Flue gas desulfurization system state prediction method based on adaptive radial basis function neural network

Abstract

The present application relates to the technical field of artificial intelligence driven complex dynamic system model representation and state prediction, in particular to a flue gas desulfurization system state prediction method based on an adaptive radial basis function neural network, comprising: obtaining real-time working parameters of a target flue gas desulfurization system; inputting the real-time working parameters into a preset state prediction model to predict the real-time flue gas desulfurization state of the target flue gas desulfurization system, wherein the state prediction model is obtained based on training of a training set, the training set comprises historical data of different desulfurization processes of the flue gas desulfurization system, and the state prediction model is constructed by introducing a weight memory mechanism and an adaptive hyperparameter into a radial basis function neural network. The present application efficiently extracts multi-step information by using a model weight information memory mechanism, a learning rate and a momentum factor adaptive update mechanism, improves the perception and prediction accuracy of the prediction model for the SO2 concentration state of the outlet gas, and reduces the model calculation complexity.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-driven complex dynamic system model representation and state prediction technology, and in particular to a state prediction method for flue gas desulfurization systems based on adaptive radial basis function neural networks. It is an innovative application technology that combines self-organizing learning with weighted memory mechanism strategies. Background Technology

[0002] Industrial sectors such as coal-fired power plants, steel smelting, and chemical production are placing higher demands on the operational efficiency and environmental performance of flue gas desulfurization (FGD) systems. As the core equipment for controlling sulfur dioxide (SO2) emissions, the stability and predictive accuracy of the FGD system's operational status directly impact pollutant reduction, equipment lifespan, and enterprise operating costs. Therefore, to improve the intelligent operation of the chemical tail gas desulfurization process and ensure compliant effluent quality, predicting and analyzing the desulfurization process status is crucial.

[0003] At the same time, the high standards and strict requirements of the state for sulfur dioxide emissions from chemical processes have brought new challenges to the existing complex chemical treatment: (1) Traditional desulfurization system state prediction mostly relies on mechanism-based mathematical models or empirical formulas. However, in actual operation, variables such as flue gas parameters, slurry pH value, and absorption tower pressure drop exhibit strong nonlinear coupling characteristics, and are affected by factors such as raw material fluctuations, equipment aging, and external environmental interference, making it difficult for fixed models to accurately depict dynamic processes, and the prediction error accumulates significantly over time. (2) Traditional neural networks lack online learning capabilities and cannot dynamically adjust parameters according to real-time operating conditions, resulting in a sharp drop in prediction accuracy under non-steady-state conditions. Furthermore, there are complex spatiotemporal correlations between variables in the desulfurization process, making it difficult for ordinary network structures to effectively extract multi-scale features.

[0004] In summary, this invention addresses the significant national need for the development of smart environmental protection and pollution prevention, and proposes a flue gas desulfurization system state prediction method based on an adaptive radial basis function neural network to solve problems such as low accuracy of flue gas desulfurization system state prediction, high computational complexity of prediction models, and poor generalization ability. Summary of the Invention

[0005] The purpose of this invention is to provide a state prediction method for flue gas desulfurization systems based on adaptive radial basis function neural networks. By constructing a prediction model through an adaptive radial basis function neural network, the dynamic characteristics of key parameters in the desulfurization process of the flue gas desulfurization system are predicted. By utilizing a weighted memory mechanism, parameter information from historical moments is fully utilized during the learning process to extract multi-step information to optimize and update key parameters in the desulfurization process, thereby improving the real-time tracking capability and state prediction accuracy of key parameters in the desulfurization process.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A state prediction method for flue gas desulfurization systems based on adaptive radial basis function neural networks includes:

[0008] Obtain the real-time operating parameters of the target flue gas desulfurization system;

[0009] The real-time operating parameters are input into a preset state prediction model to predict the real-time flue gas desulfurization state of the target flue gas desulfurization system. The state prediction model is obtained by training a training set, which includes historical data of different desulfurization processes of the flue gas desulfurization system. The state prediction model is constructed by introducing a weight memory mechanism and adaptive hyperparameters in a radial basis function neural network.

[0010] Preferably, constructing the training set includes:

[0011] Collect historical data on different desulfurization processes in the flue gas desulfurization system;

[0012] The historical data is preprocessed to obtain a task sample set;

[0013] Based on the task sample set, auxiliary variables related to SO2 concentration and pH value in the desulfurization process are determined, and the training set is constructed according to the correspondence between the auxiliary variables and key parameters in the desulfurization process.

