Aeration control method and system fusing oxygen mass transfer process and deep learning model

Abstract

The invention relates to an aeration control method and system fusing an oxygen mass transfer process and a deep learning model, and the method comprises the steps: collecting historical operation data of a sewage treatment system, the historical operation data comprising an inlet water quality parameter, an outlet water quality parameter and a dissolved oxygen concentration; generating a first training sample set based on the historical operation data; inputting the first training sample set into a pre-constructed deep learning model for training; inputting a first to-be-tested sample into the trained deep learning model for prediction to obtain a target dissolved oxygen concentration; and reversely deducing a target oxygen mass transfer coefficient based on the target dissolved oxygen concentration, and predicting a target fan control quantity by using a multi-layer perceptron model. The reverse solving process of the fan control quantity is realized; the aeration intensity can be intelligently adjusted according to the current working condition, and the energy-saving potential is higher.

Description

Technical Field

[0001] This invention relates to the fields of electronic digital data processing and energy-saving control, specifically to an aeration control method and system that integrates oxygen mass transfer processes and deep learning models. Background Technology

[0002] Biological treatment methods, represented by the A2O process, have been widely used in urban wastewater treatment systems due to their strong nitrogen and phosphorus removal capabilities, compact structure, and mature operation. The efficient operation of this process places high demands on aeration control, especially in optimizing aeration energy consumption while ensuring stable effluent quality. This has become one of the key issues in current research on intelligent wastewater treatment.

[0003] Traditional aeration control methods typically employ timed start-stop or fixed-frequency operation of the blower, supplemented by dissolved oxygen concentration (... DO Feedback-based simple closed-loop regulation is used. While this method is simple in structure and easy to implement, it struggles to adapt to dynamic fluctuations in influent water quality and load. In actual operating conditions, excessively low dissolved oxygen concentrations can lead to incomplete nitrification, affecting effluent quality; conversely, excessively high dissolved oxygen concentrations can cause over-aeration, increasing blower energy consumption and inhibiting denitrification, thus impacting nitrogen removal efficiency. Therefore, achieving precise and energy-efficient aeration control under dynamic operating conditions has become an important research direction in wastewater treatment control technology.

[0004] In recent years, deep learning, as a powerful data-driven modeling tool, has been introduced into the field of wastewater treatment for constructing complex nonlinear relationship models, predicting water quality change trends, and assisting in decision-making and control. Some studies have already attempted to use recurrent neural networks (RNNs). RNN Long Short-Term Memory Network (LSTM) LSTM )or Transformer The structure predicts changes in dissolved oxygen (DO) and uses this as a reference for fan control commands. For example, in a pond dissolved oxygen concentration prediction method based on the TCN model disclosed in patent publication number CN115166176B, it is used... TCN The model makes predictions.

[0005] In addition, oxygen mass transfer theory ( OTR This provides a clear physical basis for calculating changes in dissolved oxygen, and its core formula... It is pointed out that the rate of change depends on the mass transfer coefficient. saturated dissolved oxygen Compared with the current DO concentration The difference. However, in practical applications... It is not a constant, but a nonlinear variable affected by factors such as fan opening, water temperature, water quality, and pool type. It is difficult to measure directly, thus limiting the independent application of the OTR model. Summary of the Invention

[0006] In order to overcome the above-mentioned defects and deficiencies in the prior art, the present invention provides an aeration control method and system that integrates oxygen mass transfer process and deep learning model.

[0007] To address the aforementioned technical problems, this invention provides an aeration control method that integrates oxygen mass transfer processes and deep learning models. The method includes the following steps: Step S1: Collect historical operating data of the wastewater treatment system, wherein the historical operating data includes: influent water quality parameters, effluent water quality parameters and dissolved oxygen concentration; Step S2: Generate a first training sample set based on the historical running data; Step S3: Input the first training sample set into the pre-built deep learning model for training; Step S4: Input the first sample to be tested into the trained deep learning model for prediction to obtain the target dissolved oxygen concentration; Step S5: Based on the target dissolved oxygen concentration, deduce the target oxygen mass transfer coefficient, and use a multilayer perceptron model to predict the target fan control quantity.

[0008] Furthermore, the process of generating the first training sample set is as follows: Valid samples that meet the preset emission standards are selected from the historical operating data; Based on the influent water quality parameters, samples that meet the preset discharge standards are discretized by interval to form multiple subsets representing different operating conditions; The record with the lowest dissolved oxygen level is selected from the multiple subsets representing different working conditions obtained after grouping, and used as the minimum acceptable dissolved oxygen concentration under each working condition to form a static sample set. For each sample record in the static sample set, the multi-dimensional water inflow parameters within a fixed time window are traced back in chronological order to construct a two-dimensional input matrix with a time series structure, forming a first training sample set with temporal characteristics.

