Intelligent adding method of sewage treatment carbon source

Abstract

The invention provides a sewage treatment carbon source intelligent adding method, which comprises the following steps: collecting multi-parameter feed-forward and feedback signals of water inlet and an anoxic tank, constructing a dynamic model containing feed-forward compensation, model prediction control and feedback compensation, calculating the theoretical adding amount of a carbon source, inputting a predicted value and feedback parameters into an LSTM network for correction, and optimizing a network structure by a genetic algorithm. The adding amount is controlled in a closed-loop mode through a variable frequency pump, the LSTM weight is updated on the basis that the error is larger than 5%, and finally a control strategy is optimized by using an NSGA-II algorithm and integrating carbon source consumption, effluent total nitrogen and energy consumption. The dynamic self-adaptive carbon source adding method is constructed by fusing feedforward perception, LSTM prediction, feedback regulation and multi-objective optimization, so that quick response and accurate control on water quality fluctuation are realized, the denitrification efficiency and the carbon source utilization rate are improved, and the method has excellent engineering adaptability and popularization value.

Description

Technical Field

[0001] The present invention belongs to the technical field of sewage treatment, and in particular relates to an intelligent carbon source dosing method for sewage treatment. Background Art

[0002] Traditional carbon source addition control methods have certain technical limitations in wastewater denitrification treatment, mainly manifested in response lag, model rigidity, and disconnection between process links. Existing control strategies usually use effluent total nitrogen (TN) as the only feedback signal, and the adjustment process has obvious lag. Especially when the influent water quality load changes suddenly, conventional PID control is difficult to respond in time, which can easily lead to inaccurate carbon source addition and waste of resources. On the other hand, the fixed parameter control model lacks the ability to adapt to dynamic operating conditions such as sludge activity and water temperature. Under conditions with significant temperature changes, there will be a mismatch between carbon source addition and denitrification demand, affecting denitrification efficiency and increasing the risk of effluent water quality exceeding standards.

[0003] At the same time, carbon source addition is often controlled separately from key process parameters such as aeration and reflux, lacking overall coordinated optimization. Although some existing PID-based automatic dosing devices have basic automation capabilities, they still lack model self-learning capabilities and multi-objective optimization mechanisms, making it difficult to balance denitrification efficiency, carbon source economy, and operating energy consumption. These issues have seriously hindered the advancement of intelligent and refined operation of sewage treatment plants, and there is an urgent need to introduce intelligent dosing strategies that integrate feedforward sensing, dynamic prediction, and multi-parameter coordination to achieve breakthroughs.

[0004] Therefore, the present invention provides a method for intelligently adding carbon sources for sewage treatment. Summary of the Invention

[0005] The present invention designs an intelligent carbon source addition method for sewage treatment to solve the problems existing in the above-mentioned background technology.

[0006] A method for intelligently adding a carbon source for sewage treatment, comprising:

[0007] S1. Feedforward signal acquisition: Real-time acquisition of COD, BOD5, ammonia nitrogen and total phosphorus concentrations of influent water, using a spectrum analyzer for data acquisition, with a sampling frequency of not less than 1 time / minute;

[0008] S2. Feedback signal acquisition: Multi-parameter sensors installed at the water inlet, reaction zone and outlet of the anoxic tank continuously collect liquid level, water temperature, total nitrogen, oxidation-reduction potential (ORP), mixed liquor suspended solids concentration (MLSS), dissolved oxygen (DO) and pH value;

[0009] S3. Construct a dynamic compensation model, including a feedforward compensation module, a model predictive control module, and a feedback compensation module. Based on the changes in inlet load and water temperature, the theoretical carbon source dosage Q1 is calculated using the feedforward compensation formula:

[0010]

[0011] S4, LSTM neural network prediction and correction, including digital preprocessing, model training and deployment, and dynamic correction mechanism. Feedforward prediction values ​​and feedback parameters are input into the LSTM neural network. The network input layer contains 12-dimensional time series features, and the number of hidden layer nodes is determined by genetic algorithm.

