Method for predicting effluent quality of sewage treatment process based on recursive RBF neural network

Abstract

The present application relates to a kind of sewage treatment process effluent water quality prediction method based on recursive RBF neural network, sewage is handled, data is collected in the treatment process, relevant variable is selected, finally data nonlinear normalization processing is carried out;Dynamic PLS algorithm based on sliding window technology is used to dynamically select relevant variable at each time.The present application is aimed at the problem that the operation condition of sewage treatment process is complex, and it is difficult to establish accurate mathematical mechanism model, utilizes the advantage of RBF neural network in nonlinear system modeling to establish sewage treatment process model and predict its effluent water quality.The present application combines sewage treatment process and RBF neural network deeply, designs the network input-output relationship in line with sewage treatment process, proposes to establish a dynamic model with varying network size, data dimension and key relevant variable, and improves the adaptability of prediction model to working condition.

Description

Technical Field

[0001] This invention relates to the field of data acquisition and processing technology, specifically to a method for predicting effluent water quality in wastewater treatment processes based on a recursive RBF neural network. Background Technology

[0002] The raw data from wastewater treatment processes is enormous, making the construction of predictive models very challenging. This is mainly manifested in the following ways: 1) The data contains a large amount of redundant information, making it difficult to extract useful information and reducing the generalization ability of the predictive model;

[0003] 2) There are a large number of abnormal data in the data. Abnormal data has a great negative impact on the model and may even cause the model to fail to build; 3) The data units are not consistent. Inconsistent units will lead to incorrect prediction of the importance of the data and increase the difficulty of modeling.

[0004] Wastewater treatment is a complex biochemical reaction process, with close interconnections and continuous operation. Effluent quality is influenced not only by changes in current operating conditions but also by historical operating conditions and historical effluent quality. Due to the significant time lag in wastewater treatment, the impact of historical effluent quality and historical operating conditions on current effluent quality cannot be ignored. Summary of the Invention

[0005] In order to reduce data dimensionality, shrink data scale, reduce the impact of abnormal data, give full play to the role of each variable, and reduce the impact and lag characteristics of sudden changes in operating conditions during wastewater treatment on effluent quality, this invention provides a wastewater treatment process effluent quality prediction method based on recursive RBF neural network.

[0006] The technical solution of this invention:

[0007] A method for predicting effluent water quality in wastewater treatment process based on recursive RBF neural network is proposed. The wastewater is treated, data is collected during the treatment process, relevant variables are selected, and finally the data is nonlinearly normalized.

[0008] A dynamic PLS algorithm based on sliding window technology is used to dynamically select relevant variables at each time step. The selection method is as follows:

[0009]

[0010] Where x(t) is the collected data at time t, X(t) is the modeling dataset at time t, and L is the length of the sliding window, the value of which is set according to the operating conditions. When t≥L, the PLS algorithm is used to analyze the data from time tL to t. When t<L, the PLS algorithm is used to analyze all sample data to select the key related variables of the effluent water quality prediction index for that time period. The selected key variables are used as inputs and the corresponding effluent water quality prediction indexes are used as outputs to construct a recursive RBF neural network.

[0011] Preferably, the data compression technique employed involves performing logarithmic normalization after linear compression. The logarithmic normalization rule formula is as follows:

[0012]

[0013] Where, x min and x max and are the minimum and maximum values ​​of the normalized dataset, respectively, where a is an integer between 2 and 10.

[0014] Preferably, partial least squares method is used to filter and decompose the collected data.

[0015] Preferably, the wastewater is treated by a wastewater treatment system, and data is collected during the treatment process. There are 10 data collection points, containing 57 data points.

[0016] The wastewater treatment system includes a primary sedimentation tank, tank A, tank B, and a secondary sedimentation tank connected in sequence. The primary sedimentation tank is equipped with an inlet, and the secondary sedimentation tank is equipped with an outlet. The bottoms of the primary sedimentation tank and the secondary sedimentation tank are connected to a sludge treatment device. The bottom of the secondary sedimentation tank is also connected to the bottom of tank A. Tank A includes an anaerobic zone and an anoxic zone. Tank B is an aerobic zone, which is divided into three parts. The bottoms of the last part of the anaerobic zone of tank A and the aerobic zone of tank B are connected by an external return pipe.

