A model predictive control-based domestic waste incineration plant leachate aeration control method

Abstract

The application relates to the garbage treatment field and discloses a domestic waste incineration plant leachate aeration control method based on model predictive control, which comprises the following steps: acquiring real-time operation data of a domestic waste incineration plant leachate aeration tank; inputting the real-time operation data into a pre-trained double-time-scale prediction model to acquire a future state trajectory of the leachate aeration tank within a prediction time range; constructing a control objective function; performing sequence iterative solving on the control objective function based on the future state trajectory to determine a dissolved oxygen target curve; and generating a control instruction for driving the coordinated action of a main fan frequency converter and a branch aeration valve based on the control component of a current time step in the dissolved oxygen target curve. Through the technical scheme, gradient noise interference generated by mixed time series data is avoided when a model extracts features, and the accuracy of model prediction is improved.

Description

Technical Field

[0001] This application relates to the field of waste treatment, and in particular to a model-based predictive control method for aeration control of leachate from municipal solid waste incineration plants. Background Technology

[0002] Municipal solid waste incineration plants generate large amounts of high-concentration leachate during the waste fermentation process. This leachate contains high concentrations of chemical oxygen demand (COD) and ammonia nitrogen pollutants, requiring multi-stage biochemical treatment processes including equalization tanks, anaerobic reactors, and aerobic membrane bioreactors. In the aerobic treatment stage, the aeration system uses blowers to force aeration into the biochemical tank, maintaining the metabolic activities of aerobic microorganisms.

[0003] In related technologies, historical operating data of biological treatment ponds are collected, and a global water quality prediction model is trained using a recurrent neural network model or a long short-term memory network. A control objective function is constructed with the effluent water quality meeting the standard as the single optimization goal. The future aeration control sequence is solved by an optimization algorithm and then sent to the blower system for execution.

[0004] However, changes in dissolved oxygen concentration are rapid processes with a response time of seconds, while the degradation of pollutants by microorganisms and the evolution of sludge concentration are slow dynamic processes with a response time of several hours. Coupled with these two types of variables with drastically different properties into a prediction model with a single topology at the same time resolution, the model is easily affected by gradient noise generated by mixed time series data when extracting features, resulting in inaccurate prediction results at times.

[0005] Solving this technical problem is a technical challenge that needs to be overcome by those skilled in the art. Summary of the Invention

[0006] This application provides a model-based predictive control method for leachate aeration control in municipal solid waste incineration plants, which at least partially solves the technical problem that the model is easily affected by gradient noise generated by mixed time-series data when extracting features, resulting in inaccurate prediction results.

[0007] To achieve the above objectives, this application provides a model-based predictive control method for leachate aeration control in municipal solid waste incineration plants, comprising:

[0008] Real-time operating data of the leachate aeration tank of a municipal solid waste incineration plant is obtained; the real-time operating data includes water quality status parameters at a first sampling frequency and biochemical process parameters at a second sampling frequency; the first sampling frequency is greater than the second sampling frequency;

[0009] The real-time operating data is input into a pre-trained dual-timescale prediction model to obtain the future state trajectory of the leachate aeration tank within the prediction time range; the dual-timescale prediction model updates the state hidden vectors of the water quality state parameters and the biochemical process parameters through parallel fast recurrent neural network units and slow recurrent neural network units respectively.

[0010] Construct a control objective function; the control objective function includes a tracking error term for the target effluent quality, an energy consumption term for the aeration equipment, and a safety constraint term for sludge bulking.

[0011] Based on the future state trajectory, the control objective function is solved sequentially through iteration to determine the dissolved oxygen target curve;

[0012] Based on the control component of the current time step in the dissolved oxygen target curve, control commands are generated to drive the main fan frequency converter and the branch aeration valves to work in coordination.

[0013] The embodiments of this application have the following technical effects:

[0014] (1) By separating water quality state parameters and biochemical process parameters through a dual-timescale prediction model, the model avoids the interference of gradient noise generated by mixed time series data when extracting features, thus improving the accuracy of model prediction;

[0015] (2) Introducing a sludge bulking safety constraint ensures energy saving of the equipment and avoids the risk of sludge bulking caused by excessive aeration;

[0016] (3) By setting the action dead zone and judging the theoretical aeration increment, the frequent high-frequency jitter of the control command is suppressed, and the mechanical wear of the electromechanical equipment is reduced.

