Method and device for controlling rainwater network overflow pollution

Abstract

The application provides a rainwater pipe network overflow pollution linkage control method and device, and relates to the technical field of intelligent control, which comprises the following steps: obtaining a target integrated model of rainwater pipe network overflow pollution, and constructing a multi-objective function according to the target integrated model, wherein the multi-objective function comprises a water quality guarantee sub-target, an operation cost sub-target and a system reliability sub-target; performing normalization processing and weighted summation processing on the three-dimensional target of the multi-objective function to obtain a comprehensive performance evaluation function; taking the comprehensive performance evaluation function as an instant feedback signal, training each drainage equipment through a preset distributed facility collaborative fusion decision network model, and outputting a target control strategy corresponding to each equipment, so as to dynamically balance control the multi-target of rain season overflow pollution according to the target control strategy. The application can significantly improve the accuracy of pollution control.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a linkage control method and device for rainwater pipe network overflow pollution. Background Technology

[0002] Currently, related technologies propose using smart drainage systems to solve the problem of overflow pollution in urban drainage systems. This solution utilizes technologies such as the Internet of Things, big data, and artificial intelligence to achieve unified management of the drainage system through real-time monitoring and intelligent scheduling of facilities such as drainage pipe networks, pumping stations, and sewage treatment plants. However, the management of overflow pollution in urban drainage systems involves a large number of devices. When managing numerous devices simultaneously, the aforementioned smart drainage systems suffer from poor system coordination capabilities, resulting in low accuracy in overflow pollution control. Furthermore, existing technologies do not fully consider physical safety constraints, smooth operation, and emergency braking mechanisms, leading to a disconnect between intelligent decision-making and actual engineering operational safety, making direct application difficult. Therefore, this invention proposes a rainwater pipe network overflow pollution linkage control scheme that integrates multi-objective optimization, a dedicated network structure, and safety constraints. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a linkage control method and device for stormwater pipe network overflow pollution, which can significantly improve the accuracy of pollution control, the stability of equipment operation and the robustness of the system.

[0004] In a first aspect, embodiments of the present invention provide a method for the coordinated control of stormwater pipe network overflow pollution. The method includes: acquiring a target integration model of stormwater pipe network overflow pollution, and constructing a multi-objective function based on the target integration model, wherein the multi-objective function includes: a water quality assurance sub-objective, an operating cost sub-objective, and a system reliability sub-objective; performing normalization and weighted summation processing on the three-dimensional objectives of the multi-objective function to obtain a comprehensive performance evaluation function; using a pre-set distributed facility collaborative fusion decision network model, using the comprehensive performance evaluation function as an instantaneous feedback signal to train various drainage equipment, outputting target control strategies corresponding to each equipment, and performing safety constraint integration processing before outputting the target control strategies, so as to achieve dynamic equilibrium control of multiple objectives of stormwater overflow pollution according to the target control strategies.

[0005] In one implementation, the step of constructing a multi-objective function based on the target integration model includes: determining equipment operating characteristic quantities using the target integration model, wherein the equipment operating characteristic quantities include: overflow, energy consumption, chemical consumption, number of equipment start-ups and shutdowns, effluent and river water quality concentration; and normalizing each equipment operating characteristic quantity based on a preset benchmark value to obtain the water quality assurance sub-objective, the operating cost sub-objective and the system reliability sub-objective.

[0006] In one implementation, the step of determining equipment operating characteristic quantities using a target ensemble model includes: acquiring the driving data of each device during the rainy season overflow pollution, and inputting the driving data into the target ensemble model to obtain the equipment operating characteristic quantities.

[0007] In one implementation, the step of performing safety constraint integration processing before outputting the target control strategy includes: performing safety constraint integration processing on the following safety constraints: verifying whether the control command to be executed meets equipment capacity constraints, hydraulic safety constraints, and operational stability constraints through a physical constraint checker; if not, modifying the command or refusing execution; performing rate-of-change and acceleration constraint processing on the control command through a motion smoothing filter to prevent mechanical shock; and when the water level is detected to exceed the safety threshold, activating an emergency braking mechanism, overriding the control command generated by reinforcement learning, and executing a preset emergency control scheme.

[0008] In one implementation, the training process includes a phased pre-training and imitation learning phase, a multi-objective collaborative optimization phase, and a robust adversarial training phase. The pre-training and imitation learning phase includes: using historical expert operational data to pre-train facility decision-making units through behavioral cloning to obtain a safety baseline strategy; the multi-objective collaborative optimization phase includes: conducting centralized training in a digital twin simulation environment using priority experience playback, double-Q learning, and target network techniques; and the robust adversarial training phase includes: injecting sensor noise and rainfall forecast errors into the training environment to enhance the robustness of the strategy using adversarial perturbations.

[0009] In one implementation, the step of using the comprehensive performance evaluation function as an immediate feedback signal includes: using a deep value network model to perform cumulative performance calculation processing on the comprehensive performance evaluation function as an immediate feedback signal to obtain a cumulative performance target value.

[0010] In one implementation, the step of training each drainage device and outputting the target control strategy corresponding to each device includes: using the cumulative performance target value as a training label, performing experience replay processing on the distributed facility collaborative fusion decision network model, and obtaining an updated target distributed facility collaborative fusion decision network model.

[0011] In one implementation, the step of using the cumulative performance target value as a training label to perform experience replay processing on the distributed facility collaborative fusion decision network model to obtain an updated target distributed facility collaborative fusion decision network model includes: updating the model by using preset parameters, calculating the mean squared error loss of the current output and cumulative performance target value of the distributed facility collaborative fusion decision network model to obtain a loss function, and updating the network parameters of the distributed facility collaborative fusion decision network model by backpropagating the loss function to obtain an updated target distributed facility collaborative fusion decision network model.

[0012] In one implementation, after obtaining the updated target distributed facility collaborative fusion decision network model, the method includes: using the target distributed facility collaborative fusion decision network model to execute online decisions for each facility decision unit and outputting the target control strategy corresponding to each device.

[0013] Secondly, embodiments of the present invention provide a linkage control device for stormwater pipe network overflow pollution. The device includes: a multi-objective function construction module, which obtains an integrated target model of stormwater pipe network overflow pollution and constructs a multi-objective function based on the integrated target model, wherein the multi-objective function includes: a water quality assurance sub-objective, an operating cost sub-objective, and a system reliability sub-objective; a comprehensive performance evaluation function generation module, which performs normalization and weighted summation processing on the three-dimensional objectives of the multi-objective function to obtain a comprehensive performance evaluation function; and an intelligent control module, which uses a preset distributed facility collaborative fusion decision network model, takes the comprehensive performance evaluation function as an instantaneous feedback signal, trains various drainage equipment, and outputs target control strategies corresponding to each piece of equipment, so as to dynamically balance and control the multi-objectives of stormwater overflow pollution according to the target control strategies.

[0014] The embodiments of the present invention bring the following beneficial effects: This invention provides a method and apparatus for the coordinated control of stormwater pipe network overflow pollution. First, a target integration model of stormwater pipe network overflow pollution is obtained. Based on this model, a multi-objective function is constructed. Then, the three-dimensional objectives of the multi-objective function are normalized and weighted summed to obtain a comprehensive performance evaluation function. Finally, a pre-set distributed facility collaborative fusion decision network model is used, with the comprehensive performance evaluation function as an instantaneous feedback signal, to train various drainage devices. The resulting target control strategies for each device are then output, enabling dynamic equilibrium control of the multi-objective overflow pollution during the rainy season based on these strategies. This invention significantly improves the accuracy of pollution control.