[0012] S5, closed-loop feedback control: execute the Q2 dosing instruction through the variable frequency dosing pump, collect the actual dosage Q3 in real time, and trigger the model parameter self-calibration based on the error function E = |Q2-Q3| / Q2. When E>5%, update the LSTM weight;

[0013] S6. Multi-objective optimization: Based on carbon source consumption, effluent total nitrogen and energy consumption parameters, the NSGA-II algorithm is used to generate the optimal control strategy.

[0014] Furthermore, the feedforward compensation module in S3 monitors the change amplitude of the influent flow, COD concentration and ammonia nitrogen concentration in real time. When the change of any parameter exceeds a preset threshold, the feedforward compensation module is activated, and the hysteresis effect of the shock load on nitrate nitrogen is calculated through the delay factor, and the feedforward carbon source dosage is output;

[0015] The formula for calculating the impact load quantification in the feedforward compensation module is:

[0016]

[0017] Among them, Q 实 : Real-time water flow;

[0018] COD 实 : Actual influent COD concentration

[0019] COD 基 : baseline COD concentration;

[0020] K f : Feedforward coefficient, where the empirical value is 0.8-1.2;

[0021] The feedforward compensation module also includes a delay correction mechanism, which introduces hydraulic retention time to calibrate the impact period of the impact load.

[0022] Furthermore, the model prediction control module in S3 includes constructing a carbon source addition algorithm model, the input parameters of which include the total nitrogen concentration of the influent, the influent flow rate, the internal recirculation ratio, the historical water quality data and the nitrate nitrogen target value. The carbon source addition demand is predicted by the LSTM model to dynamically update the set value, and the multivariable nonlinear relationship is established by training the neural network through historical data.

[0023] The formula for calculating the carbon source dosage in the model prediction module is:

[0024]

[0025] Wherein, HRT is hydraulic retention time;

[0026] C 模型 The amount of carbon source added for model calculation;

[0027] f is the coefficient, reflecting the comprehensive relationship between the amount of carbon source added and other independent variables in the brackets;

[0028] TN 进水 is the total nitrogen in the influent;

[0029] Q is the water inlet flow rate;

[0030] The measured nitrate nitrogen concentration.

[0031] The model prediction control module adopts a rolling optimization strategy, updates the prediction results every 5-10 minutes, and adjusts the carbon source set value in combination with feedback parameters.

[0032] Furthermore, the feedback compensation module in S3 includes monitoring the deviation between the nitrate nitrogen concentration at the end of the anoxic tank and the target value and the total nitrogen concentration of the effluent. When the total nitrogen concentration of the effluent exceeds the set range or the nitrate nitrogen deviation continues to exceed the limit, feedback compensation is triggered; the PID parameters in the feedback compensation module are adaptively adjusted through fuzzy logic.

[0033] Furthermore, the LSTM neural network feedback correction in S4 includes:

[0034] Real-time deviation correction: Real-time deviation correction is to generate a deviation signal by comparing the actual value of the total nitrogen online instrument monitoring the effluent with the target value. The deviation value e(t) = total nitrogen target (TN 目标 )-actual total nitrogen value (TN 实际 ), deviation change rate Δe=e(t)-e(t-1);

[0035] Fuzzy-PID control: Fuzzy-PID control is to input the deviation signal into the fuzzy controller, output the carbon source addition correction amount, and superimpose it on the feedforward-model prediction value.

[0036] Furthermore, in S4, data preprocessing includes:

[0037] Eliminate abnormal instrument data: normalize multiple parameters to eliminate dimensionality effects, and select parameters with strong correlation as model input parameters;

[0038] Model training and deployment steps: Use historical data and corresponding carbon source dosage to construct a supervised learning dataset, and use mean absolute error and effluent total nitrogen constraint as joint loss indicators for training;

[0039] Dynamic correction mechanism: including feedforward-feedback fusion decision and anti-hysteresis design. The feedforward-feedback fusion decision calculation method is as follows:

[0040] Final dosage = LATM predicted value (model output) + (K p .e(t)+K d .Δe)(feedback correction term), where K p , K d is the proportional / differential coefficient output by the fuzzy controller;

[0041] Anti-hysteresis design: The feedback correction amount is calibrated based on the current effluent total nitrogen deviation and the future predicted deviation trend.

