A pipe network dynamic hydraulic balance regulation and control method based on digital twinning and model predictive control

Abstract

This invention relates to the field of intelligent pipeline network operation and control technology, and discloses a method for dynamic hydraulic balance regulation of pipeline networks based on digital twins and model predictive control. The method involves acquiring a physical topology graph network of the pipeline network, combining it with the basic state reference values ​​of the steady-state hydraulic equations, inputting the graph neural network to extract dynamic compensation parameters, and constructing a reduced-order digital twin model of the pipeline network. Real-time monitoring data is input into a state estimator for inversion, outputting a state estimation vector for the entire network and updating the reduced-order digital twin model in a closed loop. A spatiotemporal sequence network is used to process meteorological and historical energy consumption data to predict future flow demand matrices. Based on the state and demand matrix of the entire network, a global optimization problem with the goal of minimizing total pump energy consumption is solved, and pressure reference values ​​for main pipe valves are issued. A local nonlinear predictive control model is used to solve the control increment minimization problem, extracting and issuing the first-step control command to the underlying equipment for physical execution. This invention overcomes model mismatch and achieves hydraulic balance and energy saving.

Description

Technical Field

[0001] This invention relates to the field of intelligent pipeline network operation and control technology, and in particular to a method for dynamic hydraulic balance regulation of pipeline networks based on digital twin and model predictive control. Background Technology

[0002] In recent years, with the acceleration of urbanization, the scale of water supply, heating, and other pipeline systems has become increasingly large, and their operational safety and energy efficiency have become crucial for the construction of smart cities. Pipeline systems are highly nonlinear, strongly coupled, and time-varying systems. Traditional hydraulic balance control of pipeline systems mainly relies on manual experience or simple constant pressure difference / constant flow control strategies. This extensive control method cannot adapt to the spatiotemporal dynamic changes in user-side loads, easily leading to pressure imbalances at unfavorable nodes at the end of the pipeline network, and also significantly increasing the ineffective energy consumption of pumping stations and other equipment.

[0003] Existing hydraulic models based on purely physical mechanisms face numerous bottlenecks in practical applications. On the one hand, due to unmodeled dynamic characteristics such as pipe aging and internal scaling caused by long-term operation of the pipeline network, purely mechanistic models often exhibit severe prediction mismatches over time. On the other hand, solving complex nonlinear partial differential equations is extremely time-consuming, making it difficult to meet the real-time computational requirements of online rolling optimization control. Furthermore, existing technologies typically lack accurate perception of the overall state of the pipeline network and fail to effectively combine meteorological disturbances and historical energy consumption patterns to accurately predict future flow demands, resulting in lag in the control system.

[0004] Therefore, there is an urgent need for a dynamic hydraulic balance control scheme for pipeline networks that can integrate data-driven and physical mechanisms, achieve accurate state estimation and rapid prediction, and achieve a balance between minimizing global energy consumption and high-precision local tracking. Summary of the Invention

[0005] This invention provides a dynamic hydraulic balance control method for pipeline networks based on digital twins and model predictive control. By constructing a reduced-order digital twin model driven by both physical and data and combining it with hierarchical model predictive control, the method can minimize the overall energy consumption of the system while ensuring the hydraulic balance and real-time control of the entire pipeline network.

[0006] This invention provides a dynamic hydraulic balance control method for pipeline networks based on digital twins and model predictive control, applicable to pipeline networks including pumps, valves, and sensing nodes, non-sensing nodes, and user terminal nodes equipped with IoT sensors. The method includes:

[0007] S1. Obtain the physical topology data of the pipeline network to construct a graph network. Input the basic state reference value calculated based on the current operating parameters of the water pump and the current opening degree of the valve into the graph neural network to obtain the dynamic compensation parameters of the pipeline network, and construct a reduced-order digital twin model of the pipeline network based on the dynamic compensation parameters.

[0008] S2. Input the real-time monitoring data collected by the sensing node into the state estimator, perform inversion calculation in combination with the reduced-order digital twin model of the pipeline network, output the current state estimation vector of the entire network and the predicted state of the pipeline network, and update the reduced-order digital twin model of the pipeline network according to the deviation between the predicted state of the pipeline network and the real-time monitoring data.

[0009] S3. Obtain the current meteorological parameters of the area where the pipeline network is located and the historical energy consumption data of the user terminal node, input them into the pre-trained spatiotemporal sequence network, and obtain the flow demand matrix of the user terminal node within the future set prediction step size.

[0010] S4. Based on the network state estimation vector and flow demand matrix, call the updated pipeline reduced-order digital twin model to perform state evolution calculation, solve the global optimization problem with the goal of minimizing the total energy consumption of the water pump, and calculate the optimal frequency sequence of the water pump and the pressure reference value of the valve.

[0011] S5. Solve the local optimization problem with the goal of minimizing the control increment based on the pressure reference value and dynamic compensation parameters to obtain the optimal opening sequence of the valve, and extract the first control command corresponding to the optimal frequency sequence and the optimal opening sequence at the current moment and send it to the water pump and valve for execution.

[0012] Furthermore, S1 specifically includes:

[0013] Extract the geographic information system data of the pipeline network and map the physical topology data into a directed graph consisting of a set of nodes and a set of edges; wherein, for any node in the directed graph, construct a node feature vector containing the node elevation and basic water demand; and for any edge in the directed graph, construct an edge feature vector containing the pipe length, pipe diameter and design roughness coefficient.