[0017] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an illustrative system framework of an embodiment of this application.

[0020] Figure 2 This is a flowchart illustrating the steps of a model predictive control-based aeration control method for leachate from a municipal solid waste incineration plant, as provided in an exemplary embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the protection scope of this application.

[0022] Reference Figure 1 The control method described in this application operates in a distributed control system for a leachate treatment station in a municipal solid waste incineration plant. The leachate biochemical treatment unit includes multiple membrane bioreactor aerobic tanks, a main gas pipeline network, blowers equipped with frequency converters, and electrically operated regulating valves that control the air intake of each aerobic tank.

[0023] This application provides a model-based predictive control method for aeration control of leachate from municipal solid waste incineration plants. Please refer to [link to relevant documentation]. Figure 2 The present application provides a model-based predictive control method for aeration control of leachate from a municipal solid waste incineration plant, which includes the following steps:

[0024] Step S101: Obtain real-time operating data of the leachate aeration tank of the municipal solid waste incineration plant; the real-time operating data includes water quality status parameters at a first sampling frequency and biochemical process parameters at a second sampling frequency; the first sampling frequency is greater than the second sampling frequency.

[0025] Step S102: Input the real-time operation data into the pre-trained dual-timescale prediction model to obtain the future state trajectory of the leachate aeration tank within the prediction time range; the dual-timescale prediction model updates the state hidden vectors of the water quality state parameters and the biochemical process parameters through parallel fast recurrent neural network units and slow recurrent neural network units respectively.

[0026] Step S103: Construct a control objective function; the control objective function includes a tracking error term for the target effluent quality, an energy consumption term for the aeration equipment, and a safety constraint term for sludge bulking.

[0027] Step S104: Based on the future state trajectory, perform sequential iterative solution on the control objective function to determine the dissolved oxygen target curve.

[0028] Step S105: Generate control commands for driving the main fan frequency converter and branch aeration valves to work together based on the control component of the current time step in the dissolved oxygen target curve.

[0029] Specifically, water quality parameters, including dissolved oxygen concentration, oxidation-reduction potential, and mixed liquor temperature, are collected in real time by water quality probes deployed on-site. Biochemical process parameters, including oxygen demand, ammonia nitrogen concentration, and suspended solids concentration, are collected hourly by water quality analyzers. These two types of data, with different attributes and sampling frequencies, are isolated and fed into parallel fast and slow RNN units for latent vector updates, respectively. This avoids gradient noise interference caused by mixed time-series data when the model extracts features, thereby improving the accuracy of model prediction.

[0030] In some embodiments, inputting the real-time operating data into a pre-trained dual-timescale prediction model to obtain the future state trajectory of the leachate aeration tank within the prediction time range includes:

[0031] The water quality state parameters are segmented into time series according to a preset first sliding window length to construct a first feature sequence characterizing transient gas-liquid mass transfer characteristics.

[0032] The biochemical process parameters are segmented into time series according to a preset second sliding window length to construct a second feature sequence characterizing the degradation features of the biochemical substrate, wherein the second sliding window length is greater than the first sliding window length;

[0033] The first feature sequence is input into the fast recurrent neural network unit step by step according to the time sequence, and the short time memory gating structure is used to extract the short time dependence and output the first state hidden vector representing the current transient mass transfer dynamics.

[0034] The second feature sequence is input into the slow recurrent neural network unit step by step according to the time sequence, and the long-term biochemical evolution law is extracted by the long-time memory gating structure and the second state hidden vector representing the current degradation dynamic is output.

[0035] The first state latent vector and the second state latent vector are concatenated to form a joint state latent vector, and the joint state latent vector is input into the multilayer perceptron decoding module;

[0036] The multilayer perceptron decoding module performs autoregressive prediction within the prediction time range, and outputs the state prediction value for each prediction time step in sequence; the state prediction values ​​are arranged along the time axis to form the future state trajectory.