[0009] Furthermore, in step S2, the generation of the first training sample set also includes: The physical consistency of the first training sample set is verified based on the oxygen mass transfer equation, and data that does not meet the preset verification conditions are removed or downweighted. The physical consistency verification process is as follows: The dissolved oxygen change rate dC / dt for each sequence in the first training sample was measured and combined with the oxygen mass transfer equation: , in, This represents the current dissolved oxygen concentration. This represents the saturated dissolved oxygen concentration under the current operating conditions. The oxygen mass transfer coefficient; Compare the measured dissolved oxygen change rate dC / dt with the calculated value on the right side of the equation; When the oxygen mass transfer coefficient obtained by reverse calculation according to the formula exceeds the preset empirical range, or the difference exceeds the preset threshold, the sequence is removed or downweighted.

[0010] Furthermore, in step S2, the generation of the first training sample set also includes: The data from multiple subsets representing different working conditions obtained after grouping are sorted from low to high dissolved oxygen concentration, and the samples in the range of 0 to q quantile are taken as the minimum acceptable dissolved oxygen concentration for each working condition, forming a static sample set.

[0011] Furthermore, in step S3, the deep learning model construction process is as follows: Construct an input mapping layer to map the water quality feature vectors of the original input sequence at each time step. Mapped to a unified dimension of representation space The linear transformation method used in the mapping function is as follows: ; in, ; It refers to a dimension of d x The vector space represents the dimension of the input features and includes several influent water quality parameters. It refers to a dimension of d m The vector space W represents the unified feature representation space after mapping. e This refers to the linear mapping weights, which transform the original feature x... t Mapped to a unified-dimensional representation space, b e It refers to the bias vector of the input mapping layer, used to correct the offset of the linear mapping output. It refers to d m line ×d x The matrix space of columns; the result of the linear transformation constitutes a matrix of length. sequence As the encoder input, E is the input feature matrix composed of time series samples, T is the number of time steps, and e t This represents the feature representation at time step t. Build Transformer The decoder module consists of several stacked encoder layers, each of which contains two core modules: a multi-head attention mechanism and a feedforward neural network. Multi-head self-attention mechanisms are used to uncover correlations between time steps within a time series. The core calculation formula is as follows: ; in, , represents the linear mapping result of query, key, and value; Q, K, and V are the query, key, and value matrices in the attention mechanism, respectively, obtained from the input feature matrix E through linear transformation; W Q W K W V d represents the linear projection weight matrices corresponding to the query, key, and value, respectively, used to map the input E to the attention subspace; k For the feature dimension in the attention mechanism; Multi-head attention mechanisms model attention by using multiple independent attention heads in parallel: ; in, , To output the projection matrix; The feedforward neural network consists of two fully connected layers with a non-linear activation function in between. dropout layer; Integrating residual connections between the multi-head attention mechanism and the feedforward neural network modules LayerNorm operate: ; To form a stable deep network structure, in which SubLayer X represents the attention or feedforward module; X represents the tensor of the sublayer input or the previous layer output, used for residual connections and layer normalization operations. Transformer The encoder output is a context representation sequence for each time step. Finally, the output vector of the last time step is selected. The linear regression layer maps the data to scalar prediction results. ; in, , For bias terms; The target dissolved oxygen concentration is to be predicted; The model uses mean squared error as the loss function: ; in, The true label value of the sample. These are the model's predicted values. The total number of samples.

[0012] Furthermore, in step S5, the reverse calculation process for the target oxygen mass transfer coefficient is as follows: Based on the target dissolved oxygen concentration Current dissolved oxygen concentration Sampling time interval The target dissolved oxygen change rate was calculated. : ; Then, combining this with the oxygen mass transfer rate equation: ; in, Given the saturated dissolved oxygen concentration under the current operating conditions, the target oxygen mass transfer coefficient is obtained by inverse solution. : .

[0013] Furthermore, in step S5, the prediction process for the target wind turbine control quantity is as follows: The historical oxygen mass transfer coefficient at the corresponding moment is calculated based on the dissolved oxygen concentration in the historical operation data, and a second training sample set is constructed. Using the water quality parameters at the corresponding time in the second training sample set and the target oxygen mass transfer coefficient obtained by inverse solution As input, the actual wind turbine opening control value recorded at that moment is used as the corresponding supervision label to train the multilayer perceptron model; The regression function for training a multilayer perceptron model is as follows: ; in, This represents the predicted control value for the turbine opening. This represents the feature vector composed of current water quality parameters. This represents the regression function of the multilayer perceptron obtained through training; The target wind turbine control quantity is obtained by inputting the second test sample into the trained multilayer perceptron model for prediction.

