A multi-objective collaborative optimization control method for an environmental protection device and a related device

Abstract

The application discloses a multi-target collaborative optimization control method of an environmental protection device and related devices, relates to the technical field of industrial intelligent control, obtains a working condition tensor for comprehensively representing a real running state and a pollutant emission level of the environmental protection device in a previous period, inputs the working condition tensor into an energy consumption optimization strategy network, obtains a basic control instruction vector of a current period and directed at energy saving, inputs the working condition tensor and the optimization control instruction vector of the previous period into a digital twin for simulation and prediction, obtains a predicted working condition tensor of the current period, obtains a measured working condition tensor corresponding to the predicted working condition tensor, compensates and corrects the basic control instruction vector according to the component residual errors corresponding to the pollutant emission level in the predicted working condition tensor and the measured working condition tensor, obtains a compensation control instruction vector, performs limiting processing based on a physical safety boundary, and obtains an optimization control instruction vector of the current period. The application can improve the adaptive control capability of the environmental protection device under dynamic working conditions.

Description

Technical Field

[0001] This application relates to the field of industrial intelligent control technology, and in particular to a multi-objective collaborative optimization control method and related device for environmental protection equipment. Background Technology

[0002] With the deepening of the global concept of energy conservation and emission reduction and the continuous tightening of industrial environmental emission standards, environmental protection equipment has become a core link in industrial production that takes into account both pollutant treatment and energy consumption control. This type of equipment covers multiple scenarios such as flue gas desulfurization and denitrification systems, sewage treatment aeration systems, and dust removal systems.

[0003] Unlike conventional industrial processes, the operation of environmental protection equipment involves complex physicochemical reactions, generally exhibiting dynamic characteristics such as strong nonlinearity, large inertia, and time-varying hysteresis. Furthermore, actual operating conditions are often affected by factors such as load fluctuations, water quality changes, and equipment aging, resulting in high time variability and uncertainty. This necessitates that control strategies simultaneously address the conflicting objectives of emission compliance and energy minimization under dynamically changing operating conditions.

[0004] In traditional technological systems, the control of environmental protection equipment mainly relies on proportional-integral-derivative (PID) control, model predictive control, or reinforcement learning-based methods. However, these methods generally suffer from poor multi-objective equilibrium optimization performance in practical applications. PID control methods essentially employ static parameters or fixed optimization strategies, making it difficult to adjust the weights of energy saving and emissions under dynamic operating conditions. This often results in the control strategy failing to achieve a proper balance between energy saving and emissions; a conservative approach leads to higher energy consumption, while an aggressive approach easily triggers emissions exceeding standards.

[0005] Although model predictive control can achieve online rolling optimization, its control performance is heavily dependent on the accuracy of the process model. When the actual operating conditions deviate from the modeled conditions, the model error will cause the control command to deviate from the optimal setting, making it impossible to stably achieve the operating state of "just meeting emission constraints and having low energy consumption".

[0006] While existing reinforcement learning-based methods can obtain optimized strategies through training, their trial-and-error exploration mechanism is inherently contradictory to the hard constraints of excessive emissions from environmental protection equipment. Moreover, most methods focus on single objectives or simple constraint scenarios, lacking a systematic design for dynamic collaborative control of multiple objectives, making it difficult to achieve balanced optimization in practical applications.

[0007] Therefore, there is an urgent need for a multi-objective coordinated control method that can stably balance emission constraints and energy consumption optimization under dynamic operating conditions. Summary of the Invention

[0008] In view of the above problems, this application provides a multi-objective collaborative optimization control method and related device for environmental protection equipment, so as to achieve the goal of generating an optimized control command vector that is both energy-saving and takes into account emission hard constraints under dynamic operating conditions. The specific solution is as follows: The first aspect of this application provides a multi-objective collaborative optimization control method for environmental protection equipment, including: Obtain the operating condition tensor used to comprehensively characterize the actual operating status and pollutant emission level of the environmental protection equipment in the previous cycle, and input it into the pre-trained energy consumption optimization strategy network to obtain the basic control command vector for the current cycle with energy saving as the guide. The operating condition tensor and the optimized control command vector of the previous cycle are input into a pre-constructed digital twin for simulation prediction to obtain the predicted operating condition tensor for the current cycle. Obtain the measured operating condition tensor corresponding to the predicted operating condition tensor, determine the component residuals corresponding to the pollutant emission level in the predicted operating condition tensor and the measured operating condition tensor, and use the component residuals to compensate and correct the basic control command vector to obtain the compensated control command vector. The compensation control command vector is subjected to amplitude limiting processing based on physical safety boundaries to obtain the optimized control command vector for the current cycle.

[0009] In one possible implementation, the process of constructing the digital twin includes: A mechanism model is constructed based on the physical changes and chemical reaction characteristics equations of the internal operation process of environmental protection equipment; A pre-trained deep residual compensation network is obtained, and the deep residual compensation network is coupled with the mechanistic model to obtain the digital twin. The coupling method is as follows: within each simulation step, the basic state vector obtained by the integral solution of the mechanistic model and the operating condition tensor are jointly input into the deep residual compensation network to obtain a state correction quantity. The simulation state vector is generated by the state correction quantity and the basic state vector and used as the initial state of the next simulation step. The integral solution of the mechanistic model is based on the initial state of the current simulation step and the optimized control command vector of the previous cycle. The simulation state vector under the last simulation step of the current cycle is used as the predicted operating condition tensor.

[0010] In one possible implementation, the mechanism model constructed based on the physical changes and chemical reaction characteristic equations of the internal operation process of the environmental protection equipment includes: Based on the aforementioned physical change and chemical reaction characteristic equations, a set of partial differential equations is established to describe the spatiotemporal distribution of state variables at various points inside the environmental protection equipment. The partial differential equations are discretized in the spatial dimension using the finite difference method, so as to transform the spatial derivative terms in the partial differential equations into difference algebraic relationships between state variables of adjacent points, and obtain ordinary differential equations including time derivatives. A mechanistic model based on the aforementioned set of ordinary differential equations is constructed.

[0011] In one possible implementation, the training process of the deep residual compensation network includes: Acquire historical operating condition tensor samples, optimized control command vector samples executed in the period where the historical operating condition tensor samples are located, and measured operating condition tensor samples obtained by the environmental protection equipment after executing the optimized control command vector samples. The optimized control command vector sample is input into the mechanism model to obtain the basic state vector sample; A state correction sample is generated based on the basic state vector sample and the measured working condition tensor sample. The historical operating condition tensor samples and the basic state vector samples are used as training samples, and the state correction samples are used as sample labels. The deep residual compensation network is trained using the training samples and the sample labels.

[0012] In one possible implementation, training the deep residual compensation network using the training samples and the sample labels includes: The training samples are input into the deep residual compensation network to obtain the predicted residual vector; The mean squared error loss is generated based on the predicted residual vector and the sample labels; Calculate the gradient of the mean squared error loss with respect to the parameters of the deep residual compensation network, and generate a regularized loss based on the gradient; The mean squared error loss and the regularization loss are used to generate a total loss, and the total loss is used to train the deep residual compensation network.

[0013] In one possible implementation, the energy consumption optimization policy network is trained in a simulation environment provided by the digital twin, the training process in the simulation environment including: Based on the digital twin, multiple twin instances corresponding to different environmental protection equipment operating conditions are generated, and the twin instances interact with the energy consumption optimization strategy network to obtain interactive data including the current state, current action, next state, reward and constraint cost, which is stored in the priority experience replay pool. The interaction data in the priority experience replay pool is sampled, and at least one interaction trajectory is formed by the sampled interaction data in chronological order. The state sequence corresponding to the interaction trajectory is input into the reward value network and the constraint value network respectively. The reward advantage function value is calculated based on the cumulative reward estimate output by the reward value network, and the constraint advantage function value is calculated based on the cumulative cost estimate output by the constraint value network. Using the expected cumulative reward generated from the reward sequence corresponding to the interaction trajectory as the supervision objective, the parameters of the reward value network are updated by minimizing the difference between the estimated cumulative reward and the expected cumulative reward; using the expected cumulative cost generated from the constraint cost sequence corresponding to the interaction trajectory as the supervision objective, the parameters of the constraint value network are updated by minimizing the difference between the estimated cumulative cost and the expected cumulative cost. A reward pruning objective is generated based on the reward advantage function value, and a constraint pruning objective is generated based on the constraint advantage function value. The goal is to maximize the Lagrange weighted difference between the reward pruning objective and the constraint pruning objective. The energy consumption optimization strategy network is then updated, and the sampling step of the interaction data is returned until the preset training termination condition is reached.

[0014] In one possible implementation, enabling the twin instance to interact with the energy consumption optimization strategy network includes: Obtain the predicted working condition tensor generated by the twin instance as the current state; The current state is input into the energy consumption optimization strategy network to obtain the basic control command vector corresponding to the current state, which is used as the current action. Input the current action and the current state into the twin instance to obtain the next working condition tensor, which is used as the next state; Based on the next state, a composite reward function value and a constraint cost function value are generated. The composite reward function value is used to guide the energy consumption optimization strategy network to generate an energy-saving-oriented basic control command vector, and the constraint cost function value is used to constrain the energy consumption optimization strategy network to generate a basic control command vector that meets the pollutant emission constraint requirements.

[0015] In one possible implementation, the Lagrange multipliers used to calculate the Lagrange weighted difference are updated after each update of the energy consumption optimization strategy network; The update process of the Lagrange multipliers includes: Calculate the average expected value of constraint cost based on the constraint cost function values ​​in each interaction data contained in the at least one interaction trajectory; Calculate the deviation of the expected value of the average constraint cost relative to the preset emission constraint threshold; The Lagrange multipliers are updated based on the deviation.

[0016] In one possible implementation, sampling the interaction data in the priority experience replay pool includes: The current state and the next state in the interaction data are input into the reward value network to obtain the cumulative reward estimate corresponding to the current state and the cumulative reward estimate corresponding to the next state. Based on the composite reward function value in the interaction data, the cumulative reward estimate corresponding to the current state, and the cumulative reward estimate corresponding to the next state, a temporal difference error corresponding to the interaction data is generated. Based on the timing difference error corresponding to the interaction data in the priority experience replay pool, the sampling priority corresponding to the interaction data in the priority experience replay pool is determined. The interaction data in the priority experience replay pool is sampled based on the sampling priority.

[0017] In one possible implementation, the step of using the component residual to compensate and correct the basic control command vector to obtain a compensated control command vector includes: The component residuals are mapped to control correction amplitudes using a preset mapping strategy, which serve as control compensation amounts. The control compensation amount is superimposed on the basic control command vector using a proportional-integral law to generate the compensation control command vector.