[0042] The beneficial effects of the present invention are:

[0043] 1. This invention constructs a dynamic and adaptive carbon source intelligent dosing method by integrating feedforward perception, LSTM model prediction, feedback regulation and multi-objective optimization, effectively breaking through the technical limitations of traditional PID control such as response lag, model rigidity and disconnection of process links;

[0044] 2. The present invention can realize rapid identification and adjustment under the condition of fluctuating influent water quality or abnormal operating conditions, accurately control the amount of carbon source added, improve the stability and anti-interference ability of the system, and optimize the carbon source utilization rate and operating energy consumption while ensuring that denitrification meets the standard; in actual application, it significantly reduces the unit consumption of carbon source, improves the compliance rate of total nitrogen in effluent, and increases the response speed to minutes. It has good engineering adaptability and promotion value, and provides effective support for the intelligent and refined operation of sewage treatment plants. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0046] Figure 1 This is a feedforward-model-feedback data flow chart in the present invention;

[0047] Figure 2 Schematic diagram of the LSTM network structure in the present invention;

[0048] Figure 3 This is the SCADA interface diagram of a certain A2O sewage treatment plant in the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0050] Example 1

[0051] See Figures 1 to 3 As shown,

[0052] A method for intelligently adding a carbon source for sewage treatment, comprising:

[0053] S1. Feedforward signal acquisition: Real-time acquisition of COD, BOD5, ammonia nitrogen and total phosphorus concentrations of influent water, using a spectrum analyzer for data acquisition, with a sampling frequency of not less than 1 time / minute;

[0054] S2. Feedback signal acquisition: Multi-parameter sensors installed at the water inlet, reaction zone and outlet of the anoxic tank continuously collect liquid level, water temperature, total nitrogen, oxidation-reduction potential (ORP), mixed liquor suspended solids concentration (MLSS), dissolved oxygen (DO) and pH value;

[0055] S3. Construct a dynamic compensation model, including a feedforward compensation module, a model predictive control module, and a feedback compensation module. Based on the changes in inlet load and water temperature, the theoretical carbon source dosage Q1 is calculated using the feedforward compensation formula:

[0056]

[0057] S4, LSTM neural network prediction and correction, including digital preprocessing, model training and deployment, and dynamic correction mechanism. Feedforward prediction values ​​and feedback parameters are input into the LSTM neural network. The network input layer contains 12-dimensional time series features, and the number of hidden layer nodes is determined by genetic algorithm.

[0058] S5, closed-loop feedback control: execute the Q2 dosing instruction through the variable frequency dosing pump, collect the actual dosage Q3 in real time, and trigger the model parameter self-calibration based on the error function E = |Q2-Q3| / Q2. When E>5%, update the LSTM weight;

[0059] S6. Multi-objective optimization: Based on carbon source consumption, effluent total nitrogen and energy consumption parameters, the NSGA-II algorithm is used to generate the optimal control strategy.

[0060] The feedforward compensation module in S3 monitors the change amplitude of the inlet flow rate, COD concentration and ammonia nitrogen concentration in real time. When the change of any parameter exceeds the preset threshold, the feedforward compensation module is activated, and the hysteresis effect of the impact load on nitrate nitrogen is calculated by the delay factor, and the feedforward carbon source dosage is output;

[0061] The formula for calculating the impact load quantification in the feedforward compensation module is:

[0062]

[0063] Among them, Q 实 : Real-time water flow;

[0064] COD 实 : Actual influent COD concentration;

[0065] COD 基 : baseline COD concentration;

[0066] K f : Feedforward coefficient, where the empirical value is 0.8-1.2.

[0067] The feedforward compensation module also includes a delay correction mechanism, which introduces hydraulic retention time to calibrate the impact period of the impact load to avoid leading and lagging addition.