[0014] Furthermore, step S5 also includes: The target fan control quantity is corrected for errors using a PID control module based on error feedback. The error correction process is as follows: With target dissolved oxygen concentration The set value is based on the actual sensor monitoring value. For the feedback value, calculate its deviation: ; The fine-tuning control quantity is calculated based on proportional, integral, and derivative control laws. : ; in, , , These are the proportional, integral, and derivative parameters, which are set empirically based on the system response characteristics or obtained through online trial and error. By substituting the predicted target wind turbine control quantity and the fine-tuning control quantity into the following formula: u final1 =u OTR +u PID； The final fan control output is calculated, where, Indicates passage MLP-OTR The predicted wind turbine operating rate obtained by path inversion This represents the real-time error compensation term. u final1 This is the control opening degree ultimately used to drive the blower.

[0015] Furthermore, step S5 also includes: The extended state observer is used to perform one-dimensional disturbance observation on the process of error correction of the target wind turbine control quantity through the error feedback-based PID control module, and the disturbance compensation quantity is obtained. By substituting the predicted target wind turbine control quantity, fine-tuning control quantity, and disturbance compensation quantity into the following formula: ; The final fan control output is calculated, where, Indicates passage MLP-OTR The predicted wind turbine operating rate obtained by path inversion This represents the real-time error compensation term. This is the disturbance compensation amount. This is the control opening degree ultimately used to drive the blower.

[0016] This invention also provides an aeration control system that integrates oxygen mass transfer processes and deep learning models, the system comprising: The data acquisition module is used to collect historical operating data of the wastewater treatment system, wherein the historical operating data includes: influent water quality parameters, effluent water quality parameters, and dissolved oxygen concentration; The generation module is used to generate a first training sample set based on the historical running data; The training module is used to input the first training sample set into a pre-built deep learning model for training. The first prediction module is used to input the first test sample into the trained deep learning model for prediction to obtain the target dissolved oxygen concentration. The second prediction module is used to deduce the target oxygen mass transfer coefficient based on the target dissolved oxygen concentration and to predict the target fan control quantity using a multilayer perceptron model.

[0017] The beneficial technical effects achieved by this invention are as follows: This invention discloses an aeration control method and system that integrates oxygen mass transfer processes and deep learning models. It utilizes the target dissolved oxygen concentration predicted by a deep learning model combined with the oxygen mass transfer equation (…). OTR ) Back-calculate the required oxygen mass transfer coefficient And based on multilayer perceptron ( MLP By combining the oxygen mass transfer coefficient with the required fan opening, a reverse solution process for fan control variables was achieved. Compared to the traditional method of setting a fixed dissolved oxygen concentration control point, this method can intelligently adjust the aeration intensity according to the current operating conditions, thus having greater energy-saving potential. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the aeration control method that integrates oxygen mass transfer process and deep learning model provided in this embodiment of the invention. Figure 2 This is a flowchart illustrating the aeration control method that integrates oxygen mass transfer process and deep learning model, provided in a preferred embodiment of the present invention. Figure 3 This is a schematic diagram of the aeration control system that integrates the oxygen mass transfer process and the deep learning model provided in this embodiment of the invention. Detailed Implementation

[0019] The present invention will be further described below with reference to specific embodiments. These embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] like Figure 1 and 2 As shown, this invention provides an aeration control method that integrates oxygen mass transfer processes and deep learning models, comprising the following steps: Step S1: Collect historical operating data of the wastewater treatment system, wherein the historical operating data includes: influent water quality parameters, effluent water quality parameters and dissolved oxygen concentration; Historical operational data can be obtained from the online monitoring system or historical process operation records of actual wastewater treatment plants, serving as the basic data source for subsequent sample selection and model training. Sensors deployed in the wastewater treatment system can collect data such as flow rate, water quality, and dissolved oxygen concentration. The collected data is stored in the online monitoring system or used to generate historical process operation records.

[0022] Here, we primarily use influent water quality parameters, effluent water quality parameters, and dissolved oxygen concentration from historical operational data. Influent water quality parameters include, but are not limited to, total nitrogen in the influent (…). TN_in ), total phosphorus in water ( TP_in Chemical oxygen demand (COD) COD_ in ), influent flow rate, water temperature, etc.; dissolved oxygen concentration is the DO ( ) of the aerobic tank in the biological treatment tank. DO_in ); effluent water quality parameters include, but are not limited to, total nitrogen in effluent ( TN_out ), total phosphorus ( TP_out Chemical oxygen demand (COD) COD_out )wait.

[0023] Step S2: Generate a first training sample set based on the historical running data; Specifically, the process of generating the first training sample set is as follows: Valid samples meeting preset emission standards are selected from the historical operational data; these are used to construct a dataset with a minimum acceptable dissolved oxygen target value. The emission standard can be total nitrogen in the effluent (…). TN_out The total phosphorus in the effluent should not exceed 15 mg / L. TP_out The concentration of 0.5 mg / L is not higher than that of 0.5 mg / L, and the chemical oxygen demand (COD) is not higher than that of 0.5 mg / L. COD_out (Not exceeding the set threshold, etc.)