[0017] Preferably, the first sampling point is set before the inlet of the primary sedimentation tank, the second sampling point is set in the anaerobic zone of tank A, the third sampling point is set in the anoxic zone of tank A, the fourth to sixth sampling points are set in the three parts of the aerobic zone of tank B, the seventh sampling point is set in the secondary sedimentation tank, the eighth sampling point is set in the external return pipe, the ninth sampling point is set in the sludge treatment device, and the tenth sampling point is set at the sludge outlet at the bottom of the secondary sedimentation tank.

[0018] The first sampling point collects influent flow rate, influent pH, influent COD, influent ammonia nitrogen (NH3N), and total plant treated water volume;

[0019] The second collection point collected the following data: anaerobic ORP in tank A, ORP before anoxic conditions in tank A, nitrate nitrogen (NO3N) at the end of anoxic conditions in tank A, MLSS at the end of anoxic conditions in tank A, liquid level at the end of anoxic conditions in tank A, DO before aerobic conditions in tank A, DO during aerobic conditions in tank A, ammonia nitrogen (NH3N) before aerobic conditions in tank A, ammonia nitrogen at the end of aerobic conditions in tank A, temperature in tank A, and air flow rate in tank A.

[0020] The third sampling point collected data on the following: aerobic DO in pool A, orthophosphate at the end of aerobic DO in pool A, ammonia nitrogen (NH3N) at the end of aerobic DO in pool A, air flow rate 2 in pool A, air flow rate 3 in pool A, and frequency of the reflux pump 2 in pool A.

[0021] The fourth sampling point collected data on the following parameters: anaerobic ORP in tank B, pre-anoxic ORP in tank B, final nitrate nitrogen (NO3N) in tank B during anoxic conditions, final MLSS in tank B during anoxic conditions, final liquid level in tank B during anoxic conditions, final DO in tank B during aerobic conditions, final ammonia nitrogen (NH3N) in tank B during aerobic conditions, flow meter 1 for dosing pump, control command for dosing pump 1, final ammonia nitrogen in tank A of the secondary system, frequency of rotary pump 1 from deaeration to anoxic conditions in tank A of the secondary system, frequency of rotary pump 1 from deaeration to anoxic conditions in tank B of the secondary system, frequency of rotary pump 2 from deaeration to anoxic conditions in tank B of the secondary system, air flow rate 1 in tank B, and frequency of internal return pump 3 in tank B.

[0022] The fifth sampling point collects DO in aerobic tank 1, DO in aerobic tank 2, orthophosphate in aerobic tank 2, flow meter 2 of dosing pump, control command of dosing pump 2, temperature of tank B, air flow rate 2 of tank B, and frequency of return pump 4 in tank B.

[0023] The sixth sampling point collected data from the PAC2 biological pool flow meter, the final test results for ammonia nitrogen, methanol dosage, and air flow rate in pool B of the second system.

[0024] The seventh sampling point collected orthophosphorus (OP), total phosphorus (OP), pH, and total nitrogen (TN) from the effluent of the secondary sedimentation tank.

[0025] The eighth sampling point collects data from the external return flow meter;

[0026] The ninth collection point collected the remaining sludge discharge volume;

[0027] The tenth sampling point collected the backwash water volume of the sand filter.

[0028] Preferably, the recursive RBF neural network design includes the following steps:

[0029] Historical effluent water quality indicators are used as recursive inputs to predict current effluent water quality indicators. A memory factor is designed in the recursive process to adjust the role of historical effluent water quality.

[0030] The average operating condition is represented by the average distance between operating conditions. The degree of change in the operating condition at a certain moment is quantified by the distance between the operating condition at a certain moment and the average operating condition. The size of the memory factor is set according to the negative correlation of the degree of change. The specific formula is as follows:

[0031]

[0032] Where t is the current time, N is the lag period, which is adjusted according to seasonal changes and is generally taken as 2-3 weeks, and k is the adjustment parameter, which can be obtained through learning. For the average operating conditions during the time period Nt to t, x i Let Nt represent the operating conditions at the i-th time interval within the time interval Nt to t.

[0033] The preferred method for optimizing the parameters of a recursive RBF neural network is to use the instantaneous error function as the cost function for network optimization.

