Pump station operation control method based on deep Q neural network and pipe network modeling technology

Abstract

The invention provides a pump station operation control method based on a deep Q neural network and a pipe network modeling technology. The pump station operation control method comprises the following steps that the pipe diameter size, valve coordinates and water pump performance parameters of a pump station are obtained; calculating on-way loss and local head loss in combination with a steady-state hydraulic equation, establishing a pump station pipe network numerical model, and setting node pressure limiting conditions based on actual operation conditions; constructing an algorithm model framework, establishing a reward mechanism fusing equipment energy consumption, pipe network pressure deviation and flow out-of-limit penalty, and adopting a dynamic exploration strategy to complete neural network training; according to the pipe network tail end water head and flow demand parameters, a water pump control decision is obtained, and a global optimal energy efficiency scheme is output; and evaluating feasibility according to the matching degree of the global optimal energy efficiency scheme and the actual pump set operation parameters, and if the feasibility exceeds an equipment operation threshold value, searching a correction scheme in the neighborhood space of the current solution based on the pipe network simulation model. According to the invention, a scheme for realizing water circulation under the condition of lowest energy consumption of the pump station can be obtained.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control of pump stations, and in particular to a pump station operation control method based on deep Q neural network and pipe network modeling technology. Background Art

[0002] Efficient operation of industrial circulating water systems requires scientifically designed pump station fluid delivery solutions. Within industrial production systems, a systematic flow allocation mechanism, leveraging various fluid control devices, is required to achieve the orderly transfer and recycling of water resources.

[0003] Traditional operational state classification is typically based on a full-combination enumeration approach. This approach can meet basic computing requirements in small-scale systems, but when faced with the coordinated control of multiple devices in complex industrial systems, the system computing load will increase exponentially, resulting in significantly reduced operational efficiency and exponentially longer computing cycles, ultimately affecting the overall process optimization process.

[0004] At present, the commonly used methods in the field of pump station optimization mainly include intelligent algorithms such as dynamic programming, neural network and genetic algorithm. The existing technology discloses an optimization scheduling method for plain polder drainage pump stations. Combining orthogonal experiments and genetic algorithms, through the three-step strategy of intelligent screening of unit combinations, optimization of drainage sequences, and dynamic elimination of over-limit solutions, it can quickly and accurately obtain the global optimal solution for plain polder drainage, and simultaneously solve the problems of traditional scheduling that are time-consuming and have great safety hazards; the existing technology discloses a water supply pump station optimization method and system, through the collaborative optimization of real-coded genetic algorithms and artificial electric field algorithms, to construct a water pump characteristic curve model and a multi-stage solution process, combined with modular system design to achieve energy-saving control of water supply pump stations, taking into account both operating efficiency and equipment safety; the existing technology discloses a long-distance water diversion pump station system scheduling method considering multiple optimizations A method and system are proposed. This method achieves energy consumption optimization and coordinated scheduling of long-distance water diversion systems and pump stations by constructing a pump station optimization model and a machine learning cost accounting system, combining multi-objective optimization and intelligent decision-making methods, and achieving the dual benefits of efficient allocation of water resources and sustainable management. The existing technology discloses an energy-saving scheduling method for water plant intake pump stations based on multi-agent deep reinforcement learning, a water supply scheduling technology based on multi-agent deep reinforcement learning, constructs an LSTM dynamic environment through Markov game modeling, and adopts an actor-attention-critic algorithm to reduce pump station energy consumption by 12.8% while ensuring liquid level / pressure difference / switching safety thresholds, significantly improving the stability of the energy supply system and the strategy generalization ability.

[0005] After in-depth research, it was found that the existing industrial circulating water scheduling algorithms have significant technical bottlenecks: First, dynamic programming faces the problem of dimensionality explosion when dealing with multi-dimensional decision variables, especially when there are too many stage variables, the computational complexity increases geometrically; second, neural network algorithms generally have the risk of training stagnation caused by gradient vanishing, the adaptation dilemma of network topology and engineering practice, and local extreme value traps; and intelligent algorithms such as genetic algorithms are limited by the data storage requirements of high-dimensional solution space, and are prone to optimization failure phenomena such as search path occlusion.