[0068] The model prediction control module in S3 includes constructing a carbon source addition algorithm model, the input parameters of which include the total nitrogen concentration of the influent, the influent flow rate, the internal recirculation ratio, historical water quality data and the nitrate nitrogen target value, predicting the carbon source addition demand through the LSTM model, dynamically updating the set value, and training the neural network through historical data to establish a multivariate nonlinear relationship;

[0069] The formula for calculating the carbon source dosage in the model prediction module is:

[0070]

[0071] Wherein, HRT is hydraulic retention time;

[0072] C 模型 The amount of carbon source added for model calculation;

[0073] f is the coefficient, reflecting the comprehensive relationship between the amount of carbon source added and other independent variables in the brackets;

[0074] TN 进水 is the total nitrogen in the influent;

[0075] Q is the water inlet flow rate;

[0076] The measured nitrate nitrogen concentration.

[0077] The model prediction control module adopts a rolling optimization strategy, updates the prediction results every 5-10 minutes, and adjusts the carbon source set value in combination with feedback parameters.

[0078] The feedback compensation module in S3 includes monitoring the deviation between the nitrate nitrogen concentration at the end of the anoxic tank and the target value and the total nitrogen concentration of the effluent. When the total nitrogen concentration of the effluent exceeds the set range or the nitrate nitrogen deviation continues to exceed the limit, feedback compensation is triggered; the PID parameters in the feedback compensation module are adaptively adjusted through fuzzy logic.

[0079] The LSTM neural network feedback correction in S4 includes:

[0080] Real-time deviation correction: Real-time deviation correction is to generate deviation signals by comparing the actual value of the effluent total nitrogen online instrument monitoring with the target value.

[0081] Deviation value e(t) = total nitrogen target (TN 目标 )-actual total nitrogen value (TN 实际 ),

[0082] Deviation change rate Δe = e(t) - e(t-1);

[0083] Fuzzy-PID control: Fuzzy-PID control is to input the deviation signal into the fuzzy controller, output the carbon source addition correction amount, and superimpose it on the feedforward-model prediction value.

[0084] In S4, data preprocessing includes:

[0085] Eliminate abnormal instrument data: Normalize multiple parameters to eliminate dimensionality effects, and select highly correlated parameters as model input parameters to reduce overfitting risks;

[0086] Model training and deployment steps: Use historical data and corresponding carbon source dosage to construct a supervised learning dataset. Use mean absolute error and effluent total nitrogen constraint as joint loss indicators for training to ensure the economical and compliance of the predicted value.

[0087] Dynamic correction mechanism: including feedforward-feedback fusion decision and anti-hysteresis design. The feedforward-feedback fusion decision calculation method is as follows:

[0088] Final dosage = LATM predicted value (model output) + (K p .e(t)+K d .Δe)(feedback correction term), where K p , K d is the proportional / differential coefficient output by the fuzzy controller;

[0089] Anti-hysteresis design: The feedback correction is calibrated based on the current effluent total nitrogen deviation and the future predicted deviation trend to reduce system response delay.

[0090] Example 2

[0091] Model training and validation:

[0092] 1. Dataset Construction

[0093] (1) Data source: Historical operating data of a 100,000 tons / day municipal sewage treatment plant for three years (2019-2022), covering the full process parameters of the AAO process, with a sampling frequency of 5 minutes / time.

[0094] (2) Key characteristic variables:

[0095] Influent water quality: COD (100-450mg / L), NH4 + -N (20-50 mg / L), TN (30-65 mg / L), flow rate (Q), temperature (T).

[0096] Process parameters: MLSS (2000-4500 mg / L), DO (0.5-2.5 mg / L), internal reflux ratio (100%-300%).

[0097] Abnormal operating condition labels: such as sludge bulking (SVI>150mL / g), rainstorm impact (instantaneous water flow increase exceeds 30%), continuous low temperature (water temperature T<10℃, and lasting for more than 48 hours), etc.

[0098] 2. Model training and optimization:

[0099] (1) LSTM network structure:

[0100] Input layer: 12-dimensional features after standardization;

[0101] Hidden layer: two layers of LSTM (128 neurons), Dropout = 0.2 to prevent overfitting;

[0102] Output layer: Output the predicted carbon source demand value (unit: mg / L) and denitrification rate predicted value (unit: mg-N / (gMLSS·h)) in the next 2 hours.

[0103] (2) Training strategy:

[0104] Use sliding window training (window = 30 days, step size = 1 day) to dynamically update model weights;

[0105] (3) Loss function: MAE + dynamic weight adjustment, giving higher weight to abnormal working condition data (such as the error penalty coefficient during heavy rain period × 1.5).