[0034]

[0035] Preferably, to address the issue of insufficient sample learning caused by replacing the cost function, thereby reducing network performance, parameter optimization is performed based on the full gradient information. The full gradient calculation formula is as follows:

[0036]

[0037] Where, θ n φ(t) is the nth parameter to be optimized in the network, y(t) is the network output at time t, y(ti) is the network output at time ith within the time interval Nt to t, and φ k (t) is the output of the k-th neuron in the designed neural network.

[0038]

[0039] Where x(t)=[x1(t),x2(t),…,x d (t),...,x D (t)] T It is the input vector of the network, x(t)∈R D ,d=1,2,...,D,μ k (t)=[μ k,1 (t),μ k,2 (t),...,μ k,D [(t)] is the center vector of the k-th hidden neuron, μ k (t)∈R D k = 1, 2, ..., K, ||x(t) - μ k (t)|| represents x(t) and μ k The Euclidean distance between (t) and σ k (t) is the effective width of the k-th hidden neuron.

[0040] Preferably, the structural adjustment of the recurrent RBF neural network includes the following steps:

[0041] The network neurons are adjusted in real time based on their integrity and orthogonality. The definition of neuron integrity at time t is as follows:

[0042]

[0043] Among them, φ(t)=[φ1(t),φ2(t),…,φ k [(t)] is the neuron output, κ(t) is the kernel matrix of the RBF neural network, and τ 2 (x(t))∈(0,1), τ 2 A larger x(t) indicates better integrity of R(μ(t)).

[0044]

[0045] in,

[0046]

[0047] Preferably, the orthogonality of neurons at time t is defined as follows:

[0048]

[0049] Where, μ k (t)=[μ k,1 (t),μ k,2 (t),...,μ k,D [(t)] is the center vector of the k-th hidden neuron, μ k (t)∈R D k = 1, 2, ..., K, ||μ i (t)-μ j (t)|| represents the Euclidean distance between the centers of the i-th and j-th neurons, σ i (t) is the effective width of the i-th hidden neuron.

[0050] During the network construction phase, neurons are added or removed based on their integrity and orthogonality. When the integrity of neurons is insufficient, new neurons are added to supplement the integrity of the network; when there is a problem of poor orthogonality between neurons, neurons are deleted to reduce redundant information in the network.

[0051] The beneficial effects of this invention are:

[0052] (1) This invention addresses the problem of complex operating conditions in wastewater treatment processes, making it difficult to establish accurate mathematical mechanism models. It utilizes the advantages of RBF neural networks in nonlinear system modeling to establish a wastewater treatment process model and predict its effluent quality. This invention deeply integrates the wastewater treatment process with RBF neural networks, designing a network input-output relationship that conforms to the wastewater treatment process. For the first time, it proposes establishing a dynamic model where network size, data dimensions, and key related variables all change, improving the predictive model's adaptability to operating conditions. This is the main innovation of this invention.

[0053] (2) The present invention designs a relatively complete data compression and fusion technology, preprocesses the sewage treatment data, and solves problems such as data anomalies, inconsistent dimensions, and large data volume, providing data support for subsequent model design. This is the second innovation of the present invention.

[0054] (3) This invention addresses each stage of the prediction model establishment process. First, it preprocesses the data to eliminate the influence of data dimensions, outliers, etc., on the prediction model. Then, it designs a unique structure and parameter adjustment method tailored to the characteristics of the RBF neural network in the prediction model, further improving the performance of the prediction model. The completeness of the prediction model design is another innovation of this invention. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the invention's solution and technical roadmap;

[0056] Figure 2 This is a diagram illustrating the expected effect of nonlinear normalization in this invention.

[0057] Figure 3 This is a diagram of the recursive RBF neural network structure of the present invention;

[0058] Figure 4 This is a structural diagram of the device of the present invention.

[0059] Appendix Figure 4 In the middle, ①-first collection point, ②-second collection point, ③-third collection point, ④-fourth collection point, ⑤-fifth collection point, ⑥-sixth collection point, ⑦-seventh collection point, ⑧-eighth collection point, ⑨-ninth collection point, ⑩-tenth collection point. Detailed Implementation

[0060] (1) Addressing the difficulties in data processing

[0061] Based on the analysis of the correlation mechanism of various effluent water quality indicators and the collection of actual plant data, this invention employs partial least squares (PLS) to screen and decompose the collected data. Data with strong correlations to each effluent water quality indicator are extracted, while data with no or weak correlation are deleted to avoid the adverse effects of multiple couplings of variables in the modeling process. To reduce the impact of outliers and reasonably utilize the predictive value of each relevant variable, a nonlinear normalization method is used to process the original data.

