Wastewater treatment control method and device based on working condition adaptive prediction time domain, equipment and medium

Abstract

The application discloses a sewage treatment control method and device based on a working condition adaptive prediction time domain, equipment and a medium, relates to the technical field of intelligent sewage treatment control, constructs a state vector based on the operation data of the sewage treatment process, inputs the state vector into a time domain decision network, and outputs a rolling time domain length prediction value; an extended state vector is constructed, the extended state vector is input into a strategy network, and a control action is output; the extended state vector and the control action are input into a world model, a rolling prediction is performed, a state vector prediction sequence and a reward prediction sequence are obtained; based on the actual reward value, the reward prediction sequence, the state vector prediction sequence and the cumulative return prediction value, the model parameters in the world model, the network parameters in the strategy network and the network parameters in the time domain decision network are updated respectively; a target control action is determined, the sewage treatment process is controlled by using the sewage treatment equipment, and the adaptability and control effect of the complex industrial process control system are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for wastewater treatment, and in particular to a wastewater treatment control method, apparatus, equipment, and medium based on adaptive prediction time domain under operating conditions. Background Technology

[0002] Currently, reinforcement learning control methods based on world models (models that predict future states based on environmental dynamics) can make rolling predictions of future states. However, existing world model reinforcement learning methods typically use a fixed prediction time domain for future state prediction. For example, whether the system is in a stable, fluctuating, or abnormal operating condition, a pre-set prediction length is still used for rolling predictions. This fixed prediction time domain mechanism does not fully consider the characteristics of the current operating condition and changes in the operating state, resulting in a lack of dynamic coordination between the prediction module and the control decision module.

[0003] When the system is operating smoothly, the future state changes relatively gradually, and a shorter prediction time domain is sufficient to meet control requirements. However, using a longer prediction time domain for rolling simulations increases the number of inferences in the world model and the computational burden of reinforcement learning training and online decision-making, reducing the efficiency of the control system. Under complex conditions such as fluctuating influent load, equipment switching, and abnormal shocks, a shorter prediction time domain may not adequately reflect the long-term impact of control actions on the future system state, leading to short-sighted decisions in the policy network and a lack of foresight in control actions.

[0004] Furthermore, traditional methods typically treat the prediction time domain as a fixed hyperparameter pre-set by human experience. The policy network cannot access information about the current prediction time domain during the decision-making process, thus failing to adjust the control strategy according to changes in the prediction range. The length of the prediction time domain also struggles to participate in the online optimization process; the system cannot dynamically adjust the prediction depth based on historical performance, making it difficult to formulate optimal prediction strategies for different operating conditions.

[0005] As can be seen from the above, how to achieve joint decision-making between the predictive time domain and the control strategy, and improve the dynamic adaptability and control effect of complex industrial process control systems, is a problem that needs to be solved in this field. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a wastewater treatment control method, apparatus, equipment, and medium based on adaptive predictive time domain, which can realize joint decision-making of predictive time domain and control strategy, thereby improving the dynamic adaptability and control effect of complex industrial process control systems. The specific solution is as follows: In a first aspect, this application discloses a wastewater treatment control method based on adaptive predictive time-domain operation, comprising: The operation data of the wastewater treatment process is acquired, a state vector is constructed based on the operation data, and the state vector is input into a time-domain decision network to output a rolling time-domain length prediction value. Based on the state vector and the predicted value of the rolling time domain length, an extended state vector is constructed, and the extended state vector is input into the policy network to output control actions. The extended state vector and the control action are input into the world model, and the rolling prediction is performed with the rolling time domain length prediction value as the rolling step number to obtain the state vector prediction sequence and the reward prediction sequence. The actual reward value is calculated using a multi-objective reward function, and the cumulative return prediction value is determined based on the reward prediction sequence. Based on the actual reward value, the reward prediction sequence, the state vector prediction sequence, and the cumulative reward prediction value, the model parameters in the world model, the network parameters in the policy network, and the network parameters in the time domain decision network are updated respectively, and it is determined whether the preset termination condition has been triggered. If the preset termination condition is not triggered at present, the process of acquiring running data is repeated until the preset termination condition is triggered and the target control action is obtained; the target control action is the control action output by the updated policy network. The target control action is sent to the wastewater treatment equipment so that the wastewater treatment equipment can control the wastewater treatment process based on the target control action.

[0007] Optionally, acquiring operational data of the wastewater treatment process and constructing a state vector based on the operational data includes: The wastewater treatment process operation data is acquired from a data acquisition and monitoring control system, a programmable logic controller system, or a distributed control system. The operation data includes dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, chemical oxygen demand, mixed liquor suspended solids concentration, influent flow rate, water temperature, and oxidation-reduction potential. The running data is normalized to obtain the processed running data; A state vector is constructed based on the processed running data.

[0008] Optionally, inputting the state vector into a time-domain decision network to output a rolling time-domain length prediction value includes: Construct a time-domain decision network based on hierarchical neural networks; The state vector is input into a time-domain decision network so that the time-domain decision network can be used to analyze the state vector to determine the current operating condition complexity of the wastewater treatment process, and based on the operating condition complexity, output a discrete rolling time-domain length prediction value.

[0009] Optionally, constructing the extended state vector based on the state vector and the predicted value of the rolling time domain length includes: Calculate the ratio between the predicted value of the scrolling time domain length and the preset maximum scrolling length, and use the ratio as the normalized scrolling length; The normalized scroll length is concatenated with the state vector to obtain the extended state vector.

[0010] Optionally, the step of inputting the extended state vector and the control action into the world model, and performing rolling prediction with the rolling time-domain length prediction value as the rolling step number to obtain the state vector prediction sequence and the reward prediction sequence includes: The extended state vector and control action are input into the world model to output the predicted state vector and reward for the next time step. Based on the predicted state vector and reward value for the next moment, and using the predicted rolling time domain length as the number of rolling steps, continue to perform rolling prediction until the number of rolling predictions reaches the predicted rolling time domain length, then stop rolling prediction to obtain the predicted state vector sequence and the predicted reward sequence.