[0018] In one possible implementation, obtaining the operating condition tensor used to comprehensively characterize the actual operating status and pollutant emission levels of the environmental protection equipment in the previous cycle includes: Acquire multi-source operating parameters that characterize the actual operating status of environmental protection equipment in the previous cycle and emission monitoring data that characterize the actual pollutant emission level of environmental protection equipment in the previous cycle; The multi-source operating parameters and the emission monitoring data are time-series aligned to obtain the time-series aligned operating condition tensor.

[0019] The second aspect of this application provides a computer program product, including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement the multi-objective collaborative optimization control method for environmental protection equipment described in the first aspect or any implementation thereof.

[0020] A third aspect of this application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program so that the electronic device can implement the multi-objective collaborative optimization control method for environmental protection equipment described in the first aspect or any implementation thereof.

[0021] The fourth aspect of this application provides a computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the multi-objective collaborative optimization control method for environmental protection equipment described in the first aspect or any implementation thereof.

[0022] By means of the above technical solution, the multi-objective collaborative optimization control method for environmental protection equipment provided in this application takes into account that environmental protection equipment is a system with strong lag and long delay. The physical and chemical reactions that occur inside the equipment have obvious lag response, which causes the pollutant emission results of the current cycle to be affected by the operating conditions and operating behavior of at least the previous cycle. Therefore, this application obtains the operating condition tensor used to comprehensively characterize the actual operating state and pollutant emission level of the environmental protection equipment in the previous cycle, and inputs it into a pre-trained energy consumption optimization strategy network to generate the basic control command vector for the current cycle that is energy-saving oriented.

[0023] Since the basic control command vector is designed with energy conservation as its core objective, it may not meet emission hard constraints under certain dynamic operating conditions. To avoid command unavailability due to excessive emissions, this application further predicts the theoretical state that the environmental protection equipment should reach in the current cycle under the continuous action of the optimized control command vector from the previous cycle, based on the aforementioned operating condition tensor. This involves inputting the aforementioned operating condition tensor and the optimized control command vector from the previous cycle into a pre-constructed digital twin for simulation prediction, obtaining the predicted operating condition tensor for the current cycle. Next, the measured operating condition tensor corresponding to the predicted operating condition tensor is obtained, and the component residual corresponding to the pollutant emission level in the predicted and measured operating condition tensors is determined. This component residual is used to compensate and correct the basic control command vector, ensuring that the resulting compensated control command vector, while possessing energy conservation capabilities, adapts to fluctuations in dynamic operating conditions and meets emission hard constraints under fluctuating operating conditions.

[0024] Considering that the residual compensation process is affected by many factors such as the accuracy of twin predictions and the measurement accuracy of the tensor under actual operating conditions, unmodeled disturbances may still exist in actual industrial control, leading to slight exceedances in emissions. To ensure emission safety and compliance, this application adds a fallback mechanism based on physical safety boundaries. This means that the compensation control command vector is subjected to amplitude limiting processing based on physical safety boundaries, ensuring that the current cycle optimized control command vector obtained by the environmental protection equipment under all dynamic operating conditions can stably meet the hard emission constraints while maximizing energy-saving benefits, thus achieving safe and reliable implementation of the control strategy. Attached Figure Description

[0025] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0026] Figure 1 A schematic diagram of a system architecture provided for this application; Figure 2 A flowchart illustrating a multi-objective collaborative optimization control method for environmental protection equipment provided in this application; Figure 3 A schematic diagram of the structure of a multi-objective collaborative optimization control device for environmental protection equipment provided in this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0027] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0028] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0029] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0030] First, a brief introduction to the terms used in the embodiments of this application will be given.

[0031] Reinforcement learning is a paradigm in machine learning. Its core idea is that an agent learns optimal behavioral strategies through continuous interaction and trial and error with its environment. The agent performs actions in the environment, and the environment provides feedback signals of rewards and new states. The agent continuously adjusts its strategy based on the goal of maximizing cumulative rewards. Unlike supervised learning, which relies on labeled samples, reinforcement learning emphasizes autonomous learning from delayed rewards generated by interactions. It is particularly suitable for complex control problems with sequential decision-making characteristics that are difficult to model precisely, and can explore potentially better control patterns without relying on human experience rules.

[0032] Distributed Control Systems (DCS) are core control systems in industrial automation, primarily used for centralized monitoring and decentralized control of large-scale, complex industrial equipment or production processes. Their core characteristic is "centralized management and decentralized control." Through distributed controllers, sensors, actuators, and communication networks, they achieve real-time acquisition, centralized monitoring, and decentralized control of the operating status of each unit within the equipment. This effectively addresses issues such as multi-variable coupling and large time lags in industrial processes, ensuring the stability and reliability of equipment operation. In environmental protection equipment control, DCS serves as the core hardware platform for acquiring real-time equipment operating data and issuing control commands, providing fundamental support for data transmission and execution control in implementing intelligent optimization control methods.

[0033] Physical safety boundaries are rigid constraints set based on industrial equipment operation specifications, environmental emission standards, and equipment hardware limitations. They are the core thresholds for ensuring the safe and compliant operation of equipment. They cover the safe range of equipment operating parameters (such as the upper limit of fan speed and the range of pump and valve opening), the compliance threshold of pollutant emissions (such as the upper limit of pollutant concentration in flue gas after desulfurization and denitrification), and the tolerance limits of equipment hardware. They are the red lines that must be followed during the generation of control commands. Before the control strategy is executed, the initially generated control commands must be verified and limited by physical safety boundaries to ensure that the output control commands meet the optimization objectives without causing safety or compliance risks such as equipment overheating, overpressure, or excessive emissions.

[0034] This application provides a multi-objective collaborative optimization control method and related device for environmental protection equipment, which can be applied to scenarios such as flue gas denitrification, wet desulfurization, and dust removal and purification in thermal power plants.

[0035] For example, in flue gas desulfurization and denitrification scenarios, the inlet flue gas flow rate, sulfur dioxide, and nitrogen oxide concentrations fluctuate frequently with changes in boiler load and coal quality. Traditional fixed-parameter control strategies struggle to achieve a dynamic balance between removal efficiency, ammonia injection rate, and circulating pump power consumption. This method collects inlet flue gas parameters, absorber slurry state, ammonia injection regulating valve opening, and outlet emission monitoring data. After time-series alignment processing, a working condition tensor is generated. This tensor is input into an energy consumption optimization strategy network to obtain the current cycle's energy-saving-oriented basic control command vector. Simultaneously, a digital twin predicts the outlet emission response based on the working condition tensor and the previous cycle's optimized control command vector, comparing it with measured emissions to obtain the residual. A preset mapping is applied to this residual to obtain a control compensation amount, which is then superimposed on the basic control command vector and dynamically limited to generate the current cycle's optimized control command. This method, while ensuring stable compliance with outlet sulfur dioxide and nitrogen oxide concentration standards, can automatically adjust the circulating pump frequency and ammonia injection rate according to load changes, reducing absorbent consumption and power consumption.

[0036] In the control application of wastewater treatment aeration systems, the time-varying nature of influent water quality and quantity leads to continuous changes in dissolved oxygen demand in the biological treatment tank. Excessive aeration results in energy waste, while insufficient aeration affects the effluent quality compliance. Using the method described in this application, influent water quality parameters, dissolved oxygen concentration in the biological treatment tank, mixed liquor suspended solids concentration, and blower operating parameters are collected in real time. The operating condition tensor obtained through time-series alignment is input into the energy consumption optimization strategy network to infer and output the basic control command vector for blower frequency or guide vane opening. Simultaneously, a digital twin predicts the dissolved oxygen change trend under the current control action and compares it with the measured value to generate residuals. Through mapping and compensation superposition of the residuals, a compensation control command vector is generated and issued for execution after being limited by a safety constraint protection layer based on the lower limit of dissolved oxygen. This method enables the aeration intensity to match the microbial oxygen demand in real time, effectively reducing blower energy consumption while ensuring that effluent ammonia nitrogen and chemical oxygen demand meet standards, and avoiding sludge structure damage caused by over-aeration.

[0037] In industrial dust removal systems (such as electrostatic precipitators and pulse jet bag filters) in industries like metallurgy, chemicals, and building materials, dynamic synergistic optimization of dust removal efficiency and energy consumption can be achieved. These systems face challenges such as flue gas dust concentration fluctuations, changes in filter bag resistance, and pulse cleaning cycle attenuation. Traditional control methods often suffer from over- or under-cleaning due to parameter rigidity, affecting dust removal efficiency, easily causing excessive dust emissions, and increasing energy consumption for fan operation and the cleaning system. This application can process operational data such as flue gas dust concentration, filter bag pressure difference, fan current, and cleaning pulse width and interval through time-series alignment. The energy consumption optimization strategy network autonomously learns the optimal adjustment rules for control parameters such as cleaning cycle and fan negative pressure. Using a digital twin, it predicts dust removal efficiency and emission levels under different control strategies. Residual compensation generated from measured data compensates for deviations caused by factors such as filter bag aging, ensuring that control commands meet dust emission constraints while minimizing fan operation and cleaning energy consumption.

[0038] It should be noted that the above scenarios are merely examples and are not intended to limit this application.

[0039] Optionally, the multi-objective collaborative optimization control method for environmental protection equipment provided in this application can be applied to, for example... Figure 1 The system architecture shown includes a terminal 100 and a server 200. The server 200 may include one or more servers (…). Figure 1 (This example uses a server as an illustration).

[0040] Either terminal 100 or server 200 can be used independently to execute the multi-objective collaborative optimization control method for environmental protection equipment provided in the embodiments of this application. Alternatively, terminal 100 and server 200 can also be used collaboratively to execute the multi-objective collaborative optimization control method for environmental protection equipment provided in the embodiments of this application.

[0041] In a possible implementation environment, terminal 100 may be an edge computing gateway, a distributed control system operator station, a remote monitoring terminal, or an industrial control computer deployed in an industrial site. It is used to collect sensor data, run digital twin simulation models, execute energy consumption optimization strategy network inference, and output control command vectors to lower-level actuators. The sensors include, but are not limited to, sulfur dioxide concentration analyzers, nitrogen oxide concentration analyzers, oxygen content analyzers, dust concentration monitors, flue gas flow meters, temperature sensors, etc., used to monitor flue gas parameters.

[0042] Terminal 100 can implement the method of this application through a configured processor. Optionally, the processor may include, but is not limited to, a central processing unit, a graphics processing unit, a multi-core processor, a tensor processor, or an artificial intelligence acceleration chip, etc., which are not limited here.