[0024] Based on the effluent water quality parameters, samples meeting the preset discharge standards are discretized by interval to form multiple subsets representing different operating conditions. The grouping method can be based on the total nitrogen in the influent (…). TN_in ), total phosphorus ( TP_in Chemical oxygen demand (COD) COD_in Parameters such as these are discretized into intervals to form multiple subsets representing different operating conditions; for example, TN Rounding TP Round to one decimal place. COD Divide into 10 mg / L containers.

[0025] After grouping, you can select the current dissolved oxygen () in each subset. DO_in The record with the lowest dissolved oxygen concentration and effluent meeting the standards will be used as the minimum acceptable dissolved oxygen concentration (minimum acceptable) under this set of operating conditions. DO This forms the initial static sample set.

[0026] It is worth noting that after grouping, the data from multiple subsets representing different operating conditions can be sorted from low to high dissolved oxygen concentration, and the samples within the 0th to qth percentile (0 < q ≤ 0.2) can be used as the minimum acceptable dissolved oxygen concentration for that group of operating conditions, forming a static sample set. This can improve the accuracy of the model's output. The data from multiple subsets representing different operating conditions can be sorted from low to high dissolved oxygen concentration to form a static sample set.

[0027] For each sample record in the static sample set, the multi-dimensional water inflow parameters within a fixed time window are traced back in chronological order to construct a two-dimensional input matrix with a time series structure, forming a first training sample set with time series characteristics; Specifically, for each sample record, the multidimensional influent parameters within a fixed time window are traced back in chronological order (e.g., the parameters from the previous 10 minutes). TN_in , TP_in , COD_in , DO_in A two-dimensional input matrix with a time-series structure is constructed using data such as flow rate and water temperature. This matrix serves as the input to the deep learning model, with the target value corresponding to each sample being the minimum acceptable dissolved oxygen concentration for that record. This method transforms static samples into training samples with temporal characteristics, enabling the model to learn the mapping relationship between historical water quality change trends and the minimum acceptable dissolved oxygen concentration.

[0028] In this embodiment, the process of generating the first training sample set further includes: The physical consistency of the first training sample set is verified based on the oxygen mass transfer equation, and data that does not meet the preset verification conditions are removed or downweighted. The physical consistency verification process is as follows: The dissolved oxygen change rate dC / dt for each sequence in the first training sample was measured and combined with the oxygen mass transfer equation: ; Compare the measured dissolved oxygen change rate dC / dt with the calculated value on the right side of the equation; When the oxygen mass transfer coefficient obtained by reverse calculation according to the formula exceeds the preset empirical range, or the difference exceeds the preset threshold, the sequence is removed or downweighted.

[0029] Introducing the oxygen mass transfer equation to verify the physical consistency of the samples can improve the physical rationality and data quality of the training samples, without changing the selection logic of the target dissolved oxygen.

[0030] By introducing the oxygen mass transfer equation ( OTRBased on the physical characteristics of the oxygen mass transfer equation, historical operating data is filtered, grouped, and processed to generate a training sample set, thereby providing data support for model training through this sample set.

[0031] Furthermore, the completed first training sample set needs to be divided into a training set, a validation set, and a test set proportionally. Random sampling can be used for this division, with a common ratio of 7:2:1. The training set is used for fitting and learning the model parameters, the validation set is used for hyperparameter tuning and overfitting control, and the test set is used to evaluate the model's predictive performance under unseen conditions.

[0032] Step S3: Input the first training sample set into the pre-built deep learning model for training; Step S4: Input the first test sample into the trained deep learning model for prediction to obtain the target dissolved oxygen concentration.

[0033] Specifically, after forming the first training sample set, this sample set is input into a pre-built deep learning model to learn the mapping relationship between historical water quality change trends and minimum acceptable dissolved oxygen concentration in the samples. After training is completed, the first test sample (i.e., the water quality parameters measured by sensors in the scenario to be predicted) is input into the trained model to obtain the target dissolved oxygen concentration.

[0034] In this embodiment, the deep learning model is constructed as follows: Construct an input mapping layer to map the water quality feature vectors of the original input sequence at each time step. Mapped to a unified dimension of representation space The linear transformation method used in the mapping function is as follows: ; in, The transformation result constitutes a length of sequence , as encoder input; Build Transformer The decoder module consists of several stacked encoder layers, each of which contains two core modules: a multi-head attention mechanism and a feedforward neural network. Multi-head self-attention mechanisms are used to uncover correlations between time steps within a time series. The core calculation formula is as follows: ; in, This represents a linear mapping result of the query, key, and value. For attention dimension; Multi-head attention mechanisms model attention by using multiple independent attention heads in parallel: ; in, , To output the projection matrix; The feedforward neural network consists of two fully connected layers with a non-linear activation function in between. dropout layer; Integrating residual connections between the multi-head attention mechanism and the feedforward neural network modules LayerNorm operate: ; To form a stable deep network structure, in which SubLayer Indicates an attention or feedforward module; Transformer The encoder output is a context representation sequence for each time step. Finally, the output vector of the last time step is selected. The linear regression layer maps the data to scalar prediction results. ; in, , For bias terms; The target dissolved oxygen concentration is to be predicted; The model uses mean squared error as the loss function: ; in, The true label value of the sample. These are the model's predicted values. The total number of samples.