[0062] (2) Addressing the difficulty in constructing mechanistic models

[0063] Based on a full consideration of the characteristics of the sewage treatment operation mechanism, this invention designs a recursive RBF neural network with a memory factor to reflect the relationship between effluent quality and changes in operating conditions and historical effluent quality. This approach aims to leverage historical data while preventing the predictive model from being slow to respond to sudden changes in operating conditions.

[0064] (3) Addressing the difficulty in determining the structure and parameters of prediction models

[0065] To obtain a suitable network structure, this invention, based on research into RBF neural networks, designs a self-organizing network structure adjustment method based on the integrity and orthogonality of network neurons. During model construction, neurons are added or removed according to wastewater treatment operation status data to obtain a suitable network structure. To obtain optimal network parameters, a parameter optimization algorithm based on full gradients is designed to quickly acquire sample information, accelerating the model's parameter optimization.

[0066] Research plan adopted

[0067] Overall Research Methodology

[0068] This invention addresses the challenge of rapidly and accurately measuring effluent quality during wastewater treatment. Following the national guidelines for wastewater environmental quality assessment and monitoring technologies towards intelligence, automation, and networking, it designs a wastewater treatment effluent quality prediction model based on a recursive RBF neural network with memory factors. The specific implementation method is as follows... Figure 1 As shown, the wastewater undergoes three stages of treatment: primary treatment in a primary sedimentation tank, secondary treatment via anaerobic-aerobic reaction, and tertiary treatment via a secondary sedimentation tank for secondary sedimentation.

[0069] technical route

[0070] To achieve the research objectives of this invention, the technical approach adopted is summarized as follows:

[0071] Selection of key relevant variables

[0072] In order to eliminate the adverse effects of the original data and obtain the data for model construction, the preprocessing of the original data in this invention mainly includes the selection of key relevant variables and the nonlinear normalization operation of the data.

[0073] 1) Selection of Key Relevant Variables: Since wastewater treatment is a dynamic process, the operating conditions of different wastewater treatment systems vary significantly. As the operating conditions of the wastewater treatment process change, the changes in various effluent water quality indicators are not entirely the same, and the correlation between parameters at each stage and these indicators also varies. To obtain as many variables as possible that are correlated with the various effluent water quality indicators, multiple sensors and monitoring points were distributed at various stages of the wastewater treatment process, resulting in 57 process variables listed in Table 1. To address the changing operating conditions of the wastewater treatment process, this invention proposes a data selection method based on sliding window technology, dynamically changing the number and types of relevant variables at different times. Wastewater is treated by the wastewater treatment system, and data is collected during the treatment process. There are 10 data collection points (see Table 1). Figure 4 ), containing 57 data collections;

[0074] The wastewater treatment system includes a primary sedimentation tank, tank A, tank B, and a secondary sedimentation tank connected in sequence. The primary sedimentation tank is equipped with an inlet, and the secondary sedimentation tank is equipped with an outlet. The bottoms of the primary sedimentation tank and the secondary sedimentation tank are connected to a sludge treatment device. The bottom of the secondary sedimentation tank is also connected to the bottom of tank A. Tank A includes an anaerobic zone and an anoxic zone. Tank B is an aerobic zone, which is divided into three parts. The bottoms of the last part of the anaerobic zone of tank A and the aerobic zone of tank B are connected by an external return pipe.

[0075] The first sampling point was set before the inlet of the primary sedimentation tank; the second sampling point was set in the anaerobic zone of tank A; the third sampling point was set in the anoxic zone of tank A; the fourth to sixth sampling points were set in the three parts of the aerobic zone of tank B; the seventh sampling point was set in the secondary sedimentation tank; the eighth sampling point was set in the external return pipe; the ninth sampling point was set in the sludge treatment device; and the tenth sampling point was set at the sludge outlet at the bottom of the secondary sedimentation tank.