[0006] Furthermore, the current industrial circulating water scheduling technology system generally lacks closed-loop support from simulation models, directly resulting in significant application barriers for the aforementioned methods in engineering practice. Against this technical backdrop, the DQN method demonstrates unique advantages: its Q neural network mitigates the risk of dimensionality explosion; its experience replay mechanism deeply integrates simulation environment training data, enhancing project adaptability while reducing reliance on real-time data; and its learning paradigm, combined with dynamic pipeline network simulation, enables scheduling strategies to precisely align with actual projects, forming a truly intelligent scheduling solution with real-time feedback capabilities. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention provides a pump station operation control method based on deep Q neural network and pipeline network modeling technology. By adopting a DQN algorithm based on a real pipeline network environment to train a pump station scheduling optimization intelligent agent, the optimization solution of the pump station industrial circulating water pump station is realized, thereby obtaining a solution for achieving water circulation with the lowest pump station energy consumption.

[0008] The present invention achieves the above technical objectives through the following technical means.

[0009] A pump station operation control method based on deep Q neural network and pipe network modeling technology includes the following steps:

[0010] S1. Obtain the pump station pipe diameter, valve coordinates and water pump performance parameters through the original design data and manufacturing parameters of the pump station;

[0011] S2. Based on the pump station pipe diameter, valve coordinates, and pump performance parameters, the steady-state hydraulic equation is used to calculate the along-the-line loss and local head loss, establish a numerical model of the pump station network, and set node pressure constraints based on actual operating conditions.

[0012] S3. Build a DQN algorithm model framework, establish a reward mechanism that integrates equipment energy consumption, pipe network pressure deviation, and flow rate limit penalties, and use a dynamic exploration strategy to complete neural network training;

[0013] S4, load the pre-trained DQN algorithm model, according to the water head h at the end of the pipe network s t With flow q st Demand parameters, obtain water pump control decisions, and output the global optimal energy efficiency solution;

[0014] S5. Evaluate the feasibility based on the matching degree between the global optimal energy efficiency solution and the actual pump group operating parameters. If the equipment operation threshold is exceeded, search for a correction solution in the neighborhood space of the current solution based on the pipeline network simulation model.

[0015] Further, step S1 is specifically as follows:

[0016] Cross-map the pipeline direction markings and equipment layout coordinates in the two-dimensional plan of the design drawings with the floor height parameters in the three-dimensional model to form a digital skeleton of the pipeline network that includes geometric dimensions and connection relationships; extract the pump performance curve, rated power, head range and allowable working pressure based on the specification parameter table in the equipment factory materials and the water pump performance test report certification data.

[0017] Further, step S2 is specifically as follows:

[0018] A pipeline network distribution model is established based on the pump station and pipeline network topology data. The pipeline network distribution model sets up several nodes and determines the pipeline loss h between adjacent nodes i and j. ij ;

[0019] Calculate the additional loss h' of local resistance between adjacent nodes i and j ij , which is quantified by assigning a local loss coefficient to the velocity head of the fluid in the pipe, and the calculation formula is:

[0020]

[0021] in, is the local loss coefficient, v and g represent the flow velocity (m / s) and gravitational acceleration (m / s) respectively. 2 ), A ij is the cross-sectional area of the pipe between nodes i and j (m 2 );q ij is the pipeline flow between nodes i and j;

[0022] In the node constraint configuration, according to the requirements of the pump station design specifications, the pressure control range under the operating conditions is set, including the minimum service head limit at the water supply end and the maximum pressure threshold at the end of the pipeline network. At the same time, the pump head-flow characteristic curve data is input to simulate the dynamic response of the actual operating point; the model is generated, and the model accuracy is verified using historical data from on-site pressure sensors.