[0015] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

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

[0018] Figure 1 A flowchart illustrating a linkage control method for stormwater pipe network overflow pollution provided in an embodiment of the present invention; Figure 2 A schematic diagram of the specific process for a linkage control method for stormwater pipe network overflow pollution provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a linkage control device for rainwater pipe network overflow pollution provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a server provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Currently, existing technologies propose methods to address urban drainage system overflow pollution, primarily including drainage system renovation, sewage treatment plant upgrades, storage facility construction, and intelligent management. While renovations such as drainage system upgrades, sewage treatment plant upgrades, and storage facility construction are relatively easy, intelligent drainage systems utilize technologies like the Internet of Things (IoT), big data, and artificial intelligence (AI) to monitor and intelligently schedule drainage networks, pumping stations, and sewage treatment plants in real time, achieving unified management of the drainage system. By collecting data on water level, flow rate, and water quality through sensors, and automatically adjusting pumping station operation and valve switching based on data analysis, the system's operational efficiency is optimized. However, intelligent drainage systems still have shortcomings in data integration and intelligent collaboration, hindering their widespread application and in-depth development. To address these issues, this invention proposes a method based on intelligent algorithms to achieve optimized and coordinated control of pumps, gates, and adjustable weirs in a drainage system, enabling unified scheduling.

[0021] Based on this, the present invention provides a method and apparatus for the coordinated control of stormwater overflow pollution. First, it defines an integrated coordinated system including the drainage network system, sewage treatment plant, and river, and determines control variables such as pumping stations, gates, and dosing pumps. Second, it constructs a multi-objective function integrating environmental, economic, and operational considerations. Then, it sets equipment capacity, operational stability, and system state constraints. Next, it constructs a distributed facility collaborative fusion decision-making network model based on controllable facility decision-making units to achieve collaborative decision-making. This solves the problems of single objective, insufficient coordination, and poor robustness in existing technologies, and can significantly improve the environmental benefits and operational efficiency of urban drainage systems, making it suitable for sustainable urban water management scenarios.

[0022] See Figure 1 The diagram shows a process flow chart for a linkage control method for stormwater pipe network overflow pollution. The method mainly includes the following steps S102 to S106: Step S102: Obtain the target integrated model of stormwater pipe network overflow pollution, and construct a multi-objective function based on the target integrated model. The multi-objective function includes: water quality protection sub-objective, operating cost sub-objective, and system reliability sub-objective.

[0023] In one implementation, the water quality assurance sub-objective aims to reduce overflows and flooding in the drainage network system, improve the effluent quality of wastewater treatment plants, and enhance the water quality of rivers; the operating cost sub-objective aims to reduce the energy consumption of network pumping stations, the amount of phosphorus removal agents and external carbon sources used, and the dissolved oxygen setpoint of wastewater treatment plants; and the system reliability sub-objective aims to reduce the operating frequency of various facilities.

[0024] Step S104: Normalize and weighted summate the three-dimensional objectives of the multi-objective function to obtain the comprehensive performance evaluation function.

[0025] In one implementation, the overflow, pump station energy consumption, phosphorus removal agent and carbon source consumption, equipment start-up and shutdown times, effluent chemical oxygen demand, total phosphorus, ammonia nitrogen, total nitrogen concentration, and river section pollutant concentration output by the target integration model can be used as operational characteristic quantities. The operational characteristic quantities are linearly normalized with the preset benchmark values ​​of the above characteristic quantities to obtain water quality assurance sub-objectives, operating cost sub-objectives, and system reliability sub-objectives in the 0-1 interval. Then, the three sub-objectives are weighted and summed according to the set weights α, β, and γ to construct a comprehensive performance evaluation function, which can be used for subsequent neural network training labels.

[0026] Step S106: By using a pre-set distributed facility collaborative fusion decision-making network model, the comprehensive performance evaluation function is used as an instantaneous feedback signal to train various drainage devices, outputting the target control strategies corresponding to each device. Before outputting the target control strategies, safety constraint integration processing is performed to achieve dynamic equilibrium control of multiple objectives related to rainy season overflow pollution based on the target control strategies. The distributed facility collaborative fusion decision-making network model also includes: a facility decision-making unit configured for each drainage control device, and a central fusion network for integrating the outputs of each facility decision-making unit; the facility decision-making unit adopts a GRU (Gross Refrain) architecture. The MLP hybrid structure employs a monotonic hybrid network structure for the central fusion network; the training process utilizes the target network and experience replay techniques, updating the network parameters by minimizing the difference between the joint Q-value and the target Q-value.

[0027] The distributed facility collaborative fusion decision network model of the present invention adopts a centralized training-distributed execution (CTDE) architecture, which consists of two parts: facility decision-making units and a central fusion network, as detailed below: (1-1) Facility decision-making unit (edge ​​agent).

[0028] Each drainage control device (such as pumping stations, gates, dosing pumps, aeration blowers, etc.) is equipped with a separate facility decision-making unit, which acts as an edge agent responsible for local decision-making for that device. The facility decision-making unit adopts a GRU-MLP hybrid structure, and its internal structure includes: Input layer: Used to receive the system observation vector s at the current time t. t (Including pipeline water level, flow rate, water quality, rainfall, equipment status, etc.), the hidden layer state at the previous moment is h. t-1 and the control behavior a in the previous moment i,t-1 .

[0029] Gated Recurrent Unit (GRU) layer: Used to extract features from time-series observation data, capture the temporal dependencies of rainfall-runoff and equipment operation, and generate state feature vectors, overcoming the limitation of traditional MLPs in processing time-series data; the output state feature vector of the above feature extraction:

[0030] Where, θ GRU : The set of trainable parameters for a GRU network, including weights and biases.

[0031] Multilayer Perceptron (MLP) layer: Based on the state feature vector output by the GRU and the current set of valid actions, it outputs the Q-value estimate (i.e., action value, representing the long-term merit score of the action) for each selectable action; the Q-value estimation mentioned above is as follows:

[0032] Where i: the action value (Q value) output by the i-th facility decision-making unit; s t : System state vector at time t; a i,t : The control action of the i-th facility at time t; h t : The state feature vector output by the GRU; θ MLP : The set of trainable parameters for MLP networks.

[0033] The adaptive exploration strategy module is used to select exploration or exploitation behavior based on the current exploration threshold ε. The exploration threshold ε is dynamically adjusted as the device runs continuously to balance exploration and exploitation during the training process and avoid the strategy from getting stuck in local optima.

[0034] (1-2) Centralized integration network (global coordination layer).

[0035] The central fusion network is used to integrate the local Q-values ​​output by various facility decision-making units to generate a global collaborative control strategy. It adopts a monotonic hybrid network structure, specifically including: Global feature extraction subnetwork: used to generate dynamic fusion weights based on the global system state (including the state information of the entire pipe network, sewage treatment plant, and river); Based on the global system state S t Generate fusion weights: w i,t =GlobalNet(S t ;θ g ) Among them, w i,t The fusion weight of the i-th facility, satisfying w i,t ≥0; S t : Global system state vector, including the overall state of the entire pipeline network, sewage treatment plant, and river; θ g : Trainable parameters for the global feature extraction network.

