Biochemical reaction activation function-network structure simulation method for biological sewage treatment
By introducing a neural network structure with sparse masking and adaptive Monod activation function into the wastewater biological treatment model, the shortcomings of traditional models in parameter acquisition and nonlinear description are solved, achieving efficient and accurate simulation of the wastewater treatment process and improving the operational stability and reliability of wastewater treatment plants.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-23
- Publication Date
- 2026-03-27
AI Technical Summary
Existing biological wastewater treatment models have shortcomings in parameter acquisition, computational efficiency, and generalization ability, making it difficult to adapt to the rapid response requirements under dynamic operating conditions. Furthermore, traditional neural network models fail to accurately describe the nonlinear laws of microbial metabolism, leading to the accumulation of prediction errors and affecting the operational stability of wastewater treatment plants.
A biochemical reaction activation function-network structure simulation method for wastewater biological treatment is designed. Sparse masked connections and adaptive Monod activation functions are used to construct a neural network structure based on wastewater treatment processes, including anaerobic, anoxic, and aerobic feature extraction layers. By combining physical constraints and data-driven approaches, the accuracy and interpretability of the model are improved.
It improves the accuracy and efficiency of wastewater biological treatment process simulation, can more accurately capture water quality change characteristics, reduces model training time, enhances stability against noise data, and realizes the correspondence between the model and actual process units, thereby improving the operational reliability of wastewater treatment plants.
Smart Images
Figure CN121747757A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of environmental engineering and artificial intelligence, specifically to a method for simulating biochemical reaction activation functions-network structures in wastewater biological treatment. Background Technology
[0002] Biological wastewater treatment is the core component of modern wastewater treatment plants for pollutant removal. Essentially, it is a highly nonlinear, time-varying, and complexly coupled dynamic system. Multiple factors, including the metabolic activities of the microbial community, changes in substrate concentration, dissolved oxygen gradient distribution, and sludge recirculation, interact to form a complex multi-scale biochemical reaction coupling process.
[0003] Traditional simulation methods largely rely on mathematical models based on biochemical reaction mechanisms, such as the Activated Sludge Model (ASM) series. While these models can comprehensively describe the biochemical reaction mechanisms of wastewater treatment processes, they have significant limitations in practical applications. Firstly, the model parameters are numerous, requiring calibration through complex experiments or long-term operational data; parameter sensitivity analysis and optimization are time-consuming and labor-intensive. Secondly, the high complexity of the models leads to high computational resource requirements, making it difficult to meet the rapid response demands of real-time optimization control in wastewater treatment plants. Furthermore, the microbial kinetic equations involved in the models (such as the Monod equation and Haldane equation) require assumptions about specific reaction pathways and parameter ranges; if actual operating conditions deviate from the set conditions, the model's prediction accuracy will significantly decrease. Therefore, although traditional mechanistic models possess theoretical completeness, their shortcomings in parameter acquisition, computational efficiency, and generalization ability limit their engineering application value under dynamic operating conditions.
[0004] With the rise of data-driven approaches, neural network-based modeling techniques have been increasingly introduced into the simulation of wastewater biological treatment processes. Existing research often employs the standard multilayer perceptron (MLP) architecture, combined with the ReLU activation function to construct predictive models. However, the application of such general-purpose neural networks in wastewater treatment scenarios still has significant limitations:
[0005] 1) The fully connected structure of the MLP does not consider the physical flow characteristics of the wastewater treatment process. The anaerobic-anoxic-aerobic (A2-A3) process in wastewater treatment plants... 2 In the O) process flow, the biochemical reactions at each stage have clear hierarchical relationships and mass transfer paths. However, fully connected networks indiscriminately associate all input and output nodes, leading to severe parameter redundancy. This structural design is highly prone to overfitting when faced with noisy data, especially when the amount of training data is limited or the data quality is uneven, resulting in a significant reduction in the model's generalization ability.
[0006] 2) Existing data-driven models generally employ linear activation functions such as ReLU, whose inherent linear characteristics fundamentally conflict with the nonlinear metabolic processes of microorganisms in wastewater treatment. Microbial substrate uptake rates typically follow biochemical reaction kinetics (manifested as substrate saturation characteristics) or are influenced by inhibitory effects (such as the negative feedback of toxic substances). The ReLU function, however, can only capture linear relationships and cannot describe these crucial biochemical characteristics. This not only leads to slow model convergence and limited prediction accuracy but also results in a lack of biological interpretability of the model's internal parameters, making it difficult to guide process optimization and fault diagnosis.