[0035] It is important to understand that this model uses Adam The optimizer updates its parameters, with the initial learning rate set to 0.001. The number of training epochs can be set from 20 to 100 depending on the error convergence. Mini-batch stochastic gradient descent is used during training. mini-batch SGD The batch size was set to 32, and the order of the training samples was randomly shuffled before each round of training to improve the generalization ability of the model.

[0036] It is also important to understand that, in order to further improve training results, all input features are processed before model training. Z- score Standardization ensures that features across all dimensions have a uniform scale and distribution, which helps the model converge quickly and avoids feature-driven bias.

[0037] The finally trained model can be GPU or CPUThe system is deployed in the environment and supports persistent storage of model parameters for real-time loading and retrieval in the control system, enabling dynamic prediction of the minimum acceptable dissolved oxygen concentration under current water quality conditions.

[0038] By introducing the oxygen mass transfer equation ( OTR Based on the physical characteristics of the oxygen mass transfer equation, historical operating data is filtered and grouped to generate a training sample set, which is then used to train the model. In this way, the physical laws of dissolved oxygen concentration change can constrain the deep learning model, overcome the lack of physical consistency in pure "black box" prediction methods, realize the organic combination of data-driven and mechanism model, and improve the credibility of prediction results and the rationality of control strategies.

[0039] Step S5: Based on the target dissolved oxygen concentration, deduce the target oxygen mass transfer coefficient, and use a multilayer perceptron model to predict the target fan control quantity.

[0040] Specifically, the reverse calculation process for the target oxygen mass transfer coefficient is as follows: Based on the target dissolved oxygen concentration Current dissolved oxygen concentration Sampling time interval The target dissolved oxygen change rate was calculated. : ; Then, combining this with the oxygen mass transfer rate equation: ; in, Given the saturated dissolved oxygen concentration under the current operating conditions, the target oxygen mass transfer coefficient is obtained by inverse solution. : ; The target oxygen mass transfer coefficient can be obtained from the target dissolved oxygen concentration through the above calculations. This provides data support for the reverse calculation of the fan control quantity, avoiding the combined influence of factors such as fan opening degree, water temperature, water quality, and pool type in the forward calculation, and realizing the reverse measurement of the coefficient.

[0041] Specifically, in step S5, the prediction process for the target wind turbine control quantity is as follows: The historical oxygen mass transfer coefficient at the corresponding moment is calculated based on the dissolved oxygen concentration in the historical operation data, and a second training sample set is constructed. Water quality parameters (including total nitrogen) at the corresponding time points in the second training sample set TN Total phosphorus TP Chemical oxygen demand COD Dissolved oxygen concentration DO(e.g., flow rate, water temperature, etc.) and the target oxygen mass transfer coefficient obtained by reverse solution As input, the actual wind turbine opening control value recorded at that moment is used as the corresponding supervision label to train the multilayer perceptron model; The regression function for training a multilayer perceptron model is as follows: ; in, This represents the predicted control value for the turbine opening. This represents the feature vector composed of current water quality parameters. This represents the regression function of the multilayer perceptron obtained through training; The target wind turbine control quantity is obtained by inputting the second test sample into the trained multilayer perceptron model for prediction.

[0042] Using the second training sample set, the multilayer perceptron (...) can be made... MLP The model learns the mapping relationship between water quality parameters, the target oxygen mass transfer coefficient, and the actual fan operation control value. Based on this mapping relationship, the fan control quantity can be predicted for a second test sample (the oxygen mass transfer coefficient derived from the target dissolved oxygen concentration in the input test scenario, and the corresponding water quality parameters measured by sensors). This achieves the inverse solution process for the fan control quantity. Compared to a fixed system setting... DO The control point approach allows for intelligent adjustment of aeration intensity based on current operating conditions, resulting in greater energy-saving potential.

[0043] In addition, the oxygen mass transfer coefficient is also mentioned in some technologies. The calculation method mainly uses formulas related to gas holdup, apparent water flow velocity, and bubble characteristics; the calculation method is obviously fundamentally different from the process of back-calculating using the target dissolved oxygen concentration of the present invention.