[0036] Monotonic hybrid network: Used to weight and sum the local Q-values ​​output by each facility decision-making unit according to the fusion weights, generating a Q-value estimate for global joint action. The specific calculation formula is as follows:

[0037] Among them, Q total : Global joint action value, characterizing the quality of the global control strategy; I: Total number of controllable facilities; w i,t The non-negative fusion weights generated by the supernetwork satisfy the monotonicity constraint; Q i A: The local Q-value output by the i-th facility decision-making unit; t : Joint control action vector for the entire system.

[0038] The weight parameters of the monotonic hybrid network are generated through a supernetwork, and the weight coefficients are non-negative to ensure that the monotonicity constraint is met, guaranteeing the consistency between the global optimum and the local optimum, and avoiding the problem of "local optimum but global suboptimal".

[0039] (2) Definition and training mechanism of Q-value: In this invention, the Q-value (action value function) uniformly refers to the expected cumulative reward that an agent can obtain by performing a certain control action in the current system state. It is used to evaluate the long-term quality of control actions and is a core foundational concept of reinforcement learning. Wherein: Global Joint Action Q-value (Joint Q-value): This is the predicted value of the global collaborative control action output by the current central fusion network, representing the network's assessment of the quality of the global action; Target Q-value: Q target Stable value labels output by the independent target network (TargetNetwork) are used to avoid overestimation of value and accelerate training convergence.

[0040] The training process of this invention employs double-Q learning, a target network, and experience replay techniques. The specific training process is as follows: Experience replay: The samples (state, action, reward, next state) during the system operation are stored in the experience replay pool. During training, samples are randomly selected for network updates, breaking the temporal correlation of samples and improving training stability.

[0041] Dual Q-learning and target network: Two independent Q-networks (the current training network and the target network) are used for action selection and action value evaluation, respectively. By minimizing the difference between the global joint action Q-value output by the current central fusion network and the target Q-value output by the target network, the network parameters are iteratively updated to avoid overestimation of value and improve the stability of the control strategy.

[0042] in, y t : Training target Q-value labels; P syn(t): The comprehensive performance evaluation function at time t, serving as the immediate reward; γ: Discount factor, ranging from 0 to 1, with a default value of 0.95; S t+1 : Global system state at time t+1; Q target : The Q value output by the target network.

[0043] Parameter update: The loss is calculated using the mean squared error loss function, and the network parameters are updated via backpropagation. The target network parameters are periodically synchronized from the current training network to maintain training stability.

[0044] Wherein, the mean squared loss function is:

[0045] Among them, L TD B represents the temporal difference loss function; y represents the training batch size; b : The target Q value of the b-th sample; Q total : The global joint Q-value of the current network output.

[0046] Network parameter update: Where θ: the current set of network parameters; η: the learning rate, which controls the step size for parameter updates; θ L TD : The gradient of the loss function with respect to the network parameters.

[0047] (3) Three-stage phased training process: This invention employs a three-stage training approach, which differs from the single-stage training of existing technologies, further enhancing the robustness and engineering applicability of the strategy. Specifically, it includes: (3-1) Pre-training and Imitation Learning Stage: Using historical expert operation data (such as manual scheduling records and historical optimal control strategies), the facility decision-making unit is pre-trained through behavioral cloning. This allows the agent to quickly learn basic safety control logic, obtain a safety baseline strategy, avoid ineffective or dangerous control actions in the early stages of training, and shorten the training cycle. The loss function for pre-training the facility decision-making unit using historical expert operation data and behavioral cloning is as follows:

[0048] in, L clone : Imitation learning loss function; N: Total number of historical expert samples; a expert,v : The expert's manual operation action in sample v; a net,v : The predicted action of the network output.

[0049] (3-2) Multi-objective collaborative optimization stage: In the digital twin simulation environment, priority experience playback, double Q learning and target network technology are used for centralized training. The comprehensive performance evaluation function is used as the immediate reward, so that the agent can learn the optimal collaborative control strategy autonomously and achieve a multi-objective dynamic balance of water quality, cost and reliability.

[0050] (3-3) Robustness Adversarial Training Phase: Sensor noise, rainfall forecast errors, and equipment malfunctions are injected into the training environment to enhance the robustness of the strategy. This ensures that the agent can still output a stable and reliable control strategy under complex real-world conditions, addressing the problem of "good simulation results but poor field performance" in actual deployments. Sensor noise and rainfall forecast errors are injected into the training environment to enhance the strategy's robustness.

[0051] Among them, s t ′: Disturbance observation state after adding noise; δ: Sensor measurement noise, including water level, flow rate, and water quality monitoring noise; r t ′: Rainfall input after adding error; ε: Rainfall forecast error.

[0052] (4) Security constraint integrated processing mechanism: This invention performs integrated safety constraint processing before outputting the target control strategy. This is a core engineering innovation that distinguishes it from existing pure algorithmic solutions, specifically including: (4-1) Physical constraint checker: Verify whether the control command to be executed meets the equipment capacity constraints, hydraulic safety constraints, and operational stability constraints. If not, modify the command (e.g., adjust the pump station opening exceeding the rated flow to the rated value) or refuse execution to ensure the engineering feasibility of the control command. The specific verification of whether the control command meets the equipment capacity constraints is as follows: u i,min ≤ u i,t ≤ u i,max Among them, u i,t : The control command to be executed for the i-th facility; u i,min : Lower limit value of the i-th facility control command; u i,max : The upper limit of the control command for the i-th facility.

[0053] (4-2) Action smoothing filter: Control commands are subjected to rate-of-change limits and acceleration constraints to prevent frequent equipment starts and stops and mechanical shocks, thereby extending equipment lifespan and improving system operational stability. The specific calculation formula is as follows:

[0054]

[0055] Where, Δu max : Maximum permissible change in a single-step control command; : Control command change rate; : Maximum permissible fluctuation limit of the rate of change.

[0056] (4-3) Emergency braking mechanism: When emergency conditions such as water level exceeding the safety threshold or water quality seriously exceeding standards are detected, the emergency braking mechanism is activated, overriding the control commands generated by reinforcement learning and executing the preset emergency control plan (such as fully opening the pump station or closing the gate) to ensure the safe operation of the system.

[0057]

[0058]

[0059] Among them, h t Current monitored water level; h limit : Water level safety threshold; u t : Current system control command; u emergency Preset emergency control commands.

[0060] (5) Online decision-making and control execution.

[0061] After training, each facility decision-making unit makes online decisions based on real-time monitoring data, outputs local control actions, and after global coordination by the central fusion network, it is verified through safety constraint integration processing. Finally, it outputs the target control strategies corresponding to each device and sends them to the field execution terminal to realize real-time linkage control of rainwater pipe network overflow pollution.

[0062] In one implementation, the pumping stations / gates of the drainage network and the dosing pumps / aeration blowers of the sewage treatment plant in the integrated system are treated as control variables corresponding to a facility decision-making unit. Global integration is achieved through a fusion network, and intelligent control is achieved by adopting a unified training-distributed execution mechanism.