[0007] 3) Due to the lack of modeling of the coupling relationship between each stage in the process flow, existing models often ignore the dynamic interaction mechanism between anaerobic, anoxic and aerobic stages, resulting in the accumulation of prediction errors for key effluent water quality indicators (such as ammonia nitrogen, total phosphorus, COD, etc.), which affects the operational stability of wastewater treatment plants and their ability to ensure compliance with discharge standards.
[0008] In summary, there is an urgent need for a new modeling method that can integrate the constraints of process mechanisms with the advantages of data-driven approaches to solve the problem of accurate simulation of nonlinear dynamic systems in wastewater treatment. Summary of the Invention
[0009] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide a biochemical reaction activation function-network structure simulation method for wastewater biological treatment, designing feature extraction layers corresponding to the anaerobic, anoxic, and aerobic stages of wastewater biological treatment, and simultaneously achieving inter-layer connections through sparse masks and using adaptive Monod activation functions as the activation functions of neurons, thereby improving the accuracy, convergence speed, and interpretability of wastewater biological treatment process simulation (i.e., prediction of key effluent water quality data).
[0010] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0011] Biochemical reaction activation function-network structure simulation method for wastewater biological treatment, including:
[0012] S1: Obtain the wastewater treatment data to be processed, and extract key operating condition feature vectors from the wastewater treatment data;
[0013] S2: Input the key operating condition feature vectors into the trained wastewater biological treatment model and output the corresponding predicted key effluent water quality data.
[0014] The wastewater biological treatment model is a neural network structure established based on the wastewater biological treatment process, and its treatment steps include:
[0015] S201: Receives and transmits key operating condition feature vectors through the input layer;
[0016] S202: Anaerobic features are generated by extracting anaerobic features based on the key working condition feature vectors passed from the input layer through the anaerobic feature extraction layer.
[0017] S203: Hypoxia features are obtained by extracting hypoxia features based on the anaerobic features generated by the anaerobic feature extraction layer through the hypoxia feature extraction layer;
[0018] S204: Aerobic features are extracted by the aerobic feature extraction layer based on the hypoxia features generated by the hypoxia feature extraction layer to obtain aerobic features;
[0019] S205: The aerobic features generated by the aerobic feature extraction layer are mapped through the output layer to generate and output the corresponding predicted key effluent water quality data.
[0020] in:
[0021] Sparse masks are used to connect the anaerobic feature extraction layer and the hypoxic feature extraction layer, as well as the hypoxic feature extraction layer and the aerobic feature extraction layer.
[0022] The neurons in the anaerobic feature extraction layer, the hypoxic feature extraction layer, and the aerobic feature extraction layer all use the adaptive Monod activation function as the activation function.
[0023] Preferably, in step S1, the key operating condition feature vectors extracted from the wastewater treatment data include any one or more feature vectors among chemical oxygen demand, ammonia nitrogen, total phosphorus, total suspended solids, dissolved oxygen 1, dissolved oxygen 2, and dissolved oxygen 3.
[0024] Preferably, in step S201, the input layer includes neurons that correspond one-to-one with each feature vector in the key operating condition feature vector.
[0025] Preferably, in step S202, the number of neurons in the anaerobic feature extraction layer and the connections between neurons are based on the A of the wastewater treatment plant. 2 The anaerobic stage of the O process needs to be designed.
[0026] Preferably, in step S203, the number of neurons and the connections between neurons in the hypoxia feature extraction layer are based on the A of the wastewater treatment plant. 2 The process of the oxygen-deficient stage in the O process flow needs to be designed.
[0027] Preferably, in step S204, the number of neurons in the aerobic feature extraction layer and the connections between neurons are based on the A of the wastewater treatment plant. 2 The aerobic stage of the O process needs to be designed.
[0028] Preferably, in step S205, the predicted key effluent water quality data includes any one or more of the following: effluent chemical oxygen demand, effluent ammonia nitrogen, effluent total phosphorus, and effluent total nitrogen.
[0029] Preferably, in step S205, the output layer includes neurons that correspond one-to-one with each effluent data point in the predicted key effluent water quality data.
[0030] Preferably, in step S2, the sparse connection method of the sparse mask is based on the A of the wastewater treatment plant. 2 The process flow is designed.
