but also designs the sparse online learning algorithm to train the output weights of the network at different learning rates.The simulation results show that the proposed RESN-SOL model has less number of reservoir nodes and higher prediction accuracy.Movever,the soft measurement model of wastewater treatment effluent ammonia nitrogen based on RESN-SOL algorithm is established.Based on the mechanism analysis of effluent ammonia nitrogen,auxiliary variables were selected,a soft sensor model was built and optimized by using the RESN-SOL algorithm.This soft measurement model can not only delete the network redundancy weight,but also accurately reflect the mapping relationship between auxiliary variables and effluent ammonia nitrogen,so this model can realize the online accurate prediction of wastewater treatment effluent ammonia nitrogen.The effluent ammonia nitrogen of a water plant in Beijing was predicted and the validity of the soft measurement model was also verified.Finally,a prediction system for effluent ammonia nitrogen is developed based on the above theory.The system uses Matlab language,My SQL database and other technologies to achieve user information management,main interface design and visualization of prediction results,which meets the demand of online prediction of ammonia nitrogen concentration in the process of sewage treatment.