Inverse fuzzy dynamic optimal self-adaptive aeration control method based on sewage treatment

Abstract

The invention belongs to the field of urban sewage treatment, and provides an inverse fuzzy dynamic optimal self-adaptive aeration control method based on sewage treatment. Comprising the steps of dissolved oxygen concentration kinetic model construction, inverse fuzzy dynamic multi-objective optimization algorithm design, evaluation network and long-term utility function setting, execution network design, actual controller design, execution error function determination, execution network and actual controller parameter optimization and aeration pump control. According to the method, the optimal set value of the dissolved oxygen concentration is obtained through calculation of the inverse fuzzy dynamic multi-objective optimization algorithm; an execution network is designed by utilizing a judgment network and a long-term utility function, and unknown dynamic information of the dissolved oxygen concentration kinetic model is identified; optimization of an execution network and an actual controller is realized through the evaluation network and the execution error function; through the actual controller with the optimal parameters, self-adaptive control over the aeration pump is achieved, the control precision of the aeration degree is improved, and operation energy consumption is effectively reduced.

Description

Technical Field

[0001] The invention relates to the field of urban sewage treatment, and in particular to an inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment. Background Art

[0002] In recent decades, the scarcity of freshwater resources has become increasingly prominent and has evolved into a global crisis that needs to be solved urgently, which has dramatically increased the urgent need for safe and reliable drinking water. As one of the key strategies to alleviate the problem of water shortage, the importance of sewage treatment has attracted widespread attention from countries around the world. The sewage treatment process has the following characteristics: High uncertainty: The sewage treatment process is affected by uncertain interference factors such as weather, water temperature and pH, and its influent flow, composition and pollutant concentration are highly uncertain; Strong nonlinearity: During the aeration reaction, there is a strong nonlinear relationship between the respiration rate and nitrification rate of anaerobic and aerobic bacteria. Therefore, how to design the optimal control strategy under complex operating conditions to achieve high efficiency and low consumption is a key issue. Optimization control strategy is an effective means to accurately control equipment and reduce energy consumption in the production process. By obtaining the optimal set value of the control variable and tracking control, the operating energy consumption can be effectively reduced while ensuring the effluent quality. However, due to the characteristics of high nonlinearity in the sewage treatment process, traditional control methods cannot achieve accurate tracking control. Fuzzy neural networks have shown excellent performance in the face of complex processes such as sewage treatment, so their application has gradually become more widespread. In addition, the reinforcement learning method based on the execution-critic network learning structure has a strong online learning ability, and stands out in some high-dimensional and complex decision-making problems, becoming an ideal choice for optimization control. The judgment network and the execution network use the policy gradient algorithm. By combining the policy gradient and the value function update, they can effectively handle nonlinear problems in complex environments. Application in the sewage treatment control process can enhance robustness and adaptability, improve the effluent quality, and effectively reduce the energy consumption of equipment operation.

[0003] However, the prior art does not have a technical solution for sewage aeration control that simultaneously utilizes an optimization control strategy, a fuzzy neural network, a reinforcement learning method based on an execution-critic network learning structure, and a policy gradient algorithm. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide an inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment, which can achieve the acquisition of the optimal set value of dissolved oxygen concentration, the identification of unknown dynamic information of the dissolved oxygen concentration kinetic model, the optimization of the execution network and the actual controller, and the adaptive control of the aeration pump.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] An inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment, comprising:

[0007] Construct a kinetic model of dissolved oxygen concentration;

[0008] Designing an inverse fuzzy dynamic multi-objective optimization algorithm, and using the inverse fuzzy dynamic multi-objective optimization algorithm to obtain the optimal setting value;

[0009] Setting a long-term utility function and a judgment network based on a first fuzzy neural network according to the dissolved oxygen concentration kinetic model and the optimal setting value;

[0010] Designing an execution network based on a second fuzzy neural network according to the evaluation network and the long-term utility function;

[0011] The execution network is used to approximately obtain the unknown dynamic information of the dissolved oxygen concentration kinetic model; the calculation formula of the unknown dynamic information is:

[0012]

[0013] Where ε(k) = S O,set (k+1)-S O,set (k); S O,set (k) is the set value of dissolved oxygen concentration at time k; is the calculated value of the execution network; k is the time; K O is the saturation coefficient of aerobic respiration of heterotrophic bacteria; X B,H (k) is the heterotrophic biomass concentration; Y H is the heterotrophic biomass yield coefficient; μ(k) is the growth respiration coefficient; S O,5 (k) is the actual concentration of dissolved oxygen in the fifth zone of the biochemical reaction tank; Q 5 (k) is the flow rate of the fifth pool; V 5 is the volume of the biochemical reaction pool in the fifth zone; δ(k) is the tracking error of dissolved oxygen concentration; is the optimal output weight of the execution network; Θ 2 (k) is the output of the normalization layer of the execution network; θ 2 (k) is the boundary approximation value of the execution network;

[0014] The actual controller is designed according to the execution network and the calculation formula of the unknown dynamic information; the calculation formula of the actual controller is:

[0015]

[0016] Among them, K L a 5(k) is the input of the actual controller; η is the control coefficient; is the latest weight value of the execution network at time k; S O,S is the saturation value of dissolved oxygen concentration;

[0017] Determine the execution error function of the execution network; the execution error function is:

[0018]

[0019] Among them, e 2 (k) is the execution error function; is the latest output value of the evaluation network;

[0020] The parameters of the execution network and the actual controller are iteratively optimized using the evaluation network and the execution error function to obtain the actual controller with the optimal parameters; the calculation formula for iterative update of the parameters of the execution network is:

[0021]

[0022] in, E 2 (k) is the loss function of the execution network; ρ 2 is the second learning rate;

[0023] The preset aeration pump is controlled by using the actual controller with the optimal parameters.

[0024] Preferably, the dissolved oxygen concentration kinetic model is:

[0025]

[0026] Among them, K L a 5 (k) is the oxygen transfer coefficient of the last aerobic zone; μ(k) is the growth respiration coefficient; X B,H (k) is the heterotrophic biomass concentration; Y H is the heterotrophic biomass production coefficient; S O,S is the saturation value of dissolved oxygen concentration.

[0027] Preferably, the parameters of the execution network and the actual controller are iteratively optimized by using the evaluation network and the execution error function to obtain the actual controller with optimal parameters, including:

[0028] Taking the historical Pareto optimal frontier as the input of an inverse fuzzy neural network model based on a third fuzzy neural network, taking the historical Pareto optimal solution as the output of the inverse fuzzy neural network model to train the inverse fuzzy neural network model, and obtaining a trained inverse model;

[0029] According to the Pareto optimal frontier of the t-1 environment and the Pareto optimal frontier of the t-2 environment, the Pareto optimal frontier of the t environment is calculated using the central prediction method;

[0030] Mapping the Pareto optimal frontier of the t environment using the inverse model to obtain the Pareto initial solution of the t environment;

[0031] Optimizing the Pareto initial solution of the t environment by using a decomposed multi-objective optimization algorithm to obtain Pareto optimal solutions of multiple t environments, and screening the Pareto optimal solutions of the t environment by using a fuzzy membership function to obtain an optimal setting value;

[0032] Calculating the execution error function according to the optimal setting value to obtain an iterative identification error;

[0033] Iteratively optimizing the weights of the execution network according to the iterative identification error to obtain the iteratively completed execution network;

[0034] The parameters of the actual controller are updated using the weights of the iteratively executed network to obtain the actual controller with optimal parameters.

[0035] Preferably, the long-term utility function is:

[0036]

[0037] in, δ(k)=S O,5 (k)-S O,set (k); U(k) is the utility function; S O,set (k) is the optimal setting value, S O,5 (k) is the actual concentration value of dissolved oxygen in the 5th area of ​​the biochemical reaction tank; γ is the discount factor; the range of γ is 0 to 1.

[0038] Preferably, the historical Pareto optimal frontier is used as the input of an inverse fuzzy neural network model based on a third fuzzy neural network, and the historical Pareto optimal solution is used as the output of the inverse fuzzy neural network model to train the inverse fuzzy neural network model to obtain a trained inverse model, including:

[0039] Determine the calculation formula of the output of the inverse fuzzy neural network model; the calculation formula of the output of the inverse fuzzy neural network model is:

[0040]

[0041] Wherein, x(k) is the output of the inverse fuzzy neural network model; f(k) is the input of the inverse fuzzy neural network model; Ψ 3(k) is the third weight; l is the number of neurons in the input layer; p is the number of neurons in the membership layer and the rule layer; c ij (k) and σ ij (k) are the center value and width value of the membership layer respectively.

[0042] Preferably, the second weight update formula of the execution network is:

[0043]

[0044] Among them, ρ 2 is the second learning rate; is the latest weight value of the execution network.

[0045] Preferably, the Pareto optimal frontier of the t-1 environment and the Pareto optimal frontier of the t-2 environment are calculated using a central prediction method, including:

[0046] The centroids of the Pareto optimal frontier of the t-1 environment and the Pareto optimal frontier of the t-2 environment are calculated respectively to obtain a first centroid and a second centroid; the calculation formula of the first centroid is: The calculation formula of the second centroid is: C t-1 is the first centroid; POF t-1 is the Pareto optimal frontier of the t-1 environment; f i t-1 is the i-th objective function value corresponding to the Pareto optimal frontier of the t-1 environment; C t-2 is the second centroid; POF t-2 is the Pareto optimal frontier of the t-2 environment; f i t-2 is the i-th objective function value corresponding to the Pareto optimal frontier of the t-2 environment; N is the number of the objective function values ​​corresponding to the Pareto optimal frontier;

[0047] The Pareto optimal frontier of the t environment is calculated using the center prediction formula according to the first mass center and the second mass center; the center prediction formula includes: POF t =f i t-1 +C t-1 -C t-2 +G t , G t ~N(0,λ) and m is the dimension of all the objective function values.