[0031] Preferably, in step S2, the adaptive Monod activation function is expressed as follows:
[0032] ;
[0033] In the formula: The output of the neuron; The linearly weighted input to the neuron; These are parameters that can be learned through backpropagation, corresponding to the maximum specific growth rate in biodynamics; The parameters, which can be learned via backpropagation, can be different in each neuron, thereby enabling the adaptation of the Monod activation function. This corresponds to the half-saturation constant in biodynamics; These are positive constants, used to prevent the denominator from being zero; is a parameter that can be learned via the backpropagation algorithm, representing the linear slope when the input is non-positive.
[0034] Compared with existing technologies, the biochemical reaction activation function-network structure simulation method for wastewater biological treatment in this invention has the following advantages:
[0035] The wastewater biological treatment model (PM-MSCNN) designed in this invention is a sparse neural network structure based on physical constraints. Firstly, the anaerobic feature extraction layer, anoxic feature extraction layer, and aerobic feature extraction layer correspond to the functional modules of the anaerobic, anoxic, and aerobic stages of wastewater treatment. This biomimetic structural design ensures that the feature extraction process is highly consistent with the actual degradation path of pollutants, enabling more accurate capture of water quality change characteristics in different process stages. This solves the underfitting problem caused by the mismatch between structure and mechanism in general MLP models, thereby improving the accuracy of wastewater biological treatment process simulation (i.e., prediction of key effluent water quality data). Secondly, the wastewater biological treatment model structure constructed in this invention abandons the fully connected redundant mode of traditional MLPs and adopts a physically-constrained masking technique based on process mechanisms, namely sparse masking. By setting a mask in the weight matrix, non-physical connections that do not exist in the biochemical reaction are cut off, significantly compressing the parameter search space. This sparse design allows the optimizer to quickly lock the direction of the optimal solution without wasting time correcting the gradients of irrelevant weights in the early stages of training, thus improving the efficiency of wastewater biological treatment process simulation. Furthermore, addressing the pain point of high noise in wastewater treatment plant sensor data, the sparse connection characteristics of the physical model play a powerful regularization role. Fully connected MLPs, due to excessive parameters, are prone to overfitting by "rote memorizing" random noise. In contrast, the physically constrained structure of this invention forces the model to learn only trends that conform to biochemical laws, thereby maintaining extremely high output stability when facing noisy data. Finally, the network topology of PM-MSCNN designed in this invention is no longer a "black box". The anaerobic feature extraction layer, anoxic feature extraction layer and aerobic feature extraction layer of the network are physically mapped to the anaerobic tank, anoxic tank and aerobic tank of the sewage treatment plant, respectively. When the weight of a certain layer of the model is abnormal, it can be directly located to the corresponding actual process unit, realizing a one-to-one correspondence between the model structure and the physical plant, thereby improving the reliability of the simulation of the biological sewage treatment process.
[0036] The adaptive Monod activation function (biochemical activation function) designed in this invention naturally possesses a "Monod equation" form that conforms to the growth patterns of microorganisms. Compared to the ReLU function, which requires the linear superposition of multiple layers of neurons to approximate a nonlinear saturation curve, this activation function only needs to fine-tune the dynamic parameters in the formula to quickly fit data features, significantly reducing the number of training epochs required for the model to reach stability and improving the model convergence speed. Secondly, addressing the problem that the linear features of traditional ReLU cannot accurately describe the "growth saturation" effect of microorganisms under high substrate concentrations, the parameters in the adaptive Monod activation function designed in this invention... The maximum specific growth rate, corresponding to the biodynamic parameters, reflects the maximum processing activity of the microbial community in the current process stage. This parameter breaks the traditional constraint of limiting the Sigmoid / Tanh output to 1, allowing the network to adaptively learn the rate upper limit of different biochemical reactions. The half-saturation constant, corresponding to the biodynamic concept of microorganisms, reflects the affinity of the current functional unit microorganism for pollutants. This is achieved by introducing learnable biochemical parameters. and This endows neurons with the ability to simulate real-world response dynamics. Finally, the adaptive Monod activation function of this invention introduces a maximum growth rate parameter. As a numerical upper bound constraint, this saturation characteristic can automatically suppress the impact of outliers in the input data, avoiding the risk of prediction divergence under extreme conditions caused by the unbounded nature of traditional ReLU. Attached Figure Description
[0037] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:
[0038] Figure 1 This is a network structure diagram of a wastewater biological treatment model, namely a physically constrained sparse neural network structure (PM-MSCNN). Detailed Implementation
[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0040] The following detailed explanation illustrates the specific implementation methods:
[0041] Example:
[0042] This embodiment discloses a biochemical reaction activation function-network structure simulation method for biological wastewater treatment.