[0063] In a distributed facility decision-making unit system, each facility decision-making unit is responsible for regulating a controllable variable facility. The input to a facility decision-making unit is the system observation at the current moment and the control behavior at the previous moment. The output is the state-behavior value function of the effective behavior at the current moment. Subsequently, an adaptive exploration strategy is used to select a behavior, and the state-behavior value function after executing that behavior is output. The behavior space is discretized. Each facility decision-making unit has the same internal structure, consisting of an input layer, a hidden layer, and an output layer. The hidden layer contains a multilayer perceptron (MLP) and a gated recurrent unit (GRU).

[0064] The internal structure of the fusion network is also a neural network. Its inputs include the individual value functions of each facility decision-making unit and the global observations of the system at the current moment, and its output is the expected reward of the joint behavior. The fusion network dynamically updates its own weights and biases based on the global observations.

[0065] The above-mentioned linkage control method for stormwater pipe network overflow pollution provided in this invention addresses the problems of low efficiency in drainage system parameter calibration and insufficient multi-facility collaborative control in existing technologies by proposing a cascade optimization and multi-facility decision-making unit framework. First, key sensitive parameters in the pipe network, sewage treatment plant, and river models are quickly identified through HSY global sensitive parameter screening combined with three-level physical constraint checks (single-parameter boundary, multi-parameter collaboration, and model stability). Then, a two-stage adaptive genetic algorithm is used for fine optimization: in the exploration phase (first 80 generations), a dynamic high crossover / mutation rate (adjusted from 0.9 to 0.7 / 0.05 to 0.02) is used to expand the search; in the convergence phase (last 120 generations), efficient convergence is achieved through population reconstruction and Gaussian fine-tuning of mutation (±1% amplitude). Based on the calibration model, a collaborative architecture of distributed facility decision-making units and a central fusion network is constructed: each pump station / gate / dosing pump acts as an independent facility decision-making unit, outputting a local Q-value based on the GRU-MLP network; the central fusion network integrates global observations to generate a three-dimensional comprehensive performance evaluation function that integrates environmental objectives (overflow - water quality improvement), economic objectives (energy consumption - chemical cost), and operational objectives (equipment stability). Unified training is achieved through experience replay and the target network. During the execution phase, the facility decision-making units respond to local conditions in real time according to an adaptive exploration strategy. This invention can significantly improve parameter calibration efficiency (reducing computational resources by 90%), achieving a multi-objective dynamic balance between overflow pollution control, cost savings, and operational stability. NSE ≥ 0.8 verifies the model's reliability.

[0066] See Figure 2 The diagram shows a specific flow chart of a linkage control method for stormwater pipe network overflow pollution. This invention also provides an implementation method for linkage control of stormwater pipe network overflow pollution, as detailed in (A) to (B) below: (A) First, the target integration model is used to determine the equipment operation characteristics, that is, to obtain the driving data of each equipment during the overflow pollution in the rainy season, and input the driving data into the target integration model to obtain the equipment operation characteristics. Then, based on the preset benchmark value, the operation characteristics of each equipment are normalized to obtain the water quality protection sub-target, the operating cost sub-target and the system reliability sub-target. Among them, the equipment operation characteristics include: overflow, energy consumption, chemical consumption, number of equipment start-ups and shutdowns, effluent and river water quality concentration, as detailed in (1) to (3) below: (1) Water quality assurance sub-objective: The water quality assurance sub-objective comprehensively considers the river water quality, stormwater overflow water quality, surface runoff water quality, and wastewater treatment plant effluent water quality. Its calculation formula is as follows:

[0067] Where: Gqua(t) is the score of the water quality assurance sub-objective at time t. The higher the score, the better the overall water quality assurance performance of the system; λ1, λ2, λ3, and λ4 are the weight coefficients of overflow rainwater quality, surface runoff polluted water quality, river water quality, and sewage treatment plant effluent quality, respectively. The values ​​are determined by decision preferences. λ1, λ2, λ3, and λ4 ∈ [0,1] and λ1+λ2+λ3+λ4=1.

[0068]

[0069] Where N is the number of stormwater overflow outlets monitored, ξn is the contribution of the nth overflow outlet to the overall stormwater overflow pollution (weighted based on the overflow outlet's design flow rate or service area), and Z nstd Z represents the baseline concentration of pollutants at the nth overflow outlet (unit: mg / L, with baseline values ​​of 50, 1.5, 2.0, and 4.0 mg / L for COD, TP, NH3-N, and TN, respectively). nact (t) represents the measured concentration (in mg / L) of the pollutant at the nth overflow outlet at time t.

[0070]

[0071] Where M is the number of surface runoff monitoring zones, ζm is the contribution of the m-th monitoring zone to overall surface runoff pollution, and W mstd W represents the baseline concentration of pollutants in the surface runoff of the m-th zone (unit: mg / L, consistent with the baseline concentration of pollutants at the overflow outlet). mact (t) represents the measured concentration (unit: mg / L) of pollutants in the surface runoff of the m-th zone at time t.

[0072]

[0073] Where, η pqis the contribution degree of the q-th pollutant at the p-th river monitoring point to the river water quality, P is the number of river monitoring points, Q is the number of considered pollutant types, U pqstd is the reference value of the q-th pollutant at the p-th monitoring point (unit: mg / L), U pqact (t) is the concentration of the corresponding pollutant at this monitoring point at time t (unit: mg / L).

[0074] (2) Operation cost sub-objective: The operation cost sub-objective focuses on the economic optimization of the whole water quality assurance chain, covering the core costs of stormwater overflow control, surface runoff treatment, sewage treatment and supporting auxiliary systems, and its calculation formula is as follows:

[0075] Where: C eco (t) is the score of the operation cost sub-objective at time t, the higher the score (the lower the cost), C spill (t), C runoff (t), C treat (t), C aux (t) respectively represent the core cost objective scores of stormwater overflow control, surface runoff treatment, sewage treatment and supporting auxiliary systems; μ1, μ2, μ3, μ4 are the weight coefficients of stormwater overflow control cost, surface runoff treatment cost, sewage treatment cost and auxiliary system operation cost respectively, the values are determined by decision preferences, and μ1, μ2, μ3, μ4 ∈ [0,1] and satisfy μ1+μ2+μ3+μ4=1.

[0076]

[0077] Where, E pump (t) is the power consumption of the lifting pump station supporting the n-th overflow outlet at time t (unit: kWh), ρ elec is the unit electricity price (CNY / kWh); D spill (t) is the dosage of pollutant reduction agent emergently added at the n-th overflow outlet at time t (unit: kg, such as PAC, decolorizer, etc.), ρ chem1 is the unit price of the corresponding agent (CNY / kg); C spill_std is the reference cost of stormwater overflow control (unit: CNY), which is set based on the historical average operation cost, and ξn represents the contribution degree of the n-th overflow outlet to the overall stormwater overflow pollution.

[0078]

[0079] Where, P filter (t) is the operation loss cost of the ecological filter / pollution interception device in the m-th surface runoff zone at time t (unit: CNY, including filter material replacement, equipment maintenance, etc.), ρ filterCost per unit loss (yuan / m2); D runoff (t) represents the amount of reagent (in kg, such as flocculant, antibacterial agent, etc.) added for the pretreatment of surface runoff in the m-th zone at time t, ρ chem2 C represents the unit price of the corresponding drug (yuan / kg); runoff_std The baseline cost (in yuan) for surface runoff management is set based on the scale of management in each zone and historical data; ζm represents the contribution of the m-th surface runoff monitoring zone to overall surface runoff pollution.