[0014] The current operating frequency of each pump and the current opening degree of each valve are read as hydraulic boundary conditions. The steady-state hydraulic equation is solved by order reduction iteration in combination with the initial water demand of each node. The initial pressure estimate of each node and the initial flow estimate of each pipeline after the calculation converges are extracted to form the basic state reference value.

[0015] The baseline state values ​​are concatenated to the corresponding node feature vectors and edge feature vectors, and input into a graph attention network as a graph neural network. The dynamic compensation parameters of the pipeline network are output through the attention aggregation mechanism. The dynamic compensation parameters include the resistance coefficient correction value of each pipeline, the pressure residual compensation value of each node, and the flow residual compensation value of each pipeline.

[0016] The updated virtual resistance coefficient is obtained by superimposing the resistance coefficient correction value onto the initial resistance coefficient of the pipeline in the steady-state hydraulic equation, and the pressure residual compensation value and the flow residual compensation value are superimposed onto the basic state reference value as state correction terms to complete the construction of the reduced-order digital twin model of the pipeline network.

[0017] Furthermore, S2 specifically includes:

[0018] The actual pressure and flow rates at the sensing nodes in the pipeline network are read as real-time monitoring data. The real-time monitoring data is then preprocessed and used as the observation input for the state estimator.

[0019] Using an unscented Kalman filter as a state estimator, a nonlinear state-space model of the pipeline network is established. A set of Sigma points is determined through unscented transformation, and the Sigma point set is substituted into the reduced-order digital twin model of the pipeline network for nonlinear propagation to calculate the one-step predicted state and its covariance matrix at the current time.

[0020] The Kalman gain matrix and the real-time monitoring data are used to update and correct the state of the one-step predicted state, and the estimated state vector of the entire network at the current moment is output.

[0021] The prior predicted value of the observation vector is used as the output of the pipeline network prediction state. The mean relative deviation between the pipeline network prediction state and the real-time monitoring data is calculated. When the mean relative deviation exceeds a preset threshold, the error signal is backpropagated along the topological path of the physical connectivity graph. The virtual resistance coefficient of the local pipeline is finely adjusted online and overwritten into the pipeline network reduced-order digital twin model to complete the update.

[0022] Furthermore, a nonlinear state-space model of the pipeline network is established, specifically including:

[0023] Define the system state vector x as the pressure and flow of all nodes in the entire network, including sensing and non-sensing nodes, and define the observation vector. To monitor data in real time, state prediction equations and observation equations are established:

[0024]

[0025]

[0026] In the formula, This is a one-step prediction of the current state. This is the state estimate from the previous time step. The control inputs are the current pump frequency and valve opening. For the forward propagation mapping function of the reduced-order digital twin model of the pipeline network, This is process noise; Let H be the prior predicted value of the observation vector, and let H be the observation matrix. To observe noise;

[0027] The formula for calculating the network-wide state estimation vector is as follows:

[0028]

[0029] In the formula, This is the estimated vector of the entire network state at the current moment. This is the posterior state estimate. Here is the Kalman gain matrix. This represents the observation vector at the current moment;

[0030] The relative deviation mean The calculation formula is:

[0031]

[0032] In the formula, M represents the total number of sensing nodes. The actual observation data collected by the m-th sensing node Let m be the prior predicted state of the m-th sensing node.

[0033] Furthermore, S3 specifically includes:

[0034] The predictive feature data is extracted and preprocessed by outlier removal and normalization; wherein, the predictive feature data includes current meteorological parameters, historical energy consumption data and time features;

[0035] The spatiotemporal sequence network is constructed by fusing one-dimensional convolutional and long short-term memory networks, and the spatiotemporal sequence network is trained offline using the predicted feature data;

[0036] The current meteorological parameters, current time characteristics, and historical energy consumption data within a recent sliding window are spliced ​​together online in real time to form an online inference feature vector, which is then input into the pre-trained spatiotemporal sequence network for forward propagation calculation.

[0037] The traffic demand trajectories of each user terminal node are output from the spatiotemporal sequence network and combined and assembled in the spatial dimension to obtain a traffic demand matrix.

[0038] Furthermore, S4 specifically includes:

[0039] The system receives the overall network state estimation vector as the initial hydraulic state reference for the current time step, and receives the flow demand matrix as the disturbance boundary condition for changes in external loads within and outside the predicted step size in the future.

[0040] Using the initial hydraulic state baseline as the starting point, and combining the demand forecast values ​​in the flow demand matrix, forward propagation calculations are performed in the updated reduced-order digital twin model of the pipeline network to simulate and calculate the pressure evolution values ​​and flow evolution trajectories of each pump, valve and node in the pipeline network at each predicted time in the future.

[0041] Construct a global optimization objective function with the goal of minimizing the total energy consumption of the integrated operation of the pipeline network and maintaining the water pressure at the nodes;

[0042] The global optimization objective function is iteratively optimized by combining a multidimensional constraint set to find the global optimal control quantity. The optimal frequency sequence to be allocated to each water pump and the pressure reference value to be issued to the main pipe valve are then extracted from the global optimal control quantity.

[0043] Furthermore, the global optimization objective function, which aims to minimize the total energy consumption of the integrated pipeline network operation and maintain the water pressure at the nodes, is formulated as follows:

[0044]

[0045] In the formula, To optimize the objective function globally, Let k be the step size for future predictions, k be the step index of the future prediction control cycle, Pumps be the set of all pumps in the pipeline network, and Nodes be the set of all nodes in the network. Let i be the energy consumption function of the i-th pump in step k. To determine the predicted pressure value of the j-th node at step k, derived from the updated reduced-order digital twin model of the pipeline network. The minimum guaranteed pressure threshold required to maintain the most unfavorable node for hydraulic balance. The weighting coefficients for the energy consumption items are set. This represents the weighting coefficient for the penalty term for exceeding the pressure limit.