[0037] Specifically, a shorter first sliding window is used for water quality state parameters to capture instantaneous and drastic fluctuations in dissolved oxygen; a longer second sliding window is used for biochemical process parameters to cover the long lag cycle of microbial reproduction and metabolism; a fast recurrent neural network handles mass transfer dynamics; a slow recurrent neural network handles degradation dynamics; after the latent vectors generated by the two are concatenated, the multilayer perceptron rolls forward predicting step by step within the set prediction time range to generate a complete state trajectory; the combination of the two ensures that the model will not fall into a local suboptimal solution due to drastic changes in a single dimension of indicators when facing high concentration leachate shocks, thus ensuring the matching of the predicted trajectory with the actual biochemical oxygen demand cycle.

[0038] In some embodiments, constructing a control objective function includes:

[0039] Obtain a preset target dissolved oxygen value and extract a predicted dissolved oxygen trajectory from the future state trajectory;

[0040] Based on the sum of squared errors between the predicted dissolved oxygen trajectory in the future state trajectory and the preset target dissolved oxygen value within the predicted time range, a tracking error term for water quality compliance deviation is constructed.

[0041] Based on the nonlinear relationship model between the operating frequency and output power of the main fan inverter and the predicted aeration volume, the expected power consumption is calculated, and an energy consumption item is constructed.

[0042] Acquire the current sludge floc image in the leachate aeration tank; extract the sludge volume index based on the current sludge floc image and determine the critical threshold for sludge bulking; construct the index penalty function and use the penalty function value as a sludge bulking safety constraint term.

[0043] The control objective function is generated by weighting and summing the tracking error term, energy consumption term, and sludge bulking safety constraint term based on preset weighting coefficients.

[0044] Specifically, this embodiment introduces industrial vision equipment to capture the morphology of sludge flocs in real time and calculate the volume index. During the simulation, it was found that when a certain set of aeration commands would cause the aeration intensity to exceed the critical threshold that would cause sludge to break up and expand, the exponential penalty function with the natural constant as the base would generate a large penalty value, which would increase the total objective function value after weighted summation. This would effectively avoid over-aeration while ensuring energy saving of the equipment.

[0045] In some embodiments, the control objective function is solved iteratively based on the future state trajectory to determine the dissolved oxygen target curve, including:

[0046] The adjustment step size and response delay time of the actuator are obtained, wherein the actuator includes the main fan frequency converter and the branch aeration valve;

[0047] A set of candidate control sequences consisting of multiple candidate control increments is generated within the prediction time range;

[0048] Substitute each candidate control sequence in the candidate control sequence set into the dual time-scale prediction model to generate multiple sets of candidate predicted state trajectories corresponding to each candidate control sequence.

[0049] Substitute each of the candidate predicted state trajectories into the control objective function to calculate the objective function value, select the sequence with the smallest objective function value as the optimal control sequence, and take the control quantity of the first time step of the optimal control sequence as the dissolved oxygen target curve.

[0050] Specifically, the above scheme avoids the common timeout or gradient vanishing problems in deep network optimization, and ensures that every control increment output by the algorithm is absolutely executable by the field equipment, greatly improving the execution stability of the industrial PLC controller.

[0051] In some embodiments, control commands for coordinating the operation of the main fan inverter and the branch aeration valves are generated based on the control component of the current time step in the dissolved oxygen target curve, including:

[0052] Based on the dissolved oxygen target curve and the actual dissolved oxygen feedback value of each individual pool, the valve opening adjustment signal of the gas distribution valve of each individual pool is obtained.

[0053] The total pressure requirement setting value of the entire aeration network is obtained based on the valve opening adjustment signal of each individual gas distribution valve and the current network environmental pressure parameters.

[0054] The frequency adjustment signal of the main fan inverter is obtained based on the main pipe pressure demand setting value and the feedback value of the actual main pipe pressure sensor.

[0055] Specifically, the main control loop monitors the dissolved oxygen level in the water tank and uses a PID algorithm to calculate the required valve opening for that tank. The secondary control loop collects the valve opening information from all water tanks and calculates the pressure required in the main gas pipeline to meet gas supply needs. When a sudden change in the influent load causes a drastic valve movement in a particular water tank, the pressure fluctuation is quickly absorbed and suppressed by the main pipeline pressure secondary loop, preventing global gas flow oscillations caused by single-point disturbances.