[0106] 3. Verification result comparison:

[0107]

[0108] 4. Typical scenario analysis:

[0109] Heavy rain impact (2021 / 7 / 20):

[0110] The LSTM model predicted a 40% sudden increase in influent TN 30 minutes in advance, promptly increasing carbon source dosage by 25%, and keeping effluent TN below 12 mg / L.

[0111] In contrast, the ARIMA model had a response lag of 1.5 hours, causing the TN in the effluent to exceed the standard by 18 mg / L for a short period of time and a carbon source waste of 19%.

[0112] Low temperature conditions (2022 / 1 / 15):

[0113] The LSTM model automatically adjusted the carbon source addition by 12% based on the decrease in MLSS activity, and the denitrification rate only decreased by 8%;

[0114] When the fixed model was not modified, the denitrification efficiency dropped by 35%, and the TN in the effluent fluctuated between 15 and 22 mg / L.

[0115] 5.Technical advantages:

[0116]

[0117]

[0118] Improved accuracy: LSTM reduces prediction error by more than 55% compared to traditional methods, and significantly enhances adaptability to abnormal operating conditions;

[0119] Generalization ability: Verified by transfer learning, the model achieved an error of <6% in three sewage treatment plants with the same process.

[0120] Example 2

[0121] On-site application effect

[0122] After implementing the intelligent carbon source dosing system of the present invention at a 100,000-ton / day municipal sewage treatment plant (AAO process), six months of continuous operation monitoring revealed significant improvements in key performance indicators compared to the traditional PID control method. Specific data and analysis are as follows:

[0123] 1. Comparison of core indicators

[0124] 2. Comparison of Typical Operation Scenarios

[0125] Scenario 1: Heavy rain impact (July 2023)

[0126] The traditional PID responds about 2 hours later after a sudden 50% increase in the inlet TN, resulting in an instantaneous over-dosage of 35% in the carbon source and an increase in the outlet TN to 15 mg / L. The system of the present invention uses LSTM to predict load changes 30 minutes in advance, achieving a 20% gradient adjustment of the carbon source, and the outlet TN is always stable below 8 mg / L.

[0127] Scenario 2: Low-temperature operation (December 2023)

[0128] When the water temperature drops to 8°C, the traditional PID still adds water according to fixed parameters, resulting in a 40% drop in denitrification efficiency and requiring manual intervention. The system of the present invention automatically compensates the carbon source by 15% through a dynamic model and reduces the internal reflux ratio by 10%. The TN of the effluent continues to meet the standard, and the compliance rate remains 100%.

[0129] 3. Reasons for the reduction in failure rate

[0130] The reduction in failure rate is due to the coordinated optimization of software and hardware: in terms of hardware, industrial-grade edge computing terminals with IP67 protection level are used to replace traditional PLCs, significantly improving adaptability to humid and corrosive environments; in terms of software, through the abnormal data detection mechanism, it automatically switches to model prediction mode when the sensor fails (such as enabling LSTM virtual sensing when ORP is abnormal), and adopts a sliding window training strategy to update model parameters every 24 hours, effectively preventing control failure caused by data drift.

[0131] 4. Economic benefit calculation

[0132] After the system was put into operation, carbon source consumption was reduced by 0.9 kg / TN. Calculated at 2,000 yuan / ton, this directly saves approximately 2.16 million yuan in sodium acetate costs annually. Indirectly, the effluent TN compliance rate increased to 100%, avoiding approximately 500,000 yuan in fines for exceeding the standard annually. Furthermore, stable operation reduces maintenance costs, saving approximately 300,000 yuan annually. The overall system renovation investment was approximately 1.5 million yuan, with a payback period of just eight months.

[0133] 5. Promotion value

[0134] This system has been certified by the China Environmental Protection Industry Association and is suitable for municipal sewage treatment plants (especially combined sewer network areas with large water inflow fluctuations), high-nitrogen industrial wastewater treatment scenarios such as pharmaceutical and chemical industries, as well as the intelligent transformation of old sewage treatment plants. It has the application advantage of being deployable without changing the main process and only requiring the installation of intelligent edge terminals.