[0043] Biochemical reaction activation function-network structure simulation method for wastewater biological treatment, including:
[0044] S1: Obtain the wastewater treatment data to be processed, and extract key operating condition feature vectors from the wastewater treatment data;
[0045] S2: Input the key operating condition feature vectors into the trained wastewater biological treatment model and output the corresponding predicted key effluent water quality data.
[0046] like Figure 1 As shown, the wastewater biological treatment model is a neural network structure (PM-MSCNN) established based on the wastewater biological treatment process, and its treatment steps include:
[0047] The wastewater biological treatment model, i.e., the processing steps of a physically constrained sparse neural network structure, includes:
[0048] S201: Receives and transmits key operating condition feature vectors through the input layer;
[0049] S202: Anaerobic features are generated by extracting anaerobic features based on the key working condition feature vectors passed from the input layer through the anaerobic feature extraction layer.
[0050] S203: Hypoxia features are obtained by extracting hypoxia features based on the anaerobic features generated by the anaerobic feature extraction layer through the hypoxia feature extraction layer;
[0051] S204: Aerobic features are extracted by the aerobic feature extraction layer based on the hypoxia features generated by the hypoxia feature extraction layer to obtain aerobic features;
[0052] S205: The aerobic features generated by the aerobic feature extraction layer are mapped through the output layer to generate and output the corresponding predicted key effluent water quality data.
[0053] in:
[0054] The anaerobic feature extraction layer and the hypoxic feature extraction layer, as well as the hypoxic feature extraction layer and the aerobic feature extraction layer, are connected using a sparse mask.
[0055] The neurons in the anaerobic feature extraction layer, the hypoxic feature extraction layer, and the aerobic feature extraction layer all use the adaptive Monod activation function as the activation function.
[0056] In this embodiment, when training the wastewater biological treatment model (PM-MSCNN), pretreated water quality data is input into the model for network training and parameter optimization, and the backpropagation algorithm is used for training. During the process of minimizing the prediction error (MSE), the optimizer not only updates the network weights W, but also automatically adjusts the biochemical parameters in the activation function based on the data characteristics. and This enables the "identification" of dynamic parameters from data. Among these, The maximum specific growth rate corresponds to the maximum processing activity of the microbial community in the current process section. This parameter breaks the constraint of the traditional Sigmoid / Tanh output being limited to 1, allowing the network to adaptively learn the rate upper limit of different biochemical reactions. The half-saturation constant, corresponding to the biodynamic concept of microorganisms, reflects the affinity of the current functional unit microorganism for pollutants.
[0057] To better illustrate the technical solution of the present invention, this embodiment will be described in more detail through the following parts.
[0058] I. Input and Output Data of the Model
[0059] In the specific implementation process, the key operating condition feature vectors extracted from wastewater treatment data include Chemical Oxygen Demand (COD), Ammonia Nitrogen (NH4), Total Phosphorus (TP), Total Suspended Solids (TSS), Dissolved Oxygen 1 (DO1), Dissolved Oxygen 2 (DO2), and Dissolved Oxygen 3 (DO3).
[0060] The key effluent water quality data for prediction include any one or more of the following: effluent chemical oxygen demand (out_COD), effluent ammonia nitrogen (out_NH), effluent total phosphorus (out_TP), and effluent total nitrogen (out_TN).
[0061] II. Input Layer
[0062] In the specific implementation process, the input layer includes neurons that correspond one-to-one with each feature vector in the key operating condition feature vector.
[0063] In this embodiment, the key operating condition feature vector includes all seven feature vectors, so the input layer includes seven neurons that correspond one-to-one with the seven feature vectors.
[0064] III. Anaerobic Feature Extraction Layer (ANAEROBIC STAGE)
[0065] In the specific implementation process, the number of neurons in the anaerobic feature extraction layer and the connection relationships between neurons are based on the typical process of a wastewater treatment plant, namely A. 2 The process of the anaerobic stage in the O process (anaerobic-anoxic-aerobic) needs to be designed.