[0023] Further, step S3 is specifically as follows:

[0024] Based on DQN, a reinforcement learning architecture for pump station control is constructed. A two-layer hidden layer neural network model is established as a decision-making agent. The input layer receives the terminal demand pressure H of the pipeline network.D and demand flow Q D ; The hidden layer uses the ReLU activation function to implement nonlinear mapping; the output layer outputs the reward value R t , representing the current state s t Next select action a t The dimension is consistent with the number of executable actions. Each dimension represents a discrete operation instruction of the pump frequency adjustment amount. The total number of actions is defined as:

[0025]

[0026] Among them, β is the relative speed control accuracy of the motor, α max With α min Represents the relative maximum and minimum speeds allowed by the equipment;

[0027] The operating environment of the neural network is built on the basis of the pipe network simulation model, and the input is the pipe network terminal demand pressure H D and demand flow Q D , the output is the reward value R t ;

[0028] The constraints of the system's operating state are: speed range limitation, head range limitation and pipe network system pressure range limitation;

[0029] When the constraints are met, the reward value of the reward function is defined by the following formula:

[0030]

[0031] Otherwise, the reward value R t Determined by the following formula:

[0032]

[0033] Where q P 、h P and η P are the flow rate, head and efficiency of the pth pump respectively, and P is the total number of pumps;

[0034] Indicates total energy consumption; Q D With H D The product of represents useful work;

[0035] N is the total number of constraints set in the system, x n Used to evaluate the implementation of the nth constraint. If the constraint is met, it is equal to 0, otherwise it is equal to -1;

[0036] C R1 is the energy consumption reward scaling factor, C R2 is the pressure deviation penalty scaling factor, Cn Penalty scaling factor for constraints;

[0037] h m Represents the pressure of the monitoring node.

[0038] Further, step S4 is specifically as follows:

[0039] The trained DQN neural network is loaded into the industrial control system to collect the water head h of the terminal node of the pump station network within the time window t. s t With flow q s t Demand data; through the online reasoning of the neural network model, the full traversal optimization calculation of the pump frequency adjustment plan is realized within the set time, and the pump frequency configuration parameters with the lowest total energy consumption are selected as the output.

[0040] Further, step S5 is specifically as follows:

[0041] If the speed setting value of any pump in the global optimal energy efficiency solution exceeds the continuous operation threshold marked on the nameplate, or the pressure at the terminal node of the pipeline network deviates from the designed safety range, the solution will be judged as infeasible;

[0042] For infeasible solutions, based on the pump frequency combination in the global optimal energy efficiency solution, a gridded candidate solution set is constructed within the neighborhood range of ±2Hz. Then, the DQN algorithm model is used to obtain suboptimal candidate solutions with different frequencies. The suboptimal candidate solutions that simultaneously meet the constraints are screened. The solution with the smallest difference in energy consumption between the suboptimal candidate solution and the global optimal energy efficiency solution is selected as the adjustment output.

[0043] The beneficial effects of the present invention are:

[0044] 1. The pump station operation control method based on deep Q neural network and pipe network modeling technology described in the present invention uses a DQN-based reward mechanism to make the pump group operating point approach the optimal efficiency range. It has a fast calculation speed and meets the requirements of real-time optimization and scheduling, so that pump station optimization can obtain results in a short time and significantly improve energy efficiency.

[0045] 2. The pump station operation control method based on deep Q neural network and pipeline network modeling technology described in the present invention can respond to changes in water demand in real time through the coupling of pipeline network simulation model and reinforcement learning, avoid the lag of manual parameter adjustment, and has good dynamic adaptability.

[0046] 3. The pump station operation control method based on deep Q neural network and pipeline modeling technology described in the present invention and the dual verification mechanism (model warning + neighborhood search) effectively prevent equipment overload and pipeline burst risks, and the safety margin can be guaranteed.

[0047] 4. The pump station operation control method based on deep Q neural network and pipe network modeling technology described in the present invention supports multi-source data modeling from traditional CAD drawings to BIM models, and adapts to actual pump station optimization needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. The drawings described below are some embodiments of the present invention. For ordinary technicians in this field, it is obvious that other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 This is a flow chart of the pump station operation control method based on deep Q neural network and pipe network modeling technology described in the present invention.

[0050] Figure 2 This is a flowchart of a single-round training of the deep Q neural network described in the present invention.

[0051] Figure 3 This is the calling process of the deep Q neural network described in the present invention. DETAILED DESCRIPTION

[0052] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0053] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "axial", "radial", "vertical", "horizontal", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0054] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediary; or internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0055] like Figure 1 As shown, the pump station operation control method based on deep Q neural network and pipe network modeling technology of the present invention includes the following steps:

[0056] S1: Obtain the pump station pipe diameter, valve coordinates, and pump performance parameters through the original design data and manufacturing parameters of the pump station; specifically:

[0057] Parameters, including pipe diameters, geometric coordinates, pump performance curves, and valve loss coefficients, were extracted from the pump station's CAD design drawings, as-built BIM models, on-site survey data, and the manufacturer's technical manuals. By analyzing the coordinate correspondence between the 2D drawings and the 3D model, a digital skeleton of the pipe network, including geometric dimensions and connection relationships, was constructed. Based on the specifications provided in the equipment's factory materials and the certified data from the pump's performance test report, the pump's performance curve, rated power, head range, and allowable operating pressure were extracted.