[0048] After obtaining the Pareto optimal frontier of the t environment, the inverse model is used to map it to obtain the Pareto initial solution of the t environment, the decomposed multi-objective optimization algorithm is used to optimize the Pareto initial solution of the t environment to obtain the Pareto optimal solutions of multiple t environments, and the fuzzy membership function is used to screen the Pareto optimal solutions of the t environment to obtain the optimal setting value.

[0049] Preferably, it also includes:

[0050] Determine the evaluation network; the output expression of the evaluation network is:

[0051]

[0052] Define the estimation error function of the judgment network; the estimation error function is:

[0053]

[0054] in, e 1 (k) is the estimation error function;

[0055] The loss function of the evaluation network is defined according to the estimation error function; the loss function of the evaluation network is: Among them, E 1 (k) is the loss function of the evaluation network;

[0056] The first weight update formula of the evaluation network is determined according to the evaluation network, the estimation error function and the cost function; the first weight update formula is:

[0057]

[0058] Among them, ρ 1 is the first learning rate;

[0059] The evaluation network is iteratively optimized using the first weight update formula to obtain the iteratively completed evaluation network.

[0060] Preferably, it also includes:

[0061] The control effect of the actual controller with the optimal parameters is evaluated by using the comprehensive square error and the comprehensive absolute error to obtain the evaluation result; the calculation formula of the comprehensive square error is:

[0062]

[0063] The calculation formula of the comprehensive absolute error is:

[0064]

[0065] Wherein, IAE is the integrated square error; ISE is the integrated absolute error; is the total number of samples.

[0066] The present invention discloses the following technical effects:

[0067] The present invention provides an inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment. By using a judgment network and a long-term utility function to design an execution network, the problem of limited reliability caused by updating a controller based on an instantaneous tracking error in the conventional aeration control of the prior art is solved, and accurate acquisition of unknown dynamic information of a dissolved oxygen concentration kinetic model is achieved; by using a judgment network and an execution error function, the problem of low controller control accuracy is solved, and optimization of an execution network and an actual controller is achieved; by designing an inverse fuzzy dynamic optimization algorithm, the problem of optimal parameter acquisition is solved, and real-time calculation of a dynamic optimal set value of dissolved oxygen concentration in a sewage treatment process is achieved; by using an actual controller with optimal parameters, the problem of model-to-equipment correlation control is solved, and adaptive control of an aeration pump is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0069] Figure 1 A schematic diagram of the inverse fuzzy dynamic optimal adaptive aeration control process for sewage treatment provided by an embodiment of the present invention;

[0070] Figure 2 A statistical diagram of water inflow during a sewage treatment process provided by an embodiment of the present invention, Figure 2 (a) is the statistical diagram of the inflow flow of the sewage treatment process under sunny conditions. Figure 2 (b) is a statistical diagram of the inflow flow of sewage treatment process under heavy rain conditions;

[0071] Figure 3 A statistical diagram of the tracking control effect under sunny conditions provided by an embodiment of the present invention, Figure 3 (a) is the control output and set value statistics of dissolved oxygen concentration under sunny conditions. Figure 3 (b) is the tracking error statistics under sunny conditions;

[0072] Figure 4 A statistical diagram of the tracking control effect under rainstorm conditions provided by an embodiment of the present invention, Figure 3(a) is the control output and set value statistics of dissolved oxygen concentration under heavy rain conditions. Figure 3 (b) is the statistical graph of tracking error under heavy rain conditions;

[0073] Figure 5 A statistical diagram of changes in the aeration pump controller provided in an embodiment of the present invention, Figure 5 (a) is the statistical diagram of the aeration pump controller changes under sunny conditions. Figure 5 (b) is a statistical chart of changes in the aeration pump controller under heavy rain conditions. DETAILED DESCRIPTION

[0074] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] The purpose of the present invention is to provide an inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment, which can achieve the acquisition of the optimal set value of dissolved oxygen concentration, the identification of unknown dynamic information of the dissolved oxygen concentration kinetic model, the optimization of the execution network and the actual controller, and the adaptive control of the aeration pump.

[0076] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0077] Figure 1 The schematic diagram of the inverse fuzzy dynamic optimal adaptive aeration control process for sewage treatment provided by the embodiment of the present invention is as follows: Figure 1 As shown, the present invention provides an inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment, comprising:

[0078] Step 100: constructing a dissolved oxygen concentration dynamics model;

[0079] Step 200: designing an inverse fuzzy dynamic multi-objective optimization algorithm, and using the inverse fuzzy dynamic multi-objective optimization algorithm to obtain the optimal setting value;

[0080] Step 300: setting a long-term utility function and a judgment network based on a first fuzzy neural network according to a dissolved oxygen concentration kinetic model and an optimal setting value;

[0081] Step 400: designing an execution network based on a second fuzzy neural network according to the evaluation network and the long-term utility function;

[0082] Step 500: Obtain unknown dynamic information of the dissolved oxygen concentration kinetic model by using the execution network approximation; the calculation formula of the unknown dynamic information is:

[0083]

[0084] Where ε(k) = S O,set (k+1)-S O,set (k); S O,set (k) is the set value of dissolved oxygen concentration at time k; is the calculated value of the execution network; k is the time; K O is the saturation coefficient of aerobic respiration of heterotrophic bacteria; X B,H (k) is the heterotrophic biomass concentration; Y H is the heterotrophic biomass yield coefficient; μ(k) is the growth respiration coefficient; S O,5 (k) is the actual concentration of dissolved oxygen in the fifth zone of the biochemical reaction tank; Q 5 (k) is the flow rate of the fifth pool; V 5 is the volume of the biochemical reaction pool in the fifth zone; δ(k) is the tracking error of dissolved oxygen concentration; is the optimal output weight of the execution network; Θ 2 (k) is the output of the normalization layer of the execution network; θ 2 (k) is the boundary approximation value of the execution network;

[0085] Step 600: Design an actual controller according to the calculation formula of the execution network and unknown dynamic information; the calculation formula of the actual controller is:

[0086]

[0087] Among them, K L a 5 (k) is the input of the actual controller; η is the control coefficient; S is the latest weight value of the execution network at time k; O,S is the saturation value of dissolved oxygen concentration;

[0088] Step 700: Determine the execution error function of the execution network; the execution error function is:

[0089]

[0090] Among them, e 2 (k) is the execution error function; To judge the latest output value of the network;

[0091] Step 800: Optimize the parameters of the execution network and the actual controller iteratively using the judgment network and the execution error function to obtain the actual controller with the best parameters; the calculation formula for iterative update of the parameters of the execution network is:

[0092]

[0093] in, E 2 (k) is the loss function of the execution network; ρ 2 is the second learning rate;

[0094] Step 900: Using the actual controller with optimal parameters to control the preset aeration pump.

[0095] Preferably, the dissolved oxygen concentration kinetic model is:

[0096]

[0097] Among them, K L a 5 (k) is the oxygen transfer coefficient of the last aerobic zone; μ(k) is the growth respiration coefficient; X B,H (k) is the heterotrophic biomass concentration; Y H is the heterotrophic biomass production coefficient; S O,S is the saturation value of dissolved oxygen concentration.

[0098] Furthermore, the parameters of the execution network and the actual controller are iteratively optimized using the judgment network and the execution error function to obtain the actual controller with the optimal parameters, including:

[0099] Taking the historical Pareto optimal frontier as the input of the inverse fuzzy neural network model based on the third fuzzy neural network, taking the historical Pareto optimal solution as the output of the inverse fuzzy neural network model to train the inverse fuzzy neural network model, and obtaining a trained inverse model;

[0100] According to the Pareto optimal frontier of the t-1 environment and the Pareto optimal frontier of the t-2 environment, the Pareto optimal frontier of the t environment is calculated using the central prediction method;

[0101] The inverse model is used to map the Pareto optimal frontier of the t environment and obtain the Pareto initial solution of the t environment;

[0102] The Pareto initial solution of the t environment is optimized by using the decomposed multi-objective optimization algorithm to obtain the Pareto optimal solutions of multiple t environments, and the Pareto optimal solutions of the t environment are screened by using the fuzzy membership function to obtain the optimal setting value;

[0103] The execution error function is calculated according to the optimal setting value to obtain the iterative identification error;

[0104] Iteratively optimize the weights of the execution network according to the iterative identification error to obtain an iteratively completed execution network;

[0105] The weights of the iterative execution network are used to update the parameters of the actual controller to obtain the actual controller with optimal parameters.

[0106] Preferably, the long-term utility function is:

[0107]

[0108] in, δ(k)=S O,5 (k)-S O,set (k); U(k) is the utility function; S O,set (k) is the optimal setting value, S O,5 (k) is the actual concentration value of dissolved oxygen in the 5th area of ​​the biochemical reaction tank; γ is the discount factor; the range of γ is 0 to 1.

[0109] Specifically, the historical Pareto optimal frontier is used as the input of the inverse fuzzy neural network model based on the third fuzzy neural network, and the historical Pareto optimal solution is used as the output of the inverse fuzzy neural network model to train the inverse fuzzy neural network model to obtain a trained inverse model, including:

[0110] Determine the calculation formula of the output of the inverse fuzzy neural network model; the calculation formula of the output of the inverse fuzzy neural network model is:

[0111]

[0112] Among them, x(k) is the output of the inverse fuzzy neural network model; f(k) is the input of the inverse fuzzy neural network model; Ψ 3 (k) is the third weight; l is the number of neurons in the input layer; p is the number of neurons in the membership layer and the rule layer; c ij (k) and σ ij (k) are the center value and width value of the membership layer respectively.

[0113] Specifically, the second weight update formula of the execution network is:

[0114]

[0115] Among them, ρ 2 is the second learning rate; is the latest weight value of the executing network.