[0066] Among them, A 2 The O (Anaerobic-Anoxic-Oxic) process is a biological wastewater treatment technology for nitrogen and phosphorus removal. 2 The O process consists of three functional zones connected in series: 1) An anaerobic zone, where polyphosphate-accumulating bacteria release phosphorus and absorb readily degradable organic matter in the wastewater, providing energy reserves for subsequent phosphorus uptake; 2) An anoxic zone, where denitrifying bacteria use carbon sources in the influent to reduce nitrates from the aerobic zone return liquid to nitrogen gas, achieving denitrification; 3) An aerobic zone, where nitrifying bacteria oxidize ammonia nitrogen to nitrate, while polyphosphate-accumulating bacteria excessively absorb phosphorus under oxygen-rich conditions, and phosphorus removal is achieved through the discharge of excess sludge. The entire system typically includes internal recirculation (returning the mixed liquor from the aerobic zone to the anoxic zone) and external recirculation (returning sludge from the secondary sedimentation tank to the anaerobic zone) to maintain microbial concentration and denitrification and phosphorus removal efficiency.
[0067] In this embodiment, based on wastewater treatment plant A 2 The logic for designing the number of neurons and connection relationships in the anaerobic feature extraction layer of the O process is as follows:
[0068] 1) Determine the basic number of input neurons based on the core process parameters of the anaerobic stage (including anaerobic tank retention time, sludge return ratio, dissolved oxygen concentration, organic matter concentration, polyphosphate accumulation activity, etc.). The basic number of neurons can be set to 1.5-2 times the number of process parameters as needed to fully capture the complex biochemical processes of the anaerobic stage.
[0069] 2) The neuron connections are physically constrained based on the biochemical reaction mechanism of the anaerobic stage using a "sparse mask": For the synergistic relationship between anaerobic tank retention time and sludge return ratio, the mask preserves connection paths to allow the neural network to learn strongly correlated features; for the negative correlation between dissolved oxygen concentration and organic matter concentration, the mask preserves connections to allow the network to learn negative inhibition weights; for the positive correlation between phosphorus release rate by polyphosphate-accumulating bacteria and organic matter concentration, the mask preserves connection paths. The specific values of the connection weights are automatically optimized during training using a backpropagation algorithm. Furthermore, this layer of neurons employs an adaptive Monod activation function to simulate the saturation uptake characteristics of organic substrates by microorganisms under anaerobic conditions.
[0070] This invention ensures that the anaerobic feature extraction layer accurately reflects the biodynamic characteristics of the anaerobic stage, which is consistent with the biological reaction patterns in the wastewater biological treatment process, making the model output more consistent with the actual wastewater treatment process.
[0071] IV. Anaerobic Feature Extraction Layer (ANOXIC STAGE)
[0072] In the specific implementation process, the number of neurons in the hypoxia feature extraction layer and the connection relationships between neurons are based on the typical process of a wastewater treatment plant, namely A. 2 The process design for the anoxic stage in the O process (anaerobic-anoxic-aerobic) flow needs to be carried out.
[0073] In this embodiment, based on wastewater treatment plant A 2 The logic for designing the number of neurons and connection relationships in the hypoxia feature extraction layer during the hypoxia stage of the O process is as follows:
[0074] 1) Determine the basic number of input neurons based on key process parameters in the anoxic stage (such as dissolved oxygen concentration, denitrification rate, carbon source dosage, nitrate concentration, denitrification capacity of polyphosphate-accumulating bacteria, etc.), usually set to 1.2-1.5 times the number of parameters, in order to cover the nonlinear characteristics of denitrification.
[0075] 2) The connections between neurons are designed using a "sparse mask" based on the biochemical reaction mechanism during the hypoxic phase: For the negative correlation between dissolved oxygen concentration and denitrification rate (hypoxic environments require extremely low DO), the mask preserves connections to capture the inhibitory effect; for the positive correlation between carbon source concentration and denitrification efficiency, the mask preserves connections to capture the promoting effect; and for the feedback relationship between nitrate concentration and denitrification rate, the mask preserves connection paths. Connection weights are dynamically adjusted by the model using training data. Simultaneously, neurons in this layer utilize an adaptive Monod activation function to simulate the nonlinear metabolic process of denitrifying bacteria under substrate constraints.
[0076] This invention ensures that the anoxic feature extraction layer accurately captures the denitrification coupling process during the anoxic stage, matching the inhibition characteristics in the biological wastewater treatment process, thereby improving the model's prediction accuracy for carbon source allocation and denitrification efficiency.
[0077] V. Aerobic Feature Extraction Layer (AEROBIC STAGE)
[0078] In the specific implementation process, the number of neurons in the aerobic feature extraction layer and the connection relationship between neurons are designed based on the process requirements of the aerobic stage in the typical process of sewage treatment plant, namely the A2O process (anaerobic-anoxic-aerobic).