[0058] S2: Based on the data from S1, a pump station network model is established. The along-the-line losses and local head losses are calculated using the steady-state hydraulic equation. A numerical model of the pump station network is established, and node pressure constraints are set based on actual operating conditions.

[0059] A pipeline network distribution model is established based on the pump station and pipeline network topology data. The pipeline network distribution model sets up several nodes and determines the pipeline loss h between adjacent nodes i and j. ij , which is calculated using the existing Hazen-Williams formula.

[0060] Calculate the additional loss h' of local resistance between adjacent nodes i and j ij , which is quantified by assigning a local loss coefficient to the velocity head of the fluid in the pipe, and the calculation formula is:

[0061]

[0062] in, is the local loss coefficient, v and g represent the flow velocity (m / s) and gravitational acceleration (m / s) respectively. 2 ), A ij is the cross-sectional area of the pipe between nodes i and j (m 2);q ij is the pipeline flow between nodes i and j;

[0063] In node constraint configuration, according to the pump station design specifications, the pressure control range under the operating conditions is set, including the minimum service head limit at the water supply end and the maximum pressure threshold at the end of the pipe network. At the same time, the pump head-flow characteristic curve data is input to simulate the dynamic response of the actual operating point;

[0064] A model was generated and its accuracy verified using historical data from field pressure sensors. Specifically, the pressure control range was set and the pump characteristic curve was loaded. Pipe roughness and local loss coefficients were manually adjusted based on historical sensor data and empirical experience to ensure a head loss calculation error of ≤5%. The completed network model used water demand and pump operating status as outputs, generating network operational results: flow rates in individual pipelines and pressures at each node. These were then used in subsequent steps to determine constraint compliance and calculate system energy consumption.

[0065] S3. Build a DQN algorithm model framework, establish a reward mechanism that integrates equipment energy consumption, pipe network pressure deviation, and flow rate over-limit penalties, and use a dynamic exploration strategy to complete neural network training; Figure 2 As shown:

[0066] Based on DQN, a reinforcement learning architecture for pump station control is constructed, and a two-layer hidden layer neural network model is established as a decision agent, and randomly initialized. The state space is composed of the terminal demand pressure H D and demand flow Q D The environmental state at time t is defined as s t =(Q D ,H D ), as the input layer; the hidden layer uses the ReLU activation function to implement nonlinear mapping; the output layer outputs the reward value R t , representing the current state s t Next select action a t The dimension is consistent with the number of executable actions. Each dimension represents a discrete operation instruction of the pump frequency adjustment amount. The total number of actions is defined as:

[0067]

[0068] Among them, β is the relative speed control accuracy of the motor, α max With α min Represents the relative maximum and minimum speeds allowed by the equipment;

[0069] By state t =(Q D ,H D )Determine action a t, each time a decision is made, the action with the highest reward is selected by the current agent, or an action is randomly selected with a certain probability.

[0070] The operating environment of the neural network is built on the basis of the pipe network simulation model, and the input is the pipe network terminal demand pressure H D and demand flow Q D , the output is the reward value R t ;

[0071] The constraints of the system's operating state are: speed range limitation, head range limitation and pipe network system pressure range limitation;

[0072] When the constraints are met, the reward value of the reward function is defined by the following formula:

[0073]

[0074] Otherwise, the reward value R t Determined by the following formula:

[0075]

[0076] Where q P 、h P and η P are the flow rate, head and efficiency of the pth pump respectively, and P is the total number of pumps;

[0077] Indicates total energy consumption; Q D With H D The product of represents useful work;

[0078] N is the total number of constraints set in the system, x n Used to evaluate the implementation of the nth constraint. If the constraint is met, it is equal to 0, otherwise it is equal to -1. For example, N, assuming there are two pumps in the system, then according to the two constraints of speed and head for each pump, the pressure limits are set at three nodes in the pipe network, then N = 2 + 2 + 3 = 7.