[0011] Optionally, the step of calculating the actual reward value using a multi-objective reward function and determining the cumulative return prediction value based on the reward prediction sequence includes: Obtain dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, and aeration energy consumption from the operational data; Based on the dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, and aeration energy consumption, the dissolved oxygen control error item, ammonia nitrogen exceeding the standard item, total nitrogen exceeding the standard item, and aeration energy consumption item are determined respectively. Using a multi-objective reward function, the dissolved oxygen control error term, the ammonia nitrogen exceedance term, the total nitrogen exceedance term, and the aeration energy consumption term are weighted and summed to obtain the actual reward value; Based on the discount factor, the reward prediction values ​​in the reward prediction sequence are weighted and summed to obtain the cumulative return prediction value.

[0012] Optionally, updating the model parameters in the world model, the network parameters in the policy network, and the network parameters in the time-domain decision network based on the actual reward value, the reward prediction sequence, the state vector prediction sequence, and the cumulative reward prediction value respectively includes: A world model loss function is constructed based on the state vector, the state vector prediction sequence, the actual reward value, and the reward prediction sequence. The model parameters in the world model are updated with the goal of minimizing the loss function of the world model. With the goal of maximizing the cumulative return prediction, the network parameters in the policy network are updated using the policy gradient; The network parameters in the time-domain decision network are updated with the goal of maximizing the cumulative return prediction value.

[0013] Secondly, this application discloses a wastewater treatment control device based on adaptive predictive time domain of operating conditions, comprising: The rolling time-domain length prediction value determination module is used to acquire the operation data of the sewage treatment process, construct a state vector based on the operation data, and input the state vector into the time-domain decision network to output the rolling time-domain length prediction value. The control action determination module is used to construct an extended state vector based on the state vector and the predicted value of the rolling time domain length, and input the extended state vector into the policy network to output the control action; The rolling prediction module is used to input the extended state vector and the control action into the world model, and perform rolling prediction with the rolling time domain length prediction value as the rolling step number to obtain the state vector prediction sequence and the reward prediction sequence. The cumulative return prediction value determination module is used to calculate the actual reward value using a multi-objective reward function and determine the cumulative return prediction value based on the reward prediction sequence. The update module is used to update the model parameters in the world model, the network parameters in the policy network, and the network parameters in the time domain decision network based on the actual reward value, the reward prediction sequence, the state vector prediction sequence, and the cumulative reward prediction value, and to determine whether the preset termination condition has been triggered. The target control action determination module is used to repeatedly execute the process of acquiring running data if the preset termination condition is not triggered, until the preset termination condition is triggered, and then obtain the target control action; the target control action is the control action output by the updated policy network. The wastewater treatment process control module is used to send the target control action to the wastewater treatment equipment so that the wastewater treatment equipment can control the wastewater treatment process based on the target control action.

[0014] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned wastewater treatment control method based on adaptive predictive time domain of operating conditions.

[0015] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed wastewater treatment control method based on adaptive prediction time domain under operating conditions.

[0016] Beneficial Effects: This application provides a wastewater treatment control method based on adaptive prediction time domain, including acquiring operational data of the wastewater treatment process, constructing a state vector based on the operational data, and inputting the state vector into a time-domain decision network to output a rolling time-domain length prediction value. Therefore, it can map the real-time operating conditions of the wastewater treatment process into a state expression that can be used for learning and decision-making, avoiding the problem that a fixed prediction time domain cannot adapt to different operating conditions. Based on the state vector and the rolling time-domain length prediction value, an extended state vector is constructed and input into a policy network to output control actions. Therefore, the policy network can simultaneously perceive the current state and prediction time-domain information when generating control actions, improving the adaptability of the control strategy to stable, fluctuating, and abnormal operating conditions. The extended state vector and control actions are input into a world model, and rolling prediction is performed using the rolling time-domain length prediction value as the rolling step number to obtain a state vector prediction sequence and a reward prediction sequence. Therefore, it can reduce unnecessary inference calculations when the operating conditions are stable and extend the prediction range during load shocks or abnormal disturbances. This system enhances the ability to predict future state changes. It calculates actual reward values ​​using a multi-objective reward function and determines cumulative reward prediction values ​​based on the reward prediction sequence. Based on the actual reward values, reward prediction sequences, state vector prediction sequences, and cumulative reward prediction values, it updates the model parameters in the world model, network parameters in the policy network, and network parameters in the time-domain decision network, respectively, and determines whether a preset termination condition has been triggered. Therefore, it can jointly update the world model, policy network, and time-domain decision network based on real operational feedback and prediction results, enabling collaborative optimization of prediction time-domain length, state prediction capability, and control strategy within the same closed-loop learning structure. If the preset termination condition has not been triggered, the process of acquiring operational data is repeated until the preset termination condition is triggered, resulting in the target control action. This improves the continuity of the online optimization process. The target control action is then distributed to the wastewater treatment equipment, enabling the equipment to control the wastewater treatment process based on the target control action. This improves the control stability, effluent compliance rate, and operational economy of the wastewater treatment process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart of a wastewater treatment control method based on adaptive prediction in the time domain under operating conditions disclosed in this application; Figure 2This is a detailed flowchart of a wastewater treatment control based on adaptive prediction in the time domain under operating conditions disclosed in this application; Figure 3 This is a schematic diagram of a wastewater treatment control device based on adaptive prediction time domain under operating conditions disclosed in this application; Figure 4 This is a schematic diagram of the structure of an electronic device disclosed in this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Currently, reinforcement learning control methods based on world models can make rolling predictions of future states. However, existing world model reinforcement learning methods typically use a fixed prediction time domain for future state extrapolation. For example, whether the system is in a stable, fluctuating, or abnormal operating condition, a pre-set prediction length is still used for rolling prediction. This fixed prediction time domain mechanism does not fully consider the characteristics of the current operating condition and changes in the operating state, resulting in a lack of dynamic coordination between the prediction module and the control decision module. When the system is operating smoothly, the trend of future state changes is relatively gradual, and a shorter prediction time domain can meet the control requirements. If a longer prediction time domain is still used for rolling extrapolation, it will increase the number of inferences of the world model and increase the computational burden of reinforcement learning training and online decision-making, reducing the operating efficiency of the control system. Under complex operating conditions such as fluctuating influent load, equipment switching, and abnormal shocks, a shorter prediction time domain may not be able to fully reflect the long-term impact of control actions on the future system state, and the policy network is prone to making short-sighted decisions, resulting in a lack of foresight in control actions. Furthermore, traditional methods typically treat the prediction time domain as a fixed hyperparameter pre-set by human experience. The policy network cannot access current prediction time domain information during the decision-making process, thus failing to adjust the control strategy according to changes in the prediction range. The length of the prediction time domain also struggles to participate in the online optimization process; the system cannot dynamically adjust the prediction depth based on historical operating results, making it difficult to formulate optimal prediction strategies for different operating conditions. Therefore, how to achieve joint decision-making between the prediction time domain and the control strategy to improve the dynamic adaptability and control performance of complex industrial process control systems is a problem that needs to be solved in this field.