[0135] Example 3

[0136] Extreme working condition test: System response analysis under high concentration COD shock load

[0137] 1. Test scenario design

[0138] (1) Test conditions: In a normally operating AAO process (processing capacity of 100,000 tons / day), the influent COD is instantly increased by 50% (from 300 mg / L to 450 mg / L) due to the mixing of simulated industrial wastewater.

[0139] (2) Test parameters:

[0140] The total nitrogen in the influent increased by 35%; the water temperature was stabilized at 22±1℃; and the mixed liquor suspended solids (MLSS) concentration was maintained at approximately 3500 mg / L.

[0141] Other process parameters (DO, reflux ratio, etc.) remain in automatic operation;

[0142] 2. System response process (time series analysis)

[0143]

[0144] 3. Control effect comparison:

[0145] (1) Traditional PID control:

[0146] The response delay is long, usually taking 45 to 60 minutes to wait for the total nitrogen feedback of the effluent; the TN fluctuation range is large, and the instantaneous peak can reach 18.3 mg / L; the system recovery time exceeds 4 hours; the carbon source waste rate is high, about 22% to 25%.

[0147] (2) This system controls:

[0148] The response delay is less than 8 minutes (combining feedforward perception and model prediction); TN fluctuation is controlled within ±2.1 mg / L, and the maximum value does not exceed 11.2 mg / L; the system recovery time is about 1.2 hours; and the carbon source waste rate is reduced to 7.3%.

[0149] 4.Technical implementation mechanism:

[0150] (1) Feedforward rapid response: Using the sliding window variance analysis (SWVA) algorithm, an early warning is triggered immediately when a COD change rate >15% / 5min is detected, and a pre-stored library of 12 impact load response solutions is called.

[0151] (2) Dynamic model adjustment: real-time fine-tuning of LSTM network weights (learning rate increased by 3 times);

[0152] (3) Combined with the ASM2d model, the temperature-sludge activity compensation coefficient was corrected.

[0153] (4) Feedback precision control: three parameters coordinated monitoring (ORP derivative + NO3 - -N+MLSS), the carbon source cutoff mechanism is initiated when dORP / dt>0.5mV / s.

[0154] 5. Engineering verification data:

[0155] In three consecutive rounds of repeated tests, the system operated stably, as shown by: the parameter adjustment time was stable at about 7.8 minutes, the fluctuation of total nitrogen in the effluent was controlled within 2.3 mg / L, and the carbon source utilization rate was maintained at more than 92%, with a small fluctuation range.

[0156] 6. Industry application value:

[0157] Test results show that the system performs excellently under the following typical working conditions: such as the inrush of industrial wastewater such as food processing wastewater, water quality impact on combined sewer networks during the rainy season, and disturbance of reflux liquid from sludge treatment systems. It can realize the transformation of carbon source addition from "passive response" to "active defense."

Claims

1. A method for intelligently adding carbon sources to sewage treatment, characterized in that: include: S1. Feedforward signal acquisition: Real-time acquisition of COD, BOD5, ammonia nitrogen and total phosphorus concentrations of influent water, using a spectrum analyzer for data acquisition, with a sampling frequency of not less than 1 time / minute; S2. Feedback signal acquisition: Multi-parameter sensors installed at the water inlet, reaction zone and outlet of the anoxic tank continuously collect liquid level, water temperature, total nitrogen, oxidation-reduction potential (ORP), mixed liquor suspended solids concentration (MLSS), dissolved oxygen (DO) and pH value; S3. Construct a dynamic compensation model, including a feedforward compensation module, a model predictive control module, and a feedback compensation module. Based on the changes in inlet load and water temperature, the theoretical carbon source dosage Q1 is calculated using the feedforward compensation formula: S4, LSTM neural network prediction and correction, including digital preprocessing, model training and deployment, and dynamic correction mechanism. Feedforward prediction values ​​and feedback parameters are input into the LSTM neural network. The network input layer contains 12-dimensional time series features, and the number of hidden layer nodes is determined by genetic algorithm. S5, closed-loop feedback control: execute the Q2 dosing instruction through the variable frequency dosing pump, collect the actual dosage Q3 in real time, and trigger the model parameter self-calibration based on the error function E = |Q2-Q3| / Q2. When E>5%, update the LSTM weight; S6. Multi-objective optimization: Based on carbon source consumption, effluent total nitrogen and energy consumption parameters, the NSGA-II algorithm is used to generate the optimal control strategy.