[0043] To enable those skilled in the art to better understand this application, the multi-objective collaborative optimization control method for environmental protection equipment according to embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0044] Reference Figure 2 , Figure 2 A flowchart illustrating a multi-objective collaborative optimization control method for environmental protection equipment provided in this application embodiment is shown below. Figure 2 As shown, the multi-objective collaborative optimization control method for this environmental protection equipment may include: Step S201: Obtain the operating condition tensor used to comprehensively characterize the actual operating status and pollutant emission level of the environmental protection equipment in the previous cycle, and input it into the pre-trained energy consumption optimization strategy network to obtain the basic control command vector for the current cycle with energy saving as the guide.

[0045] It is understandable that environmental protection equipment involves complex physical changes and chemical reactions, and its overall operating condition is determined by multiple variables such as load, flue gas conditions, process parameters, and emission indicators. A single parameter cannot fully characterize the actual operating condition of the equipment. In order to accurately characterize the operating conditions of environmental protection equipment, it is necessary to obtain the operating condition tensor, which represents the actual operating state of the environmental protection equipment and the pollutant emission level.

[0046] Furthermore, considering that environmental protection equipment is a system with strong hysteresis and long time delay, the physicochemical reactions occurring within the equipment exhibit significant lag responses. This means that the pollutant emission results for the current cycle are not determined by the instantaneous parameters at the current moment, but rather by the cumulative effects of the operating conditions and behaviors of at least the previous cycle. Therefore, in order to accurately assess the basic control command vector acting on the actuators of the environmental protection equipment in the current cycle, this application needs to obtain an operating condition tensor that comprehensively characterizes the actual operating state and pollutant emission levels of the environmental protection equipment in the previous cycle.

[0047] Optionally, the process of “obtaining the operating condition tensor for comprehensively characterizing the actual operating status and pollutant emission level of the environmental protection equipment in the previous cycle” in the embodiments of this application may include: obtaining multi-source operating parameters characterizing the actual operating status of the environmental protection equipment in the previous cycle and emission monitoring data characterizing the actual pollutant emission level of the environmental protection equipment in the previous cycle; performing time-series alignment on the multi-source operating parameters and emission monitoring data to obtain the time-series aligned operating condition tensor.

[0048] Here, the multi-source operating parameters are derived from real-time operating data collected by the equipment's sensor network. These can include physical quantities that reflect the actual operating status of the environmental protection equipment, such as fan speed, pump and valve opening, reaction chamber temperature, pressure, flow rate, and energy consumption. Of course, the multi-source operating parameters can also be other types, which are not limited here.

[0049] The emission monitoring data comes from pollutant concentration data collected by compliant emission monitoring equipment.

[0050] It should be noted that both types of data obtained in this embodiment carry timestamp information, and the collection frequency can be consistent with the control cycle.

[0051] Because different sensors have different installation locations and data transmission delays, directly using raw data can lead to lag bias between variables, affecting the accuracy of control decisions. Therefore, this embodiment can perform time-series alignment of multi-source operating parameters and emission monitoring data.

[0052] Optionally, the alignment process can use a unified control cycle timestamp as a benchmark, employing linear interpolation, sliding window synchronization, or dynamic time warping algorithms to complete missing data and calibrate time-series offset data. This maps multi-source operating parameters and emission monitoring data into a tensor structure within the same time dimension, resulting in a time-aligned operating condition tensor. This operating condition tensor can fully preserve the time-series correlation characteristics of each parameter and the synchronization relationships between variables, providing a high-quality data input foundation for subsequent control strategy inference.

[0053] In order to obtain the basic control command vector for the current cycle, this embodiment pre-trains an energy consumption optimization strategy network. Optionally, the energy consumption optimization strategy network can be a deep neural network pre-trained offline. The training process can be completed in a simulation environment constructed by a digital twin using a reinforcement learning algorithm with safety constraints. For details of the training process, please refer to the embodiment below.

[0054] Optionally, the energy consumption optimization strategy network may include multiple hidden layers and nonlinear activation functions, so that after the operating condition tensor is input into the energy consumption optimization strategy network, the network can generate the current period's energy-saving oriented basic control command vector at the output layer through the multiplication of the weight matrices of each hidden layer and the layer-by-layer transformation of the nonlinear activation function.

[0055] In this embodiment, each element of the basic control command vector corresponds to the adjustment target value of a continuously controllable actuator, such as the target opening degree of the ammonia injection regulating valve, the target frequency of the circulating pump frequency converter, and the target opening degree of the aeration blower guide vanes.

[0056] For example, for a flue gas denitrification system, the basic control command vector may include a reference value for ammonia injection flow rate; for a wastewater treatment aeration system, the basic control command vector may include a reference value for blower operating frequency.

[0057] In this embodiment, the basic control command vector directly reflects the energy-saving-oriented control decision, realizing intelligent mapping of multi-objective trade-offs under complex operating conditions, and avoiding the decision rigidity problem caused by the traditional static weighting method.

[0058] Step S202: Input the operating condition tensor and the optimized control command vector of the previous cycle into the pre-constructed digital twin for simulation prediction to obtain the predicted operating condition tensor for the current cycle.

[0059] In this embodiment, the digital twin constructs a simulation environment that can simulate the dynamic response process of the environmental protection equipment. Thus, when the operating condition tensor and the optimized control command vector of the previous cycle are input into the digital twin, the response can be simulated in the simulation environment, and the predicted operating condition tensor for the next cycle (i.e., the current cycle) can be output. This predicted operating condition tensor represents the theoretical state that the environmental protection equipment should reach in the current cycle under the continuous action of the optimized control command vector of the previous cycle.

[0060] Step S203: Obtain the measured operating condition tensor corresponding to the predicted operating condition tensor, determine the component residuals corresponding to the pollutant emission level in the predicted operating condition tensor and the measured operating condition tensor, and use the component residuals to compensate and correct the basic control command vector to obtain the compensated control command vector.

[0061] Here, the measured operating condition tensor corresponding to the predicted operating condition tensor refers to the measured operating condition tensor collected by the data acquisition module of the distributed control system at the time when the predicted operating condition tensor is located. The dimensions of this measured operating condition tensor are completely consistent with those of the predicted operating condition tensor and the operating condition tensor after time-series alignment mentioned above.

[0062] Although the predicted operating condition tensor obtained from the digital twin can approximate the characteristics of the actual equipment to a certain extent, factors such as unmodeled disturbances, sensor zero-point drift, and slow time-varying equipment performance (e.g., equipment aging) due to issues in the actual operating conditions, such as the prediction accuracy of the twin and the measurement accuracy of the measured operating condition tensor, can still lead to accumulated deviations between the predicted and measured operating condition tensors over time. To compensate for the impact of this deviation on emission hard constraints, this embodiment obtains the measured operating condition tensor corresponding to the predicted operating condition tensor. Then, based on the residual components corresponding to pollutant emission levels in the predicted and measured operating condition tensors, the basic control command vector is compensated and corrected to obtain the compensated control command vector.

[0063] For example, if the period from k-1 to k is in the previous cycle, and the period from k to k+1 is in the current cycle, then this embodiment uses the operating condition tensor sampled from the previous cycle and the optimized control command vector (the previous cycle may have issued multiple optimized control command vectors to the actuator, so the optimized control command vector of the previous cycle can be selected as the last optimized control command vector issued in the previous cycle) to predict the predicted operating condition tensor at time k using a digital twin. At time k, the measured operating condition tensor is also collected, and the two are subtracted element-wise to obtain a state residual vector, which is used to quantify the deviation between the predictive capability of the digital twin and the actual physical process. Finally, the component residual corresponding to the pollutant emission level is obtained from the state residual vector. This component residual directly reflects the deviation between the prediction of the digital twin and the actual emission monitoring value of the equipment. The existence of this deviation means that the basic control command vector may not fully meet the emission constraints or there is room for energy consumption optimization. Therefore, the basic control command vector at time k is further compensated and corrected using the component residual to obtain the compensated control command vector.

[0064] In other words, this embodiment can determine the state residual vector that can characterize the impact of factors such as dynamic changes in operating conditions, equipment aging, and unmodeled disturbances on the control effect, and extract the component residuals that affect the pollutant emission level from it, providing a direct basis for online correction of the control strategy. This breaks through the limitation of traditional control methods that are difficult to perceive model deviations in real time, and ensures that control decisions can adapt to changes in equipment operating status in a timely manner.

[0065] Step S204: Perform amplitude limiting processing on the compensation control command vector based on physical safety boundaries to obtain the optimized control command vector for the current cycle.

[0066] Referring to the previous introduction on physical safety boundaries, in order to ensure emission safety compliance, before the compensation control command vector is finally issued to the actuator, it can be subjected to a limit processing based on physical safety boundaries through a safety constraint protection layer that is independent of the energy consumption optimization strategy network and compensation logic, so as to prevent the compensation control command vector from touching the red line.

[0067] Considering that the specific range of the physical safety boundary may vary under different operating conditions, in order to adapt to the current operating conditions, the limiting process in this application embodiment can be a dynamic limiting process adapted to the current operating conditions.

[0068] In this embodiment, key parameters affecting the physical safety boundary range, such as inlet flue gas flow rate, current ammonia escape monitoring value, and real-time dissolved oxygen value in the biological treatment tank, can be continuously monitored through a safety constraint protection layer. The safe operating range of each control channel under the current operating conditions can be dynamically calculated according to the process safety specifications. For example, the instantaneous upper limit of ammonia injection can be calculated based on flue gas flow rate and maximum allowable ammonia escape concentration, and the lower limit of aeration fan frequency can be calculated based on the dissolved oxygen safety lower limit.

[0069] If a component in the compensation control command vector exceeds the corresponding safe action range, that component is forcibly clamped to the range boundary value. The command vector after dynamic limiting is used as the optimized control command vector for the current cycle. Therefore, in this embodiment, the optimized control command vector for the current cycle can be sent to the actuator through the output register of the distributed control system, so that it can play a control role in the time period from k to k+1, completing the closed loop of one control cycle.

[0070] The multi-objective collaborative optimization control method for environmental protection equipment provided in this application takes into account that environmental protection equipment is a system with strong lag and long delay. The physical and chemical reactions that occur inside the equipment have obvious lag responses, which causes the pollutant emission results of the current cycle to be affected by the operating conditions and operating behavior of at least the previous cycle. Therefore, this application obtains the operating condition tensor used to comprehensively characterize the actual operating state and pollutant emission level of the environmental protection equipment in the previous cycle, and inputs it into a pre-trained energy consumption optimization strategy network to generate the basic control command vector for the current cycle that is energy-saving oriented.