[0076] The first sampling point collects influent flow rate, influent pH, influent COD, influent ammonia nitrogen (NH3N), and total plant treated water volume;

[0077] The second collection point collected the following data: anaerobic ORP in tank A, ORP before anoxic conditions in tank A, nitrate nitrogen (NO3N) at the end of anoxic conditions in tank A, MLSS at the end of anoxic conditions in tank A, liquid level at the end of anoxic conditions in tank A, DO before aerobic conditions in tank A, DO during aerobic conditions in tank A, ammonia nitrogen (NH3N) before aerobic conditions in tank A, ammonia nitrogen at the end of aerobic conditions in tank A, temperature in tank A, and air flow rate in tank A.

[0078] The third sampling point collected data on the following: aerobic DO in pool A, orthophosphate at the end of aerobic DO in pool A, ammonia nitrogen (NH3N) at the end of aerobic DO in pool A, air flow rate 2 in pool A, air flow rate 3 in pool A, and frequency of the reflux pump 2 in pool A.

[0079] The fourth sampling point collected data on the following parameters: anaerobic ORP in tank B, pre-anoxic ORP in tank B, final nitrate nitrogen (NO3N) in tank B during anoxic conditions, final MLSS in tank B during anoxic conditions, final liquid level in tank B during anoxic conditions, final DO (dissolved oxygen) in tank B before aerobic conditions, final ammonia nitrogen (NH3N) in tank B before aerobic conditions, flow meter 1 for dosing pump, control command for dosing pump 1, final ammonia nitrogen in tank A of the secondary system, frequency of rotary pump 1 from deaeration to anoxic conditions in tank A of the secondary system, frequency of rotary pump 1 from deaeration to anoxic conditions in tank B of the secondary system, frequency of rotary pump 2 from deaeration to anoxic conditions in tank B of the secondary system, air flow rate 1 in tank B, and frequency of internal return pump 3 in tank B. The fifth sampling point collected data on the following parameters: DO during aerobic conditions in tank B, DO during aerobic conditions in tank B, final orthophosphate in tank B during aerobic conditions, flow meter 2 for dosing pump 2, control command for dosing pump 2, temperature in tank B, air flow rate 2 in tank B, and frequency of internal return pump 4 in tank B.

[0080] The sixth sampling point collected data from the PAC2 biological pool flow meter, the final test results for ammonia nitrogen, methanol dosage, and air flow rate in pool B of the second system.

[0081] The seventh sampling point collected orthophosphorus (OP), total phosphorus (OP), pH, and total nitrogen (TN) from the effluent of the secondary sedimentation tank.

[0082] The eighth sampling point collects data from the external return flow meter;

[0083] The ninth collection point collected the remaining sludge discharge volume;

[0084] The tenth sampling point collected the backwash water volume of the sand filter.

[0085] Table 1 Variables in the wastewater treatment process

[0086]

[0087]

[0088] To select key relevant variables, this invention employs a dynamic PLS algorithm based on sliding window technology to dynamically select key relevant variables at each time step. The selection method is as follows:

[0089]

[0090] Here, x(t) represents the collected data at time t, X(t) represents the modeling dataset at time t, and L is the length of the sliding window, the value of which is set according to the operating conditions. Simply put, the PLS algorithm is used to analyze the data from time tL to time t to select key related variables for predicting effluent water quality during this time period. The selected key variables are used as input, and the corresponding effluent water quality prediction indicators are used as output to construct an RBF neural network.

[0091] 2) Nonlinear normalization processing of data

[0092] Given the complexity of wastewater treatment processes, to reduce the impact of outliers during normalization, a nonlinear logarithmic compression method is proposed for data normalization. This ensures that even if significant outliers exist in the original data, the compressed data will be evenly distributed across a smaller area. The expected results are shown in [link to expected results]. Figure 2 The data compression technique employed involves performing a logarithmic normalization step after the commonly used linear compression. The formula for logarithmic normalization is as follows:

[0093]

[0094] Where, x min and x max These are the minimum and maximum values ​​of the normalized dataset, respectively.

[0095] Design of Recurrent RBF Neural Networks with Memory Factor

[0096] 1) Recursive RBF Neural Network Structure Design

[0097] Because the wastewater treatment process is continuous, this invention designs a recursive RBF neural network for model construction. The recursive structure is as follows: Figure 3 As shown, historical effluent water quality indicators are used as recursive inputs to predict current effluent water quality indicators. Considering that the impact of historical effluent water quality on current effluent water quality varies with operating conditions during wastewater treatment, a memory factor is designed in the recursive stage to adjust the role of historical effluent water quality, thus addressing the problem of the prediction model being insensitive or overly sensitive to sudden changes in operating conditions.