[0116] Preferably, in order to obtain the Pareto initial solution of the t environment, the Pareto optimal frontier of the t environment is mapped using an inverse model. Then, the Pareto initial solution of the t environment is optimized using a decomposed multi-objective optimization algorithm to obtain the Pareto optimal solution of the t environment, and the optimal setting value is screened out from the Pareto optimal solution of the t environment using a fuzzy membership function.

[0117] Furthermore, according to the Pareto optimal frontier of the t-1 environment and the Pareto optimal frontier of the t-2 environment, the Pareto optimal frontier of the t environment is calculated using the central prediction method, including:

[0118] Calculate the centroids of the Pareto optimal frontier of the t-1 environment and the Pareto optimal frontier of the t-2 environment respectively, and obtain the first centroid and the second centroid; the calculation formula of the first centroid is: The calculation formula for the second centroid is: C t-1 is the first centroid; POF t-1 is the Pareto optimal frontier of the t-1 environment; f i t-1 is the i-th objective function value corresponding to the Pareto optimal frontier of the t-1 environment; C t-2 is the second centroid; POF t-2 is the Pareto optimal frontier of the t-2 environment; f i t-2 is the i-th objective function value corresponding to the Pareto optimal frontier of the t-2 environment; N is the number of objective function values ​​corresponding to the Pareto optimal frontier;

[0119] The Pareto optimal frontier of the t environment is calculated using the center prediction formula based on the first and second centroids; the center prediction formula includes: POF t =f i t-1 +C t-1 -C t-2 +G t , G t ~N(0,λ) and m is the dimension of all objective function values.

[0120] After obtaining the Pareto optimal frontier of the t environment, the inverse model is used to map it to obtain the Pareto initial solution of the t environment. The decomposed multi-objective optimization algorithm is used to optimize the Pareto initial solution of the t environment to obtain the Pareto optimal solutions of multiple t environments. The fuzzy membership function is used to screen the Pareto optimal solutions of the t environment to obtain the optimal setting value.

[0121] Specifically, it also includes:

[0122] Determine the judgment network; the long-term utility function is approximated by the judgment network, and the output expression of the judgment network is:

[0123]

[0124] Define the estimated error function of the judgment network; the estimated error function is:

[0125]

[0126] in, e 1 (k) is the estimation error function;

[0127] The loss function of the evaluation network is defined according to the estimated error function; the loss function of the evaluation network is: Among them, E 1 (k) is the loss function of the evaluation network;

[0128] The first weight update formula of the judgment network is determined according to the judgment network, the estimated error function and the cost function; the first weight update formula is:

[0129]

[0130] Among them, ρ 1 is the first learning rate;

[0131] The first weight update formula is used to iteratively optimize the judgment network to obtain an iteratively completed judgment network.

[0132] Furthermore, it also includes:

[0133] The comprehensive square error and comprehensive absolute error are used to evaluate the control effect of the actual controller with the optimal parameters, and the evaluation results are obtained; the calculation formula of the comprehensive square error is:

[0134]

[0135] The calculation formula of comprehensive absolute error is:

[0136]

[0137] Among them, IAE is the integrated square error; ISE is the integrated absolute error; is the total number of samples.

[0138] Specifically, the kinetic model for determining dissolved oxygen concentration is:

[0139]

[0140] Among them, k is the time, S O,m(k), m = 4, 5 is the concentration of the mth area of ​​the biochemical reaction pool, X B,H (k) is the heterotrophic biomass concentration, Y H =0.67 represents the heterotrophic biomass yield coefficient, μ(k) is the growth respiration coefficient, K L a 5 (k) is the oxygen transfer coefficient of the last aerobic zone, Q n (k), n = 4, 5 is the flow rate of the nth pool, V 5 =1333 is the volume of the fifth zone biochemical reaction pool, k O =0.2 is the saturation coefficient of aerobic respiration of heterotrophic bacteria, S O,S =8 is the saturation value of dissolved oxygen concentration.

[0141] Furthermore, the optimal set value of dissolved oxygen concentration is dynamically obtained: In the dynamic multi-objective optimization algorithm, the generation of the initial solution is crucial when the operating environment changes. In order to generate an initial population with good convergence and diversity, an inverse fuzzy dynamic multi-objective optimization algorithm is designed, which uses the Pareto optimal solution POS under environment t-1 t-1 and Pareto Optimal Frontier POF t-1 And the inverse fuzzy neural network modeling method is used to predict the initial population of the new environment t.

[0142] Specifically, in environment t-1, and is used as the input and output value training of the inverse fuzzy neural network model, where is the dissolved oxygen concentration in the environment t-1, is the objective function value of operating energy consumption and effluent water quality, N is POS t-1 and POF t-1 Then, in environment t, by directly predicting the individual Generate POF t In order to improve the prediction accuracy of the population and reduce the computational complexity, a center prediction-based method is designed to generate POF t , as follows:

[0143] 1) Calculate POF separately t-1 The centroid C t-1 and POF t-2 The centroid C t-2 :

[0144]

[0145] 2) POF t It can be predicted by the following formula:

[0146] POF t=f i t-1 +C t-1 -C t-2 +G t

[0147] Among them G t ~N(0,λ) is a Gaussian perturbation with mean 0 and standard deviation λ, which can be defined as:

[0148]

[0149] Among them, m=2.