[0071] Since the basic control command vector is designed with energy conservation as its core objective, it may not meet emission hard constraints under certain dynamic operating conditions. To avoid command unavailability due to excessive emissions, this application further predicts, based on the aforementioned operating condition tensor, the theoretical state that the environmental protection equipment should reach in the current cycle under the continued action of the optimized control command vector from the previous cycle. This involves inputting the aforementioned operating condition tensor and the optimized control command vector from the previous cycle into a pre-constructed digital twin for simulation prediction, obtaining the predicted operating condition tensor for the current cycle. Next, the measured operating condition tensor corresponding to the predicted operating condition tensor is obtained, and the component residual corresponding to the pollutant emission level in the predicted and measured operating condition tensors is determined. This component residual is then used to compensate and correct the basic control command vector, ensuring that the resulting compensated control command vector, while possessing energy-saving capabilities, adapts to fluctuations in dynamic operating conditions and meets emission hard constraints under fluctuating operating conditions.

[0072] Considering that the residual compensation process relies on factors such as the accuracy of twin predictions and the measurement accuracy of the measured operating condition tensor, unmodeled disturbances may still exist in actual industrial control, leading to slight emissions exceeding the limits. To ensure emission safety and compliance, this application adds a fallback mechanism based on physical safety boundaries. This means that the compensation control command vector is subjected to amplitude limiting processing based on physical safety boundaries, ensuring that the current-cycle optimized control command vector obtained by the environmental protection equipment under all dynamic operating conditions can stably meet the hard emission constraints while maximizing energy-saving benefits, thus achieving the safe and reliable implementation of the control strategy.

[0073] In some embodiments of this application, the process of constructing the digital twin in step S202 above is described.

[0074] To describe the basic dynamic characteristics of environmental protection equipment operation, embodiments of this application construct a digital twin based on a mechanistic model. Here, the mechanistic model can be constructed based on the physical changes and chemical reaction characteristic equations of the internal operation process of the environmental protection equipment. More specifically, it is derived and established based on the physicochemical reaction laws, fluid dynamic characteristics, and equipment structural parameters of the environmental protection equipment.

[0075] Optionally, in order to construct a mechanism model, we can first establish a set of partial differential equations based on the characteristic equations of physical changes and chemical reactions to describe the spatiotemporal distribution of state variables at various points inside the environmental protection equipment, so as to characterize the main laws of physical and chemical reactions inside the environmental protection equipment from the perspective of first principles.

[0076] Taking the gas-liquid two-phase contact mass transfer and accompanying chemical reaction process inside environmental protection equipment as an example, a set of partial differential equations as described above can be established based on the two-mode theory and the reaction kinetic rate equation. In this case, the state variables in the set of partial differential equations can include the concentration distribution of gaseous pollutants, the concentration distribution of active components of liquid absorbent, and the concentration distribution of reaction products.

[0077] For example, in a wet flue gas desulfurization absorption tower, sulfur dioxide in the flue gas diffuses from the gas phase to the gas-liquid interface, passes through the liquid film and enters the liquid phase, where it undergoes an acid-base neutralization reaction with the dissolved limestone slurry. This process exhibits significant spatial distribution characteristics, with state variables such as sulfur dioxide concentration, slurry pH, and calcium-sulfur ratio all showing gradient changes along the tower height.

[0078] Therefore, this embodiment can describe the gas-liquid interphase mass transfer flux based on the two-film theory, describe the internal chemical reaction rate of the liquid phase based on the reaction kinetic rate equation, and combine the principles of material conservation and energy conservation to establish a set of partial differential equations for each micro-element segment along the axial direction of the absorption tower. The state variables in the equation set include the distribution of gas phase sulfur dioxide concentration along the tower height, the distribution of liquid phase absorbent active component concentration along the tower height, and the distribution of reaction product concentration along the tower height.

[0079] For example, the partial differential equation for the variation of gaseous pollutant concentration along the tower height can be established by the material balance of gaseous sulfur dioxide in the infinitesimal segment, and is expressed by the following formula (1): Formula (1); in, This represents the concentration of gaseous sulfur dioxide at time t at position z on the tower. Indicates the flue gas velocity. Indicates the gas phase diffusion coefficient. Represents the volumetric mass transfer coefficient. Represents Henry's constant, This represents the concentration of liquid sulfur dioxide at time t at position z on the tower.

[0080] In the above formula (1), the left side of the equation represents the rate of change of gaseous sulfur dioxide concentration with time, and the first term on the right side of the equation ( The term in question is the convection term, describing the migration effect of sulfur dioxide carried by flue gas flow. The second term on the right side of the equation ( The term in question is the diffusion term, describing molecular diffusion and turbulent diffusion caused by the concentration gradient. The third term on the right side of the equation ( The term (in which the term is located) is an interphase mass transfer source and sink term, describing the transfer rate of sulfur dioxide from the gas phase to the liquid phase.

[0081] The partial differential equation for the variation of the concentration of the active component of the liquid absorbent along the column height can be expressed by the following formula (2): Formula (2); in, This represents the concentration of the active component of the liquid-phase absorbent at time t at position z in the tower. Indicates the slurry flow rate. Indicates the liquid phase diffusion coefficient. It represents the rate constant of a chemical reaction.

[0082] In the above formula (2), the third term on the right side of the equation ( The item in question is the chemical reaction consumption item, which describes the consumption rate of the absorbent in the neutralization reaction with dissolved sulfur dioxide.

[0083] Similarly, for wastewater treatment aeration systems, a set of partial differential equations can be established based on the activated sludge model to determine organic matter degradation, ammonia nitrogen nitrification, and dissolved oxygen transfer; for dust removal systems, a corresponding set of partial differential equations can be established based on the filtration pressure drop model and dust deposition mechanism. Of course, other forms of partial differential equations are also possible, depending on the application scenario, and will not be described in detail in this application.

[0084] Considering that partial differential equations contain spatial partial derivatives, it is difficult to directly solve them numerically and perform real-time simulations. In order to solve them efficiently using numerical integration methods, this embodiment can use the finite difference method to discretize the above partial differential equations in the spatial dimension. Specifically, the spatial derivative terms in the partial differential equations are transformed into difference algebraic relationships between state variables of adjacent points, resulting in a system of ordinary differential equations including time derivatives.

[0085] Taking the above formula (1) as an example, this embodiment can use the finite difference method to divide the continuous spatial domain into a finite number of discrete nodes along the preset equipment space axis. Let N nodes be divided along the tower height direction, and the node index be j, where j=1,2,...,N, and the distance between adjacent nodes is... Set according to the process accuracy requirements.

[0086] For the spatial first derivative term in formula (1), the upwind approximation is adopted according to the flow direction, and it is discretized by the following formula (3): Formula (3); in, This represents the concentration of gaseous sulfur dioxide at the j-th (discrete value of z) spatial node.

[0087] For the spatial second derivative term in formula (1), the central difference scheme can be used for approximation, and it can be discretized by the following formula (4): Formula (4).

[0088] After spatial discretization, the partial differential equations describing the state distribution inside the entire device are transformed into a set of ordinary differential equations containing only time derivative terms at each spatial node. For the j-th spatial node corresponding to equation (1), the ordinary differential equation for the concentration of gaseous sulfur dioxide is expressed by the following equation (5): Formula (5); in, This represents the concentration of liquid sulfur dioxide at the j-th spatial node.

[0089] In this embodiment, each ordinary differential equation describes the algebraic relationship between the rate of change of the local state variable of the node with time and the states and boundary conditions of adjacent nodes. The inlet boundary node is assigned by the inlet flue gas parameters (i.e., at the spatial node at the bottom of the absorber or the flue gas inlet, the gas phase sulfur dioxide concentration is directly given by the actual operating condition measurement value, without the need for calculation through difference equations). The outlet boundary node is treated by extrapolation boundary conditions (i.e., at the spatial node at the top of the absorber or the flue gas outlet, since this position is at the end of the computational domain, there is no downstream node to provide the adjacent node information required for difference, so it is necessary to assume that the concentration gradient tends to zero near the outlet, or to use linear extrapolation, parabolic extrapolation, etc. to extrapolate the outlet concentration from the internal nodes to ensure the closure of the numerical calculation). This method approximates the infinite-dimensional distributed parameter system as a finite-dimensional lumped parameter system, laying the foundation for subsequent numerical integration solutions.

[0090] After generating the system of ordinary differential equations in the manner described above, a mechanistic model based on the system of ordinary differential equations can be constructed.

[0091] To enable internal computation of the mechanistic model, this embodiment employs a fourth-order Runge-Kutta numerical integrator for the ordinary differential equation system. This integrator approximates the true integral curve through four trials within each simulation step (which can also be referred to as the integration step in this embodiment), exhibiting high numerical accuracy and stability.

[0092] For example, the recursive process of the fourth-order Runge-Kutta method is represented by the following formulas (6)-(10): Formula (6); Formula (7); Formula (8); Formula (9); Formula (10); in, This represents the function on the right-hand side of an ordinary differential equation, such as the function on the right-hand side of formula (5) above. Indicates the simulation step size. to This represents four intermediate slope estimates. This represents the base state vector output at the current simulation step size. This base state vector reflects the evolution trend of the internal state of the device under the current control action.

[0093] This represents the optimized control command vector of the previous cycle. This optimized control command vector of the previous cycle indirectly affects the coefficients in the system of ordinary differential equations through the mapping relationship between control actions and equipment characteristics. For example, the frequency command component of the circulating pump inverter in the optimized control command vector of the previous cycle is converted into slurry flow rate through the pump characteristic curve. This, in turn, affects the system dynamics.

[0094] This represents the initial state of the current simulation step. This initial state is a high-dimensional state vector containing a complete set of all spatial nodes and all state variables, such as all j values. A set of.

[0095] As shown in formulas (6)-(10) above, the fourth-order Runge-Kutta numerical integrator receives two sets of inputs at the beginning of each simulation step: one is the optimized control command vector of the previous cycle. Secondly, it is the simulation state vector output at the end of the previous simulation. (For details on the determination process, please refer to the coupling method section below). The fourth-order Runge-Kutta numerical integrator first processes... Transform it into the correlation coefficient in a system of ordinary differential equations, and then use... As the initial values ​​for the integrals of each state variable, the state variable values ​​of each spatial node in the next simulation step are solved recursively according to the fourth-order Runge-Kutta scheme.