[0021] See Figure 1 As shown in the figure, this invention discloses a wastewater treatment control method based on adaptive prediction time domain under operating conditions, which may specifically include: Step S11: Obtain the operation data of the sewage treatment process, construct a state vector based on the operation data, and input the state vector into the time domain decision network to output the rolling time domain length prediction value.

[0022] In this embodiment, operational data of the wastewater treatment process is acquired from a data acquisition and monitoring control system, a programmable logic controller system, or a distributed control system; the operational data is normalized to obtain processed operational data; a state vector is constructed based on the processed operational data; a time-domain decision network is constructed based on a hierarchical neural network; the state vector is input into the time-domain decision network so that the time-domain decision network can be used to analyze the state vector to determine the current operational complexity of the wastewater treatment process, and based on the operational complexity, a discrete rolling time-domain length prediction value is output; the operational data includes dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, chemical oxygen demand, mixed liquor suspended solids concentration, influent flow rate, water temperature, and oxidation-reduction potential.

[0023] Specifically, online monitoring data for the wastewater treatment process can come from on-site SCADA (Supervisory Control and Data Acquisition) systems, PLC (Programmable Logic Controller) systems, or DCS (Distributed Control System) systems.

[0024] Runtime data may include (Dissolved oxygen) (ammonia nitrogen concentration) (Total nitrogen concentration) (Chemical Oxygen Demand) (Concentration of suspended solids in the mixture) (Inlet flow rate) (Water temperature) and (Oxidation-reduction potential). The above operational data reflects information such as water quality status, process load, and environmental conditions during wastewater treatment. By normalizing the operational data, the influence of different dimensions and numerical ranges on subsequent network inputs can be reduced, enabling the state vector to serve as a unified state input for the reinforcement learning system.

[0025] In this embodiment, the state vector It can be represented as: ; In this embodiment, the temporal decision network can output a predicted rolling temporal length based on the current state vector. This predicted rolling temporal length represents the number of steps the world model needs to perform subsequent rolling predictions. Temporal Decision Network H - Net The input-output relationship can be represented as: ; in, H For time-domain decision networks, This represents the predicted scroll length at the current moment.

[0026] The time-domain decision network can dynamically output the prediction time domain length based on the complexity of the operating conditions. When the wastewater treatment process is under stable operating conditions, it outputs a shorter prediction time domain; when the wastewater treatment process experiences load fluctuations, influent impacts, or abnormal operating conditions, it outputs a longer prediction time domain.

[0027] In this embodiment, the predicted rolling time-domain length can be in discrete form, enabling the time-domain decision network to determine the currently applicable rolling prediction step from multiple candidate rolling step numbers. The discrete form of the predicted rolling time-domain length can be expressed as: .

[0028] Step S12: Based on the state vector and the predicted value of the rolling time domain length, construct an extended state vector and input the extended state vector into the policy network to output control actions.

[0029] In this embodiment, the ratio between the predicted value of the scrolling time domain length and the preset maximum scrolling length is calculated, and the ratio is used as the normalized scrolling length; the normalized scrolling length is concatenated with the state vector to obtain the extended state vector.

[0030] To enable the policy network to perceive the prediction depth that the world model will adopt when outputting control actions, this application extends the rolling time-domain length prediction value as part of the state variable to the reinforcement learning state space. Since there is a difference in dimensions between the rolling time-domain length prediction value and operational data such as water quality, flow rate, and temperature, the rolling time-domain length prediction value can be normalized first.

[0031] Normalized roll length It can be represented as: ; in, The maximum value among the discrete rolling time domain length prediction values ​​or the pre-set maximum number of rolling prediction steps.

[0032] Extended state vector It can be represented as: ; The extended state vector is obtained by concatenating the state vector and the normalized scroll length.

[0033] Through this extended state vector, the policy network can not only obtain the current operating state of the wastewater treatment process when making decisions, but also obtain the prediction time-domain information of the world model under the current operating conditions for subsequent rolling predictions.

[0034] In this embodiment, the policy network Control actions can be output based on extended state vectors. Its input-output relationship can be expressed as: ; .

[0035] In wastewater treatment scenarios, control actions may include aeration rate / blower frequency. Aeration valve opening degree, internal reflux ratio External reflux ratio Carbon source addition amount And the dosage of the medicine.

[0036] Since the prediction time domain has already been input into the policy network as part of the extended state, the policy network can automatically adjust its control behavior based on the future prediction depth.

[0037] Step S13: Input the extended state vector and the control action into the world model, and use the rolling time domain length prediction value as the rolling step number to perform rolling prediction, thereby obtaining the state vector prediction sequence and the reward prediction sequence.