[0079] C R1 is the energy consumption reward scaling factor, C R2 is the pressure deviation penalty scaling factor, C n is the constraint penalty scaling factor; the three are calculated based on the R t The size of the debugger is finally adjusted to make R t In the range of (-10, 10), the effect of excessively large or small reward values on neural network fitting is reduced.

[0080] h m Represents the pressure of the monitoring node.

[0081] Different reward values R t The significance of determination: when the pipeline pressure is within a safe range and the flow rate does not exceed the threshold, energy efficiency is optimized, and the reward value is weighted by the difference between the real-time pump power consumption and the benchmark value; when the node pressure deviates from the set interval or the instantaneous flow rate exceeds the critical value, a penalty term is introduced for negative feedback, and its intensity is adjusted using a scaling factor.

[0082] Based on existing technologies, the intelligent agent is trained for multiple rounds, and finally a neural network is obtained that can accurately obtain the corresponding rewards for all actions in any state.

[0083] S04, load the pre-trained DQN algorithm model, according to the water head h at the end of the pipe network s t With flow q s t Demand parameters, get the pump control decision, and output the global optimal energy efficiency solution, such as Figure 3 As shown, specifically:

[0084] The trained DQN model is integrated into the host computer of the pump station control system. t When the neural network is called, a full traversal optimization calculation is performed within 0.5 seconds, and the pump frequency combination with the lowest energy consumption is output (that is, the action with the highest reward determined by the trained intelligent agent), such as starting two pumps with an operating frequency of 85Hz.

[0085] S05: Evaluate the feasibility based on the matching degree between the global optimal energy efficiency solution and the actual pump group operating parameters. If the equipment operation threshold is exceeded, search for a correction solution in the neighborhood space of the current solution based on the pipe network simulation model. Specifically:

[0086] The pipeline network model checks whether the pump speed exceeds the rated value and the pipeline terminal pressure deviates from the safe range under the S4 decision. Exceeding any of these parameters triggers a correction procedure. A candidate solution is generated every 0.2 Hz within a ±2 Hz range, centered around the pump operating frequency of the currently infeasible solution. Each candidate solution is then applied to the pipeline network simulation model to determine its corresponding pipeline network operation status. Solutions that do not meet the constraints are filtered out, and the feasible solution with the lowest energy consumption is selected as the correction output. The adjusted result is used as the final pump operation decision, and the mechanical execution instructions are generated simultaneously.

[0087] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0088] The series of detailed descriptions listed above are only specific descriptions of feasible embodiments of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent embodiments or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A pump station operation control method based on deep Q neural network and pipe network modeling technology, characterized in that: The steps include: S1. Obtain the pump station pipe diameter, valve coordinates and water pump performance parameters through the original design data and manufacturing parameters of the pump station; S2. Based on the pump station pipe diameter, valve coordinates, and pump performance parameters, the steady-state hydraulic equation is used to calculate the along-the-line loss and local head loss, establish a numerical model of the pump station network, and set node pressure constraints based on actual operating conditions. S3. Build a DQN algorithm model framework, establish a reward mechanism that integrates equipment energy consumption, pipe network pressure deviation, and flow rate limit penalties, and use a dynamic exploration strategy to complete neural network training; S4, load the pre-trained DQN algorithm model, according to the water head h at the end of the pipe network s t With flow q s t Demand parameters, obtain water pump control decisions, and output the global optimal energy efficiency solution; S5. Evaluate the feasibility based on the matching degree between the global optimal energy efficiency solution and the actual pump group operating parameters. If the equipment operation threshold is exceeded, search for a correction solution in the neighborhood space of the current solution based on the pipeline network simulation model.

2. The method for efficient operation control of a pumping station with known demand side based on DQN and pipe network modeling technology according to claim 1 is characterized in that: Step S1 is specifically as follows: Cross-map the pipeline direction markings and equipment layout coordinates in the two-dimensional plan of the design drawings with the floor height parameters in the three-dimensional model to form a digital skeleton of the pipeline network that includes geometric dimensions and connection relationships; extract the pump performance curve, rated power, head range and allowable working pressure based on the specification parameter table in the equipment factory materials and the water pump performance test report certification data.