[0098] 1) Memory Factor Design Method

[0099] The memory factor is primarily used to adjust the predictive effect of historical effluent water quality on current effluent water quality. Effluent water quality generally changes with operating conditions, but these changes lag behind changes in operating conditions. To address the insensitivity of effluent water quality to operating conditions, the memory factor is adjusted according to these changes. The proposed memory factor design uses the average distance between operating conditions to represent the average operating condition, quantifies the degree of change in operating conditions at a given moment by the distance between the current operating condition and the average operating condition, and sets the memory factor value based on a negative correlation with the degree of change. The specific formula is as follows:

[0100]

[0101] Where N is the lag period and k is the adjustment parameter.

[0102] Structure and parameter optimization methods of recurrent RBF neural networks

[0103] 1) Parameter optimization of recurrent RBF neural networks

[0104] Wastewater treatment processes generate a large amount of data, which remains substantial even after preprocessing. This large dataset poses a challenge for model training. To reduce the difficulty of parameter optimization, this invention uses an instantaneous error function in the parameter optimization process of the recursive RBF neural network.

[0105]

[0106] Replace the cumulative error function

[0107]

[0108] This serves as the cost function for learning. Where e i This represents the error at time i. Replacing the cost function essentially reduces the model's repeated learning process on samples, which can lead to insufficient learning and thus reduced network performance. To address this issue, the parameter optimization in this invention aims to adjust the parameters based on the full gradient information. The full gradient calculation formula is:

[0109]

[0110] Where, θ n y(t) is the nth parameter to be optimized in the network, and y(t) is the network output at time t.

[0111] 2) Structural adjustment of recursive RBF neural networks

[0112] The network structure adjustment adopts the previous research results of the inventor, namely, real-time adjustment based on the integrity and orthogonality of network neurons. This adjustment method has been proven to be scientific and effective.

[0113] The definition of neuron integrity at time t is:

[0114]

[0115] Where x(t)=[x1(t),x2(t),…,x d (t),...,x D (t)] T It is the input vector of the network, x(t)∈R D ,d=1,2,...,D,φ(t)=[φ1(t),φ2(t),…,φ k [τ(t)] is the neuron output, and κ(t) is the kernel matrix of the RBF neural network. According to the definition, τ 2 (x(t))∈(0,1), τ 2A larger x(t) indicates better integrity of ψ(μ(t)).

[0116]

[0117] in,

[0118]

[0119] Where, μ k (t)=[μ k,1 (t),μ k,2 (t),...,μ k,D [(t)] is the center vector of the k-th hidden neuron, μ k (t)∈R D k = 1, 2, ..., K, ||μ i (t)-μ j (t)|| represents the Euclidean distance between the centers of the i-th and j-th neurons, σ i (t) is the effective width of the i-th hidden neuron;

[0120] The orthogonality of neurons at time t is defined as follows:

[0121]

[0122] Where, μ k (t)=[μ k,1 (t),μ k,2 (t),...,μ k,D [(t)] is the center vector of the k-th hidden neuron, μ k (t)∈R D k = 1, 2, ..., K, ||x(t) - μ k (t)|| represents x(t) and μ k The Euclidean distance between (t), σ k (t) is the effective width of the k-th hidden neuron.

[0123] During the network construction phase, neurons are added or removed based on their integrity and orthogonality. When the integrity of neurons is insufficient, new neurons are added to supplement the integrity of the network; when there is a problem of poor orthogonality between neurons, neurons are deleted to reduce redundant information in the network.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for predicting effluent water quality in wastewater treatment processes based on a recursive RBF neural network, characterized in that, Wastewater is treated, data is collected during the treatment process, relevant variables are selected, and finally the data is subjected to nonlinear normalization. A dynamic PLS algorithm based on sliding window technology is used to dynamically select relevant variables at each time step. The selection method is as follows: Where x(t) is the collected data at time t, X(t) is the modeling dataset at time t, and L is the length of the sliding window, the value of which is set according to the operating conditions. When t≥L, the PLS algorithm is used to analyze the data from time tL to t. When t<L, the PLS algorithm is used to analyze all sample data to select the key related variables of the effluent water quality prediction index for that time period. The selected key variables are used as inputs and the corresponding effluent water quality prediction indexes are used as outputs to construct a recursive RBF neural network.