2. The method for intelligently adding carbon sources to sewage treatment according to claim 1, characterized in that: The feedforward compensation module in S3 monitors the change amplitude of the inlet flow rate, COD concentration and ammonia nitrogen concentration in real time. When the change of any parameter exceeds the preset threshold, the feedforward compensation module is activated, and the hysteresis effect of the impact load on nitrate nitrogen is calculated by the delay factor, and the feedforward carbon source dosage is output; The formula for calculating the impact load quantification in the feedforward compensation module is: Among them, Q 实 : Real-time water flow; COD 实 : Actual influent COD concentration; COD 基 : baseline COD concentration; K f : Feedforward coefficient, where the empirical value is 0.8-1.2; The feedforward compensation module also includes a delay correction mechanism, which introduces hydraulic retention time to calibrate the impact period of the impact load.

3. The method for intelligently adding carbon sources to sewage treatment according to claim 1, characterized in that: The model prediction control module in S3 includes constructing a carbon source addition algorithm model, the input parameters of which include the total nitrogen concentration of the influent, the influent flow rate, the internal recirculation ratio, historical water quality data and the nitrate nitrogen target value, predicting the carbon source addition demand through the LSTM model, dynamically updating the set value, and training the neural network through historical data to establish a multivariate nonlinear relationship; The formula for calculating the carbon source dosage in the model prediction module is: Wherein, HRT is hydraulic retention time; C 模型 The amount of carbon source added for model calculation; f is the coefficient, reflecting the comprehensive relationship between the amount of carbon source added and other independent variables in the brackets; TN 进水 is the total nitrogen in the influent; Q is the water inlet flow rate; The measured nitrate nitrogen concentration. The model prediction control module adopts a rolling optimization strategy, updates the prediction results every 5-10 minutes, and adjusts the carbon source set value in combination with feedback parameters.

4. The method for intelligently adding carbon sources to sewage treatment according to claim 1, characterized in that: The feedback compensation module in S3 includes monitoring the deviation between the nitrate nitrogen concentration at the end of the anoxic tank and the target value and the total nitrogen concentration of the effluent. When the total nitrogen concentration of the effluent exceeds the set range or the nitrate nitrogen deviation continues to exceed the limit, feedback compensation is triggered; the PID parameters in the feedback compensation module are adaptively adjusted through fuzzy logic.

5. The method for intelligently adding carbon sources for sewage treatment according to claim 1, characterized in that: The LSTM neural network feedback correction in S4 includes: Real-time deviation correction: Real-time deviation correction is to generate a deviation signal by comparing the actual value of the total nitrogen online instrument monitoring the effluent with the target value. The deviation value e(t) = total nitrogen target (TN 目标 )-actual total nitrogen value (TN 实际 ), deviation change rate Δe=e(t)-e(t-1); Fuzzy-PID control: Fuzzy-PID control is to input the deviation signal into the fuzzy controller, output the carbon source addition correction amount, and superimpose it on the feedforward-model prediction value.

6. The method for intelligently adding carbon sources for sewage treatment according to claim 1, characterized in that: In S4, data preprocessing includes: Eliminate abnormal instrument data: normalize multiple parameters to eliminate dimensionality effects, and select parameters with strong correlation as model input parameters; Model training and deployment steps: Use historical data and corresponding carbon source dosage to construct a supervised learning dataset, and use mean absolute error and effluent total nitrogen constraint as joint loss indicators for training; Dynamic correction mechanism: including feedforward-feedback fusion decision and anti-hysteresis design, among which, Feedforward-feedback fusion decision calculation method: Final dosage = LATM predicted value (model output) + (K p .e(t)+K d .Δe)(feedback correction term), where K p , K d is the proportional / differential coefficient output by the fuzzy controller; Anti-hysteresis design: The feedback correction amount is calibrated based on the current effluent total nitrogen deviation and the future predicted deviation trend.

Images