[0046] Furthermore, S5 specifically includes:

[0047] The pressure reference values ​​of each valve are received as the target trajectory for local tracking, and the rated flow-pressure drop physical characteristic curves of each valve are adaptively corrected using the dynamic compensation parameters to establish a local nonlinear predictive control model.

[0048] Construct a local optimization cost function with the objectives of minimizing the control increment and minimizing the trajectory tracking error;

[0049] Set physical action constraints for the equipment, perform rolling optimization on the local optimization cost function under the physical action constraints, calculate the control increment sequence, and accumulate the initial opening degree of the current cycle with the control increment sequence to obtain the optimal opening degree sequence of the valve.

[0050] Redundant control quantities in subsequent time steps of the prediction sequence are discarded. The first-step frequency control command and the first-step opening control command corresponding to the current time step are extracted from the optimal frequency sequence and the optimal opening sequence, respectively, to form the first-step control command, which is then sent to the underlying controller for physical execution.

[0051] Furthermore, the local optimization cost function, which aims to minimize the control increment and the trajectory tracking error, is calculated as follows:

[0052]

[0053] In the formula, For local optimization cost function, For the prediction step size of local control, This is the predicted valve outlet pressure at step k, calculated based on a local nonlinear predictive control model. This is the pressure reference value corresponding to step k. Let Q be the control increment at step k, Q be the semi-definite error weight matrix, and R be the positive definite control increment penalty weight matrix.

[0054] This invention also provides a dynamic hydraulic balance control device for pipeline networks based on digital twins and model predictive control, applicable to pipeline networks including pumps, valves, and sensing nodes, non-sensing nodes, and user terminal nodes equipped with IoT sensors. Based on the dynamic hydraulic balance control method for pipeline networks based on digital twins and model predictive control described above, the device includes:

[0055] The construction module is used to acquire the physical topology data of the pipeline network to construct a graph network. The basic state reference value calculated based on the current operating parameters of the water pump and the current opening degree of the valve is input into the graph neural network to obtain the dynamic compensation parameters of the pipeline network, so as to construct a reduced-order digital twin model of the pipeline network based on the dynamic compensation parameters.

[0056] The update module is used to input the real-time monitoring data collected by the sensing node into the state estimator, perform inversion calculation in combination with the reduced-order digital twin model of the pipeline network, output the current state estimation vector of the entire network and the predicted state of the pipeline network, and update the reduced-order digital twin model of the pipeline network according to the deviation between the predicted state of the pipeline network and the real-time monitoring data.

[0057] The prediction module is used to obtain the current meteorological parameters of the area where the pipeline network is located and the historical energy consumption data of the user terminal node, input them into a pre-trained spatiotemporal sequence network, and obtain the flow demand matrix of the user terminal node within a set prediction step size in the future.

[0058] The calculation module is used to call the updated reduced-order digital twin model of the pipeline network to perform state evolution calculation based on the overall network state estimation vector and flow demand matrix, solve the global optimization problem with the goal of minimizing the total energy consumption of the water pump, and calculate the optimal frequency sequence of the water pump and the pressure reference value of the valve.

[0059] The execution module is used to solve a local optimization problem with the goal of minimizing the control increment based on the pressure reference value and dynamic compensation parameters, obtain the optimal opening sequence of the valve, and extract the first control command corresponding to the optimal frequency sequence and the optimal opening sequence at the current moment and send it to the water pump and valve for execution.

[0060] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0061] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0062] The beneficial effects of this invention are as follows:

[0063] This invention constructs a reduced-order digital twin model of a pipeline network by fusing graph neural networks and mechanistic models. This transforms the complex hydraulic extrapolation of the pipeline network into a network forward propagation process, improving computational speed while ensuring fluid dynamics conservation constraints. It also solves the technical problems of traditional mechanistic models being computationally intensive and prone to model mismatch due to pipeline aging. Simultaneously, it combines an unscented Kalman filter to construct a virtual sensor network covering the entire network, achieving high-precision inversion of the hydraulic state of the entire pipeline network and closed-loop adaptive fine-tuning. Furthermore, it utilizes a spatiotemporal sequence network to accurately capture load evolution patterns under meteorological and historical operating conditions to generate a flow demand matrix. Finally, based on a hierarchical distributed model predictive control architecture, the upper layer achieves global dynamic optimization with the goal of minimizing the overall energy consumption of water pumps, while the lower layer achieves local nonlinear predictive control and precise trajectory tracking of regulating valves. Thus, while ensuring the safety baseline of hydraulic balance at unfavorable nodes at the end of the pipeline network, it achieves the minimization of the overall network's operating energy consumption and the smooth execution of control commands. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0065] Figure 2 This is a schematic diagram of the device structure according to an embodiment of the present invention.

[0066] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.

[0067] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0068] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0069] like Figure 1 As shown, this invention provides a dynamic hydraulic balance control method for pipeline networks based on digital twins and model predictive control. This method is applied to pipeline networks containing pumps, valves, and sensing nodes, non-sensing nodes, and user terminal nodes equipped with IoT sensors. The method includes the following steps:

[0070] S1. Obtain the physical topology data of the pipeline network to construct a graph network. Input the basic state reference value calculated based on the current operating parameters of the water pump and the current opening degree of the valve into the graph neural network to obtain the dynamic compensation parameters of the pipeline network, and construct a reduced-order digital twin model of the pipeline network based on the dynamic compensation parameters.