[0014] Preferably, constructing the state prediction model includes:

[0015] Initialize the radial basis function neural network;

[0016] Based on the correspondence between auxiliary variables and key parameters in the desulfurization process, an initial state prediction model is constructed based on the initialized radial basis function neural network.

[0017] A weight memory mechanism and adaptive hyperparameters are added to the initial state prediction model. The model is trained using the training set to obtain the state prediction model. The weight memory mechanism is used to extract multi-step information from historical data during training to optimize and update key parameters in the desulfurization process. The adaptive hyperparameters are used to dynamically adjust the learning rate and momentum factor according to gradient change characteristics during training.

[0018] Preferably, the radial basis function neural network includes an input layer, a hidden layer, and an output layer connected in sequence, wherein the node activation function of the hidden layer is represented by a radial basis function, and the output of the hidden layer is:

[0019]

[0020] Among them, h i x(t) is the output of the hidden layer, x(t) is the input vector, and c i(t) is the center vector, b i Let ||x(t) - c| represent the width vector, where t represents the current time step, i represents the current node in the hidden layer, m represents the number of nodes in the hidden layer, and ||x(t) - c| represents the width vector. i (t)|| is c i The Euclidean distance between x(t) and x(t).

[0021] Preferably, the output of the radial basis function neural network is:

[0022]

[0023] Among them, y i (t) represents the output of the radial basis function neural network, w ij is the output weight between the i-th hidden node and the j-th output node; p represents the number of output nodes.

[0024] Preferably, the use of the weighted memory mechanism to extract multi-step information from historical data during training to optimize and update key parameters in the desulfurization process includes:

[0025]

[0026]

[0027] Where t represents the current time, η and γ are the learning rate and momentum factor, respectively, J(t) is the model learning performance evaluation value, and c, b, and w are the parameters of the input layer, hidden layer, and output layer, respectively.

[0028] Preferably, dynamically adjusting the learning rate and momentum factor based on gradient change characteristics during training using the adaptive hyperparameters includes:

[0029] Variables are defined based on the similarity between the historical gradient and the current gradient direction at each iteration during training, and the learning rate and momentum factor are dynamically adjusted based on the changing trend of these variables.

[0030] Preferably, the variable is:

[0031]

[0032] Among them, GDC t Indicates the current variable, Indicates the current gradient g t With historical gradient g r The cosine similarity between them, where Θ is the state prediction model parameter, k represents the current iteration step, t represents the current time, and r represents the historical time.

[0033] Preferably, dynamically adjusting the learning rate and momentum factor based on the changing trends of the variables includes:

[0034] When the variable approaches 1, the learning rate and momentum factor will increase; when the variable approaches 0, the learning rate and momentum factor will decrease.

[0035] η ada =η0(1+λ·GDC) t );

[0036] γ ada =γ0(1+ρ·GDC) t );

[0037] Where, η ada and γ ada λ and ρ represent the adjusted learning rate and momentum factor, respectively; η0 and γ0 represent the direction coefficients of the learning rate and momentum factor, respectively; and η0 and γ0 represent the initial values ​​of the learning rate and momentum factor, respectively.

[0038] Preferably, the adaptive parameter update of the state prediction model during training includes:

[0039]

[0040] Where Θ represents the parameters of the state prediction model.

[0041] The beneficial effects of this invention are as follows:

[0042] (1) In view of the problems that existing neural network prediction models have poor generalization ability, cannot learn accurately due to the use of a single learning strategy, and have low operation efficiency, this invention proposes an adaptive radial basis function neural network state prediction model. By making full use of historical parameter information through the weight memory mechanism, correction signals are obtained from multiple dimensions. At the same time, adaptive hyperparameters are designed to dynamically adjust the learning rate and momentum factor according to the gradient direction, which accelerates the convergence speed of the model, improves the generalization performance of the prediction model, and realizes accurate state prediction of the flue gas desulfurization process.