[0038] In this embodiment, the extended state vector and control action are input into the world model to output the state vector prediction value and reward prediction value for the next moment. Based on the state vector prediction value and reward prediction value for the next moment, and with the rolling time domain length prediction value as the rolling step number, the rolling prediction continues to be executed until the number of rolling predictions reaches the rolling time domain length prediction value. Then, the rolling prediction stops, and the state vector prediction sequence and reward prediction sequence are obtained.

[0039] A world model can be represented as an environmental dynamic model used to predict future states and rewards based on an extended state vector and control actions. Its inputs are the extended state vector and control actions, and its outputs are the predicted state vector and reward values ​​for the next time step. The input-output relationship of the world model can be represented as: ; in, M For the world model, The state of the wastewater treatment process at the next moment, as predicted by the world model. The reward prediction value is the reward corresponding to the current control action predicted by the world model.

[0040] Specifically, at the current moment, the world model first outputs the predicted state vector for the next moment and the predicted current reward based on the extended state vector and control actions. Then, the predicted state for the next moment can be used as the basis for subsequent rolling predictions, and the next prediction can continue. This process is repeated according to the number of rolling steps determined by the predicted rolling time-domain length, until the number of rolling predictions reaches the predicted rolling time-domain length.

[0041] Execution using world model The step-by-step rolling prediction state vector prediction sequence can be represented as: ; in, for Based on the rolling prediction of the wastewater treatment process status at the next moment, that is, from the current moment... t The wastewater treatment process status corresponding to the second rolling prediction step, From the current moment t Starting from the first The wastewater treatment process status corresponding to each rolling prediction step.

[0042] The reward prediction sequence can be represented as: ; in, The reward prediction value is the value predicted by the world model in the next rolling prediction step. That is, after predicting the state at the next time step, the reward prediction value is obtained based on the predicted state and subsequent control process. World model predicted values ​​based on rolling time domain length In continuous rolling forecasting, the predicted reward value corresponding to the last rolling forecast step.

[0043] Understandably, fixed-time-domain prediction methods perform rolling predictions with the same time values ​​under all operating conditions, making it difficult to simultaneously satisfy computational efficiency under stable operating conditions and prediction depth under complex operating conditions. This application controls the stopping condition of the world model's rolling predictions by using the predicted value of the rolling time-domain length, enabling the prediction process to dynamically scale according to changes in operating conditions, thereby making the future state assessment results more suitable for the current operating state of the wastewater treatment process.

[0044] Step S14: Calculate the actual reward value using a multi-objective reward function, and determine the cumulative return prediction value based on the reward prediction sequence.

[0045] In this embodiment, dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, and aeration energy consumption are obtained from operational data. Based on the dissolved oxygen, ammonia nitrogen, total nitrogen concentration, and aeration energy consumption, dissolved oxygen control error, ammonia nitrogen exceedance, total nitrogen exceedance, and aeration energy consumption are determined respectively. Using a multi-objective reward function, the dissolved oxygen control error, ammonia nitrogen exceedance, total nitrogen exceedance, and aeration energy consumption are weighted and summed to obtain the actual reward value. Based on a discount factor, the reward prediction values ​​in the reward prediction sequence are weighted and summed to obtain the cumulative reward prediction value.

[0046] In this embodiment, the multi-objective reward function can simultaneously consider both effluent water quality compliance and energy-saving operation. The dissolved oxygen control error term is used to constrain the deviation between the actual dissolved oxygen value and the set value; the ammonia nitrogen exceedance term is used to constrain situations where the ammonia nitrogen concentration exceeds the limit; the total nitrogen exceedance term is used to constrain situations where the total nitrogen concentration exceeds the limit; and the aeration energy consumption term is used to constrain the energy consumption generated during the aeration process.

[0047] The multi-objective reward function can be expressed as: ; in, This is the actual reward value. The weighting coefficients corresponding to the dissolved oxygen control error term. The dissolved oxygen at the current moment, To preset the target dissolved oxygen value, The weighting coefficients for the ammonia nitrogen exceeding the standard are as follows: The ammonia nitrogen concentration at the current moment. To preset the ammonia nitrogen concentration limit, The weighting coefficients for the total nitrogen exceeding the standard are as follows: The total nitrogen concentration at the current moment. To preset the total nitrogen concentration limit, The weighting coefficient corresponding to the aeration energy consumption item. This represents the aeration energy consumption at the current moment.

[0048] A multi-objective reward function is used to achieve comprehensive optimization of effluent quality and energy-saving operation.

[0049] In this embodiment, the cumulative return prediction value can be obtained by weighted summing of the reward prediction values ​​in the reward prediction sequence based on the discount factor. It can be represented as: ; in, This is a discount factor used to represent the degree of decay of future rewards relative to current rewards.

[0050] The cumulative return forecast value reflects the long-term benefits of the current control action over a future forecast period by applying a discounted weighted average of the reward forecast value within the current rolling forecast time domain.

[0051] This application determines the cumulative return prediction value based on the reward prediction sequence, enabling the policy network and time-domain decision network to be optimized around long-term returns, thereby improving the foresight and stability of control actions under complex operating conditions.

[0052] Step S15: Update the model parameters in the world model, the network parameters in the policy network, and the network parameters in the time-domain decision network based on the actual reward value, the reward prediction sequence, the state vector prediction sequence, and the cumulative reward prediction value, and determine whether the preset termination condition has been triggered.

[0053] In this embodiment, a world model loss function is constructed based on the state vector, the state vector prediction sequence, the actual reward value, and the reward prediction sequence; the model parameters in the world model are updated with the goal of minimizing the world model loss function; the network parameters in the policy network are updated using the policy gradient with the goal of maximizing the cumulative reward prediction value; and the network parameters in the time-domain decision network are updated with the goal of maximizing the cumulative reward prediction value.