3. The method for efficient operation control of a pumping station with known demand side based on DQN and pipe network modeling technology according to claim 1 is characterized in that: Step S2 is specifically as follows: A pipeline network distribution model is established based on the pump station and pipeline network topology data. The pipeline network distribution model sets up several nodes and determines the pipeline loss h between adjacent nodes i and j. ij ; Calculate the additional loss h' of local resistance between adjacent nodes i and j ij , which is quantified by assigning a local loss coefficient to the velocity head of the fluid in the pipe, and the calculation formula is: in, is the local loss coefficient, v and g represent the flow velocity (m / s) and gravitational acceleration (m / s) respectively. 2 ), A ij is the cross-sectional area of the pipe between nodes i and j (m 2 );q ij is the pipeline flow between nodes i and j; In the node constraint configuration, according to the requirements of the pump station design specifications, the pressure control range under the operating conditions is set, including the minimum service head limit at the water supply end and the maximum pressure threshold at the end of the pipeline network. At the same time, the pump head-flow characteristic curve data is input to simulate the dynamic response of the actual operating point; the model is generated, and the model accuracy is verified using historical data from on-site pressure sensors.

4. The method for efficient operation control of a pumping station with known demand side based on DQN and pipe network modeling technology according to claim 1 is characterized in that: Step S3 is specifically as follows: Based on DQN, a reinforcement learning architecture for pump station control is constructed. A two-layer hidden layer neural network model is established as a decision-making agent. The input layer receives the terminal demand pressure H of the pipeline network. D and demand flow Q D ; The hidden layer uses the ReLU activation function to implement nonlinear mapping; the output layer outputs the reward value R t , representing the current state s t Next select action a t The dimension is consistent with the number of executable actions. Each dimension represents a discrete operation instruction of the pump frequency adjustment amount. The total number of actions is defined as: Among them, β is the relative speed control accuracy of the motor, α max With α min Represents the relative maximum and minimum speeds allowed by the equipment; The operating environment of the neural network is built on the basis of the pipe network simulation model, and the input is the pipe network terminal demand pressure H D and demand flow Q D , the output is the reward value R t ; The constraints of the system's operating state are: speed range limitation, head range limitation and pipe network system pressure range limitation; When the constraints are met, the reward value of the reward function is defined by the following formula: Otherwise, the reward value R t Determined by the following formula: Where q P 、h P and η P are the flow rate, head and efficiency of the pth pump respectively, and P is the total number of pumps; Indicates total energy consumption; Q D With H D The product of represents useful work; N is the total number of constraints set in the system, x n Used to evaluate the implementation of the nth constraint. If the constraint is met, it is equal to 0, otherwise it is equal to -1; C R1 is the energy consumption reward scaling factor, C R2 is the pressure deviation penalty scaling factor, C n Penalty scaling factor for constraints; h m Represents the pressure of the monitoring node.

5. The method for controlling efficient operation of a pumping station with known demand side based on DQN and pipe network modeling technology according to claim 1 is characterized in that: Step S4 is specifically as follows: The trained DQN neural network is loaded into the industrial control system to collect the water head h of the terminal node of the pump station network within the time window t. s t With flow q s t Demand data; through the online reasoning of the neural network model, the full traversal optimization calculation of the pump frequency adjustment plan is realized within the set time, and the pump frequency configuration parameters with the lowest total energy consumption are selected as the output.

6. The method for controlling efficient operation of a pumping station with known demand side based on DQN and pipe network modeling technology according to claim 1, characterized in that: Step S5 is specifically as follows: If the speed setting value of any pump in the global optimal energy efficiency solution exceeds the continuous operation threshold marked on the nameplate, or the pressure at the terminal node of the pipeline network deviates from the designed safety range, the solution will be judged as infeasible; For infeasible solutions, based on the pump frequency combination in the global optimal energy efficiency solution, a gridded candidate solution set is constructed within the neighborhood range of ±2Hz. Then, the DQN algorithm model is used to obtain suboptimal candidate solutions with different frequencies. The suboptimal candidate solutions that simultaneously meet the constraints are screened. The solution with the smallest difference in energy consumption between the suboptimal candidate solution and the global optimal energy efficiency solution is selected as the adjustment output.

Images