[0044] What needs to be understood is that MLP The model consists of an input layer, several hidden layers, and an output layer, where the hidden layers employ... ReLU The activation function, with linear regression nodes in the output layer, is used to predict continuous values ​​of wind turbine operating rates. The loss function uses mean squared error (MSE). MSE ), optimizer is Adam The learning rate was set to 0.001, and training was performed using a mini-batch method with a batch size of 32. The number of training epochs was set to 50-100 epochs based on the convergence of the validation set error. The input data was pre-processed. Z-score Standardization ensures that all input dimensions have a uniform numerical scale, improving training stability and model convergence speed.

[0045] In this embodiment, step S5 further includes: using error feedback...PID The control module performs error correction on the target fan control quantity.

[0046] Specifically, the error correction process is as follows: With target dissolved oxygen concentration The set value is based on the actual sensor monitoring value. For the feedback value, calculate its deviation: ; The fine-tuning control quantity is calculated based on proportional, integral, and derivative control laws. : ; in, , , These are the proportional, integral, and derivative parameters, which are set empirically based on the system response characteristics or obtained through online trial and error. By substituting the predicted target wind turbine control quantity and the fine-tuning control quantity into the following formula: u final1 =u OTR +u PID； The final fan control output is calculated, where, Indicates passage MLP-OTR The predicted wind turbine operating rate obtained by path inversion This represents the real-time error compensation term. u final1 This is the control opening degree ultimately used to drive the blower.

[0047] Using the above PID The feedback adjustment mechanism corrects errors in the target wind turbine control quantities predicted by the model, achieving a complementary integration of deep modeling and traditional control logic. Even in cases of model prediction deviation or sensor drift, closed-loop adjustment can still maintain system operational stability, improving the feasibility of engineering implementation.

[0048] Furthermore, to further improve the robustness and prediction accuracy of the system under complex disturbance conditions, this invention also introduces an Extended State Observer (ESO) into the PID-based error compensation mechanism. This observer can uniformly estimate the modeling errors and external disturbances during system operation and treat them as a "total disturbance" for real-time compensation, thereby enhancing the adaptability to non-ideal operating conditions.

[0049] Specifically, suppose the dynamic equation for the dissolved oxygen concentration in the system can be expressed as: ; in, Dissolved oxygen concentration, For known ideal system model terms, This refers to the total disturbance caused by external disturbances and modeling uncertainties. The extended state observer adds a one-dimensional disturbance observation to the conventional state estimation, obtaining a real-time estimate of the disturbance. .

[0050] An extended state observer is used to perform one-dimensional disturbance observation on the process of error correction of the target wind turbine control quantity through the error feedback-based PID control module, and the disturbance compensation quantity is obtained. ; By substituting the predicted target wind turbine control quantity, fine-tuning control quantity, and disturbance compensation quantity into the following formula: ; The final fan control output is calculated, where, Indicates passage MLP-OTR The predicted wind turbine operating rate obtained by path inversion This represents the real-time error compensation term. This is the disturbance compensation amount. This is the control opening degree ultimately used to drive the blower.

[0051] This invention organically combines the oxygen mass transfer equation with a deep learning model. By predicting the minimum acceptable dissolved oxygen concentration and then inferring the blower control strategy, it achieves intelligent control of aeration energy consumption in wastewater treatment. Compared with traditional aeration control methods based on fixed dissolved oxygen concentration setpoints or purely data-driven approaches, this invention is based on the oxygen mass transfer equation... OTR The introduction of the physical mechanism of the formula improves the interpretability and controllability of the model; at the same time, through Transformer Collaborative modeling with a multilayer sensor significantly enhances the adaptability to changes in nonlinear operating conditions. It not only dynamically adjusts the oxygen dissolved concentration setpoint to achieve energy-saving operation while ensuring effluent quality meets standards, but also introduces… PID The feedback mechanism provides a safety net, enhancing the system's robustness and engineering feasibility. The overall solution constructs a closed-loop intelligent control architecture of "target prediction - physical inverse solution - feedback calibration," providing a scalable technical path for achieving energy-efficient operation of wastewater treatment processes, demonstrating significant practical value and innovative potential.

[0052] Furthermore, in actual operating conditions, changes in dissolved oxygen concentration are also affected by microbial oxygen consumption. Although this term is not included in the oxygen mass transfer equation, the multilayer perceptron model constructed in this invention implicitly compensates for this factor during the training process based on the joint feature learning of historical water quality parameters and oxygen mass transfer coefficients, thereby ensuring the accuracy and applicability of the prediction results.