[0071] In one specific embodiment, step S1 includes the following sub-steps:

[0072] S101. Obtain the physical topology data of the pipeline network and map the physical topology structure of the pipeline network into a directed graph network.

[0073] Specifically, the geographic information system (GIS) data of the pipeline network is extracted, and the physical topology of the pipeline network is mapped to a directed graph G=(V,E), where the node set V represents the pipeline intersection point, water source point and user terminal node, and the edge set E represents the pipe, water pump and valve.

[0074] For any node in the graph network Construct its node feature vector The node feature vector contains the node elevation and basic water demand; for any edge in the graph network... Construct its edge feature vector The edge feature vector includes pipe length, pipe diameter, and design roughness coefficient.

[0075] S102. Obtain the current operating parameters of the water pump and the current opening degree of the valve, substitute them into the steady-state hydraulic equation for pre-calculation, and obtain the basic state reference value.

[0076] The steady-state hydraulic equations are conventional equations in the field of fluid mechanics. Their core includes the continuity equation describing the conservation of flow at nodes and the energy equation describing the relationship between pressure drop and flow in a pipeline (such as the Hassen-Williams equation or the Darcy-Weisbach equation).

[0077] At the beginning of the control cycle, the current operating frequency of each pump and the current opening degree of each valve are read as hydraulic boundary conditions. Combined with the initial water demand of each node, the steady-state hydraulic equations are solved using the conventional Newton-Raphson method with a basic reduced-order iterative solution. After the calculation converges, the initial pressure estimates of each node in the pipeline network are extracted. and the initial flow rate estimates for each pipeline The above estimates constitute the basic state reference values.

[0078] S103. Input the baseline state values ​​into the pre-trained graph neural network and output the dynamic compensation parameters of the pipeline network.

[0079] Specifically, in order to eliminate model mismatch caused by physical pipeline aging and dynamic nonlinear features not modeled by pure mechanism models, this invention introduces graph attention network (GAT) to extract nonlinear spatial features.

[0080] The basic state reference value obtained in S102 and Concatenate to the corresponding node feature vector and edge eigenvectors The input is fed into a multi-layer graph attention network. The graph attention network calculates the attention weight coefficients between node i and its neighbor node j using the following formula. :

[0081]

[0082]

[0083] in, , , Let be a learnable weight matrix, and ‖ denote the vector concatenation operation. Let i represent the set of first-order neighboring nodes of node i.

[0084] Through an attention-based aggregation mechanism, the network outputs dynamic compensation parameters under the current operating conditions. These parameters include: correction values ​​for the resistance coefficients of each pipeline. Pressure residual compensation value at each node and the flow residual compensation value of each pipeline .

[0085] To ensure the graph neural network is fully public and can be implemented in engineering, the specific steps for its model training phase are as follows: extract long-term historical SCADA operation data from the pipeline network and store the data in a database; use historical operating data as input and actual collected pressure and flow rate as labels for supervised training. To ensure that the model output does not violate objective fluid laws, this embodiment constructs a joint loss function that includes data-driven loss and physical mechanism loss. :

[0086]

[0087] in, This is the balance coefficient; The mean squared error loss is the error between the network's predicted state and the actual sensor data. The penalty term for violating the physical equations is based on Kirchhoff's first law (conservation of mass) to constrain nodal flows. The specific calculation formula is as follows:

[0088]

[0089] In the formula, and Let i represent the sets of connected pipes flowing into and out of node i, respectively. The loss function, representing the actual flow consumed by node i, forces the network to learn the physical conservation laws of the network during iterative training.

[0090] S104. Construct a reduced-order digital twin model of the pipeline network based on the dynamic compensation parameters.

[0091] Specifically, the resistance coefficient output by the graph attention network is corrected. The updated virtual resistance coefficient is obtained by superimposing it onto the initial pipe resistance coefficient in the original physical equation; simultaneously, the pressure residual compensation value is... and flow residual compensation value As a state correction term, it is added to the basic state reference value. and superior.

[0092] This model construction method, while ensuring physical constraints (mass conservation), transforms the solution of complex nonlinear partial differential equations into a forward propagation process of a graph neural network, compressing the calculation time of the hydraulic state of the entire pipeline network from minutes to milliseconds in traditional mechanistic models. This allows for the construction of a reduced-order digital twin model of the pipeline network that meets the requirements of rapid online rolling optimization of model predictive control (MPC).

[0093] S2. Input the real-time monitoring data collected by the sensing node into the state estimator, perform inversion calculation in combination with the reduced-order digital twin model of the pipeline network, output the current state estimation vector of the entire network and the predicted state of the pipeline network, and update the reduced-order digital twin model of the pipeline network according to the deviation between the predicted state of the pipeline network and the real-time monitoring data.

[0094] In one specific embodiment, step S2 includes the following sub-steps:

[0095] S201. Acquire real-time monitoring data collected by the sensing nodes and perform preprocessing.

[0096] Specifically, in each control cycle (e.g., time interval) When the arrival time is set to 1 minute, the real-time data transmitted from IoT sensors at a limited number of critical nodes in the pipeline network (such as pump station outlets and typical unfavorable ends of the pipeline network) is read through the Supervisory Control and Data Acquisition (SCADA) system. The IoT sensor data includes, but is not limited to, the actual pressure value P and the actual flow rate Q at the sensing node. To ensure the reliability of subsequent state estimation, the collected real-time monitoring data undergoes preprocessing operations such as coarse error removal and filtering / noise reduction. The preprocessed data is then entered into a database or stored as the observation input for the state estimator.