[0079] In this embodiment, the logic for designing the number of neurons and connection relationships in the aerobic feature extraction layer based on the process requirements of the aerobic stage in the A2O process of a wastewater treatment plant is as follows:
[0080] 1) Determine the basic number of input neurons based on the core parameters of the aerobic stage (such as dissolved oxygen concentration, nitrification rate, ammonia nitrogen oxidation rate, sludge concentration, heterotrophic bacteria activity, etc.), usually set to 1.3-1.8 times the number of parameters, in order to fully characterize the nitrification-denitrification synergistic process.
[0081] 2) The connections between neurons are designed using a "sparse mask" based on the biochemical reaction patterns of the aerobic stage: For the positive correlation between dissolved oxygen concentration and nitrification rate (DO ≥ 2 mg / L), the mask preserves connection paths; for the Monod kinetic relationship between ammonia nitrogen concentration and nitrification rate, the mask preserves connections and performs nonlinear fitting using an adaptive Monod activation function; for the positive correlation between sludge concentration and nitrification efficiency, the mask preserves connections. Connection weights are adjusted online based on actual operating data from the aerobic tank.
[0082] This invention enables the aerobic feature extraction layer to accurately simulate the nitrification kinetics of the aerobic stage, which is consistent with the saturation characteristics in the biological wastewater treatment process, thereby improving the model's ability to fit ammonia nitrogen removal and nitrate nitrogen formation.
[0083] VI. Output Layer
[0084] In the specific implementation process, the output layer includes neurons that correspond one-to-one with each effluent data point in the prediction of key effluent water quality data.
[0085] In this embodiment, the key effluent water quality data for prediction includes all four effluent data points, therefore the output layer includes four neurons that correspond one-to-one with the four effluent data points.
[0086] 7. Sparsity Mask
[0087] In practice, the sparse connection method of the sparse mask is based on the typical process of a wastewater treatment plant, namely A. 2 The O process (anaerobic-anoxic-aerobic) flow was designed.
[0088] In this embodiment, the sparse mask connection method needs to be based on A. 2 Design of the material transfer and energy flow patterns throughout the O process:
[0089] 1) According to A 2 The coupling relationship between each stage (anaerobic-anoxic-aerobic) in the O process is analyzed, and the correlation matrix between process nodes is constructed: the release of organic matter and the metabolism of polyphosphate-accumulating bacteria in the anaerobic stage directly determine the substrate composition for entering the anoxic stage, while the denitrification effect in the anoxic stage directly determines the remaining carbon source and ammonia nitrogen load for entering the aerobic stage.
[0090] 2) Key connectivity paths are extracted using a process knowledge graph. Unnecessary connections are assigned zero weight, retaining only those with significant process relevance. For example, the "dissolved oxygen-polyphosphate activity" connection in the anaerobic stage only has a weak correlation with the "carbon source-denitrification efficiency" connection in the anoxic stage (due to differences in biochemical mechanisms); while the "denitrification-substrate consumption" state node output in the anoxic stage needs to have a strong correlation with the "nitrification rate-dissolved oxygen" node in the aerobic stage. Finally, a sparse mask is generated using a process mechanism-driven pruning strategy, reducing model complexity (pruning approximately 90% of redundant connections, retaining only effective connections conforming to biochemical fluid dynamics) while preserving A... 2 The core coupling logic of the O process improves the model's computational efficiency while maintaining prediction accuracy superior to that of fully connected models.
[0091] 8. Adaptive Monod Activation Function
[0092] In practical implementation, the formula for the adaptive Monod activation function is expressed as follows:
[0093] ;
[0094] In the formula: The output of the neuron; The linearly weighted input to the neuron represents the substrate concentration; The parameter is a parameter that can be learned through backpropagation. It corresponds to the maximum specific growth rate in biodynamics and reflects the maximum treatment activity of the microbial community in the current process section. This parameter breaks the constraint of the traditional Sigmoid / Tanh output being limited to 1, allowing the network to adaptively learn the rate upper limit of different biochemical reactions. The parameters, which can be learned via backpropagation, can be different in each neuron, thereby enabling the adaptation of the Monod activation function. The half-saturation constant in biodynamics can explain the corresponding biochemical reaction characteristics to a certain extent and reflect the affinity of the current functional unit microorganism for pollutants. Its absolute value is taken in the formula to ensure that the physical meaning is positive. It is a very small positive number used to prevent the denominator from being zero and to ensure the stability of numerical calculations; is a parameter that can be learned via the backpropagation algorithm, representing the linear slope when the input is non-positive.