2. The method for predicting effluent quality in wastewater treatment processes based on recurrent RBF neural networks according to claim 1, characterized in that, The data compression technique involves performing a logarithmic normalization step after linear compression. The logarithmic normalization rule formula is as follows: Where, x min and x max and are the minimum and maximum values ​​of the normalized dataset, respectively, where a is an integer between 2 and 10.

3. The method for predicting effluent quality in wastewater treatment processes based on a recursive RBF neural network according to claim 1, characterized in that, The partial least squares method was used to filter and decompose the collected data.

4. The method for predicting effluent quality in wastewater treatment processes based on recurrent RBF neural networks according to claim 3, characterized in that, Wastewater is treated by a wastewater treatment system. Data is collected during the treatment process at 10 data collection points, containing 57 data points. The wastewater treatment system includes a primary sedimentation tank, tank A, tank B, and a secondary sedimentation tank connected in sequence. The primary sedimentation tank is equipped with an inlet, and the secondary sedimentation tank is equipped with an outlet. The bottoms of the primary sedimentation tank and the secondary sedimentation tank are connected to a sludge treatment device. The bottom of the secondary sedimentation tank is also connected to the bottom of tank A. Tank A includes an anaerobic zone and an anoxic zone. Tank B is an aerobic zone, which is divided into three parts. The bottoms of the last part of the anaerobic zone of tank A and the aerobic zone of tank B are connected by an external return pipe.

5. The method for predicting effluent quality in wastewater treatment processes based on a recurrent RBF neural network according to claim 4, characterized in that, The first sampling point was set before the inlet of the primary sedimentation tank; the second sampling point was set in the anaerobic zone of tank A; the third sampling point was set in the anoxic zone of tank A; the fourth to sixth sampling points were set in the three parts of the aerobic zone of tank B; the seventh sampling point was set in the secondary sedimentation tank; the eighth sampling point was set in the external return pipe; the ninth sampling point was set in the sludge treatment device; and the tenth sampling point was set at the sludge outlet at the bottom of the secondary sedimentation tank. The first sampling point collects influent flow rate, influent pH, influent COD, influent ammonia nitrogen (NH3N), and total plant treated water volume; The second collection point collected the following data: anaerobic ORP in tank A, ORP before anoxic conditions in tank A, nitrate nitrogen (NO3N) at the end of anoxic conditions in tank A, MLSS at the end of anoxic conditions in tank A, liquid level at the end of anoxic conditions in tank A, DO before aerobic conditions in tank A, DO during aerobic conditions in tank A, ammonia nitrogen (NH3N) before aerobic conditions in tank A, ammonia nitrogen at the end of aerobic conditions in tank A, temperature in tank A, and air flow rate in tank A. The third sampling point collected data on the following: aerobic DO in pool A, orthophosphate at the end of aerobic DO in pool A, ammonia nitrogen (NH3N) at the end of aerobic DO in pool A, air flow rate 2 in pool A, air flow rate 3 in pool A, and frequency of the reflux pump 2 in pool A. The fourth sampling point collected data on the following parameters: anaerobic ORP in tank B, pre-anoxic ORP in tank B, final nitrate nitrogen (NO3N) in tank B during anoxic conditions, final MLSS in tank B during anoxic conditions, final liquid level in tank B during anoxic conditions, final DO in tank B during aerobic conditions, final ammonia nitrogen (NH3N) in tank B during aerobic conditions, flow meter 1 for dosing pump, control command for dosing pump 1, final ammonia nitrogen in tank A of the secondary system, frequency of rotary pump 1 from deaeration to anoxic conditions in tank A of the secondary system, frequency of rotary pump 1 from deaeration to anoxic conditions in tank B of the secondary system, frequency of rotary pump 2 from deaeration to anoxic conditions in tank B of the secondary system, air flow rate 1 in tank B, and frequency of internal return pump 3 in tank B. The fifth sampling point collects DO in aerobic tank 1, DO in aerobic tank 2, orthophosphate in aerobic tank 2, flow meter 2 of dosing pump, control command of dosing pump 2, temperature of tank B, air flow rate 2 of tank B, and frequency of return pump 4 in tank B. The sixth sampling point collected data from the PAC2 biological pool flow meter, the final test results for ammonia nitrogen, methanol dosage, and air flow rate in pool B of the second system. The seventh sampling point collected the following parameters: orthophosphorus (OP), total phosphorus (OP), pH, and total nitrogen (TN) at the effluent of the secondary sedimentation tank. The eighth sampling point collects data from the external return flow meter; The ninth collection point collected the remaining sludge discharge volume; The tenth sampling point collected the backwash water volume of the sand filter.