[0097] S202. Input the real-time monitoring data into the state estimator and perform inversion calculations in conjunction with the reduced-order digital twin model of the pipeline network.

[0098] Specifically, since it is impossible to install sensors at all nodes in a real physical pipeline network, and the reduced-order digital twin model of the pipeline network constructed in this embodiment contains a highly nonlinear graph attention network, the state estimator preferably adopts an unscented Kalman filter.

[0099] First, a nonlinear state-space model of the pipeline network is established. The system state vector is defined. Define observation vectors to measure the pressure and traffic of all nodes in the network (including sensing and non-sensing nodes). This refers to the real-time monitoring data obtained in S201.

[0100] The state prediction equation and the observation equation are as follows:

[0101]

[0102]

[0103] in, The forward propagation mapping function for the pipeline network reduced-order digital twin model constructed in the aforementioned steps; For the current control inputs (pump frequency and valve opening); This is the observation matrix, used to extract the state of sensing nodes equipped with sensors from the overall network status; and These are process noise and observation noise, respectively.

[0104] During the inversion calculation, UKF determines a set of Sigma points through Unscented Transform (UT), and substitutes this set of Sigma points into the reduced-order digital twin model of the pipeline network. Nonlinear propagation is performed to calculate the one-step predicted state (i.e., prior state estimate) at the current time. and its covariance matrix.

[0105] S203. Output the current state estimation vector of the entire network and the predicted state of the pipeline network.

[0106] Specifically, in obtaining the predicted state at one step Then, UKF utilizes the Kalman gain matrix. and real-time monitoring data State update and correction (i.e., posterior estimation) of prior states:

[0107]

[0108] The above formula outputs This is the current state estimation vector for the entire network. This vector not only includes the smoothed values ​​of the nodes where the sensors are located, but also accurately reflects the hydraulic state of nodes in the pipeline network without installed sensors, thus forming a "virtual sensor" network covering the entire network. Simultaneously, the prior prediction values ​​of the observed vector... As the output of the pipeline network prediction status, it is used for subsequent model fidelity evaluation.

[0109] S204. When the deviation between the predicted state of the pipeline network and the real-time monitoring data meets the set conditions, update the reduced-order digital twin model of the pipeline network.

[0110] Specifically, physical pipe networks undergo unmodeled dynamic phenomena such as pipe aging and internal scaling during long-term operation. This leads to severe prediction biases in the digital twin model over time (i.e., model mismatch). To eliminate model mismatch, a closed-loop adaptive fine-tuning mechanism is introduced.

[0111] Define the predicted state of the pipeline network With real-time monitoring data The mean of the relative deviations between them is :

[0112]

[0113] Where M is the total number of sensing nodes.

[0114] The setting condition is that the average deviation exceeds a preset threshold (e.g.) When this condition is met, the parameter update mechanism of the digital twin model is automatically triggered. The error signal is backpropagated along the topological path of the physical connectivity graph, and the virtual resistance coefficient S of the local pipeline is fine-tuned online using adaptive filtering or gradient descent. The updated resistance coefficient is then overwritten and stored in the reduced-order digital twin model of the pipeline network, ensuring that the model can track the slow time-varying characteristics of the physical pipeline network in real time, thereby providing a highly faithful prediction engine for the MPC rolling optimization of the next control cycle.

[0115] S3. Obtain the current meteorological parameters of the area where the pipeline network is located and the historical energy consumption data of the user terminal node, input them into the pre-trained spatiotemporal sequence network, and obtain the flow demand matrix of the user terminal node within the future set prediction step size.

[0116] In one specific embodiment, step S3 includes the following sub-steps:

[0117] S301. Obtain and preprocess the predicted feature data.

[0118] Specifically, the current meteorological parameters of the area where the pipeline network is located are obtained through a meteorological data interface. These parameters include temperature and humidity forecasts for the current time and the future forecast period. Simultaneously, historical energy consumption data of user terminal nodes are acquired through the underlying data acquisition link. This historical energy consumption data is specifically represented as historical water or heat flow trajectories. Furthermore, time features are extracted synchronously, including calendar features such as holiday markers, weekday markers, and the current time period.

[0119] To ensure the prediction accuracy of subsequent network models, the acquired raw data undergoes outlier removal and normalization. The normalized data is then stored in a database as a structured time series sequence, serving as offline training samples and online inference input for the spatiotemporal sequence network.

[0120] S302. Construct a spatiotemporal sequence network model that integrates one-dimensional convolution and long short-term memory networks, and train the spatiotemporal sequence network offline.

[0121] Specifically, since the load of user-side terminal nodes is affected by both short-term weather changes and long-term work-rest cycles, the spatiotemporal sequence network is preferably constructed using a one-dimensional convolutional network combined with a long short-term memory network (1D-CNN-LSTM) architecture.

[0122] Among them, the one-dimensional convolutional layer (1D-CNN) serves as the front end for feature extraction. It extracts local features from the input multivariate time series through sliding convolution kernels to capture short-term, high-frequency spatial local fluctuation features such as sudden weather changes or sudden load increases. The long short-term memory network (LSTM) is cascaded after the one-dimensional convolutional layer. It is used to receive the sequence features after convolution mapping and extract the long-term temporal evolution logic of user-side load cycles (such as day-night patterns and weekend patterns) through its internal forget gate, input gate, and output gate mechanisms.