[0096] Furthermore, this embodiment pre-trains a data-driven model based on the mechanistic model. This data-driven model can use long short-term memory networks, graph neural networks, etc., to compensate for the difficulty of the mechanistic model in representing nonlinear and time-varying characteristics (such as efficiency decay caused by equipment aging).

[0097] Specifically, the data-driven model can be a deep residual compensation network.

[0098] Optionally, the deep residual compensation network can adopt a convolutional neural network architecture containing multiple residual blocks. Each residual block alleviates the gradient vanishing problem in deep network training through a skip connection mechanism. The input layer dimension (i.e., the number of neurons) of the network matches the concatenation dimension of the basic state vector and the time-aligned working condition tensor output by the mechanism model, and the output layer dimension is consistent with the dimension of the basic state vector output by the mechanism model, ensuring that the correction amount can be directly superimposed and applied.

[0099] Optionally, the deep residual compensation network can also employ a temporal convolutional network structure with a gating mechanism. This structure constructs a deep network by stacking multiple residual modules. Each residual module contains dilated convolutional layers, gated linear unit activation layers, and skip connections. The dilated convolutional layers expand the receptive field by inserting holes between convolutional kernel elements, enabling top-level neurons to directly perceive input information from tens or even hundreds of time steps ago, thereby effectively capturing the large-latency dynamic characteristics of environmental protection equipment. The gated linear units perform element-wise weighting of the convolutional output through a learnable gating vector, selectively allowing feature components that contribute to residual prediction while suppressing irrelevant noise. The forward propagation process of the network starts from receiving the concatenated feature vector at the input layer, sequentially passes through multiple residual blocks for feature extraction and transformation, and finally generates the state correction value by the fully connected output layer.

[0100] Based on this, this embodiment can couple the deep residual compensation network with the mechanism model to obtain a digital twin. The coupling method is as follows: within each simulation step, the basic state vector obtained by integral solution of the mechanism model and the working condition tensor (preferably the time-aligned working condition tensor) in step S201 are jointly input into the deep residual compensation network to obtain the state correction amount. The simulation state vector is generated by the state correction amount and the basic state vector and used as the initial state of the next simulation step. The integral solution of the mechanism model is based on the initial state of the current simulation step and the optimized control command vector of the previous cycle.

[0101] Optionally, the basic state vector obtained by integral solution of the mechanism model can be concatenated with the working condition tensor to form a composite input vector, which is then input into the deep residual compensation network. The network performs forward propagation to calculate and output the corresponding state correction amount, which is used to compensate for the prediction deviation caused by simplification assumptions, parameter uncertainty, nonlinearity, time-varying characteristics, etc. in the mechanism model.

[0102] Optionally, the process of "generating the simulation state vector from the state correction and the basic state vector" can be as follows: add the state correction and the basic state vector element by element to obtain the simulation state vector.

[0103] In this embodiment, multiple simulations can be performed in each cycle. Therefore, the simulation state vector under the last simulation step of the current cycle is the predicted operating condition tensor in step S202 above.

[0104] In this embodiment, a digital twin is constructed by coupling the mechanistic model with a deep residual compensation network, which enables the digital twin to more accurately reproduce the internal physicochemical reactions and flue gas transmission lag of environmental protection equipment, achieving higher accuracy compared to a single mechanistic model or a pure data model.

[0105] In an optional embodiment, in order to train a deep residual compensation network as described above, this embodiment can adopt a supervised learning approach, using historical data to train a deep residual compensation network capable of predicting the residuals of the mechanism model, such as a temporal convolutional network with a gating mechanism.

[0106] Specifically, training data is first acquired, including: historical operating condition tensor samples, optimized control command vector samples executed in the same period as the historical operating condition tensor samples, and measured operating condition tensor samples obtained from feedback by the environmental protection equipment after executing the optimized control command vector samples. Among them, the measured operating condition tensor samples are one period later than the historical operating condition tensor samples.

[0107] Optionally, the massive amounts of historical operating condition tensor samples, optimized control command vector samples, and measured operating condition tensor samples accumulated by environmental protection equipment during long-term operation can be stored in a historical operation database, which can be retrieved when needed.

[0108] Optionally, the training data mentioned above can be stored in the historical operation database in chronological order to form a complete data link. In order to enable the training data to cover the operating modes under different load ranges, different seasonal operating conditions, and different equipment aging stages, and to enable the deep residual compensation network to have sufficient operating condition generalization ability, this embodiment can continuously extract multiple data segments from the data link of the historical operation database. Each data segment includes more than one training data, and the time length of each data segment covers the complete dynamic response process of the environmental protection equipment transitioning from one steady state to another.

[0109] For example, multiple data segments covering typical stages such as high-load summer periods, low-load winter periods, and the initial and final stages of catalyst commissioning can be extracted to fully train the deep residual compensation network using the training data contained in these data segments.

[0110] Specifically, the optimized control command vector samples from the training data are input into the mechanism model to obtain the basic state vector samples. Then, based on the basic state vector samples and the measured working condition tensor samples from the training data, state correction samples are generated. The historical working condition tensor samples and the basic state vector samples are used as training samples, and the state correction samples are used as sample labels. The deep residual compensation network is trained using the training samples and sample labels.

[0111] Specifically, since the task of the deep residual compensation network is to learn the mapping relationship between the prediction bias of the mechanism model and the operating conditions, the input of the training samples must contain information reflecting the previous operating condition and information reflecting the prediction results of the mechanism model under the current control action. That is, historical operating condition tensor samples and basic state vector samples are needed as training samples. Correspondingly, the measured operating condition tensor samples at the same time as the basic state vector samples are subtracted element by element to obtain the state correction samples. These state correction samples directly reflect the prediction bias of the mechanism model at each stage, and therefore can be used as sample labels (i.e., training labels) for the deep residual compensation network.

[0112] Optionally, the sample labels can be constructed using the following formula (11): Formula (11); in, This represents the state correction sample corresponding to the i-th training data. This represents the tensor sample of the measured working condition corresponding to the i-th training data. This represents the basic state vector sample corresponding to the i-th training data.

[0113] In this embodiment, the weight parameters of the deep residual compensation network are iteratively optimized using the backpropagation algorithm in conjunction with the adaptive moment estimation optimizer until the loss function converges to a preset range or no longer decreases significantly for multiple consecutive rounds. The network weight parameters at this time are saved to obtain the trained deep residual compensation network.

[0114] Optionally, the loss function can consist of a mean squared error loss term and an L2 (Euclidean norm) regularization term for the network weight gradients, and the expression for the loss function is as follows: Formula (12); in, This represents the total loss during training of the deep residual compensation network. The smaller this value, the higher the fitting accuracy of the deep residual compensation network to the residuals. This indicates the number of samples in the training batch, i.e., the training data sent to the deep residual compensation network for gradient calculation each time. This represents the predicted residual vector output by the deep residual compensation network for the training sample in the i-th training data. The mean squared error term represents the square of the Euclidean distance between the state correction sample and the predicted residual vector. This represents the regularization penalty coefficient; The weight parameters of the deep residual compensation network The gradient of the loss function without regularization, i.e., the partial derivative of the loss function with respect to the network weights when only the mean squared error term is considered; This represents the square of the L2 norm of the gradient.

[0115] Based on this, the optional process of "training the deep residual compensation network using training samples and sample labels" may include: inputting training samples into the deep residual compensation network to obtain the predicted residual vector; generating a mean squared error loss based on the predicted residual vector and sample labels (i.e., state correction samples); calculating the gradient of the mean squared error loss with respect to the parameters of the deep residual compensation network; generating a regularization loss based on the gradient; generating a total loss using the mean squared error loss and the regularization loss; and training the deep residual compensation network using the total loss.

[0116] In this embodiment, the regularization term imposes a penalty on the L2 norm of the loss function gradient, constraining the variation of network weights during training, preventing the deep residual compensation network from overfitting to noise and abnormal samples in the training data, and improving the generalization ability of the deep residual compensation network under unseen conditions.

[0117] As described above, the energy optimization strategy network can be trained in a simulation environment provided by a digital twin. In further embodiments of this application, the training process in the simulation environment is described.

[0118] In order to enable the energy consumption optimization strategy network to implicitly learn the complex mapping relationship between minimizing energy consumption (i.e. energy saving) and meeting emission standards under different operating conditions, this embodiment can first generate multiple twin instances corresponding to different environmental protection equipment operating conditions based on digital twins.

[0119] Specifically, in order to improve the robustness and generalization ability of the energy consumption optimization strategy network under the complex and variable operating conditions of real devices, this embodiment can start multiple independent digital twin instances to run in parallel during the training phase, with each instance corresponding to an independent copy of the simulation environment.

[0120] Different equipment aging degradation coefficients are configured for different twin instances, such as catalyst activity degradation factor, heat exchanger fouling thermal resistance increment, and aerator blockage coefficient, so that each twin instance can simulate the dynamic characteristics of environmental protection equipment at different service stages. Optionally, the equipment aging degradation coefficient can be set according to typical aging mechanisms such as catalyst activity degradation, membrane fouling degree, and mechanical component wear. Its value range is determined based on the equipment design life and historical aging data, and this embodiment does not impose specific limitations.

[0121] Different inlet load disturbance curves can be configured for different twin instances. Optionally, the inlet load disturbance curve is generated by piecewise linear interpolation or a stochastic process based on the actual operating conditions such as flue gas flow fluctuations and influent water quality changes in industrial production. Furthermore, the parameter configurations of different twin instances cover the entire operating range from the initial state of new equipment to the late stage of aging, and from low load to high load.

[0122] Optionally, this embodiment can implement reinforcement learning training of the energy consumption strategy optimization network based on each twin instance. To this end, each twin instance can interact with the energy consumption optimization strategy network to obtain interactive data including the current state, current action, next state, reward, and constraint cost.

[0123] In this embodiment, the twin instance can generate a predicted operating condition tensor according to the process described above. The predicted operating condition tensor generated by the twin instance can be used as the current state, and the current state can be input into the energy consumption optimization strategy network to obtain the basic control command vector corresponding to the current state, which is used as the current action.

[0124] Optionally, the energy consumption optimization strategy network can generate action samples conforming to a Gaussian distribution based on the predicted operating condition tensor fed back from the twin instance. For example, action samples can be generated using the following formula (13): Formula (13); in, This represents the generated action sampling vector, i.e., the current action. Indicates a Gaussian distribution. This represents the action mean vector output by the energy optimization strategy network. This represents the action standard deviation vector output by the energy optimization strategy network. The weight parameters represent the network weights for the energy optimization strategy. The tensor representing the predicted load condition feedback from the twin instance. Represents the identity matrix.