[0056] In some embodiments, before concatenating the first state latent vector and the second state latent vector into a joint state latent vector, the method further includes:

[0057] Calculate the ratio of the time step size of the first feature sequence to the time step size of the second feature sequence;

[0058] Based on the ratio, a transposed convolution kernel is constructed to upsample the second state latent vector in the time dimension to obtain a slow-scale state latent vector with the same time resolution as the first state latent vector.

[0059] The first state latent vector and the slow-scale state latent vector are multiplied by a preset query weight matrix, key weight matrix, and value weight matrix, respectively; the first state latent vector is linearly mapped to a query vector, and the slow-scale state latent vector is linearly mapped to a key vector and a value vector, respectively.

[0060] The inner product of the query vector and the transpose of the key vector is calculated and processed by a normalized exponential function to obtain the attention weight matrix; the attention weight matrix is ​​multiplied by the value vector to complete the cross-scale information fusion of the fast and slow scale state latent vectors.

[0061] The concatenation of the first state latent vector and the second state latent vector into a joint state latent vector includes: concatenating the first state latent vector after cross-scale information fusion with the slow-scale state latent vector in the channel dimension to form the joint state latent vector.

[0062] Specifically, within the model, the vectors output by the fast and slow RNN units have different lengths on the time axis. The slow signal is stretched to the same time granularity as the fast signal through transposed convolution. An attention mechanism is used to use the first state latent vector representing transient oxygen demand as the query vector to evaluate which feature segments in the upsampled slow-scale state latent vector can best support the current oxygen demand. Different weights are then assigned for weighted fusion, enabling the prediction model to allocate the physical proportion of fast and slow features under different load impacts.

[0063] In some embodiments, the training method for the dual-timescale prediction model includes:

[0064] Historical operation records of leachate aeration and water quality test data were collected, and a training sample set was constructed after outlier removal and normalization.

[0065] Construct biochemical consistency constraints to characterize the correlation between substrate degradation rate and oxygen consumption rate;

[0066] The mean squared error of the predicted trajectory of the training sample set in the dual-timescale prediction model is added to the biochemical consistency constraint term to form a composite loss function;

[0067] The gradient of the composite loss function is calculated by backpropagation along the time dimension, and the model weight parameters are iteratively updated by limiting the magnitude of the gradient vector.

[0068] In some embodiments, a set of candidate control sequences consisting of multiple candidate control increments is generated within the prediction time range, including:

[0069] Extract the optimal control sequence from the previous control cycle and shift the optimal control sequence forward by one control step along the time axis;

[0070] The forward-shifted sequence is used as the reference sequence for the current control cycle; a search radius is applied near the reference sequence according to the adjustment step size of the actuator to generate the candidate control sequence set;

[0071] Calculate the theoretical aeration increment before applying the search radius; if the theoretical aeration increment is less than the start / stop threshold set by the actuator dead zone, then lock the baseline sequence of the current dimension and skip generating candidate control increments for that dimension.

[0072] Specifically, valve stepper motors and fan frequency converters commonly suffer from dead zones due to mechanical friction, meaning that even minute changes in control commands result in no mechanical action. Before each round of rolling prediction, the theoretically required increase in aeration amplitude is calculated. Once this theoretical change is found to be less than the dead zone threshold, the search for the surrounding area of ​​that control dimension is immediately stopped, and the command from the previous step is directly reused. This effectively suppresses frequent high-frequency jitter in control commands, saves the controller's computing resources, reduces mechanical wear on electromechanical equipment, and extends the service life of the field actuators.

[0073] In some embodiments, before generating the control objective function by weighting and summing the tracking error term, energy consumption term, and sludge bulking safety constraint term based on preset weighting coefficients, the method further includes:

[0074] Calculate the actual tracking error of the effluent water quality at the current moment;

[0075] If the actual tracking error is greater than the preset water quality warning threshold, then the weighting coefficient of the tracking error term is increased and the weighting coefficient of the energy consumption term is decreased.

[0076] If the actual tracking error is less than or equal to the water quality warning threshold, then the weighting coefficient of the energy consumption term is increased and the weighting coefficient of the tracking error term is decreased.

[0077] In some embodiments, the method further includes:

[0078] Obtain the actual state observation value of the leachate aeration tank at the current time step;

[0079] Extract the predicted state value output by the dual-time-scale prediction model for the current time step in the previous control cycle and calculate the prediction deviation between the actual state observation value and the predicted state value.