[0095] The adaptive Monod activation function used in this invention enables the neuron's output to... With input The increase shows a "rapid at first, then slow" growth pattern, eventually approaching a saturation point. Instead of growing indefinitely, this is entirely consistent with the monod equation for microbial growth kinetics. The maximum specific growth rate corresponds to the maximum processing activity of the microbial community in the current process section. This parameter breaks the constraint of the traditional Sigmoid / Tanh output being limited to 1, allowing the network to adaptively learn the rate upper limit of different biochemical reactions. The half-saturation constant, corresponding to biodynamics, can explain the corresponding biochemical reaction characteristics to a certain extent, reflecting the affinity of the current functional unit microorganisms for pollutants.
[0096] The wastewater biological treatment model (PM-MSCNN) designed in this invention is a sparse neural network structure based on physical constraints. Firstly, the anaerobic feature extraction layer, anoxic feature extraction layer, and aerobic feature extraction layer correspond to the functional modules of the anaerobic, anoxic, and aerobic stages of wastewater treatment. This biomimetic structural design ensures that the feature extraction process is highly consistent with the actual degradation path of pollutants, enabling more accurate capture of water quality change characteristics in different process stages. This solves the underfitting problem caused by the mismatch between structure and mechanism in general MLP models, thereby improving the accuracy of wastewater biological treatment process simulation (i.e., prediction of key effluent water quality data). Secondly, the wastewater biological treatment model structure constructed in this invention abandons the fully connected redundant mode of traditional MLPs and adopts a physically-constrained masking technique based on process mechanisms, namely sparse masking. By setting a mask in the weight matrix, non-physical connections that do not exist in the biochemical reaction are cut off, significantly compressing the parameter search space. This sparse design allows the optimizer to quickly lock the direction of the optimal solution without wasting time correcting the gradients of irrelevant weights in the early stages of training, thus improving the efficiency of wastewater biological treatment process simulation. Furthermore, addressing the pain point of high noise in wastewater treatment plant sensor data, the sparse connection characteristics of the physical model play a powerful regularization role. Fully connected MLPs, due to excessive parameters, are prone to overfitting by "rote memorizing" random noise. In contrast, the physically constrained structure of this invention forces the model to learn only trends that conform to biochemical laws, thereby maintaining extremely high output stability when facing noisy data. Finally, the network topology of PM-MSCNN designed in this invention is no longer a "black box". The anaerobic feature extraction layer, anoxic feature extraction layer and aerobic feature extraction layer of the network are physically mapped to the anaerobic tank, anoxic tank and aerobic tank of the sewage treatment plant, respectively. When the weight of a certain layer of the model is abnormal, it can be directly located to the corresponding actual process unit, realizing a one-to-one correspondence between the model structure and the physical plant, thereby improving the reliability of the simulation of the biological sewage treatment process.
[0097] The adaptive Monod activation function (biochemical activation function) designed in this invention naturally possesses a "Monod equation" form that conforms to the growth patterns of microorganisms. Compared to the ReLU function, which requires the linear superposition of multiple layers of neurons to approximate a nonlinear saturation curve, this activation function only needs to fine-tune the dynamic parameters in the formula to quickly fit data features, significantly reducing the number of training epochs required for the model to reach stability and improving the model convergence speed. Secondly, addressing the problem that the linear features of traditional ReLU cannot accurately describe the "growth saturation" effect of microorganisms under high substrate concentrations, the parameters in the adaptive Monod activation function designed in this invention... The maximum specific growth rate, corresponding to the biodynamic parameters, reflects the maximum processing activity of the microbial community in the current process stage. This parameter breaks the traditional constraint of limiting the Sigmoid / Tanh output to 1, allowing the network to adaptively learn the rate upper limit of different biochemical reactions. The half-saturation constant, corresponding to the biodynamic concept of microorganisms, reflects the affinity of the current functional unit microorganism for pollutants. This is achieved by introducing learnable biochemical parameters. and This endows neurons with the ability to simulate real-world response dynamics. Finally, the adaptive Monod activation function of this invention introduces a maximum growth rate parameter. As a numerical upper bound constraint, this saturation characteristic can automatically suppress the impact of outliers in the input data, avoiding the risk of prediction divergence under extreme conditions caused by the unbounded nature of traditional ReLU.