[0054] In this embodiment, the world model, policy network, and temporal decision network constitute a closed-loop learning structure for joint optimization. The world model is used to predict future states and rewards, the policy network is used to output control actions, and the temporal decision network is used to output the predicted value of the rolling time-domain length. The three are interconnected through the actual reward value, the reward prediction sequence, the state vector prediction sequence, and the cumulative reward prediction value.

[0055] In this embodiment, the world model loss function can be constructed based on the deviation between the actual next-time state and the predicted state vector, as well as the deviation between the actual reward value and the predicted reward value. It can be represented as: ; Updating the model parameters in the world model with the goal of minimizing the world model loss function can improve the world model's ability to express the dynamic changes in the wastewater treatment process.

[0056] In this embodiment, the policy network parameters can be updated based on the policy gradient, with the update direction aiming to maximize the cumulative return prediction value. It can be represented as: ; in, This represents the policy gradient.

[0057] The parameters of the time-domain decision network can also be updated with the objective of maximizing the cumulative return prediction value. It can be represented as: .

[0058] In this embodiment, determining whether a preset termination condition has been triggered can be understood as determining whether the current cyclic learning and online control process has reached a preset stop requirement. If the preset termination condition has not been triggered, new running data is acquired and subsequent steps are executed cyclically; if the preset termination condition is triggered, the control action output by the updated policy network is determined as the target control action.

[0059] Through the above update process, the world model can correct its predictive ability of dynamic changes in the wastewater treatment process based on real operational data, the policy network can optimize control actions based on cumulative return predictions, and the time-domain decision network can optimize the rolling prediction steps based on cumulative return predictions. Thus, the prediction time-domain length and the control strategy are jointly and adaptively adjusted within the same closed-loop framework.

[0060] Step S16: If the preset termination condition is not triggered, the process of acquiring running data is repeated until the preset termination condition is triggered and the target control action is obtained; the target control action is the control action output by the updated policy network.

[0061] In this embodiment, if the preset termination condition is not triggered, it means that the control system still needs to continue state construction, rolling time-domain length prediction, control action generation, world model rolling prediction, reward calculation, and network parameter update based on new operating data. At this time, the process of obtaining operating data of the sewage treatment process can be returned to, and the loop can continue according to the set control cycle.

[0062] Specifically, within each control cycle, the system can obtain the latest operational data from the field system and construct a new state vector based on the latest operational data. The time-domain decision network outputs a new rolling time-domain length prediction value based on the new state vector; the policy network outputs a new control action based on the new extended state vector; the world model executes a new rolling prediction based on the new extended state vector and control action; the parameter update process continues based on the new actual reward value, reward prediction sequence, state vector prediction sequence, and cumulative reward prediction value.

[0063] When the preset termination condition is triggered, the control action output by the updated policy network in the current extended state can be determined as the target control action. Since this target control action is obtained after joint updating of the world model, policy network, and time-domain decision network, it can reflect the current operating conditions of the wastewater treatment process and predict the time-domain optimization results.

[0064] This embodiment enables the control system to continuously adjust the prediction depth according to changes in operating conditions during operation by repeatedly executing the above process, and allows the strategy network to generate control actions based on the extended state. Compared with methods that rely solely on a fixed prediction time domain, the cyclic learning structure can reduce computational overhead under stable operating conditions and enhance the ability to extrapolate future states under complex operating conditions, thereby improving the adaptability of the online control process.

[0065] Step S17: Send the target control action to the wastewater treatment equipment so that the wastewater treatment equipment can control the wastewater treatment process based on the target control action.

[0066] In this embodiment, the target control action can be sent to the wastewater treatment field equipment. The wastewater treatment equipment may include blowers, aeration valves, return pumps, and a dosing system. Control actions may include blower frequency, aeration valve opening degree, internal return ratio, external return ratio, carbon source dosage, and reagent dosage. By sending the target control action to the aforementioned field equipment, the wastewater treatment equipment can execute control based on the updated strategy network output.

[0067] Using the above methods, the system can dynamically adjust the prediction depth of the world model according to changes in operating conditions and directly incorporate the prediction depth into the decision state space, enabling the control strategy to have the ability to perceive the future prediction range, thereby improving the control stability, water quality compliance rate and operating economy under complex operating conditions.

[0068] In this embodiment, the specific process of wastewater treatment control based on adaptive predictive time domain is as follows: Figure 2 As shown, a state space is first constructed using online monitoring data from the wastewater treatment process, and an H-Net is designed to perform real-time analysis of the current operating conditions, dynamically determining the prediction roll length of the world model based on the complexity of the operating conditions. Subsequently, the determined roll length is used as part of the state variables, forming an extended state input reinforcement learning policy network together with the process state, enabling the policy decision to perceive changes in the future prediction range. During the control process, the policy network generates control actions based on the extended state, and the world model performs multi-step state prediction and reward estimation according to the dynamically determined roll length, feeding the prediction results back to the policy optimization module. Through the collaborative optimization of the time-domain decision network, the world model, and the reinforcement learning policy network, the joint adaptive adjustment of the prediction time domain length and the control policy is achieved. Specifically, a shorter prediction time domain is automatically adopted when the operating conditions are stable to reduce computational overhead; under complex operating conditions such as load shocks, abnormal influent, and equipment fluctuations, the prediction time domain is automatically extended to improve the ability to predict future state changes and control robustness. Ultimately, the goals of stable effluent quality, reduced energy consumption, and improved adaptive capability of the control system are achieved in the wastewater treatment process.

[0069] Unlike existing world model reinforcement learning methods with fixed prediction time domains, this invention incorporates dynamically generated prediction roll lengths as part of the reinforcement learning state space, enabling the policy network to perceive and utilize prediction time domain information for decision-making, thereby achieving joint optimization of the prediction time domain and control policy.