[0053] like Figure 3 As shown, this invention also provides an aeration control system integrating an oxygen mass transfer process and a deep learning model. The system includes: a data acquisition module 10 for acquiring historical operating data of a wastewater treatment system, wherein the historical operating data includes influent water quality parameters, effluent water quality parameters, and dissolved oxygen concentration; a generation module 20 for generating a first training sample set based on the historical operating data; a training module 30 for inputting the first training sample set into a pre-constructed deep learning model for training; a first prediction module 40 for inputting a first test sample into the trained deep learning model for prediction to obtain a target dissolved oxygen concentration; and a second prediction module 50 for deriving a target oxygen mass transfer coefficient based on the target dissolved oxygen concentration and predicting a target blower control quantity using a multilayer perceptron model. The aeration control system integrating an oxygen mass transfer process and a deep learning model of this invention can achieve the same technical effects as the aeration control method integrating an oxygen mass transfer process and a deep learning model.

[0054] The present invention has been disclosed above with reference to preferred embodiments, but it is not intended to limit the present invention. All technical solutions obtained by adopting equivalent substitutions or equivalent transformations fall within the protection scope of the present invention.

Claims

1. An aeration control method integrating oxygen mass transfer processes and deep learning models, characterized in that, Includes the following steps: Step S1: Collect historical operating data of the wastewater treatment system, wherein the historical operating data includes: influent water quality parameters, effluent water quality parameters and dissolved oxygen concentration; Step S2: Generate a first training sample set based on the historical running data; Step S3: Input the first training sample set into the pre-built deep learning model for training; Step S4: Input the first sample to be tested into the trained deep learning model for prediction to obtain the target dissolved oxygen concentration; Step S5: Based on the target dissolved oxygen concentration, deduce the target oxygen mass transfer coefficient, and use a multilayer perceptron model to predict the target fan control quantity.

2. The aeration control method based on the fusion of oxygen mass transfer process and deep learning model according to claim 1, characterized in that, In step S2, the process of generating the first training sample set is as follows: Valid samples that meet the preset emission standards are selected from the historical operating data; Based on the influent water quality parameters, samples that meet the preset discharge standards are discretized by interval to form multiple subsets representing different operating conditions; The record with the lowest dissolved oxygen level is selected from the multiple subsets representing different working conditions obtained after grouping, and used as the minimum acceptable dissolved oxygen concentration under each working condition to form a static sample set. For each sample record in the static sample set, the multi-dimensional water inflow parameters within a fixed time window are traced back in chronological order to construct a two-dimensional input matrix with a time series structure, forming a first training sample set with temporal characteristics.

3. The aeration control method based on the fusion of oxygen mass transfer process and deep learning model according to claim 2, characterized in that, Step S2, the process of generating the first training sample set also includes: The physical consistency of the first training sample set is verified based on the oxygen mass transfer equation, and data that does not meet the preset verification conditions are removed or downweighted. The physical consistency verification process is as follows: The dissolved oxygen change rate dC / dt for each sequence in the first training sample was measured and combined with the oxygen mass transfer equation: ， in, This represents the current dissolved oxygen concentration. This represents the saturated dissolved oxygen concentration under the current operating conditions. The oxygen mass transfer coefficient; Compare the measured dissolved oxygen change rate dC / dt with the calculated value on the right side of the equation; When the oxygen mass transfer coefficient obtained by reverse calculation according to the formula exceeds the preset empirical range, or the difference exceeds the preset threshold, the sequence is removed or downweighted.

4. The aeration control method based on the fusion of oxygen mass transfer process and deep learning model according to claim 2, characterized in that, Step S2, the process of generating the first training sample set also includes: The data from multiple subsets representing different working conditions obtained after grouping are sorted from low to high dissolved oxygen concentration, and the samples in the range of 0 to q quantile are taken as the minimum acceptable dissolved oxygen concentration for each working condition, forming a static sample set.

5. The aeration control method based on the fusion of oxygen mass transfer process and deep learning model according to claim 1, characterized in that, In step S3, the deep learning model is constructed as follows: Construct an input mapping layer to map the water quality feature vectors of the original input sequence at each time step. Mapped to a unified dimension of representation space The linear transformation method used in the mapping function is as follows: ； in, ; It refers to a dimension of d x The vector space represents the dimension of the input features and includes several influent water quality parameters. It refers to a dimension of d m The vector space W represents the unified feature representation space after mapping. e This refers to the linear mapping weights, which transform the original feature x... t Mapped to a unified-dimensional representation space, b e It refers to the bias vector of the input mapping layer, used to correct the offset of the linear mapping output. It refers to d m line ×d x The matrix space of columns; the result of the linear transformation constitutes a matrix of length. sequence As the encoder input, E is the input feature matrix composed of time series samples, T is the number of time steps, and e t This represents the feature representation at time step t. Build Transformer The decoder module consists of several stacked encoder layers, each of which contains two core modules: a multi-head attention mechanism and a feedforward neural network. Multi-head self-attention mechanisms are used to uncover correlations between time steps within a time series. The core calculation formula is as follows: ； in, , represents the linear mapping result of query, key, and value; Q, K, and V are the query, key, and value matrices in the attention mechanism, respectively, obtained from the input feature matrix E through linear transformation; W Q W K W V d represents the linear projection weight matrices corresponding to the query, key, and value, respectively, used to map the input E to the attention subspace; k For the feature dimension in the attention mechanism; Multi-head attention mechanisms model attention by using multiple independent attention heads in parallel: ； in, , To output the projection matrix; The feedforward neural network consists of two fully connected layers with a non-linear activation function in between. dropout layer; Integrating residual connections between the multi-head attention mechanism and the feedforward neural network modules LayerNorm operate: ； To form a stable deep network structure, in which SubLayer X represents the attention or feedforward module; X represents the tensor of the sublayer input or the previous layer output, used for residual connections and layer normalization operations. Transformer The encoder output is a context representation sequence for each time step. Finally, the output vector of the last time step is selected. The linear regression layer maps the data to scalar prediction results. ； in, , For bias terms; The target dissolved oxygen concentration is to be predicted; The model uses mean squared error as the loss function: ； in, The true label value of the sample. These are the model's predicted values. The total number of samples.