[0125] In this embodiment, a balance between strategy exploration and utilization can be achieved by controlling probabilistic action sampling within the feasible region of control parameters.

[0126] Next, the current action and current state can be input into the twin instance to obtain the next operating condition tensor, which serves as the next state. Then, based on the next state, a composite reward function value and a constraint cost function value are generated. The composite reward function value is used to guide the energy consumption optimization strategy network to generate an energy-saving-oriented basic control command vector, while the constraint cost function value is used to constrain the energy consumption optimization strategy network to generate a basic control command vector that meets the pollutant emission constraint requirements.

[0127] Optionally, the composite reward function value can be calculated using the following formula (14): Formula (14); in, This represents the value of the composite reward function at time t. This represents the energy consumption reward weighting coefficient (with a value range of 0-1). This represents the energy consumption reward at time t (which is negatively correlated with the actual energy consumption of the device; the lower the energy consumption, the greater the reward value). This represents the operational efficiency bonus at time t (which is positively correlated with pollutant removal efficiency).

[0128] Optionally, the constraint cost function value can be calculated using the following formula (15): Formula (15); in, This represents the value of the constraint cost function at time t. This represents the simulated pollutant emission concentration at time t. This represents the preset pollutant emission constraint threshold. When the simulated pollutant emission concentration does not exceed the standard, the constraint cost function value is 0. When it exceeds the standard, the cost is quantified according to the extent of the exceedance, providing a clear basis for rewards and penalties for subsequent strategy updates.

[0129] Among them, the above-mentioned , , All can be obtained based on the next state.

[0130] Considering that not all interaction data may be high-value experience data, in order to make the training process of the energy consumption optimization strategy network more focused on high-value experience data, this embodiment can store the interaction data generated by the interaction between the twin instance and the energy consumption optimization strategy network in a priority experience replay pool. Then, in each training round of the training process, the interaction data in the priority experience replay pool can be sampled, and the sampled interaction data can be arranged in chronological order to form at least one interaction trajectory. The state sequence corresponding to the interaction trajectory is then input into the reward value network and the constraint value network respectively. The reward advantage function value is calculated based on the cumulative reward estimate output by the reward value network, and the constraint advantage function value is calculated based on the cumulative cost estimate output by the constraint value network.

[0131] In one possible implementation, in order to sample high-value experience data from the priority experience replay pool, this embodiment can input the current state and the next state in the interaction data into the reward value network to obtain the cumulative reward estimate corresponding to the current state and the cumulative reward estimate corresponding to the next state. Then, based on the composite reward function value in the interaction data, the cumulative reward estimate corresponding to the current state, and the cumulative reward estimate corresponding to the next state, a temporal difference error corresponding to the interaction data is generated. Based on the temporal difference error corresponding to each interaction data in the priority experience replay pool, the sampling priority corresponding to each interaction data in the priority experience replay pool is determined, and the interaction data in the priority experience replay pool is sampled based on the sampling priority.

[0132] Optionally, the timing difference error can be calculated using the following formula (16): Formula (16); in, This represents the temporal difference error of the interactive data at time t. This represents the reward discount factor (with a value ranging from 0 to 1, used to weigh the importance of immediate rewards against future rewards). This represents the reward value network's state at time t+1. The cumulative reward estimate, This represents the reward value network for the state at time t. The cumulative reward estimate, This represents the weight parameters of the reward value network.

[0133] In this embodiment, the larger the absolute value of the timing difference error, the more valuable the policy optimization information contained in the interaction data is, and the higher the probability that the interaction data will be sampled.

[0134] Optionally, the sampling probability of the interactive data can be determined by the following formula (17): Formula (17); in, This represents the sampling probability of the interactive data at time t. , β represents a preset minimum positive number used to avoid a sampling probability of 0; β represents the priority weight coefficient (ranging from 0 to 1, used to adjust the intensity of priority sampling). This represents the sum of the priority weights of all interactive data.

[0135] In this embodiment, the higher the sampling probability of an interaction data, the higher the effective value of the interaction data for the energy consumption optimization strategy network, and thus the higher the sampling priority of the interaction data. Based on the sampling priority, high-value interaction data can be sampled from the priority experience replay pool.

[0136] In this embodiment, sampling high-value interaction data for training the energy consumption optimization strategy network can effectively improve training efficiency and convergence speed.

[0137] Optionally, the aforementioned priority experience replay pool can use a circular buffer structure to store interactive data. When the amount of data reaches the preset upper limit, the old data is replaced according to the first-in-first-out principle to ensure that the data in the pool always maintains a certain timeliness and diversity, and to avoid the training process from getting stuck in local optima.

[0138] Optionally, the process of "calculating the reward advantage function value based on the cumulative reward estimate output by the reward value network" mentioned above can be specifically as follows: calculate the time series difference error as described above based on the cumulative reward estimate output by the reward value network, and then obtain the reward generalized advantage estimate based on the time series difference error, which is used as the reward advantage function value.

[0139] Optionally, the formula for calculating the reward generalized advantage estimate is as follows: Formula (18); in, Let t represent the generalized advantage estimate of the reward at time t, which is used to quantify the degree of advantage of the current action relative to the average action level; This represents the reward advantage estimation decay coefficient (ranging from 0 to 1, used to balance the bias and variance of the reward advantage estimation). This represents the time step offset, used to mark the first time step after the current time t. One cycle; Indicates the combined discount attenuation weight, superscript This indicates the offset of the joint discount decay weight over the time step. The increase in decays exponentially, gradually reducing the impact of errors at longer time points on the estimation of current reward advantage; T represents the total duration of a single interaction trajectory. This represents the time-series difference error at time t+1.

[0140] In this embodiment, by introducing a reward discount factor and a reward advantage estimation decay coefficient, the immediate reward and future reward are weighted and summed, which can effectively reduce the variance of the reward advantage estimation and improve the estimation accuracy, thereby enabling a more accurate quantification of the advantage of the current action relative to the average action level.

[0141] Optionally, the constraint advantage function value in this embodiment can be a constraint generalized advantage estimate, calculated using the following formula: Formula (19); in, This represents the constraint generalized advantage estimate at time t, used to quantify the advantage of the current action in meeting pollutant emission constraints; This represents the attenuation coefficient of the constraint-side advantage estimate, used to adjust the bias and variance of the constraint advantage estimate; Indicates the joint discount decay weight in the constraint direction, superscript. This represents the offset of the joint discount decay weight in the constraint direction with respect to the time step. The increase in decreases exponentially.

[0142] Let represent the constrained time series difference error at time t+1. The formula for calculating this constrained time series difference error is: ,in, This represents the value of the constraint cost function at time t+1. Represents the constraint value network on the t+th time. +1 time state The cumulative cost estimate, This represents the constrained value network with respect to the state at time t+1. The cumulative cost estimate, This represents the weight coefficients that constrain the value network.

[0143] As mentioned above, in this embodiment, the constraint generalized advantage estimate adopts a similar calculation logic to the reward generalized constraint estimate, focusing on the cumulative effect of the constraint cost sequence, quantifying the performance difference of the current action in meeting emission constraints, and the two types of advantage function values ​​together provide dual optimization guidance for policy updates, so that the energy consumption optimization policy network can better take into account both energy consumption and emission targets.

[0144] Optionally, the reward value network and constraint value network mentioned above can adopt a deep neural network structure similar to the energy consumption optimization strategy network, but only output a single scalar value in the output layer, which represents the expected cumulative discount reward and the expected cumulative discount constraint cost that can be obtained in the future from the current state, respectively.

[0145] To optimize the training of the reward value network, this embodiment can optionally use the expected cumulative reward generated based on the reward sequence corresponding to the interaction trajectory as the supervision objective, and update the parameters of the reward value network by minimizing the difference between the estimated cumulative reward and the expected cumulative reward.

[0146] Optionally, the expected cumulative reward can be calculated using the following formula (20): Formula (20); in, This represents the expected cumulative reward at time t. This represents the value of the composite reward function at time t+1.

[0147] Optionally, the difference between the above-mentioned cumulative reward estimate and the expected cumulative reward can be specifically represented as the mean squared error loss, calculated using the following formula: Formula (21); in, This represents the mean squared error loss of the reward value network. This indicates the number of interaction data points in the interaction trajectory. This represents the expected cumulative reward for the i-th interaction data at time t in the interaction trajectory. This represents the state of the reward value network at time t for the i-th interaction data. The estimated cumulative reward.

[0148] Similarly, in this embodiment, the cumulative cost expectation generated based on the constraint cost sequence corresponding to the interaction trajectory can be used as the supervision objective. By minimizing the difference between the cumulative cost estimate and the cumulative cost expectation, the parameters of the constraint value network can be updated.

[0149] Optionally, the expected cumulative cost can be calculated using the following formula (22): Formula (22); in, Let t represent the expected cumulative cost at time t.

[0150] Optionally, the difference between the above-mentioned cumulative cost estimate and the expected cumulative cost can be specifically the mean squared error loss, calculated using the following formula: Formula (23); in, This represents the mean squared error loss of the constrained value network. Let $t$ represent the expected cumulative cost of the $i$-th interaction data point at time $t$ in the interaction trajectory. This represents the state of the constrained value network at time t for the i-th interaction data. The cumulative cost estimate.

[0151] In this embodiment, the core function of the reward value network is to accurately predict the future cumulative reward in the current state, while the constraint value network focuses on predicting the future cumulative constraint cost in the current state. The update processes of the two are carried out independently and iterate synchronously. By minimizing the mean square error between the predicted value and the actual cumulative value, the accuracy of the network estimation is ensured, providing a reliable value assessment basis for subsequent policy optimization.

[0152] Given the reward advantage function value and the constraint advantage function value, the energy consumption optimization strategy network can be updated based on these two function values ​​to complete one round of training.

[0153] Optionally, this embodiment can use a constrained proximal policy optimization algorithm to update the policy network parameters. That is, a reward pruning objective is generated based on the reward advantage function value, and a constraint pruning objective is generated based on the constraint advantage function value. The energy consumption optimization policy network is updated with the goal of maximizing the Lagrange weighted difference between the reward pruning objective and the constraint pruning objective.

[0154] Optionally, the formula for calculating the reward pruning target (also known as the near-end policy optimization reward pruning target) is as follows: Formula (24); in, This indicates a reward for the cut-off target. This represents the estimate of the generalized advantage of rewards. Expectations This represents the ratio of the probability of actions under the new and old strategies. This indicates the current energy consumption optimization strategy network is in state. Next generation action The probability, This indicates the state of the network before the energy optimization strategy update. Next generation action The probability, This represents the cropping factor (usually between 0.1 and 0.2). This represents the clipping function, used to restrict the ratio of action probabilities to a certain value. Within a certain range, avoid excessively large policy updates that could lead to training instability.