[0150] 3) Use the trained fuzzy neural network model to transform POF t Mapped to the decision space to obtain the initial solution POS of environment t init At the same time, a multi-objective optimization algorithm based on decomposition is selected to obtain the POS init Optimize to find the Pareto optimal solution POS for environment t t Finally, the fuzzy membership function method is used to obtain the POS t Select the optimal setting value, that is, S of the aeration controller in the sewage treatment process O,set (k) value.

[0151] Preferably, the evaluation network is designed: first, the tracking performance is evaluated by a reinforcement learning strategy, and then the control decision is determined by analyzing the evaluation results. Define the dissolved oxygen concentration tracking error δ(k):

[0152] δ(k)=S O,5 (k)-S O,set (k)

[0153] Where S O,5 (k) is the actual concentration of dissolved oxygen, S O,set (k) is the optimal setting value.

[0154] Then, the derivative of the dissolved oxygen concentration tracking error formula Δδ(k) is:

[0155] Δδ(k)=-Q 5 (k)S O,5 (k) / V 5 +Q 4 (k)S O,4 (k) / V 5

[0156] -K O μ(k)X B,H (k) / Y H +ε(k)

[0157] -K L a5 (k)(S O,S -S O,5 (k))

[0158] During the iterative learning phase, the current controller policy can be adjusted by providing rewards or penalties to the critic network. The utility function U(k) is defined as:

[0159]

[0160] The long-term utility function J(k) is as follows:

[0161]

[0162] Where γ∈(0,1) is the discount factor.

[0163] Furthermore, a fuzzy neural network is chosen to approximate the long-term cost function:

[0164]

[0165] in is the optimal weight of the first fuzzy neural network, Θ 1 (k) is the output value of the first fuzzy neural network normalization layer, θ 1 (k) is the boundary approximation value of the first fuzzy neural network. Therefore, the current output value of the judgment network is is the weight value of the current network.

[0166] Function estimation error e 1 (k) is defined as:

[0167]

[0168] Define the cost function of the critic network Therefore, the update rule of the weights in the network is designed as:

[0169]

[0170] Where 0 < ρ 1 <1 is the learning rate.

[0171] Specifically, the controller design based on the execution network: the control strategy is formulated by gradually accumulating control system experience, and the execution network is composed of fuzzy neural networks. Consider the following Lyapunov function V of the execution network 1 (k):

[0172] V 1 (k) = δ 2 (k)

[0173] Then the derivative of the Lyapunov function of the execution network ΔV 1 (k) is:

[0174] ΔV 1 (k) = -K O μ(k)X B,H (k)ξ(k) / Y H +∈(k)ξ(k)

[0175] +Q 4 (k)S O,4 (k)ξ(k) / V 5 -Q 5 (k)S O,5 (k)ξ(k) / V

[0176] +K L a 5 (k)(S O,S -S O,5 (k))ξ(k)

[0177] Since there is unknown dynamic information in the sewage treatment process, the execution network is introduced to approximate the system uncertainty. The execution network is introduced to approximate the unknown dynamic information B(k) of the system in the sewage treatment process, which is expressed as:

[0178]

[0179] in is the optimal output weight of the second fuzzy neural network, Θ 2 (k) is the output of the second fuzzy neural network normalization layer, θ 2 (k) is the second fuzzy neural network boundary approximation value.

[0180] Then enter the desired controller into Designed for:

[0181]

[0182] is the optimal weight expected by the second fuzzy neural network;

[0183] Where η is the control coefficient and the weight change is Therefore, the actual controller input K L a 5 (k) Designed to:

[0184]

[0185] Then the error e of the execution network will be 2 (k) is defined as:

[0186]

[0187] Define the cost function of the execution network as Then the update rule of the weights in the execution network is designed as:

[0188]

[0189] where ρ 2 >0 is the learning rate.

[0190] Further, system stability analysis: the Lyapunov function V of the evaluation network 2 (k) Designed to:

[0191]

[0192] Then the derivative of the Lyapunov function of the evaluation network ΔV 2 (k) is:

[0193]

[0194] in, definition Then ΔV 2 (k) Satisfy:

[0195]

[0196] Execute the network Lyapunov function V 3 (k) designed to:

[0197]

[0198] Then the derivative ΔV of the network Lyapunov function is executed 3 (k) is:

[0199]

[0200] in definition Then ΔV 3 (k) Satisfy:

[0201]

[0202] Introduce the Lyapunov candidate function:

[0203]

[0204] In summary, the derivative ΔV(k) of the Lyapunov candidate function can be obtained as:

[0205]

[0206] According to Young's inequality, we have:

[0207]

[0208]

[0209] can be further rewritten as:

[0210]

[0211] in

[0212] Γ 1 =||Θ 1 (k)|| 2 -2ρ 1 ||Θ 1 (k)|| 4

[0213]

[0214] Furthermore, considering the dynamic time-varying conditions of the sewage treatment process, the inverse fuzzy dynamic optimal aeration adaptive tracking controller and two network weight updating laws can ensure the ideal control performance.