[0070] As can be seen, this application proposes a closed-loop feedback world model, enabling the model to continuously correct its internal latent states during operation, rather than relying solely on a fixed predictor trained offline. Introducing the world model into the wastewater treatment biochemical process control scenario allows for the prediction of future evolution of dissolved oxygen, ammonia nitrogen, total nitrogen, sludge concentration, and system operating status, improving control foresight. A latent state feedback correction mechanism is introduced, continuously updating the world model's understanding based on the deviation between actual water quality and energy consumption measurements and model predictions, reducing model drift issues caused by long-term wastewater system operation. A wastewater control strategy is constructed by combining reinforcement learning strategy gradients, enabling the system to dynamically optimize aeration, recirculation, and carbon source addition actions based on future operating conditions, rather than passively adjusting based solely on current errors. A closed-loop control structure of "time-domain adaptive decision-making - world model feedback correction - multi-step future prediction - joint optimization of strategy and time-domain network" is constructed, elevating wastewater treatment control from "passive response control" to "integrated prediction-correction-cooperative decision-making control." This structure can adapt to complex dynamic conditions such as influent load shocks, water quality fluctuations, seasonal water temperature changes, and sludge aging. Achieve multi-objective synergistic optimization of effluent quality, precise dissolved oxygen control, and reduced aeration energy consumption, thereby improving the operational stability and economy of the biochemical system.

[0071] The feedback correction in this invention does not directly compensate for the output of the ammonia injection controller, but rather performs online correction on the potential states and state transition representations of the closed-loop world model. The system encodes the deviation between the measured field state and the state projected by the world model as a feedback signal, and inputs this feedback signal into the state update unit of the world model to correct the model's internal understanding of SCR dynamics. Based on this, the updated potential states are used to re-project future operating conditions and generate the next round of ammonia injection optimization strategy. Therefore, this invention achieves closed-loop correction at the "model cognition layer," rather than error compensation at the "controller parameter layer."

[0072] As can be seen, this application provides a wastewater treatment control method based on adaptive prediction time domain, including acquiring operational data of the wastewater treatment process, constructing a state vector based on the operational data, and inputting the state vector into a time-domain decision network to output a rolling time-domain length prediction value. Therefore, it can map the real-time operating conditions of the wastewater treatment process into a state expression that can be used for learning and decision-making, avoiding the problem that a fixed prediction time domain cannot adapt to different operating conditions. Based on the state vector and the rolling time-domain length prediction value, an extended state vector is constructed, and the extended state vector is input into a policy network to output control actions. Therefore, the policy network can simultaneously perceive the current state and prediction time-domain information when generating control actions, improving the adaptability of the control strategy to stable operating conditions, fluctuating operating conditions, and abnormal operating conditions. The extended state vector and control actions are input into a world model, and rolling prediction is performed with the rolling time-domain length prediction value as the rolling step number to obtain a state vector prediction sequence and a reward prediction sequence. Therefore, it can reduce unnecessary inference calculations when the operating conditions are stable, and extend the prediction range during load shocks or abnormal disturbances. It possesses a high predictive ability for future state changes; it calculates actual reward values ​​using a multi-objective reward function and determines cumulative reward prediction values ​​based on reward prediction sequences; it updates model parameters in the world model, network parameters in the policy network, and network parameters in the time-domain decision network based on actual reward values, reward prediction sequences, state vector prediction sequences, and cumulative reward prediction values, and determines whether a preset termination condition has been triggered. Therefore, it can jointly update the world model, policy network, and time-domain decision network based on real operational feedback and prediction results, enabling collaborative optimization of prediction time-domain length, state prediction capability, and control strategy within the same closed-loop learning structure; if the preset termination condition has not been triggered, the process of acquiring operational data is repeated until the preset termination condition is triggered, resulting in the target control action, thus improving the continuity of the online optimization process; the target control action is then distributed to the wastewater treatment equipment so that the wastewater treatment equipment can control the wastewater treatment process based on the target control action, thereby improving the control stability, effluent compliance rate, and operational economy of the wastewater treatment process.

[0073] See Figure 3 As shown in the figure, an embodiment of the present invention discloses a wastewater treatment control device based on adaptive prediction time domain under operating conditions, which may specifically include: The rolling time domain length prediction value determination module 11 is used to acquire the operation data of the sewage treatment process, construct a state vector based on the operation data, and input the state vector into the time domain decision network to output the rolling time domain length prediction value. The control action determination module 12 is used to construct an extended state vector based on the state vector and the predicted value of the rolling time domain length, and input the extended state vector into the policy network to output the control action; The rolling prediction module 13 is used to input the extended state vector and the control action into the world model, and perform rolling prediction with the rolling time domain length prediction value as the rolling step number to obtain the state vector prediction sequence and the reward prediction sequence. The cumulative return prediction value determination module 14 is used to calculate the actual reward value using a multi-objective reward function and determine the cumulative return prediction value based on the reward prediction sequence. The update module 15 is used to update the model parameters in the world model, the network parameters in the policy network, and the network parameters in the time domain decision network based on the actual reward value, the reward prediction sequence, the state vector prediction sequence, and the cumulative reward prediction value, and to determine whether the preset termination condition has been triggered. The target control action determination module 16 is used to repeatedly execute the process of acquiring running data if the preset end condition is not triggered until the preset end condition is triggered, and then obtain the target control action; the target control action is the control action output by the updated policy network. The wastewater treatment process control module 17 is used to send the target control action to the wastewater treatment equipment so that the wastewater treatment equipment can control the wastewater treatment process based on the target control action.

[0074] In this embodiment, the rolling time-domain length prediction value determination module 11 is used to perform operational data acquisition, state vector construction, and rolling time-domain length prediction value output. The control action determination module 12 is used to construct an extended state vector based on the state vector and the rolling time-domain length prediction value, and output the control action through the policy network. The rolling prediction module 13 is used to perform dynamic rolling prediction through the world model. The cumulative reward prediction value determination module 14 is used to calculate the actual reward value and determine the cumulative reward prediction value. The update module 15 is used to update the parameters of the world model, policy network, and time-domain decision network. The target control action determination module 16 is used to determine the target control action when the preset termination condition is met. The wastewater treatment process control module 17 is used to send the target control action to the wastewater treatment equipment.