[0123] Historical meteorological parameters, historical time features, and historical energy consumption data are extracted from the database, aligned by time steps, and multimodal concatenation is performed to construct a historical feature sequence vector. This historical feature sequence vector is then input into a pre-constructed 1D-CNN-LSTM network model, and supervised training is performed using actual user-side terminal node traffic consumption as the ground truth label. During training iterations, mean squared error (MSE) is used as the loss function to measure the deviation between the predicted trajectory and the ground truth label. The network's weight parameters are updated using a backpropagation time-based (BPTT) optimization algorithm until the loss function converges, thus obtaining a pre-trained spatiotemporal sequence network.

[0124] S303, online rolling forecasting generates a traffic demand matrix.

[0125] Specifically, during the online dynamic control phase, current meteorological parameters, current calendar features, and historical energy consumption data within recent sliding windows are spliced ​​in real time to form an online inference feature vector, which is then input into a pre-trained spatiotemporal sequence network for forward propagation calculation.

[0126] Network output future prediction step size The traffic demand trajectory of each user terminal node within a control cycle. Assuming the network contains a total of M independent user terminal nodes, the traffic demand trajectories of each node are combined and assembled spatially to form a dimension of... Traffic demand matrix Elements in the traffic demand matrix This matrix explicitly represents the predicted water or heat demand of the i-th user terminal node within the k-th prediction step. This will then be used as an external disturbance boundary condition, input into the subsequent Model Predictive Control (MPC) optimization solution module for state evolution deduction.

[0127] S4. Based on the overall network state estimation vector and flow demand matrix, call the updated reduced-order digital twin model of the pipeline network to perform state evolution calculation, solve the global optimization problem with the goal of minimizing the total energy consumption of the water pump, and calculate the optimal frequency sequence of the water pump and the pressure reference value of the valve.

[0128] In one specific embodiment, step S4 includes the following sub-steps:

[0129] S401. Initialize the boundary conditions and disturbance inputs of the global model predictive control, and call the updated pipeline network reduced-order digital twin model to perform multi-step state evolution deduction.

[0130] Specifically, the current network state estimation vector output from step S2 is received as the initial hydraulic state reference for global model predictive control at the current time step. Simultaneously, the flow demand matrix output from step S3 is received and used as the basis for setting the future prediction step size. Perturbation boundary conditions for changes in internal and external loads. By integrating the current initial state with dynamic disturbances in future time periods, accurate solution boundaries are provided for subsequent global multi-step evolution.

[0131] At the upper layer of the Hierarchical Distributed Model Predictive Control (MPC) architecture (i.e., the global optimization level), using the network-wide state estimation vector as the starting point and combining the demand prediction values ​​of each user terminal node in the traffic demand matrix, forward propagation calculations are performed in the updated reduced-order digital twin model of the pipeline network. This simulates and calculates the future state of the pipeline network. The state evolution trajectory within each time step yields the pressure evolution value and flow evolution trajectory of each pump, valve and node in the pipeline network at each predicted time in the future.

[0132] S402. Construct a global optimization objective function with the goal of minimizing the total energy consumption of the integrated operation of the pipeline network and maintaining the water pressure at the nodes.

[0133] Specifically, to achieve global collaborative optimization among multiple pumping stations and main pipeline valves, and to reduce system energy consumption while ensuring hydraulic balance at all user terminals, a nonlinear global optimization objective function is constructed. The specific formula is as follows:

[0134]

[0135] In the formula, k represents the step index of the future predictive control cycle; Pumps represents the set of all water pumps in the pipeline network; and Nodes represents the set of all nodes in the network. This represents the specific energy consumption function of the i-th pump in the k-th step, which is usually a nonlinear polynomial function of the pump's operating frequency and head. This represents the predicted stress value of the j-th node at step k, derived through the evolution of a reduced-order digital twin model. The minimum guaranteed pressure threshold (i.e., safety red line) required to maintain the most unfavorable node for hydraulic balance. and These are the weighting coefficients for the energy consumption item and the pressure exceeding the limit penalty item, respectively. The left half of the objective function aims to minimize the overall physical energy consumption of the water pump, while the right half uses a penalty function mechanism to force the pressure at any node to not fall below the lower limit of demand, effectively eliminating the redundant head in the pipeline network caused by the traditional constant pressure difference control strategy.

[0136] S403. Solve the global optimization problem by combining the constraints, and output the optimal frequency sequence and the pressure reference value sent down to the local control level.

[0137] Specifically, a multidimensional constraint set is set for the global optimization problem. The constraint set includes physical extreme value constraints of the control equipment (such as the upper and lower limits of the pump operating frequency and the ramp rate limit of the operating frequency) and equality constraints of the network fluid dynamics.

[0138] Under the constraints set described above, the global optimization objective function is optimized using the Sequential Quadratic Programming (SQP) algorithm or other nonlinear programming solvers. Perform iterative optimization. After convergence, calculate the corresponding future control step size. (in The global optimal control quantity is determined within this range. From this global optimal control quantity, the optimal frequency sequence allocated to each pump is extracted, along with the target pressure reference value issued to each key main valve (this reference value will serve as the set trajectory for the local tracking level). This step achieves millisecond-level prediction calculations through a reduced-order digital twin model, avoiding the computational explosion defects of traditional fluid dynamics models in MPC online optimization and ensuring the real-time issuance of optimization commands.

[0139] S5. Solve the local optimization problem with the goal of minimizing the control increment based on the pressure reference value and dynamic compensation parameters to obtain the optimal opening sequence of the valve, and extract the first control command corresponding to the optimal frequency sequence and the optimal opening sequence at the current moment and send it to the water pump and valve for execution.

[0140] In one specific embodiment, step S5 includes the following sub-steps:

[0141] S501. Construct a valve local predictive control model based on dynamic compensation parameters and pressure reference values, and set a local optimization cost function.