[0080] A state compensation vector is generated based on the prediction deviation, and the state compensation vector is superimposed on the state latent vector at the current time step to perform benchmark compensation for the future state trajectory within the prediction time range.

[0081] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0082] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0083] The embodiments, implementation methods, and related technical features of this application can be combined and substituted for each other without conflict.

[0084] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this application without departing from the scope of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. A model-based predictive control method for aeration control of leachate from a municipal solid waste incineration plant, characterized in that, include: Obtain real-time operational data of the leachate aeration tank in a municipal solid waste incineration plant; The real-time operating data includes water quality status parameters at a first sampling frequency and biochemical process parameters at a second sampling frequency; the first sampling frequency is greater than the second sampling frequency. The real-time operating data is input into a pre-trained dual-timescale prediction model to obtain the future state trajectory of the leachate aeration tank within the prediction time range. The dual-timescale prediction model updates the state hidden vectors of the water quality state parameters and the biochemical process parameters through parallel fast recurrent neural network units and slow recurrent neural network units, respectively. Construct a control objective function; the control objective function includes a tracking error term for the target effluent quality, an energy consumption term for the aeration equipment, and a safety constraint term for sludge bulking. Based on the future state trajectory, the control objective function is solved sequentially through iteration to determine the dissolved oxygen target curve; Based on the control component of the current time step in the dissolved oxygen target curve, control commands are generated to drive the main fan frequency converter and the branch aeration valves to work in coordination.

2. The method according to claim 1, characterized in that, The real-time operating data is input into a pre-trained dual-timescale prediction model to obtain the future state trajectory of the leachate aeration tank within the predicted time range, including: The water quality state parameters are segmented into time series according to a preset first sliding window length to construct a first feature sequence characterizing transient gas-liquid mass transfer characteristics. The biochemical process parameters are segmented into time series according to a preset second sliding window length to construct a second feature sequence characterizing the degradation features of the biochemical substrate, wherein the second sliding window length is greater than the first sliding window length; The first feature sequence is input into the fast recurrent neural network unit step by step according to the time sequence, and the short time memory gating structure is used to extract the short time dependence and output the first state hidden vector representing the current transient mass transfer dynamics. The second feature sequence is input into the slow recurrent neural network unit step by step according to the time sequence, and the long-term biochemical evolution law is extracted by the long-time memory gating structure and the second state hidden vector representing the current degradation dynamic is output. The first state latent vector and the second state latent vector are concatenated to form a joint state latent vector, and the joint state latent vector is input into the multilayer perceptron decoding module; The multilayer perceptron decoding module performs autoregressive prediction within the prediction time range, and sequentially outputs the system state prediction values ​​for each prediction time step; the system state prediction values ​​are arranged along the time axis to form the future state trajectory.

3. The method according to claim 2, characterized in that, Construct the control objective function, including: Obtain a preset target dissolved oxygen value and extract a predicted dissolved oxygen trajectory from the future state trajectory; Based on the sum of squared errors between the predicted dissolved oxygen trajectory in the future state trajectory and the preset target dissolved oxygen value within the predicted time range, a tracking error term for water quality compliance deviation is constructed. Based on the nonlinear relationship model between the operating frequency and output power of the main fan inverter and the predicted aeration volume, the expected power consumption is calculated, and an energy consumption item is constructed. Acquire the current sludge floc image in the leachate aeration tank; extract the sludge volume index based on the current sludge floc image and determine the critical threshold for sludge bulking; construct the index penalty function and use the penalty function value as a sludge bulking safety constraint term. The control objective function is generated by weighting and summing the tracking error term, energy consumption term, and sludge bulking safety constraint term based on preset weighting coefficients.

4. The method according to claim 3, characterized in that, Based on the future state trajectory, the control objective function is solved iteratively to determine the dissolved oxygen target curve, including: The adjustment step size and response delay time of the actuator are obtained, wherein the actuator includes the main fan frequency converter and the branch aeration valve; A set of candidate control sequences consisting of multiple candidate control increments is generated within the prediction time range; Substitute each candidate control sequence in the candidate control sequence set into the dual time-scale prediction model to generate multiple sets of candidate predicted state trajectories corresponding to each candidate control sequence. Substitute each of the candidate predicted state trajectories into the control objective function to calculate the objective function value, select the sequence with the smallest objective function value as the optimal control sequence, and take the control quantity of the first time step of the optimal control sequence as the dissolved oxygen target curve.