[0080]

[0081] Among them, E treat (t) represents the total power consumption of the k-th wastewater treatment plant at time t (unit: kWh, including aeration blowers, return pumps, dewatering machines, etc.); D treat (t) represents the total amount of core treatment agents used by the k-th wastewater treatment plant at time t (unit: kg, such as phosphorus removal agent, nitrifying bacteria agent, carbon source, etc.), ρ chem3 The corresponding unit price of the drug (yuan / kg); S sludge (t) represents the sludge production (in t) of the k-th wastewater treatment plant at time t, ρ sludge The unit cost of sludge disposal (RMB / t, including transportation, harmless treatment, etc.); C treat_std The baseline cost (unit: yuan) for wastewater treatment is set based on the treatment scale and Class A standard operating requirements, θ k This represents the weight of the contribution of the k-th wastewater treatment plant to the overall wastewater treatment cost.

[0082]

[0083] Among them, E monitor (t) represents the total power consumption (in kWh) of all water quality monitoring equipment (overflow outlet, runoff zone, sewage treatment plant, and river monitoring point) at time t; M maintain (t) represents the maintenance cost (in yuan, including calibration, repair, etc.) of all system control equipment (valves, sensors, intelligent terminals) at time t; C aux_std The baseline cost (unit: yuan) for assisting system operation is set based on the number of monitoring points and equipment operation and maintenance standards.

[0084] (3) System reliability sub-objective: The system reliability sub-objective focuses on the stable operation capability of all equipment and systems in the water quality assurance chain, covering the operational stability, fault risk prevention and control, and service life guarantee of core control facilities, treatment facilities, processing facilities and auxiliary systems. It ensures the continuous and reliable operation of the system during cost optimization and water quality compliance. The calculation formula is as follows:

[0085] Where: R rel (t) represents the score of the system reliability sub-objective at time t. The higher the score, the stronger the overall system stability and the lower the failure risk. ν1, ν2, ν3, and ν4 are the weight coefficients of the reliability of control facilities, the reliability of treatment facilities, the reliability of processing facilities, and the reliability of auxiliary systems, respectively. Their values ​​are determined by decision preferences. ν1, ν2, ν3, and ν4 ∈ [0,1] and satisfy ν1+ν2+ν3+ν4=1.

[0086]

[0087] Among them, R control (t) corresponds to the core facility in the stormwater overflow control process, T pump,ref The baseline continuous operating time (in hours, based on equipment design life) for the overflow outlet booster pump station, T pump,act (t) represents the cumulative continuous running time at time t (in hours); F valve,ref The reference allowable start / stop frequency (unit: times / day) for overflow control valves, F valve,act (t) represents the actual start-stop frequency at time t (unit: times / day), and ξn represents the contribution of the nth overflow outlet to the overall rainwater overflow pollution.

[0088]

[0089] Among them, R govern (t) corresponds to the core facilities in the surface runoff management process, I filter,ref The baseline pollution load tolerance of the ecological filter / sewage interception device (unit: kg / m2), I filter,act (t) represents the actual pollution load at time t (unit: kg / m2), and ζm represents the contribution of the m-th surface runoff monitoring zone to the overall surface runoff pollution.

[0090] W runoff,ref W is the baseline permissible operating time (in hours / day) for surface runoff pretreatment equipment. runoff,act (t) represents the actual runtime at time t (unit: h / day).

[0091]

[0092] Among them, R process (t) corresponds to the core facility in the wastewater treatment process, A blower,ref A is the baseline continuous operating time (in hours) for the aeration blower. blower,act (t) represents the cumulative continuous running time at time t (in hours); B dosage,ref B is the baseline allowable start / stop frequency for the dosing pump (unit: times / day).dosage,act (t) represents the actual start / stop frequency at time t (unit: times / day), θ k This represents the weight of the contribution of the k-th wastewater treatment plant to the overall wastewater treatment cost.

[0093] C sludge,ref C is the baseline load threshold (unit: t / h) for sludge treatment equipment. sludge,act (t) represents the actual processing load at time t (unit: t / h).

[0094]

[0095] Among them, R support (t) corresponds to the auxiliary support facilities of the entire system, D monitor,ref D is the baseline fault interval (in hours) for water quality monitoring equipment. monitor,act (t) represents the actual fault interval time at time t (unit: h, the baseline value is taken when there is no fault); E control,ref E is the baseline response delay threshold (in seconds) for the intelligent control terminal. control,act (t) represents the actual response delay at time t (in seconds).

[0096] Finally, the various objectives are integrated into a comprehensive performance evaluation function using a weighted method:

[0097] P syn R(t) is the comprehensive performance evaluation function value at time t, with a value range of [0,1]. The higher the value, the better the comprehensive performance of the system in the three dimensions of "water quality compliance - cost optimization - operational reliability". α, β, and γ are the global weight coefficients of the water quality assurance sub-objective, the operating cost sub-objective, and the system reliability sub-objective, respectively. Their values ​​are determined by the decision preferences of the actual application scenario. α, β, and γ ∈ [0,1] and satisfy α + β + γ = 1. Among them, R(t) is the comprehensive performance evaluation function, and α, β, and γ are the weight coefficients related to the three sub-objectives.

[0098] The mathematical expressions for the above constraints are as follows:

[0099]

[0100]

[0101] in, u i,min and u i,max These are the control behaviors of the i-th facility. u iThe lower and upper limits of (t); the adjustment of controllable facilities should not be too frequent, that is, the number of state switching of the control variable should not exceed 1 time within its minimum control interval N; : Hydrodynamic and water quality equations; p: The set of parameters for the above equations.

[0102] (B) First, the comprehensive performance evaluation function is used as an immediate feedback signal for cumulative performance calculation. That is, through the deep value network model, the comprehensive performance evaluation function is used as an immediate feedback signal for cumulative performance calculation to obtain the cumulative performance target value. Then, the cumulative performance target value is used as a training label to perform experience playback processing on the distributed facility collaborative fusion decision network model to obtain the updated target distributed facility collaborative fusion decision network model. That is, by updating the model through preset parameters, the mean squared error loss is calculated on the current output and cumulative performance target value of the distributed facility collaborative fusion decision network model to obtain the loss function. By backpropagating the loss function, the network parameters of the distributed facility collaborative fusion decision network model are updated to obtain the updated target distributed facility collaborative fusion decision network model. The target distributed facility collaborative fusion decision network model is then used to execute the online decision-making of each facility decision unit and output the target control strategy corresponding to each device. In one embodiment, the calculation expression of the performance function at the facility decision unit level is as follows:

[0103] Wherein: S i (ω i ,b t,i Let ω be the state-behavioral performance function of the i-th facility decision-making unit, representing its state-behavioral performance on its historical operating trajectory ω. i Next, execute the current action b. t,i The expected overall performance after that.

[0104] The first equation is the local state mapping submodule: through the time-series transformation operator SeqTrans, the observation vector s at time t is transformed... t (System status + rainfall forecast) and the hidden layer status c at the previous moment t 1,i Fusion, outputting the current state feature vector. i κ i,1 These are the trainable parameters for this submodule.

[0105] The second module is the behavior value assessment submodule: it uses the behavior assessment operator ActEval, based on the state feature vector. i Previous moment control behavior b t 1,i The current set of valid behaviors b t,i,perm Output the performance value corresponding to the current behavior, κ i,2 These are the trainable parameters for this submodule.