[0075] In some specific embodiments, the rolling time domain length prediction value determination module 11 may specifically include: The operation data acquisition module is used to acquire operation data of the wastewater treatment process from the data acquisition and monitoring control system, programmable logic controller system or distributed control system; the operation data includes dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, chemical oxygen demand, mixed liquor suspended solids concentration, influent flow rate, water temperature and oxidation-reduction potential; The normalization processing module is used to normalize the running data to obtain processed running data; A state vector construction module is used to construct a state vector based on the processed running data.

[0076] In some specific embodiments, the rolling time domain length prediction value determination module 11 may further include: The temporal decision network construction module is used to construct temporal decision networks based on hierarchical neural networks; The rolling time-domain length prediction module is used to input the state vector into the time-domain decision network so that the time-domain decision network can be used to analyze the state vector to determine the current operating condition complexity of the wastewater treatment process, and output a discrete rolling time-domain length prediction value based on the operating condition complexity.

[0077] In some specific embodiments, the control action determination module 12 may specifically include: The normalized scroll length determination module is used to calculate the ratio between the predicted value of the scroll time domain length and the preset maximum scroll length, and to use the ratio as the normalized scroll length. An extended state vector concatenation module is used to concatenate the normalized scroll length with the state vector to obtain an extended state vector.

[0078] In some specific embodiments, the rolling prediction module 13 may specifically include: The next-moment prediction module is used to input the extended state vector and control action into the world model to output the predicted state vector and reward value for the next moment. The sequence determination module is used to perform rolling prediction based on the state vector prediction value and reward prediction value at the next time moment, and with the rolling time domain length prediction value as the rolling step number, continue to perform rolling prediction until the number of rolling predictions reaches the rolling time domain length prediction value, then stop rolling prediction, and obtain the state vector prediction sequence and reward prediction sequence.

[0079] In some specific embodiments, the cumulative return prediction value determination module 14 may specifically include: The target data acquisition module is used to obtain dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, and aeration energy consumption from the operational data; The reward item determination module is used to determine the dissolved oxygen control error item, ammonia nitrogen exceedance item, total nitrogen exceedance item, and aeration energy consumption item based on the dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, and aeration energy consumption, respectively. The actual reward value determination module is used to use a multi-objective reward function to perform a weighted summation of the dissolved oxygen control error item, the ammonia nitrogen exceedance item, the total nitrogen exceedance item, and the aeration energy consumption item to obtain the actual reward value; The cumulative return prediction value determination submodule is used to perform a weighted summation of the reward prediction values ​​in the reward prediction sequence based on the discount factor to obtain the cumulative return prediction value.

[0080] In some specific embodiments, the update module 15 may specifically include: The world model loss function construction module is used to construct the world model loss function based on the state vector, the state vector prediction sequence, the actual reward value, and the reward prediction sequence. The world model parameter update module is used to update the model parameters in the world model with the goal of minimizing the world model loss function; The policy network parameter update module is used to update the network parameters in the policy network with the goal of maximizing the cumulative return prediction value by using the policy gradient. The time-domain decision network parameter update module is used to update the network parameters in the time-domain decision network with the goal of maximizing the cumulative return prediction value.

[0081] Figure 4 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 20 specifically includes: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the wastewater treatment control method based on adaptive predictive time domain of operating conditions disclosed in any of the foregoing embodiments.

[0082] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0083] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0084] The operating system 221 manages and controls the various hardware devices on the electronic device 20 and the computer program 222 to enable the processor 21 to perform calculations and processing on the data 223 in the memory 22. It can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the wastewater treatment control method based on adaptive predictive time domain of operating conditions as disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the wastewater treatment control device based on adaptive predictive time domain of operating conditions transmitted from external devices, and may also include data collected by its own input / output interface 25.

[0085] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0086] Furthermore, this application also discloses a computer-readable storage medium storing a computer program. When the computer program is loaded and executed by a processor, it implements the steps of the wastewater treatment control method based on adaptive prediction time domain of operating conditions disclosed in any of the foregoing embodiments.

[0087] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0088] The above provides a detailed description of a wastewater treatment control method, apparatus, equipment, and medium based on adaptive predictive time domain under operating conditions provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A wastewater treatment control method based on adaptive predictive time domain under operating conditions, characterized in that, include: The operation data of the wastewater treatment process is acquired, a state vector is constructed based on the operation data, and the state vector is input into a time-domain decision network to output a rolling time-domain length prediction value. Based on the state vector and the predicted value of the rolling time domain length, an extended state vector is constructed, and the extended state vector is input into the policy network to output control actions. The extended state vector and the control action are input into the world model, and the rolling prediction is performed with the rolling time domain length prediction value as the rolling step number to obtain the state vector prediction sequence and the reward prediction sequence. The actual reward value is calculated using a multi-objective reward function, and the cumulative return prediction value is determined based on the reward prediction sequence. Based on the actual reward value, the reward prediction sequence, the state vector prediction sequence, and the cumulative reward prediction value, the model parameters in the world model, the network parameters in the policy network, and the network parameters in the time domain decision network are updated respectively, and it is determined whether the preset termination condition has been triggered. If the preset termination condition is not triggered at present, the process of acquiring running data is repeated until the preset termination condition is triggered and the target control action is obtained; the target control action is the control action output by the updated policy network. The target control action is sent to the wastewater treatment equipment so that the wastewater treatment equipment can control the wastewater treatment process based on the target control action; The process of inputting the state vector into a time-domain decision network to output a rolling time-domain length prediction value includes: constructing a time-domain decision network based on a hierarchical neural network; inputting the state vector into the time-domain decision network so as to use the time-domain decision network to analyze the state vector, determine the current operating condition complexity of the wastewater treatment process, and output a discrete rolling time-domain length prediction value based on the operating condition complexity. Based on the state vector and the predicted value of the scrolling time domain length, an extended state vector is constructed, including: calculating the ratio between the predicted value of the scrolling time domain length and the preset maximum scrolling length, and using the ratio as the normalized scrolling length; concatenating the normalized scrolling length with the state vector to obtain the extended state vector; The model parameters in the world model, the network parameters in the policy network, and the network parameters in the temporal decision network are updated based on the actual reward value, the reward prediction sequence, the state vector prediction sequence, and the cumulative reward prediction value, respectively. This includes: constructing a world model loss function based on the state vector, the state vector prediction sequence, the actual reward value, and the reward prediction sequence; updating the model parameters in the world model with the objective of minimizing the world model loss function; updating the network parameters in the policy network using the policy gradient with the objective of maximizing the cumulative reward prediction value; and updating the network parameters in the temporal decision network with the objective of maximizing the cumulative reward prediction value.