6. The aeration control method based on the fusion of oxygen mass transfer process and deep learning model according to claim 1, characterized in that, In step S5, the reverse calculation process for the target oxygen mass transfer coefficient is as follows: Based on the target dissolved oxygen concentration Current dissolved oxygen concentration Sampling time interval The target dissolved oxygen change rate was calculated. : ； Then, combining this with the oxygen mass transfer rate equation: ； in, Given the saturated dissolved oxygen concentration under the current operating conditions, the target oxygen mass transfer coefficient is obtained by inverse solution. : 。 7. The aeration control method based on the fusion of oxygen mass transfer process and deep learning model according to claim 1, characterized in that, In step S5, the prediction process for the target wind turbine control quantity is as follows: The historical oxygen mass transfer coefficient at the corresponding moment is calculated based on the dissolved oxygen concentration in the historical operation data, and a second training sample set is constructed. Using the water quality parameters at the corresponding time in the second training sample set and the target oxygen mass transfer coefficient obtained by inverse solution As input, the actual wind turbine opening control value recorded at that moment is used as the corresponding supervision label to train the multilayer perceptron model; The regression function for training a multilayer perceptron model is as follows: ； in, This represents the predicted control value for the turbine opening. This represents the feature vector composed of current water quality parameters. This represents the regression function of the multilayer perceptron obtained through training; The target wind turbine control quantity is obtained by inputting the second test sample into the trained multilayer perceptron model for prediction.

8. The aeration control method based on the fusion of oxygen mass transfer process and deep learning model according to claim 1, characterized in that, Step S5 also includes: The target fan control quantity is corrected for errors using a PID control module based on error feedback. The error correction process is as follows: With target dissolved oxygen concentration The set value is based on the actual sensor monitoring value. For the feedback value, calculate its deviation: ； The fine-tuning control quantity is calculated based on proportional, integral, and derivative control laws. : ； in, , , These are the proportional, integral, and derivative parameters, which are set empirically based on the system response characteristics or obtained through online trial and error. By substituting the predicted target wind turbine control quantity and the fine-tuning control quantity into the following formula: u final1 =u OTR +u PID； The final fan control output is calculated, where, Indicates passage MLP-OTR The predicted wind turbine operating rate obtained by path inversion This represents the real-time error compensation term. u final1 This is the control opening degree ultimately used to drive the blower.

9. The aeration control method based on the fusion of oxygen mass transfer process and deep learning model according to claim 8, characterized in that, Step S5 also includes: The extended state observer is used to perform one-dimensional disturbance observation on the process of error correction of the target wind turbine control quantity through the error feedback-based PID control module, and the disturbance compensation quantity is obtained. By substituting the predicted target wind turbine control quantity, fine-tuning control quantity, and disturbance compensation quantity into the following formula: ； The final fan control output is calculated, where, Indicates passage MLP-OTR The predicted wind turbine operating rate obtained by path inversion This represents the real-time error compensation term. This is the disturbance compensation amount. This is the control opening degree ultimately used to drive the blower.

10. An aeration control system integrating oxygen mass transfer processes and deep learning models, characterized in that, include: The data acquisition module is used to collect historical operating data of the wastewater treatment system, wherein the historical operating data includes: influent water quality parameters, effluent water quality parameters, and dissolved oxygen concentration; The generation module is used to generate a first training sample set based on the historical running data; The training module is used to input the first training sample set into a pre-built deep learning model for training. The first prediction module is used to input the first test sample into the trained deep learning model for prediction to obtain the target dissolved oxygen concentration. The second prediction module is used to deduce the target oxygen mass transfer coefficient based on the target dissolved oxygen concentration and to predict the target fan control quantity using a multilayer perceptron model.

Images