[0043] (2) This invention addresses the problem that the operating status of flue gas desulfurization equipment is affected by various factors, including inlet flue gas temperature, sulfur dioxide concentration, and absorbent flow rate. These factors have complex nonlinear relationships and fluctuating operating conditions, making it difficult to achieve real-time tracking and accurate state prediction of the desulfurization process. Therefore, this invention proposes a flue gas desulfurization system state prediction method based on an adaptive radial basis function neural network. Using a self-organizing radial basis function neural network model as the prediction model, the dynamic characteristics of key parameters in the flue gas desulfurization system are predicted. A weighted memory mechanism is utilized to fully leverage parameter information from historical moments during the learning process, extracting multi-step information to optimize and update key parameters in the desulfurization process, thereby achieving real-time tracking and accurate state prediction of key parameters in the flue gas desulfurization process. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the RBFNN-WMM model structure according to an embodiment of the present invention;

[0046] Figure 2 This is a comparison chart of the prediction results of SO2 concentration by the RBFNN-WMM model and the ordinary RBF model in this embodiment of the invention;

[0047] Figure 3 This is a comparison chart of the prediction errors of SO2 concentration between the RBFNN-WMM model and the ordinary RBF model in this embodiment of the invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] This embodiment provides a method for predicting the state of a flue gas desulfurization system based on an adaptive radial basis function neural network, including:

[0051] Obtain the real-time operating parameters of the target flue gas desulfurization system;

[0052] The real-time operating parameters are input into a preset state prediction model to predict the real-time flue gas desulfurization state of the target flue gas desulfurization system. The state prediction model is obtained by training a training set, which includes historical data of different desulfurization processes of the flue gas desulfurization system. The state prediction model is constructed by introducing a weight memory mechanism and adaptive hyperparameters in a radial basis function neural network.

[0053] Furthermore, constructing the training set includes:

[0054] Collect historical data on different desulfurization processes in the flue gas desulfurization system;

[0055] The historical data is preprocessed to obtain a task sample set;

[0056] Based on the task sample set, auxiliary variables related to SO2 concentration and pH value in the desulfurization process are determined, and the training set is constructed according to the correspondence between the auxiliary variables and key parameters in the desulfurization process.

[0057] Furthermore, constructing the state prediction model includes:

[0058] Initialize the radial basis function neural network;

[0059] Based on the correspondence between auxiliary variables and key parameters in the desulfurization process, an initial state prediction model is constructed based on the initialized radial basis function neural network.

[0060] A weight memory mechanism and adaptive hyperparameters are added to the initial state prediction model. The model is trained using the training set to obtain the state prediction model. The weight memory mechanism is used to extract multi-step information from historical data during training to optimize and update key parameters in the desulfurization process. The adaptive hyperparameters are used to dynamically adjust the learning rate and momentum factor according to gradient change characteristics during training.

[0061] Specifically, this embodiment first analyzes the existing problems in state prediction during flue gas desulfurization (FGD). To improve the real-time tracking of key parameters and the accuracy of state prediction during FGD, a state prediction method for FGD systems based on an adaptive radial basis function (RBF) neural network is proposed. Raw data is obtained from the FGD plant and preprocessed to obtain a task sample set to be processed. Auxiliary variables with strong correlations to sulfur dioxide concentration and pH value during the FGD process are identified. Based on the correspondence between the auxiliary variables and key parameters during the FGD process, a state prediction model based on a RBF neural network is established.

[0062] Then, a weighted memory mechanism is introduced into the state prediction model. During the learning process, the parameter information of historical moments is fully utilized to extract multi-step information to optimize and update the key parameters of the desulfurization process. At the same time, the hyperparameters of the independent variables are designed, and the learning rate and momentum factor are dynamically adjusted according to the gradient change characteristics during the learning process to improve learning efficiency and thus realize model prediction. This provides a static model foundation for realizing real-time tracking and accurate state prediction of key parameters in the desulfurization process.

[0063] Finally, based on the dynamic characteristics of the desulfurization process at different times, the aforementioned data acquisition, data processing, and model training tasks were performed to provide a dynamic model foundation for accurate prediction of key parameters in the desulfurization process. The trained radial basis function neural network was used to predict the state of the desulfurization process, and the performance of the proposed prediction model was evaluated using evaluation metrics.