2. The wastewater treatment control method based on adaptive predictive time domain according to claim 1, characterized in that, The acquisition of operational data from the wastewater treatment process, and the construction of a state vector based on the operational data, includes: The wastewater treatment process operation data is acquired from a data acquisition and monitoring control system, a programmable logic controller system, or a distributed control system. The operation data includes dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, chemical oxygen demand, mixed liquor suspended solids concentration, influent flow rate, water temperature, and oxidation-reduction potential. The running data is normalized to obtain the processed running data; A state vector is constructed based on the processed running data.

3. The wastewater treatment control method based on adaptive predictive time domain according to claim 1, characterized in that, The process of inputting the extended state vector and the control action into the world model, and performing rolling prediction with the rolling time domain length prediction value as the rolling step number, to obtain a state vector prediction sequence and a reward prediction sequence includes: The extended state vector and control action are input into the world model to output the predicted state vector and reward for the next time step. Based on the predicted state vector and reward value for the next moment, and using the predicted rolling time domain length as the number of rolling steps, continue to perform rolling prediction until the number of rolling predictions reaches the predicted rolling time domain length value, then stop rolling prediction to obtain the predicted state vector sequence and the predicted reward sequence.

4. The wastewater treatment control method based on adaptive prediction in the time domain according to claim 1, characterized in that, The step of calculating the actual reward value using a multi-objective reward function and determining the cumulative return prediction value based on the reward prediction sequence includes: Obtain dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, and aeration energy consumption from the operational data; Based on the dissolved oxygen, ammonia nitrogen concentration, total nitrogen concentration, and aeration energy consumption, the dissolved oxygen control error item, ammonia nitrogen exceeding the standard item, total nitrogen exceeding the standard item, and aeration energy consumption item are determined respectively. Using a multi-objective reward function, the dissolved oxygen control error term, the ammonia nitrogen exceedance term, the total nitrogen exceedance term, and the aeration energy consumption term are weighted and summed to obtain the actual reward value; Based on the discount factor, the reward prediction values ​​in the reward prediction sequence are weighted and summed to obtain the cumulative return prediction value.

5. A wastewater treatment control device based on adaptive predictive time domain under operating conditions, characterized in that, include: The rolling time-domain length prediction value determination module is used to acquire the operation data of the sewage treatment process, construct a state vector based on the operation data, and input the state vector into the time-domain decision network to output the rolling time-domain length prediction value. The control action determination module is used to construct an extended state vector based on the state vector and the predicted value of the rolling time domain length, and input the extended state vector into the policy network to output the control action; The rolling prediction module is used to input the extended state vector and the control action into the world model, and perform rolling prediction with the rolling time domain length prediction value as the rolling step number to obtain the state vector prediction sequence and the reward prediction sequence. The cumulative return prediction value determination module is used to calculate the actual reward value using a multi-objective reward function and determine the cumulative return prediction value based on the reward prediction sequence. The update module is used to update the model parameters in the world model, the network parameters in the policy network, and the network parameters in the time domain decision network based on the actual reward value, the reward prediction sequence, the state vector prediction sequence, and the cumulative reward prediction value, and to determine whether the preset termination condition has been triggered. The target control action determination module is used to repeatedly execute the process of acquiring running data if the preset termination condition is not triggered, until the preset termination condition is triggered, and then obtain the target control action; the target control action is the control action output by the updated policy network. The wastewater treatment process control module is used to send the target control action to the wastewater treatment equipment so that the wastewater treatment equipment can control the wastewater treatment process based on the target control action; The process of inputting the state vector into a time-domain decision network to output a rolling time-domain length prediction value includes: constructing a time-domain decision network based on a hierarchical neural network; inputting the state vector into the time-domain decision network so as to use the time-domain decision network to analyze the state vector, determine the current operating condition complexity of the wastewater treatment process, and output a discrete rolling time-domain length prediction value based on the operating condition complexity. Based on the state vector and the predicted value of the scrolling time domain length, an extended state vector is constructed, including: calculating the ratio between the predicted value of the scrolling time domain length and the preset maximum scrolling length, and using the ratio as the normalized scrolling length; concatenating the normalized scrolling length with the state vector to obtain the extended state vector; The model parameters in the world model, the network parameters in the policy network, and the network parameters in the temporal decision network are updated based on the actual reward value, the reward prediction sequence, the state vector prediction sequence, and the cumulative reward prediction value, respectively. This includes: constructing a world model loss function based on the state vector, the state vector prediction sequence, the actual reward value, and the reward prediction sequence; updating the model parameters in the world model with the objective of minimizing the world model loss function; updating the network parameters in the policy network using the policy gradient with the objective of maximizing the cumulative reward prediction value; and updating the network parameters in the temporal decision network with the objective of maximizing the cumulative reward prediction value.

6. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the wastewater treatment control method based on adaptive predictive time domain as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the wastewater treatment control method based on the adaptive prediction time domain of operating conditions as described in any one of claims 1 to 4.

Images