[0142] Specifically, in the lower layer (i.e., the local tracking level) of the hierarchical distributed model predictive control (MPC) architecture, its core task is to ensure that each main control valve smoothly and quickly tracks the set target issued by the upper level. First, it receives the pressure reference values ​​of each valve issued in step S4. This is used as the target trajectory for local tracking. Simultaneously, the dynamic compensation parameters extracted in step S1 (including the pipeline resistance coefficient correction value) are received. and node pressure residual compensation value By using dynamic compensation parameters to adaptively correct the rated flow-pressure drop physical characteristic curves of each valve, a local nonlinear predictive control model is established.

[0143] Subsequently, to avoid mechanical wear and water hammer effect in the pipeline caused by frequent valve operation, a local optimization cost function is constructed with the objectives of minimizing the control increment and minimizing the trajectory tracking error. The calculation formula is as follows:

[0144]

[0145] In the formula, The prediction step size for local control; This is the predicted valve outlet pressure value calculated based on the local nonlinear predictive control model at step k. This is the pressure reference value corresponding to step k; Let be the control increment at step k, which is the change in valve opening between adjacent time steps; Q and R are the semi-definite error weight matrix and the positive definite control increment penalty weight matrix, respectively. By adjusting the value of R, the control increment can be minimized as the control objective.

[0146] S502. Solve the local optimization cost function by combining the physical operation constraints of the equipment to obtain the optimal valve opening sequence.

[0147] Specifically, physical motion constraints of the equipment are set during the local optimization process, including absolute opening boundary constraints of the regulating valve (such as limiting the opening range to a certain value). And valve action rate constraints (i.e., the limit value of single-step control increment) (to limit the maximum torque of the physical mechanical actuator).

[0148] Under the aforementioned physical constraints, the local optimization cost function is optimized using sequential quadratic programming (SQP) or the interior-point method. A rolling optimization process is performed. After convergence, the future... The control increment sequence within each control step By accumulating the initial opening degree of the current cycle with the control increment sequence, the optimal opening degree sequence of the valve, which includes the action trajectories of multiple future time steps, can be obtained.

[0149] S503: Extract the initial control command, send it to the underlying controller for physical execution, and complete the data entry operation for the closed loop.

[0150] Specifically, based on the rolling time-domain execution principle of model predictive control, redundant control quantities in subsequent time steps of the prediction sequence are discarded. The first-step frequency control command corresponding to the current time step is extracted from the optimal frequency sequence of the water pump output in step S4, and the first-step opening control command corresponding to the current time step is extracted from the optimal opening sequence output in step S502.

[0151] The initial control commands are converted into standard communication message formats and sent via industrial Ethernet to the frequency converters of each pump and the programmable logic controllers (PLCs) of each valve in the field for physical execution. After the physical actuators complete their actions, the hydraulic conditions of the pipeline network undergo a physical evolution. The new operating status data is collected by the SCADA system, triggering the data storage and entry process, and then enters the preprocessing and calculation loop of the next control cycle, thus forming a complete physical closed loop of "prediction-optimization-execution-feedback". This implementation method, through the decoupling and collaboration of the upper and lower MPCs, not only provides a clear data processing flow, but also meets the requirements for dynamic optimization and adaptive control of the pipeline network system.

[0152] like Figure 2 As shown, the present invention also provides a dynamic hydraulic balance control device for pipeline networks based on digital twins and model predictive control, applicable to pipeline networks including pumps, valves, and sensing nodes, non-sensing nodes, and user terminal nodes equipped with IoT sensors. Based on the dynamic hydraulic balance control method for pipeline networks based on digital twins and model predictive control described above, the device includes:

[0153] Module 1 is used to acquire the physical topology data of the pipeline network to construct a graph network. The basic state reference value calculated based on the current operating parameters of the water pump and the current opening degree of the valve is input into the graph neural network to obtain the dynamic compensation parameters of the pipeline network, so as to construct a reduced-order digital twin model of the pipeline network based on the dynamic compensation parameters.

[0154] Update module 2 is used to input the real-time monitoring data collected by the sensing node into the state estimator, perform inversion calculation in combination with the reduced-order digital twin model of the pipeline network, output the current state estimation vector of the entire network and the predicted state of the pipeline network, and update the reduced-order digital twin model of the pipeline network according to the deviation between the predicted state of the pipeline network and the real-time monitoring data.

[0155] Prediction module 3 is used to obtain the current meteorological parameters of the area where the pipeline network is located and the historical energy consumption data of the user terminal node, input them into a pre-trained spatiotemporal sequence network, and obtain the flow demand matrix of the user terminal node within a set prediction step size in the future.

[0156] Calculation module 4 is used to perform state evolution calculations by calling the updated reduced-order digital twin model of the pipeline network based on the overall network state estimation vector and flow demand matrix, to solve the global optimization problem with the goal of minimizing the total energy consumption of the water pump, and to calculate the optimal frequency sequence of the water pump and the pressure reference value of the valve.

[0157] The execution module 5 is used to solve a local optimization problem with the goal of minimizing the control increment based on the pressure reference value and dynamic compensation parameters, obtain the optimal opening sequence of the valve, and extract the first control command corresponding to the optimal frequency sequence and the optimal opening sequence at the current moment and send it to the water pump and valve for execution.

[0158] Each of the above modules is used to perform the respective steps in the above-mentioned dynamic hydraulic balance control method for pipeline networks based on digital twins and model predictive control. The specific implementation methods are as described in the above-mentioned method embodiments, and will not be repeated here.