[0106] The facility decision-making unit uses the performance function as the basis for decision-making, employs an adaptive exploration strategy to select the control behavior at the current moment, and outputs an estimated value corresponding to the selected behavior, providing support for the evaluation of the joint behavior performance function. The decision-making logic of the adaptive exploration strategy is as follows:

[0107] Where: b t,i S: The control behavior of the i-th facility decision-making unit at time t; i (ω i b): The state-behavioral performance function of the i-th facility decision-making unit; b t,i,perm : The set of effective actions of the i-th facility decision-making unit at time t; τ: Exploration threshold (a real number between 0 and 1, replacing the original ε), used to balance "utilization (selecting the optimal action)" and "exploration (randomly selecting actions)"; p: a random number uniformly distributed between 0 and 1; This indicates that an action is randomly selected from the set of valid actions; b t,total : The joint control behavior of the entire system at time t (the set of behaviors of all facility decision-making units); S total (Ω,b t,total ): Joint behavioral performance function of the entire system; Ω: Historical operating trajectory of the entire system (replacing the original τ); I is the total number of facility decision-making units.

[0108] The algorithm constructed above is trained in three stages: human experience collection, interactive experience collection, and network collaborative update.

[0109] The historical data collected for human experience includes time-series monitoring data collected during actual rainfall events (such as water level, flow rate, influent and effluent water quality, and operational behavior), as well as corresponding system response data obtained from calibrated urban drainage model simulations (used to calculate the comprehensive performance evaluation function). This data constitutes human operational experience and is pre-loaded into a buffer before training begins.

[0110] During training, updates to the unit subnetworks and the fusion network of the distributed facility collaborative fusion decision network depend on the training data stored in a centralized experience replay buffer. Experience is stored as normalized tuples to ensure data consistency and facilitate reuse. The structure of each record's experience tuple is as follows:

[0111] Where O: observation vector, including partial system state, rainfall intensity, and historical information; a: behavior vector; r: feedback signal; a avail : The set of currently available behaviors; tag is a binary event termination flag, where 0 indicates that the event is in progress and 1 indicates that the event has terminated.

[0112] The network collaborative update mechanism is based on the deep value network algorithm and employs two techniques: experience replay and target network. Experience replay updates parameters by extracting small batches of samples from the buffer, combining experience at different time steps to improve efficiency and reduce bias. The target network addresses training instability by providing stable target computation, the calculation formula of which is as follows:

[0113] Where y is the target output of the value network; This is the overall performance feedback value for the current time step; Discount factor; These represent the joint behavior, trajectory, and state at the next time step; The parameter set of the target network is updated synchronously with the parameters of the current value network every certain number of steps; the parameter update mechanism of the value network is calculated using the following formula:

[0114]

[0115] Where i is the i-th sample drawn from the experience replay; Batch is the batch size; Loss(κ) is the mean squared error loss function of the value network (κ is the current value network parameter); κ′ is the updated value network parameter; α is the learning rate; κ Loss(κ) is the gradient of the loss function with respect to the parameter κ (used for backpropagation).

[0116] Therefore, the aforementioned distributed collaborative decision-making reinforcement learning architecture can solve the multi-objective dynamic optimization problem of drainage systems by constructing a collaborative control framework of distributed facility decision-making units and a central fusion network. Each facility (pump station / gate / dosing pump) acts as an independent facility decision-making unit, outputting local performance values ​​based on the temporal feature extraction unit—multilayer sensing unit network. The central fusion network integrates global observations and individual performance values ​​to generate a three-dimensional comprehensive performance evaluation function that integrates water quality assurance objectives (rainwater overflow quality—surface runoff pollution quality—sewage treatment plant effluent quality—river water quality), operating cost objectives (overflow control cost—runoff treatment cost—sewage treatment cost—auxiliary system cost), and system reliability objectives (control facility—treatment facility—treatment facility—auxiliary system stability). Through a unified training-distributed execution mechanism: the training phase uses a target network and an experience playback pool to stabilize the learning process; during the execution phase, each facility decision-making unit responds to its local state according to an adaptive exploration strategy, achieving a global dynamic balance between overflow pollution control, cost savings, and operational stability.

[0117] In summary, this invention significantly improves the accuracy of overflow pollution control and the system's collaborative capabilities: by integrating regional sensitivity analysis (HSY algorithm) with a phased genetic algorithm, it achieves high-precision modeling and parameter optimization of the integrated system of drainage networks, sewage treatment plants, and receiving rivers. The HSY algorithm efficiently identifies sensitive parameters through Latin hypercube sampling and a double-layer verification mechanism, avoiding the dimensionality curse problem caused by optimizing all parameters in traditional methods. The phased genetic algorithm extensively searches the solution space during the exploration phase and performs fine optimization during the convergence phase, significantly reducing system simulation errors (NSE≥0.85, WRE≤5%). Furthermore, by constructing multi-objective functions (environment, economy, operation) and designing a collaborative facility decision-making unit control framework, it achieves global linkage and optimized scheduling of facilities such as pumping stations, gates, and dosing pumps, solving the problems of single objectives and insufficient collaboration in existing smart drainage systems, and improving overall control accuracy and system response consistency.

[0118] Furthermore, this invention enhances the system's robustness, economy, and real-time response capabilities: by introducing a multi-objective comprehensive performance evaluation function, it organically combines environmental benefits (such as minimizing overflow and achieving water quality standards), economic operating costs (optimizing energy and chemical consumption), and equipment operational stability (frequency control), achieving multi-objective collaborative optimization through weight adjustment. Based on this, a deep value network-based algorithm is employed, using a unified training and distributed execution mechanism to achieve real-time control decisions, significantly improving the system's adaptability and robustness in complex rainfall scenarios. This mechanism effectively handles rainfall forecast errors and monitoring noise, with short single-step decision-making time, meeting real-time control requirements. Simultaneously, experience playback and target network technology ensure training stability and convergence, ultimately significantly reducing system operating costs and equipment wear while ensuring water environmental quality, thus improving overall economy and practicality.

[0119] Regarding the linkage control method for stormwater pipe network overflow pollution provided in the foregoing embodiments, this invention provides a linkage control device for stormwater pipe network overflow pollution, see [link / reference]. Figure 3 The diagram shows a structural schematic of a linkage control device for stormwater pipe network overflow pollution. The device includes the following parts: The multi-objective function construction module 302 obtains the target integration model of stormwater pipe network overflow pollution, and constructs multi-objective functions based on the target integration model. The multi-objective functions include: water quality protection sub-objective, operating cost sub-objective, and system reliability sub-objective. The comprehensive performance evaluation function generation module 304 performs normalization and weighted summation on the three-dimensional objectives of the multi-objective function to obtain the comprehensive performance evaluation function. The intelligent control module 306 uses a pre-set distributed facility collaborative fusion decision network model and takes the comprehensive performance evaluation function as an instant feedback signal to train various drainage equipment and output the target control strategy corresponding to each equipment. Based on the target control strategy, it performs dynamic equilibrium control of multiple objectives of overflow pollution during the rainy season.

[0120] The linkage control device for rainwater pipe network overflow pollution provided in this application embodiment can significantly improve the accuracy of pollution control.