[0064] This embodiment proposes an adaptive radial basis function neural network (RBFNN-WMM model) state prediction model. It fully utilizes historical parameter information through a weight memory mechanism to obtain correction signals from multiple dimensions. Simultaneously, it designs adaptive hyperparameters to dynamically adjust the learning rate and momentum factor based on the gradient direction, accelerating model convergence, improving the generalization performance of the prediction model, and achieving accurate state prediction of the flue gas desulfurization process. The construction of the state prediction model specifically includes: First, obtaining raw data from the flue gas desulfurization plant to obtain the task sample set to be processed, and identifying auxiliary variables with strong correlations to sulfur dioxide concentration and pH value during the desulfurization process; second, designing independent variable hyperparameters, dynamically adjusting the learning rate and momentum factor based on gradient change characteristics during the learning process, improving learning efficiency, and effectively reducing the prediction time for key parameters of the desulfurization process; finally, designing a prediction model that introduces a weight memory mechanism, fully utilizing parameter information from historical moments during the learning process, extracting multi-step information to optimize and update key parameters of the desulfurization process, thereby achieving model prediction and improving the real-time tracking capability and state prediction accuracy of key parameters during the desulfurization process. Specifically, it includes the following steps:

[0065] Step 1: First, initialize a Radial Basis Function Neural Network (RBFNN) model, which includes an input layer, hidden layers, and an output layer. The activation functions of the nodes in the hidden layers are represented by radial basis functions. In the hidden layers, each node contains a center vector c with the same dimension as the input vector x(t). i (t), c i The Euclidean distance between x(t) and x(t) is defined as ||x(t) - c i (t)||. The output of the hidden layer comes from the output of the non-linear activation function, expressed as:

[0066]

[0067] Among them, h i (t) represents the output of the hidden layer, c i =[c i1 ,c i2 ,...,c in ] and b i =[b1,b2,...,b m ] T These are the center vector and width vector of the Gaussian function, respectively; m and n represent the number of nodes in the hidden layer and the input dimension, respectively. The output description of RBFNN is:

[0068]

[0069] Among them, y i (t) represents the output of the radial basis function neural network, w ijis the output weight between the i-th hidden node and the j-th output node; p represents the number of output nodes.

[0070] Step 2: Add a weight memory mechanism to the prediction model. First, given the data samples: {x(t), y d (t)} N ,in, For the target output, x = [x1(t), x2(t), ..., x n (t)] T The target input is defined as follows. A cost function for evaluating the learning performance of the RBFNN model is defined as follows:

[0071]

[0072] Where N is the number of data samples, y(t) and y d (t) represent the target output and the actual output of the prediction model, respectively. The training method that improves and replaces the learning process of RBFNN using a weighted memory mechanism strategy is described as follows:

[0073]

[0074] Where η and γ are the learning rate and momentum factor, respectively, and c, b, and w are the parameters of the input layer, hidden layer, and output layer, respectively.

[0075] Step 3: Based on the characteristics of gradient changes during the learning process, an adaptive method for hyperparameters η and γ is proposed. First, a variable GDC is defined to measure the similarity between the historical gradient and the current gradient direction in k iterations:

[0076]

[0077] Among them, GDC t Indicates the current variable, Indicates the current gradient g t With historical gradient g r The cosine similarity between them, where Θ is the state prediction model parameter, k represents the current iteration step, t represents the current time, and r represents the historical time.

[0078] The hyperparameters increase when GDC → 1 and decrease when GDC → 0. The adaptive hyperparameters are as follows:

[0079] η ada =η0(1+λ·GDC) t (8);

[0080] γ ada =γ0(1+ρ·GDC) t (9);

[0081] Where, η ada and γ ada λ and ρ represent the adjusted learning rate and momentum factor, respectively; η0 and γ0 represent the direction coefficients of the learning rate and momentum factor, respectively; and η0 and γ0 represent the initial values ​​of the learning rate and momentum factor, respectively.

[0082] Step 4: The parameters of RBFNN-WMM can be adaptively updated as follows:

[0083]

[0084] Where Θ represents the parameters of the state prediction model.

[0085] RBFNN-WMM can not only utilize multi-step information, but also adjust its parameters using adaptive learning rate and momentum factor.

[0086] Model training and testing:

[0087] To evaluate the performance of the self-organizing deep fuzzy neural network prediction model, the following performance metrics are defined:

[0088]

[0089] Where N is the number of data samples, y(t) and y d (t) represents the target output and the actual output of the prediction model, respectively.

[0090] First, the designed state prediction model was pre-trained. The RBFNN-WMM structure size was set to 2-6-1, and the value ranges of w, b, and c were initialized. The adaptive hyperparameters λ = 0.7, ρ = 0.9, η0 = 0.5, and γ0 = 0.6 were determined. The training samples consisted of 2500 sets of data, and the test samples consisted of 100 sets of data.