[0215] Optionally, control performance evaluation: In order to fairly evaluate the control performance, the integrated square error (ISE) and the integrated absolute error (IAE) are used as evaluation indicators.

[0216]

[0217] Where S O,5 (k) is the actual concentration of dissolved oxygen, which is measured by the sensor during the sewage treatment process. O,set (k) is the desired dissolved oxygen concentration, is the total number of samples.

[0218] Specifically, Figure 2 The inflow rate of the wastewater treatment process under two conditions, sunny day and rainy day, is shown. Figure 3 The tracking control results under sunny conditions are shown. Figure 3 (a) shows that the control output of dissolved oxygen concentration is closely matched to the set value. Figure 3 (b) shows that the tracking control error remains in the range of [-0.035, 0.055] mg / L. The results show that the designed controller can accurately track the dynamic optimal setpoint in real time. Figure 4 The tracking control results under heavy rain conditions are shown. Figure 4 (a) shows that the control output of dissolved oxygen concentration is closely matched to the set value. Figure 4(b) shows that the tracking control error remains in the range of [-0.1, 0.16] mg / L. The results show that the designed controller can accurately identify the current state and accurately track the dynamic optimal set value even under heavy rainfall disturbance. Figure 5 The aeration pump controller shows the variation of K under sunny and rainy conditions. L a 5 All of them float between [0, 300].

[0219] The beneficial effects of the present invention are as follows:

[0220] The present invention realizes the acquisition of the optimal set value of dissolved oxygen concentration by calculating the inverse fuzzy dynamic multi-objective optimization algorithm; designs the execution network by utilizing the judgment network and the long-term utility function, realizes the identification of unknown dynamic information of the dissolved oxygen concentration kinetic model; realizes the optimization of the execution network and the actual controller by the judgment network and the execution error function; realizes the adaptive control of the aeration pump by the actual controller with the optimal parameters, improves the control accuracy of the aeration degree, and effectively reduces the operating energy consumption.

[0221] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0222] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. An inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment, characterized in that: include: Construct a kinetic model of dissolved oxygen concentration; Designing an inverse fuzzy dynamic multi-objective optimization algorithm, and using the inverse fuzzy dynamic multi-objective optimization algorithm to obtain the optimal setting value; Setting a long-term utility function and a judgment network based on a first fuzzy neural network according to the dissolved oxygen concentration kinetic model and the optimal setting value; Designing an execution network based on a second fuzzy neural network according to the evaluation network and the long-term utility function; The execution network is used to approximately obtain the unknown dynamic information of the dissolved oxygen concentration kinetic model; the calculation formula of the unknown dynamic information is: Where ε(k) = S O,set (k+1)-S O,set (k); S O,set (k) is the set value of dissolved oxygen concentration at time k; B(k) is the calculated value of the execution network; k is the time; K O is the saturation coefficient of aerobic respiration of heterotrophic bacteria; X B,H (k) is the heterotrophic biomass concentration; Y H is the heterotrophic biomass yield coefficient; μ(k) is the growth respiration coefficient; S O,5 (k) is the actual concentration of dissolved oxygen in the fifth zone of the biochemical reaction tank; Q5(k) is the flow rate of the fifth tank; V5 is the volume of the fifth zone biochemical reaction tank; δ(k) is the tracking error of dissolved oxygen concentration; is the optimal output weight of the execution network; θ2(k) is the output of the normalization layer of the execution network; θ2(k) is the boundary approximation value of the execution network; The actual controller is designed according to the execution network and the calculation formula of the unknown dynamic information; the calculation formula of the actual controller is: Among them, K L a5(k) is the input of the actual controller; η is the control coefficient; is the latest weight value of the execution network at time k; S O,S is the saturation value of dissolved oxygen concentration; Determine the execution error function of the execution network; the execution error function is: Wherein, e2(k) is the execution error function; is the latest output value of the evaluation network; The parameters of the execution network and the actual controller are iteratively optimized using the evaluation network and the execution error function to obtain the actual controller with the optimal parameters; the calculation formula for iterative update of the parameters of the execution network is: in, E2(k) is the loss function of the execution network; ρ2 is the second learning rate; The preset aeration pump is controlled by using the actual controller with the optimal parameters.

2. The inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment according to claim 1 is characterized in that: The dissolved oxygen concentration kinetic model is: Among them, K L a5(k) is the oxygen transfer coefficient of the last aerobic zone; μ(k) is the growth respiration coefficient; X B,H (k) is the heterotrophic biomass concentration; Y H is the heterotrophic biomass production coefficient; S O,S is the saturation value of dissolved oxygen concentration.