[0121] In one embodiment, when performing the step of constructing a multi-objective function based on the target integration model, the multi-objective function construction module 302 is further configured to: determine the equipment operation characteristic quantities using the target integration model, wherein the equipment operation characteristic quantities include: overflow, energy consumption, chemical consumption, number of equipment start-ups and shutdowns, effluent and river water quality concentration; and normalize each equipment operation characteristic quantity based on a preset benchmark value to obtain the water quality assurance sub-objective, the operating cost sub-objective and the system reliability sub-objective.

[0122] In one embodiment, when performing the safety constraint integration processing step before implementing the output target control strategy, the intelligent control module 306 is further configured to: verify whether the control command to be executed meets the equipment capacity constraints, hydraulic safety constraints, and operational stability constraints through a physical constraint checker; if not, modify the command or refuse to execute it; perform rate-of-change and acceleration constraint processing on the control command through an action smoothing filter to prevent mechanical shock; and when the water level is detected to exceed the safety threshold, activate the emergency braking mechanism, overwrite the control command generated by reinforcement learning, and execute the preset emergency control scheme.

[0123] In one embodiment, the intelligent control module 306 is further configured to: perform training processing including a phased pre-training and imitation learning phase, a multi-objective collaborative optimization phase, and a robust adversarial training phase, wherein the pre-training and imitation learning phase includes: using historical expert operation data to pre-train the facility decision-making unit through behavior cloning to obtain a safety baseline strategy; the multi-objective collaborative optimization phase includes: performing centralized training in a digital twin simulation environment using priority experience playback, double-Q learning, and target network technology; the robust adversarial training phase includes: injecting sensor noise and rainfall forecast errors into the training environment and using adversarial perturbations to enhance the robustness of the strategy.

[0124] In one embodiment, when performing the step of determining equipment operating characteristic quantities using the target integration model, the multi-objective function construction module 302 is further used to: acquire the driving data of each device during the rainy season overflow pollution, and input the driving data into the target integration model to obtain the equipment operating characteristic quantities.

[0125] In one embodiment, when performing the step of using the comprehensive performance evaluation function as an instantaneous feedback signal, the intelligent control module 306 is further configured to: use the comprehensive performance evaluation function as an instantaneous feedback signal through a deep value network model to perform performance accumulation calculation processing to obtain a cumulative performance target value.

[0126] In one embodiment, when performing the step of training each drainage device and outputting the target control strategy corresponding to each device, the intelligent control module 306 is further configured to: use the cumulative performance target value as a training label to perform experience playback processing on the distributed facility collaborative fusion decision network model to obtain the updated target distributed facility collaborative fusion decision network model.

[0127] In one embodiment, when performing the step of using the cumulative performance target value as a training label to perform experience replay processing on the distributed facility collaborative fusion decision network model to obtain an updated target distributed facility collaborative fusion decision network model, the aforementioned intelligent control module 306 is further configured to: update the model by using preset parameters, calculate the mean squared error loss of the current output and cumulative performance target value of the distributed facility collaborative fusion decision network model to obtain a loss function, and update the network parameters of the distributed facility collaborative fusion decision network model by backpropagating the loss function to obtain an updated target distributed facility collaborative fusion decision network model.

[0128] In one embodiment, after obtaining the updated target distributed facility collaborative fusion decision network model, the intelligent control module 306 is further configured to: execute online decisions of each facility decision unit using the target distributed facility collaborative fusion decision network model, and output the target control strategy corresponding to each device.

[0129] The device provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0130] This invention provides a server, specifically, the server includes a processor and a storage device; the storage device stores a computer program, which, when run by the processor, executes the method described in any of the above embodiments.

[0131] Figure 4 This is a schematic diagram of the structure of a server provided in an embodiment of the present invention. The server 100 includes: a processor 40, a memory 41, a bus 42 and a communication interface 43. The processor 40, the communication interface 43 and the memory 41 are connected through the bus 42. The processor 40 is used to execute executable modules, such as computer programs, stored in the memory 41.

[0132] The memory 41 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 43 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.

[0133] Bus 42 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0134] The memory 41 is used to store programs. After receiving an execution instruction, the processor 40 executes the program. The method executed by the device for defining the flow process disclosed in any of the foregoing embodiments of the present invention can be applied to the processor 40 or implemented by the processor 40.

[0135] Processor 40 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 40 or by instructions in software form. Processor 40 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 41, and the processor 40 reads the information from memory 41 and, in conjunction with its hardware, completes the steps of the above method.

[0136] The computer program product of the readable storage medium provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the foregoing method embodiments. For specific implementation, please refer to the foregoing method embodiments, which will not be repeated here.

[0137] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0138] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for coordinated control of overflow pollution in a rainwater pipe network, characterized in that, The method includes: A target integration model for stormwater pipe network overflow pollution is obtained, and a multi-objective function is constructed based on the target integration model. The multi-objective function includes: water quality protection sub-objective, operating cost sub-objective, and system reliability sub-objective. The three-dimensional objectives of the multi-objective function are normalized and weighted summation is performed to obtain the comprehensive performance evaluation function; By using a pre-set distributed facility collaborative fusion decision network model, the comprehensive performance evaluation function is used as an instantaneous feedback signal to train various drainage equipment, output the target control strategy corresponding to each equipment, and perform safety constraint integration processing before outputting the target control strategy, so as to achieve dynamic equilibrium control of multiple objectives of rainy season overflow pollution according to the target control strategy. The step of performing safety constraint integration processing before outputting the target control strategy includes: performing safety constraint integration processing on the following safety constraints: verifying whether the control command to be executed meets the equipment capacity constraints, hydraulic safety constraints, and operational stability constraints through a physical constraint checker; if not, modifying the command or refusing execution; performing rate-of-change and acceleration constraint processing on the control command through a motion smoothing filter to prevent mechanical shock; and when the water level is detected to exceed the safety threshold, activating an emergency braking mechanism, overriding the control command generated by reinforcement learning, and executing a preset emergency control scheme. The training process includes a phased pre-training and imitation learning phase, a multi-objective collaborative optimization phase, and a robust adversarial training phase. The pre-training and imitation learning phase includes: using historical expert operational data to pre-train facility decision-making units through behavioral cloning to obtain a safety baseline strategy; the multi-objective collaborative optimization phase includes: conducting centralized training in a digital twin simulation environment using priority experience replay, double-Q learning, and target network techniques; the robust adversarial training phase includes: injecting sensor noise and rainfall forecast errors into the training environment and using adversarial perturbations to enhance the robustness of the strategy. The calculation formula for the water quality assurance sub-target is as follows: Among them, G qua (t) represents the score of the water quality assurance sub-objective at time t, F spill F(t) represents the score function for the overflow stormwater quality. runoff F(t) represents the score function for surface runoff pollution of water quality. stream F(t) represents the score function for river water quality. effluent (t) represents the score function for the effluent quality of the wastewater treatment plant; λ1, λ2, λ3, and λ4 are the weight coefficients for the overflow rainwater quality, surface runoff pollution quality, river water quality, and wastewater treatment plant effluent quality, respectively, and their values ​​are determined by decision preferences. λ1, λ2, λ3, and λ4 ∈ [0,1] and λ1+λ2+λ3+λ4=1; The formula for calculating the operating cost sub-objective is as follows: Among them, C eco (t) represents the score of the cost sub-objective at time t, C spill (t), C runoff (t), C treat (t), C aux (t) represents the core cost target scores for stormwater overflow control, surface runoff management, sewage treatment and supporting auxiliary systems, respectively; μ1, μ2, μ3, and μ4 are the weight coefficients of stormwater overflow control cost, surface runoff management cost, sewage treatment cost and auxiliary system operation cost, respectively, μ1, μ2, μ3, and μ4 ∈ [0,1] and satisfy μ1+μ2+μ3+μ4=1; The calculation formula for the system reliability sub-objective is as follows: Among them, R rel (t) represents the score of the system reliability sub-objective at time t, R control (t) corresponds to the core facility in the stormwater overflow control system, R govern (t) corresponds to the core facility in the surface runoff management process, R process (t) corresponds to the core facilities in the wastewater treatment process, R support (t) corresponds to the auxiliary support facilities of the whole system; ν1, ν2, ν3, and ν4 are the weight coefficients of the reliability of control facilities, the reliability of treatment facilities, the reliability of processing facilities, and the reliability of auxiliary systems, respectively. ν1, ν2, ν3, and ν4 ∈ [0,1] and satisfy ν1+ν2+ν3+ν4=1.