5. The method according to claim 4, characterized in that, Based on the control component of the current time step in the dissolved oxygen target curve, control commands are generated for coordinating the operation of the main fan inverter and the branch aeration valves, including: Based on the dissolved oxygen target curve and the actual dissolved oxygen feedback value of each individual pool, the valve opening adjustment signal of the gas distribution valve of each individual pool is obtained. The total pressure requirement setting value of the entire aeration network is obtained based on the valve opening adjustment signal of each individual gas distribution valve and the current network environmental pressure parameters. The frequency adjustment signal of the main fan inverter is obtained based on the main pipe pressure demand setting value and the feedback value of the actual main pipe pressure sensor.

6. The method according to claim 5, characterized in that, Before concatenating the first state latent vector and the second state latent vector into a joint state latent vector, the method further includes: Calculate the ratio of the time step size of the first feature sequence to the time step size of the second feature sequence; Based on the ratio, a transposed convolution kernel is constructed to upsample the second state latent vector in the time dimension to obtain a slow-scale state latent vector with the same time resolution as the first state latent vector. The first state latent vector and the slow-scale state latent vector are multiplied by a preset query weight matrix, key weight matrix, and value weight matrix, respectively; the first state latent vector is linearly mapped to a query vector, and the slow-scale state latent vector is linearly mapped to a key vector and a value vector, respectively. The inner product of the query vector and the transpose of the key vector is calculated and processed by a normalized exponential function to obtain the attention weight matrix; the attention weight matrix is ​​multiplied by the value vector to complete the cross-scale information fusion of the fast and slow scale state latent vectors. The concatenation of the first state latent vector and the second state latent vector into a joint state latent vector includes: concatenating the first state latent vector after cross-scale information fusion with the slow-scale state latent vector in the channel dimension to form the joint state latent vector.

7. The method according to claim 6, characterized in that, The training method for the dual-timescale prediction model includes: Historical operation records of leachate aeration and water quality test data were collected, and a training sample set was constructed after outlier removal and normalization. Construct biochemical consistency constraints to characterize the correlation between substrate degradation rate and oxygen consumption rate; The mean squared error of the predicted trajectory of the training sample set in the dual-timescale prediction model is added to the biochemical consistency constraint term to form a composite loss function; The gradient of the composite loss function is calculated by backpropagation along the time dimension, and the model weight parameters are iteratively updated by limiting the magnitude of the gradient vector.

8. The method according to claim 7, characterized in that, Within the prediction time range, a set of candidate control sequences consisting of multiple candidate control increments is generated, including: Extract the optimal control sequence from the previous control cycle and shift the optimal control sequence forward by one control step along the time axis; The shifted sequence is used as the reference sequence for the current control cycle; The set of candidate control sequences is generated by applying a search radius near the reference sequence based on the adjustment step size of the actuator. Calculate the theoretical aeration increment before applying the search radius; if the theoretical aeration increment is less than the start / stop threshold set by the actuator dead zone, then lock the baseline sequence of the current dimension and skip generating candidate control increments for the corresponding dimension.

9. The method according to claim 8, characterized in that, Before generating the control objective function by weighting and summing the tracking error term, energy consumption term, and sludge bulking safety constraint term based on preset weighting coefficients, the method further includes: Calculate the actual tracking error of the effluent water quality at the current moment; If the actual tracking error is greater than the preset water quality warning threshold, then the weighting coefficient of the tracking error term is increased and the weighting coefficient of the energy consumption term is decreased. If the actual tracking error is less than or equal to the water quality warning threshold, then the weighting coefficient of the energy consumption term is increased and the weighting coefficient of the tracking error term is decreased.

10. The method according to claim 9, characterized in that, The method further includes: Obtain the actual state observation value of the leachate aeration tank at the current time step; Extract the predicted state value output by the dual-time-scale prediction model for the current time step in the previous control cycle and calculate the prediction deviation between the actual state observation value and the predicted state value. A state compensation vector is generated based on the prediction deviation, and the state compensation vector is superimposed on the state latent vector at the current time step to perform benchmark compensation for the future state trajectory within the prediction time range.

Images