3. The inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment according to claim 1 is characterized in that: Iteratively optimizing the parameters of the execution network and the actual controller by using the evaluation network and the execution error function to obtain the actual controller with optimal parameters includes: Taking the historical Pareto optimal frontier as the input of an inverse fuzzy neural network model based on a third fuzzy neural network, taking the historical Pareto optimal solution as the output of the inverse fuzzy neural network model to train the inverse fuzzy neural network model, and obtaining a trained inverse model; According to the Pareto optimal frontier of the t-1 environment and the Pareto optimal frontier of the t-2 environment, the Pareto optimal frontier of the t environment is calculated using the central prediction method; Mapping the Pareto optimal frontier of the t environment using the inverse model to obtain the Pareto initial solution of the t environment; Optimizing the Pareto initial solution of the t environment by using a decomposed multi-objective optimization algorithm to obtain Pareto optimal solutions of multiple t environments, and screening the Pareto optimal solutions of the t environment by using a fuzzy membership function to obtain an optimal setting value; Calculating the execution error function according to the optimal setting value to obtain an iterative identification error; Iteratively optimizing the weights of the execution network according to the iterative identification error to obtain the iteratively completed execution network; The parameters of the actual controller are updated using the weights of the iteratively executed network to obtain the actual controller with optimal parameters.

4. The inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment according to claim 1 is characterized in that: The long-term utility function is: in, δ(k)=S O,5 (k)-S O,set (k); U(k) is the utility function; S O , set(k) is the optimal setting value, S O,5 (k) is the actual concentration value of dissolved oxygen in the 5th area of ​​the biochemical reaction tank; γ is the discount factor; the range of γ is 0 to 1.

5. The inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment according to claim 3 is characterized in that: The method uses the historical Pareto optimal frontier as the input of an inverse fuzzy neural network model based on the third fuzzy neural network, uses the historical Pareto optimal solution as the output of the inverse fuzzy neural network model to train the inverse fuzzy neural network model, and obtains a trained inverse model, including: Determine the calculation formula of the output of the inverse fuzzy neural network model; the calculation formula of the output of the inverse fuzzy neural network model is: Wherein, x(k) is the output of the inverse fuzzy neural network model; f(k) is the input of the inverse fuzzy neural network model; Ψ3(k) is the third weight; l is the number of neurons in the input layer; p is the number of neurons in the membership layer and the rule layer; c ij (k) and σ ij (k) are the center value and width value of the membership layer respectively.

6. The inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment according to claim 3 is characterized in that: The second weight update formula of the execution network is: Among them, ρ2 is the second learning rate; is the latest weight value of the execution network.

7. The inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment according to claim 4 is characterized in that: According to the Pareto optimal frontier of the t-1 environment and the Pareto optimal frontier of the t-2 environment, the Pareto optimal frontier of the t environment is calculated using the central prediction method, including: The centroids of the Pareto optimal frontier of the t-1 environment and the Pareto optimal frontier of the t-2 environment are calculated respectively to obtain a first centroid and a second centroid; the calculation formula of the first centroid is: The calculation formula of the second centroid is: C t-1 is the first centroid; POF t-1 is the Pareto optimal frontier of the t-1 environment; f i t-1 is the i-th objective function value corresponding to the Pareto optimal frontier of the t-1 environment; C t-2 is the second centroid; POF t-2 is the Pareto optimal frontier of the t-2 environment; f i t-2 is the i-th objective function value corresponding to the Pareto optimal frontier of the t-2 environment; N is the number of the objective function values ​​corresponding to the Pareto optimal frontier; The Pareto optimal frontier of the t environment is calculated using the center prediction formula according to the first mass center and the second mass center; the center prediction formula includes: POF t =f i t-1 +C t-1 -C t-2 +G t , G t ~N(0,λ) and m is the dimension of all the objective function values. After obtaining the Pareto optimal frontier of the t environment, the inverse model is used to map it to obtain the Pareto initial solution of the t environment, the decomposed multi-objective optimization algorithm is used to optimize the Pareto initial solution of the t environment to obtain the Pareto optimal solutions of multiple t environments, and the fuzzy membership function is used to screen the Pareto optimal solutions of the t environment to obtain the optimal setting value.

8. The inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment according to claim 4 is characterized in that: Also includes: Determine the evaluation network; the output expression of the evaluation network is: Define the estimation error function of the judgment network; the estimation error function is: in, e1(k) is the estimation error function; The loss function of the evaluation network is defined according to the estimation error function; the loss function of the evaluation network is: Wherein, E1(k) is the loss function of the evaluation network; The first weight update formula of the evaluation network is determined according to the evaluation network, the estimation error function and the cost function; the first weight update formula is: Among them, ρ1 is the first learning rate; The evaluation network is iteratively optimized using the first weight update formula to obtain the iteratively completed evaluation network.

9. The inverse fuzzy dynamic optimal adaptive aeration control method based on sewage treatment according to claim 4, characterized in that: Also includes: The control effect of the actual controller with the optimal parameters is evaluated by using the comprehensive square error and the comprehensive absolute error to obtain the evaluation result; the calculation formula of the comprehensive square error is: The calculation formula of the comprehensive absolute error is: Wherein, IAE is the integrated square error; ISE is the integrated absolute error; is the total number of samples.

Images