6. The method for predicting effluent quality in wastewater treatment processes based on a recurrent RBF neural network according to claim 1, characterized in that, The design of a recursive RBF neural network includes the following steps: Historical effluent water quality indicators are used as recursive inputs to predict current effluent water quality indicators. A memory factor is designed in the recursive process to adjust the role of historical effluent water quality. The average operating condition is represented by the average distance between operating conditions. The degree of change in the operating condition at a certain moment is quantified by the distance between the operating condition at a certain moment and the average operating condition. The size of the memory factor is set according to the negative correlation of the degree of change. The specific formula is as follows: Where t is the current time, N is the lag period, which is adjusted according to seasonal changes and is generally taken as 2-3 weeks, and k is the adjustment parameter, which can be obtained through learning. For the average operating conditions during the time period Nt to t, x i Let Nt represent the operating conditions at the i-th time interval within the time interval Nt to t.

7. The method for predicting effluent quality in wastewater treatment processes based on a recurrent RBF neural network according to claim 6, characterized in that, The parameter optimization method for recursive RBF neural networks is as follows: The instantaneous error function is used as the cost function for network optimization.

8. The method for predicting effluent quality in wastewater treatment processes based on a recurrent RBF neural network according to claim 7, characterized in that, Parameter optimization is based on the full gradient information, and the full gradient calculation formula is: Where, θ n φ(t) is the nth parameter to be optimized in the network, y(t) is the network output at time t, y(ti) is the network output at time ith within the time interval Nt to t, and φ k (t) is the output of the k-th neuron in the designed neural network. Where x(t)=[x1(t),x2(t),…,x d (t),...,x D [t]T is the input vector of the network, x(t)∈R D ,d=1,2,...,D,μ k (t)=[μ k,1 (t),μ k,2 (t),...,μ k,D [(t)] is the center vector of the k-th hidden neuron, μ k (t)∈R D k = 1, 2, ..., K, ||x(t) - μ k (t)|| represents x(t) and μ k The Euclidean distance between (t) and σ k (t) is the effective width of the k-th hidden neuron.

9. The method for predicting effluent quality in wastewater treatment processes based on a recurrent RBF neural network according to claim 8, characterized in that, The structural adjustment of a recursive RBF neural network includes the following steps: The network neurons are adjusted in real time based on their integrity and orthogonality. The definition of neuron integrity at time t is as follows: Among them, φ(t)=[φ1(t),φ2(t),…,φ k [(t)] is the neuron output, κ(t) is the kernel matrix of the RBF neural network, and τ 2 (x(t))∈(0,1), τ 2 The larger (x(t)) is, the better the integrity of ψ(μ(t)), where in, Where, μ k (t)=[μ k,1 (t),μ k,2 (t),...,μ k,D [(t)] is the center vector of the k-th hidden neuron, μ k (t)∈R D k = 1, 2, ..., K, ||μ i (t)-μ j (t)|| represents the Euclidean distance between the centers of the i-th and j-th neurons, σ i (t) is the effective width of the i-th hidden neuron.

10. The method for predicting effluent quality in wastewater treatment processes based on a recurrent RBF neural network according to claim 9, characterized in that, The orthogonality of neurons at time t is defined as follows: Where, μ k (t)=[μ k,1 (t),μ k,2 (t),...,μ k,D [(t)] is the center vector of the k-th hidden neuron, μ k (t)∈R D k = 1, 2, ..., K, ||μ i (t)-μ j (t)|| represents the Euclidean distance between the centers of the i-th and j-th neurons, σ i (t) is the effective width of the i-th hidden neuron; During the network construction phase, neurons are added or removed based on their integrity and orthogonality. When the integrity of neurons is insufficient, new neurons are added to supplement the integrity of the network; when there is a problem of poor orthogonality between neurons, neurons are deleted to reduce redundant information in the network.

Images