[0155] Optionally, the formula for calculating the constraint pruning objective (also known as the near-end policy optimization constraint pruning objective term) is as follows: Formula (25); in, This indicates the constraint on the clipping target. This represents the estimate of the constrained generalized advantage. The expectation.

[0156] Optionally, the Lagrange weighted difference between the reward pruning objective and the constraint pruning objective can be calculated using the following formula (26): Formula (26); in, This represents the Lagrange weighted difference between the reward pruning objective and the constraint pruning objective, which is the objective function for policy update; This represents the Lagrange multiplier, used to adjust the weight ratio between reward optimization and constraint satisfaction, thereby achieving a dynamic balance between energy consumption optimization and emission constraints.

[0157] In this embodiment, the gradient ascent update process of the energy consumption optimization strategy network can employ an adaptive moment estimation optimizer, which calculates the weight parameters of the objective function on the energy consumption optimization strategy network. The gradient is used to iteratively update the parameters according to the preset learning rate, so that the policy network gradually learns an optimal control strategy that takes into account both energy consumption and emissions.

[0158] The above process completes one round of training for the energy consumption optimization strategy network. At this point, the process can return to sampling the interaction data in the priority experience replay pool until the preset training termination condition is met.

[0159] Optionally, the training termination condition can be that the performance improvement rate of the energy consumption optimization strategy network updates for X consecutive rounds is less than a preset improvement rate threshold, where X is a preset value.

[0160] In a preferred implementation, in order to train the energy consumption optimization strategy network to balance the energy-saving optimization and emission hard constraints, this embodiment can also use the Lagrange multipliers used to calculate the Lagrange weighted difference to be updated after each update of the energy consumption optimization strategy network.

[0161] Optionally, the update process of the Lagrange multiplier can be as follows: calculate the average constraint cost expectation value based on the constraint cost function values ​​in each interaction data contained in at least one interaction trajectory, calculate the deviation of the average constraint cost expectation value relative to the preset emission constraint threshold, and update the Lagrange multiplier based on the deviation.

[0162] Optionally, the expected value of the average constraint cost is calculated using the following formula (27): Formula (27); in, This represents the expected average constraint cost, used to comprehensively assess the level at which the current strategy meets emission constraints. Indicates the number of interaction trajectories. Let T represent the constraint cost function value of the i-th interaction trajectory at time t, and let T represent the total duration of a single interaction trajectory.

[0163] Therefore, the adaptive adjustment of the Lagrange multipliers can be achieved through the following formula: Formula (28); in, Denotes the Lagrange multiplier in the (q+1)th iteration. Let q represent the Lagrange multiplier in the q-th iteration, and b represent the multiplier adjustment step size (ranging from 0 to 1). This represents the maximum permissible average constraint cost threshold corresponding to the preset emission constraints.

[0164] As shown in formula (28), when the expected value of the average constraint cost is... Exceed When the emission constraints increase, the Lagrange multiplier increases, strengthening the guiding role of emission constraints in strategy updates; conversely, when the emission constraints decrease, the Lagrange multiplier remains stable or decreases moderately, making the strategy more focused on energy consumption optimization.

[0165] Optionally, to prevent the Lagrange multiplier from decreasing indefinitely and causing emission constraints to fail, this embodiment can also preset a lower limit value for the Lagrange multiplier, within the expected value of the average constraint cost. Below However, when the value is greater than or equal to the lower limit, the strategy focuses more on energy consumption optimization. This adaptive adjustment mechanism ensures that the strategy network always prioritizes meeting the hard emission constraints under different operating conditions, while continuously approaching the energy consumption optimum.

[0166] The energy consumption optimization strategy network obtained in this embodiment has internalized the dynamic trade-off between energy consumption and emissions under complex operating conditions. It has the ability to generate safe and energy-saving basic control command vectors under unseen operating conditions. The training process relies on a digital twin to simulate the operating characteristics under different loads, equipment aging, inlet disturbances, and extreme operating conditions offline. It makes full use of the safe, controllable, and parallelizable rich simulation environment provided by the digital twin. A reliable initial strategy can be obtained without any trial and error exploration on real equipment. This solves the contradiction between safety and strategy optimization in the process of applying reinforcement learning in the field of environmental protection equipment. The trained energy consumption optimization strategy network can be directly deployed in online controllers, making it more practical.

[0167] In some other embodiments of this application, the process of “using component residuals to compensate and correct the basic control command vector to obtain a compensated control command vector” described above is introduced.

[0168] Understandably, the residual components of the predicted operating condition tensor and the measured operating condition tensor corresponding to the pollutant emission level (i.e., pollutant emission concentration) are themselves expressed in terms of physical dimensions as state deviations. They need to be converted into control compensation quantities for the corresponding control channels before they can be superimposed on the basic control command vector.

[0169] Therefore, in this embodiment, the component residuals can be mapped to control correction amplitudes using a preset mapping strategy, which can then be used as control compensation amounts.

[0170] Optionally, the mapping strategy can be a piecewise linear function, a saturated exponential function, a preset nonlinear mapping function, or a small feedforward neural network trained with lightweight methods.

[0171] Taking a preset nonlinear mapping function (such as the Sigmoid function, piecewise linear function, or exponential function) as an example, the nonlinear mapping function can be designed according to the emission response characteristics of the environmental protection equipment. This allows for precise matching between the residual and the control correction magnitude through nonlinear mapping, avoiding excessive control correction magnitude that could cause system oscillation when the residual is small, while ensuring that sufficient correction magnitude can be output quickly when the residual is large to suppress the expansion of the deviation.

[0172] For example, a nonlinear mapping function can be: Formula (29); in, This represents the output value of the nonlinear mapping function (i.e., the control correction magnitude). denoted as the component residual corresponding to the pollutant emission level, and d represents the nonlinear adjustment coefficient.

[0173] For example, a nonlinear mapping function is designed for the residual component of emission concentration. When the residual is small and within the safety margin, the mapping output is close to zero, avoiding frequent jitter of control commands due to small fluctuations. When the residual approaches or exceeds the preset threshold direction, the mapping output increases rapidly in an exponential or piecewise linear form, providing sufficient compensation.

[0174] By setting a nonlinear mapping function, the residual emission concentration can be mapped to a correction range that conforms to the control characteristics of the equipment according to a nonlinear law, thus ensuring the smoothness and effectiveness of control compensation.

[0175] After obtaining the control compensation amount, the control compensation amount can be further superimposed on the basic control command vector according to the proportional integral law to generate the compensation control command vector.

[0176] Optionally, the superposition process here can use the following formula: Formula (30); in, This represents the compensation control command vector, whose dimension is the same as the basic control command vector. It can cover the corrected set values ​​of all control parameters such as fan speed, pump and valve opening, and chemical dosing flow rate. Represents the basic control command vector; This represents the preset proportional gain coefficient, used to adjust the instantaneous response speed of the control compensation. This represents the preset integral gain coefficient, which is used to accumulate the control correction magnitude over the time dimension, eliminate static deviations, and avoid the accumulation of emission deviations during long-term operation. This represents the integral accumulation of the control compensation over time, calculated over each control cycle. The value is continuously integrated to achieve gradual correction of persistent deviations. This superposition process responds quickly to the current emission deviation through the proportional term and compensates for long-term accumulated deviations through the integral term. This allows the compensation control command vector to retain the energy-saving optimization characteristics of the basic control command vector while ensuring stable compliance of emission indicators through dynamic correction, effectively balancing the dual needs of energy consumption optimization and emission constraints.

[0177] The above embodiments transform the state residual between digital twin prediction and real device feedback into an effective correction of reinforcement learning basic instructions. The design of the preset nonlinear mapping function takes into account both control stability and response timeliness. The dead zone mechanism avoids instruction fluctuations caused by sensor noise or small prediction errors. The exponential or piecewise linear growth mechanism ensures rapid correction capability under large deviation conditions. The proportional-integral superposition structure organically combines instantaneous correction and steady-state error elimination, so that the entire control closed loop has both rapid response capability to operating condition fluctuations and can maintain accurate and stable emission concentrations during long-term operation. The compensation control command vector will also be subject to dynamic amplitude limiting protection based on physical safety boundaries in subsequent steps to obtain a safe, energy-saving, and adaptive optimized control command vector.

[0178] In summary, the multi-objective collaborative optimization control method for energy consumption optimization and emission hard constraints of environmental protection equipment based on reinforcement learning training provided in this application eliminates the lag differences between variables by processing multi-source operating parameters and emission monitoring data through time-series alignment to form a high-quality time-series aligned operating condition tensor. Then, using an energy consumption optimization strategy network, a basic control command vector oriented towards energy saving is generated by forward inference based on this operating condition tensor, thereby realizing intelligent decision-making for complex multi-objective trade-offs.

[0179] Simultaneously, leveraging a digital twin of the fusion mechanism model and deep residual compensation network, the system performs simulation prediction by inputting the operating condition tensor and the optimized control command vector from the previous cycle. Combined with the actual measured conditions of real equipment, it calculates the component residuals corresponding to pollutant emission levels, accurately capturing dynamic deviations caused by changes in operating conditions and equipment aging. Subsequently, through a pre-set nonlinear mapping and proportional-integral law, the component residuals are transformed into control compensation quantities and superimposed on the basic control command vector. Finally, the optimized control command vector is generated by dynamic limiting at physical safety boundaries.

[0180] Furthermore, during the training process of the energy consumption optimization strategy network, multi-condition parallel simulation, priority experience playback, dual-value network collaborative estimation, and Lagrange adaptive weighting mechanism are used to ensure that the network continuously learns the optimal control strategy while meeting the hard emission constraints.

[0181] The above approach deeply integrates the online learning capabilities of reinforcement learning with the predictive verification capabilities of digital twins and the residual-driven closed-loop compensation mechanism. This enables the control strategy to evolve online adaptively with changing operating conditions. It strictly adheres to the hard constraints of pollutant emissions while continuously approaching the optimal energy consumption operating point. This effectively addresses the pain points of traditional control strategies, such as rigidity, insufficient multi-objective trade-off capabilities, and reliance on manual updates. It significantly improves the adaptive control capabilities and multi-objective collaborative optimization level of environmental protection equipment under dynamic operating conditions, providing reliable technical support for the efficient and low-consumption operation of environmental protection equipment throughout its entire life cycle.