[0091] Figure 1 It is the RBFNN-WMM model structure; Figure 2 This is a comparison of the SO2 concentration prediction results between the RBFNN-WMM model and the ordinary RBF model. Figure 3 This table compares the prediction errors of the RBFNN-WMM model and the ordinary RBF model for SO2 concentration. Table 1 shows a comparison of the prediction results of various prediction models for SO2 concentration. It can be seen that the RBFNN-WMM prediction model proposed in this embodiment can achieve more efficient prediction of the SO2 concentration in the exhaust gas with higher accuracy.

[0092] Table 1

[0093] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A flue gas desulfurization system state prediction method based on an adaptive radial basis function neural network, characterized by, The method comprises: acquiring real-time working parameters of a target flue gas desulfurization system; inputting the real-time working parameters into a preset state prediction model to predict a real-time flue gas desulfurization state of the target flue gas desulfurization system, wherein the state prediction model is obtained based on a training set, the training set comprises historical data of different desulfurization processes of a flue gas desulfurization system, and the state prediction model is constructed by introducing a weight memory mechanism and an adaptive hyperparameter into a radial basis function neural network; constructing the state prediction model comprises: initializing the radial basis function neural network; constructing an initial state prediction model based on the initialized radial basis function neural network according to a correspondence between auxiliary variables and key parameters in a desulfurization process; adding the weight memory mechanism and the adaptive hyperparameter to the initial state prediction model, training the training set to obtain the state prediction model, wherein the weight memory mechanism is used to extract multi-step information from historical data to realize optimization and update of key parameters in a desulfurization process during training, and the adaptive hyperparameter is used to dynamically adjust a learning rate and a momentum factor according to gradient change characteristics during training; the radial basis function neural network comprises an input layer, a hidden layer and an output layer connected in sequence, wherein a node activation function of the hidden layer is represented by a radial basis function, and an output of the hidden layer is: ； wherein, is the output of the hidden layer, is the input vector, is the center vector, is the width vector, denotes the current time instant, denotes the current node in the hidden layer, denotes the number of nodes in the hidden layer, is and the Euclidean distance between and an output of the radial basis function neural network is: ； wherein, is an output of the radial basis function neural network, is an output weight between the th hidden node and the th output node; denotes the number of output nodes. In the prediction model, the weight memory mechanism is added. First, given the data sample: wherein, is the target output, is the target input, and a cost function for evaluating the learning performance of the RBFNN model is defined as follows: ； where N is the number of data samples, and are the target output and the predicted model actual output, respectively; extracting multi-step information from historical data to realize optimization and update of key parameters in a desulfurization process during training by using the weight memory mechanism comprises: ； ； ； wherein, denotes the current time, and are the learning rate and the momentum factor, respectively, is the model learning performance evaluation value, , , are the parameters of the input layer, the hidden layer and the output layer, respectively; dynamically adjusting a learning rate and a momentum factor according to gradient change characteristics during training by using the adaptive hyperparameter comprises: defining a variable according to similarity of historical gradient and current gradient direction of each iteration during training, and dynamically adjusting the learning rate and the momentum factor based on a change trend of the variable; the variable is: ； wherein, denotes the current variable, denotes the current gradient the cosine similarity between the historical gradient, is the state prediction model parameter, denotes the current iteration step, denotes the current time instant, denotes the historical time instant;​ dynamically adjusting the learning rate and the momentum factor based on the change trend of the variable comprises: when the variable tends to 1, the learning rate and the momentum factor will increase; when the variable tends to 0, the learning rate and the momentum factor will decrease, and are: ； ； wherein, and denote the adjusted learning rate and momentum factor, respectively, and denote the directional coefficients of the learning rate and momentum factor, respectively, and denote the initial values of the learning rate and momentum factor, respectively; parameter adaptive update of the state prediction model during training comprises: ； wherein, are state prediction model parameters.

2. The method of claim 1, wherein the method is a method of predicting a state of a flue gas desulfurization system based on an adaptive radial basis function neural network, the method comprising: constructing the training set comprises: collecting historical data of different desulfurization processes of a flue gas desulfurization system; preprocessing the historical data to obtain a task sample set; determining auxiliary variables related to SO2 concentration and pH value in a desulfurization process based on the task sample set, and constructing the training set according to a correspondence between the auxiliary variables and key parameters in the desulfurization process.

Images