2. The linkage control method for rainwater pipe network overflow pollution according to claim 1, characterized in that, The step of constructing a multi-objective function based on the target ensemble model includes: The target integration model is used to determine the equipment operation characteristics, which include: overflow, energy consumption, chemical consumption, number of equipment start-ups and shutdowns, and concentration of effluent and river water. Based on preset benchmark values, the various equipment operation characteristic quantities are normalized to obtain the water quality assurance sub-objective, the operating cost sub-objective, and the system reliability sub-objective.

3. The linkage control method for rainwater pipe network overflow pollution according to claim 2, characterized in that, The step of determining the device operating characteristic quantities using the target ensemble model includes: The driving data of each device during the overflow pollution during the rainy season is obtained, and the driving data is input into the target integrated model to obtain the operating characteristic quantities of the device.

4. The linkage control method for rainwater pipe network overflow pollution according to claim 1, characterized in that, The step of using the comprehensive performance evaluation function as an instantaneous feedback signal includes: Using a deep value network model, the comprehensive performance evaluation function is used as an instantaneous feedback signal to perform cumulative performance calculations and obtain the cumulative performance target value.

5. The linkage control method for rainwater pipe network overflow pollution according to claim 1, characterized in that, The step of training and processing each drainage device and outputting the target control strategy corresponding to each device includes: Using the cumulative performance target value as the training label, the distributed facility collaborative fusion decision network model is subjected to experience replay processing to obtain the updated target distributed facility collaborative fusion decision network model.

6. The linkage control method for rainwater pipe network overflow pollution according to claim 5, characterized in that, The step of using the cumulative performance target value as training label to perform experience replay processing on the distributed facility collaborative fusion decision network model to obtain the updated target distributed facility collaborative fusion decision network model includes: By updating the model with preset parameters, the mean squared error loss is calculated on the current output of the distributed facility collaborative fusion decision network model and the cumulative performance target value to obtain the loss function. Then, by backpropagating the loss function, the network parameters of the distributed facility collaborative fusion decision network model are updated to obtain the updated target distributed facility collaborative fusion decision network model.

7. The linkage control method for rainwater pipe network overflow pollution according to claim 6, characterized in that, Following the step of obtaining the updated target distributed facility collaborative fusion decision network model, the following is included: The target distributed facility collaborative fusion decision network model is used to execute online decisions of each facility decision unit and output the target control strategy corresponding to each device.

8. A linkage control device for rainwater pipe network overflow pollution, characterized in that, The device includes: A multi-objective function construction module obtains an integrated target model of stormwater pipe network overflow pollution and constructs a multi-objective function based on the integrated target model. The multi-objective function includes: water quality assurance sub-objective, operating cost sub-objective, and system reliability sub-objective. The comprehensive performance evaluation function generation module performs normalization and weighted summation on the three-dimensional objectives of the multi-objective function to obtain the comprehensive performance evaluation function. The intelligent control module, through a preset distributed facility collaborative fusion decision network model, uses the comprehensive performance evaluation function as an instantaneous feedback signal to train various drainage devices, outputs the target control strategies corresponding to each device, and performs safety constraint integration processing before outputting the target control strategies, so as to achieve dynamic equilibrium control of multiple objectives of overflow pollution during the rainy season according to the target control strategies. The step of performing safety constraint integration processing before outputting the target control strategy includes: performing safety constraint integration processing on the following safety constraints: verifying whether the control command to be executed meets the equipment capacity constraints, hydraulic safety constraints, and operational stability constraints through a physical constraint checker; if not, modifying the command or refusing execution; performing rate-of-change and acceleration constraint processing on the control command through a motion smoothing filter to prevent mechanical shock; and when the water level is detected to exceed the safety threshold, activating an emergency braking mechanism, overriding the control command generated by reinforcement learning, and executing a preset emergency control scheme. The training process includes a phased pre-training and imitation learning phase, a multi-objective collaborative optimization phase, and a robust adversarial training phase. The pre-training and imitation learning phase includes: using historical expert operational data to pre-train facility decision-making units through behavioral cloning to obtain a safety baseline strategy; the multi-objective collaborative optimization phase includes: conducting centralized training in a digital twin simulation environment using priority experience replay, double-Q learning, and target network techniques; the robust adversarial training phase includes: injecting sensor noise and rainfall forecast errors into the training environment and using adversarial perturbations to enhance the robustness of the strategy. The calculation formula for the water quality assurance sub-target is as follows: Among them, G qua (t) represents the score of the water quality assurance sub-objective at time t, F spill F(t) represents the score function for the overflow stormwater quality. runoff F(t) represents the score function for surface runoff pollution of water quality. stream F(t) represents the score function for river water quality. effluent (t) represents the score function for the effluent quality of the wastewater treatment plant; λ1, λ2, λ3, and λ4 are the weight coefficients for the overflow rainwater quality, surface runoff pollution quality, river water quality, and wastewater treatment plant effluent quality, respectively, and their values ​​are determined by decision preferences. λ1, λ2, λ3, and λ4 ∈ [0,1] and λ1+λ2+λ3+λ4=1; The formula for calculating the operating cost sub-objective is as follows: Among them, C eco (t) represents the score of the cost sub-objective at time t, C spill (t), C runoff (t), C treat (t), C aux (t) represents the core cost target scores for stormwater overflow control, surface runoff management, sewage treatment and supporting auxiliary systems, respectively; μ1, μ2, μ3, and μ4 are the weight coefficients of stormwater overflow control cost, surface runoff management cost, sewage treatment cost and auxiliary system operation cost, respectively, μ1, μ2, μ3, and μ4 ∈ [0,1] and satisfy μ1+μ2+μ3+μ4=1; The calculation formula for the system reliability sub-objective is as follows: Among them, R rel (t) represents the score of the system reliability sub-objective at time t, R control (t) corresponds to the core facility in the stormwater overflow control system, R govern (t) corresponds to the core facility in the surface runoff management process, R process (t) corresponds to the core facilities in the wastewater treatment process, R support (t) corresponds to the auxiliary support facilities of the whole system; ν1, ν2, ν3, and ν4 are the weight coefficients of the reliability of control facilities, the reliability of treatment facilities, the reliability of processing facilities, and the reliability of auxiliary systems, respectively. ν1, ν2, ν3, and ν4 ∈ [0,1] and satisfy ν1+ν2+ν3+ν4=1.

Images