[0182] The above describes a multi-objective collaborative optimization control method for environmental protection equipment provided by the embodiments of this application. The following will describe the apparatus for implementing the above-described multi-objective collaborative optimization control method for environmental protection equipment.

[0183] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a multi-objective collaborative optimization control device for environmental protection equipment provided in an embodiment of this application. Figure 3 As shown, the multi-objective collaborative optimization control device for this environmental protection equipment may include: The instruction prediction unit 301 is used to obtain the operating condition tensor that comprehensively characterizes the actual operating status and pollutant emission level of the environmental protection equipment in the previous cycle, and input it into the pre-trained energy consumption optimization strategy network to obtain the basic control instruction vector of the current cycle with energy saving as the guide. The operating condition tensor prediction unit 302 is used to input the operating condition tensor and the optimized control command vector of the previous cycle into a pre-constructed digital twin for simulation prediction to obtain the predicted operating condition tensor of the current cycle. The instruction compensation unit 303 is used to obtain the measured operating condition tensor corresponding to the predicted operating condition tensor, determine the component residuals corresponding to the pollutant emission level in the predicted operating condition tensor and the measured operating condition tensor, and use the component residuals to compensate and correct the basic control instruction vector to obtain the compensated control instruction vector. The instruction optimization unit 304 is used to perform amplitude limiting processing on the compensation control instruction vector based on physical safety boundaries to obtain the optimized control instruction vector for the current period.

[0184] The modules in the multi-objective collaborative optimization control device of the aforementioned environmental protection equipment can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0185] This application also provides an electronic device, which may include at least one processor and a memory connected to the processor, wherein: Memory is used to store computer programs; The processor is used to execute computer programs to enable electronic devices to implement the multi-objective collaborative optimization control method for any of the environmental protection devices provided in the embodiments of this application.

[0186] refer to Figure 4 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 4 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0187] like Figure 4 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0188] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, memory cards, hard drives, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0189] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement a multi-objective collaborative optimization control method for any of the environmental protection devices provided in this application.

[0190] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the multi-objective collaborative optimization control methods for environmental protection equipment provided in this application.

[0191] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0192] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0193] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0194] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A multi-objective collaborative optimization control method for environmental protection equipment, characterized in that, include: Obtain the operating condition tensor used to comprehensively characterize the actual operating status and pollutant emission level of the environmental protection equipment in the previous cycle, and input it into the pre-trained energy consumption optimization strategy network to obtain the basic control command vector for the current cycle with energy saving as the guide. The operating condition tensor and the optimized control command vector of the previous cycle are input into a pre-constructed digital twin for simulation prediction to obtain the predicted operating condition tensor for the current cycle. Obtain the measured operating condition tensor corresponding to the predicted operating condition tensor, determine the component residuals corresponding to the pollutant emission level in the predicted operating condition tensor and the measured operating condition tensor, and use the component residuals to compensate and correct the basic control command vector to obtain the compensated control command vector. The compensation control command vector is subjected to amplitude limiting processing based on physical safety boundaries to obtain the optimized control command vector for the current cycle; The energy consumption optimization strategy network is trained in the simulation environment provided by the digital twin, and the training process in the simulation environment includes: Based on the digital twin, multiple twin instances corresponding to different environmental protection equipment operating conditions are generated, and the twin instances interact with the energy consumption optimization strategy network to obtain interactive data including the current state, current action, next state, reward and constraint cost, which is stored in the priority experience replay pool. The interaction data in the priority experience replay pool is sampled, and at least one interaction trajectory is formed by the sampled interaction data in chronological order. The state sequence corresponding to the interaction trajectory is input into the reward value network and the constraint value network respectively. The reward advantage function value is calculated based on the cumulative reward estimate output by the reward value network, and the constraint advantage function value is calculated based on the cumulative cost estimate output by the constraint value network. Using the expected cumulative reward generated from the reward sequence corresponding to the interaction trajectory as the supervision objective, the parameters of the reward value network are updated by minimizing the difference between the estimated cumulative reward and the expected cumulative reward; using the expected cumulative cost generated from the constraint cost sequence corresponding to the interaction trajectory as the supervision objective, the parameters of the constraint value network are updated by minimizing the difference between the estimated cumulative cost and the expected cumulative cost. A reward pruning objective is generated based on the reward advantage function value, and a constraint pruning objective is generated based on the constraint advantage function value. The goal is to maximize the Lagrange weighted difference between the reward pruning objective and the constraint pruning objective. The energy consumption optimization strategy network is then updated, and the sampling step of the interaction data is returned until the preset training termination condition is reached.

2. The multi-objective collaborative optimization control method for environmental protection equipment according to claim 1, characterized in that, The process of constructing the digital twin includes: A mechanism model is constructed based on the physical changes and chemical reaction characteristics equations of the internal operation process of environmental protection equipment; A pre-trained deep residual compensation network is obtained, and the deep residual compensation network is coupled with the mechanistic model to obtain the digital twin. The coupling method is as follows: within each simulation step, the basic state vector obtained by the integral solution of the mechanistic model and the operating condition tensor are jointly input into the deep residual compensation network to obtain a state correction quantity. The simulation state vector is generated by the state correction quantity and the basic state vector and used as the initial state of the next simulation step. The integral solution of the mechanistic model is based on the initial state of the current simulation step and the optimized control command vector of the previous cycle. The simulation state vector under the last simulation step of the current cycle is used as the predicted operating condition tensor.

3. The multi-objective collaborative optimization control method for environmental protection equipment according to claim 2, characterized in that, The mechanism model is constructed based on the physical changes and chemical reaction characteristic equations of the internal operation process of environmental protection equipment, including: Based on the aforementioned physical change and chemical reaction characteristic equations, a set of partial differential equations is established to describe the spatiotemporal distribution of state variables at various points inside the environmental protection equipment. The partial differential equations are discretized in the spatial dimension using the finite difference method, so as to transform the spatial derivative terms in the partial differential equations into difference algebraic relationships between state variables of adjacent points, and obtain ordinary differential equations including time derivatives. A mechanistic model based on the aforementioned set of ordinary differential equations is constructed.

4. The multi-objective collaborative optimization control method for environmental protection equipment according to claim 2, characterized in that, The training process of the deep residual compensation network includes: Acquire historical operating condition tensor samples, optimized control command vector samples executed in the period where the historical operating condition tensor samples are located, and measured operating condition tensor samples obtained by the environmental protection equipment after executing the optimized control command vector samples. The optimized control command vector sample is input into the mechanism model to obtain the basic state vector sample; A state correction sample is generated based on the basic state vector sample and the measured working condition tensor sample. The historical operating condition tensor samples and the basic state vector samples are used as training samples, and the state correction samples are used as sample labels. The deep residual compensation network is trained using the training samples and the sample labels.

5. The multi-objective collaborative optimization control method for environmental protection equipment according to claim 4, characterized in that, The step of training the deep residual compensation network using the training samples and the sample labels includes: The training samples are input into the deep residual compensation network to obtain the predicted residual vector; The mean squared error loss is generated based on the predicted residual vector and the sample labels; Calculate the gradient of the mean squared error loss with respect to the parameters of the deep residual compensation network, and generate a regularized loss based on the gradient; The mean squared error loss and the regularization loss are used to generate a total loss, and the total loss is used to train the deep residual compensation network.

6. The multi-objective collaborative optimization control method for environmental protection equipment according to claim 1, characterized in that, The step of enabling the twin instance to interact with the energy consumption optimization strategy network includes: Obtain the predicted working condition tensor generated by the twin instance as the current state; The current state is input into the energy consumption optimization strategy network to obtain the basic control command vector corresponding to the current state, which is used as the current action. Input the current action and the current state into the twin instance to obtain the next working condition tensor, which is used as the next state; Based on the next state, a composite reward function value and a constraint cost function value are generated. The composite reward function value is used to guide the energy consumption optimization strategy network to generate an energy-saving-oriented basic control command vector, and the constraint cost function value is used to constrain the energy consumption optimization strategy network to generate a basic control command vector that meets the pollutant emission constraint requirements.

7. The multi-objective collaborative optimization control method for environmental protection equipment according to claim 6, characterized in that, The Lagrange multipliers used to calculate the Lagrange weighted difference are updated after each update of the energy consumption optimization strategy network; The update process of the Lagrange multipliers includes: Calculate the average expected value of constraint cost based on the constraint cost function values ​​in each interaction data contained in the at least one interaction trajectory; Calculate the deviation of the expected value of the average constraint cost relative to the preset emission constraint threshold; The Lagrange multipliers are updated based on the deviation.

8. The multi-objective collaborative optimization control method for environmental protection equipment according to claim 1, characterized in that, The sampling of interaction data in the priority experience replay pool includes: The current state and the next state in the interaction data are input into the reward value network to obtain the cumulative reward estimate corresponding to the current state and the cumulative reward estimate corresponding to the next state. Based on the composite reward function value in the interaction data, the cumulative reward estimate corresponding to the current state, and the cumulative reward estimate corresponding to the next state, a temporal difference error corresponding to the interaction data is generated. Based on the timing difference error corresponding to the interaction data in the priority experience replay pool, the sampling priority corresponding to the interaction data in the priority experience replay pool is determined. The interaction data in the priority experience replay pool is sampled based on the sampling priority.

9. The multi-objective collaborative optimization control method for environmental protection equipment according to claim 1, characterized in that, The step of using the component residual to compensate and correct the basic control command vector to obtain the compensated control command vector includes: The component residuals are mapped to control correction amplitudes using a preset mapping strategy, which serve as control compensation amounts. The control compensation amount is superimposed on the basic control command vector using a proportional-integral law to generate the compensation control command vector.

10. The multi-objective collaborative optimization control method for environmental protection equipment according to claim 1, characterized in that, The acquisition of the operating condition tensor used to comprehensively characterize the actual operating status and pollutant emission level of the environmental protection equipment in the previous cycle includes: Acquire multi-source operating parameters that characterize the actual operating status of environmental protection equipment in the previous cycle and emission monitoring data that characterize the actual pollutant emission level of environmental protection equipment in the previous cycle; The multi-source operating parameters and the emission monitoring data are time-series aligned to obtain the time-series aligned operating condition tensor.

11. A computer program product, characterized in that, It includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the multi-objective collaborative optimization control method for environmental protection equipment as described in any one of claims 1 to 10.

12. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the multi-objective collaborative optimization control method for environmental protection equipment as described in any one of claims 1 to 10.

13. A computer storage medium, characterized in that, The storage medium carries one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the multi-objective collaborative optimization control method for environmental protection equipment as described in any one of claims 1 to 10.

Images