[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.
Claims
1. A biochemical reaction activation function-network structure simulation method for biological wastewater treatment, characterized in that, include: S1: Obtain the wastewater treatment data to be processed, and extract key operating condition feature vectors from the wastewater treatment data; S2: Input the key operating condition feature vectors into the trained wastewater biological treatment model and output the corresponding predicted key effluent water quality data. The wastewater biological treatment model is a neural network structure established based on the wastewater biological treatment process, and its treatment steps include: S201: Receives and transmits key operating condition feature vectors through the input layer; S202: Anaerobic features are generated by extracting anaerobic features based on the key working condition feature vectors passed from the input layer through the anaerobic feature extraction layer. S203: Hypoxia features are obtained by extracting hypoxia features based on the anaerobic features generated by the anaerobic feature extraction layer through the hypoxia feature extraction layer; S204: Aerobic features are extracted by the aerobic feature extraction layer based on the hypoxia features generated by the hypoxia feature extraction layer to obtain aerobic features; S205: The aerobic features generated by the aerobic feature extraction layer are mapped through the output layer to generate and output the corresponding predicted key effluent water quality data. in: Sparse masks are used to connect the anaerobic feature extraction layer and the hypoxic feature extraction layer, as well as the hypoxic feature extraction layer and the aerobic feature extraction layer. The neurons in the anaerobic feature extraction layer, the hypoxic feature extraction layer, and the aerobic feature extraction layer all use the adaptive Monod activation function as the activation function.
2. The biochemical reaction activation function-network structure simulation method for wastewater biological treatment as described in claim 1, characterized in that: In step S1, the key operating condition feature vectors extracted from the wastewater treatment data include any one or more feature vectors from chemical oxygen demand, ammonia nitrogen, total phosphorus, total suspended solids, dissolved oxygen 1, dissolved oxygen 2, and dissolved oxygen 3.
3. The biochemical reaction activation function-network structure simulation method for wastewater biological treatment as described in claim 2, characterized in that: In step S201, the input layer includes neurons that correspond one-to-one with each feature vector in the key working condition feature vector.
4. The biochemical reaction activation function-network structure simulation method for wastewater biological treatment as described in claim 1, characterized in that: In step S202, the number of neurons and the connections between neurons in the anaerobic feature extraction layer are based on the A of the wastewater treatment plant. 2 The anaerobic stage of the O process needs to be designed.
5. The biochemical reaction activation function-network structure simulation method for wastewater biological treatment as described in claim 1, characterized in that: In step S203, the number of neurons and the connections between neurons in the hypoxia feature extraction layer are based on the A of the wastewater treatment plant. 2 The process of the oxygen-deficient stage in the O process flow needs to be designed.
6. The biochemical reaction activation function-network structure simulation method for biological wastewater treatment as described in claim 1, characterized in that: In step S204, the number of neurons in the aerobic feature extraction layer and the connections between neurons are based on the A of the wastewater treatment plant. 2 The aerobic stage of the O process needs to be designed.
7. The biochemical reaction activation function-network structure simulation method for wastewater biological treatment as described in claim 1, characterized in that: In step S205, the predicted key effluent water quality data includes any one or more of the following: effluent chemical oxygen demand, effluent ammonia nitrogen, effluent total phosphorus, and effluent total nitrogen.
8. The biochemical reaction activation function-network structure simulation method for wastewater biological treatment as described in claim 7, characterized in that: In step S205, the output layer includes neurons that correspond one-to-one with each effluent data point in the predicted key effluent water quality data.
9. The biochemical reaction activation function-network structure simulation method for biological wastewater treatment as described in claim 1, characterized in that: In step S2, the sparse connection method of the sparse mask is based on the A of the sewage treatment plant. 2 The process flow is designed.
10. The biochemical reaction activation function-network structure simulation method for wastewater biological treatment as described in claim 9, characterized in that: In step S2, the formula for the adaptive Monod activation function is expressed as follows: ； In the formula: The output of the neuron; The linearly weighted input to the neuron; These are parameters that can be learned through backpropagation, corresponding to the maximum specific growth rate in biodynamics; The parameters, which can be learned via backpropagation, can be different in each neuron, thereby enabling the adaptation of the Monod activation function. This corresponds to the half-saturation constant in biodynamics; These are positive constants, used to prevent the denominator from being zero; is a parameter that can be learned via the backpropagation algorithm, representing the linear slope when the input is non-positive.