[0159] like Figure 3 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores all the data required for the process of the dynamic hydraulic balance control method for pipeline networks based on digital twins and model predictive control. The network interface is used for communication with external terminals via a network connection. The computer program is executed by the processor to implement the dynamic hydraulic balance control method for pipeline networks based on digital twins and model predictive control.

[0160] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0161] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described methods for dynamic hydraulic balance control of a pipeline network based on digital twins and model predictive control.

[0162] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).

[0163] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0164] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for dynamic hydraulic balance control of a pipeline network based on digital twin and model predictive control, applied to a pipeline network including pumps, valves, and sensing nodes, non-sensing nodes, and user terminal nodes equipped with IoT sensors, characterized in that, The method includes: S1. Obtain the physical topology data of the pipeline network to construct a graph network. Input the basic state reference value calculated based on the current operating parameters of the water pump and the current opening degree of the valve into the graph neural network to obtain the dynamic compensation parameters of the pipeline network, and construct a reduced-order digital twin model of the pipeline network based on the dynamic compensation parameters. S2. Input the real-time monitoring data collected by the sensing node into the state estimator, perform inversion calculation in combination with the reduced-order digital twin model of the pipeline network, output the current state estimation vector of the entire network and the predicted state of the pipeline network, and update the reduced-order digital twin model of the pipeline network according to the deviation between the predicted state of the pipeline network and the real-time monitoring data. S3. Obtain the current meteorological parameters of the area where the pipeline network is located and the historical energy consumption data of the user terminal node, input them into the pre-trained spatiotemporal sequence network, and obtain the flow demand matrix of the user terminal node within the future set prediction step size. S4. Based on the network state estimation vector and flow demand matrix, call the updated pipeline reduced-order digital twin model to perform state evolution calculation, solve the global optimization problem with the goal of minimizing the total energy consumption of the water pump, and calculate the optimal frequency sequence of the water pump and the pressure reference value of the valve. S5. Solve the local optimization problem with the goal of minimizing the control increment based on the pressure reference value and dynamic compensation parameters to obtain the optimal opening sequence of the valve, and extract the first control command corresponding to the optimal frequency sequence and the optimal opening sequence at the current moment and send it to the water pump and valve for execution.

2. The method for dynamic hydraulic balance regulation of pipeline networks based on digital twin and model predictive control according to claim 1, characterized in that, S1 specifically includes: Extract the geographic information system data of the pipeline network and map the physical topology data into a directed graph consisting of a set of nodes and a set of edges; wherein, for any node in the directed graph, construct a node feature vector containing the node elevation and basic water demand; and for any edge in the directed graph, construct an edge feature vector containing the pipe length, pipe diameter and design roughness coefficient. The current operating frequency of each pump and the current opening degree of each valve are read as hydraulic boundary conditions. The steady-state hydraulic equation is solved by order reduction iteration in combination with the initial water demand of each node. The initial pressure estimate of each node and the initial flow estimate of each pipeline after the calculation converges are extracted to form the basic state reference value. The baseline state values ​​are concatenated to the corresponding node feature vectors and edge feature vectors, and input into a graph attention network as a graph neural network. The dynamic compensation parameters of the pipeline network are output through the attention aggregation mechanism. The dynamic compensation parameters include the resistance coefficient correction value of each pipeline, the pressure residual compensation value of each node, and the flow residual compensation value of each pipeline. The updated virtual resistance coefficient is obtained by superimposing the resistance coefficient correction value onto the initial resistance coefficient of the pipeline in the steady-state hydraulic equation, and the pressure residual compensation value and the flow residual compensation value are superimposed onto the basic state reference value as state correction terms to complete the construction of the reduced-order digital twin model of the pipeline network.

3. The method for dynamic hydraulic balance control of pipeline networks based on digital twin and model predictive control according to claim 1, characterized in that, S2 specifically includes: The actual pressure and flow rates at the sensing nodes in the pipeline network are read as real-time monitoring data. The real-time monitoring data is then preprocessed and used as the observation input for the state estimator. Using an unscented Kalman filter as a state estimator, a nonlinear state-space model of the pipeline network is established. A set of Sigma points is determined through unscented transformation, and the Sigma point set is substituted into the reduced-order digital twin model of the pipeline network for nonlinear propagation to calculate the one-step predicted state and its covariance matrix at the current time. The Kalman gain matrix and the real-time monitoring data are used to update and correct the state of the one-step predicted state, and the estimated state vector of the entire network at the current moment is output. The prior predicted value of the observation vector is used as the output of the pipeline network prediction state. The mean relative deviation between the pipeline network prediction state and the real-time monitoring data is calculated. When the mean relative deviation exceeds a preset threshold, the error signal is backpropagated along the topological path of the physical connectivity graph. The virtual resistance coefficient of the local pipeline is finely adjusted online and overwritten into the pipeline network reduced-order digital twin model to complete the update.

4. The method for dynamic hydraulic balance regulation of pipeline networks based on digital twin and model predictive control according to claim 3, characterized in that, Establishing a nonlinear state-space model of the pipeline network specifically includes: Define the system state vector x as the pressure and flow of all nodes in the entire network, including sensing and non-sensing nodes, and define the observation vector. To monitor data in real time, state prediction equations and observation equations are established: In the formula, This is a one-step prediction of the current state. This is the state estimate from the previous moment. The control inputs are the current pump frequency and valve opening. For the forward propagation mapping function of the reduced-order digital twin model of the pipeline network, This is process noise; Let H be the prior predicted value of the observation vector, and let H be the observation matrix. To observe noise; The formula for calculating the network-wide state estimation vector is as follows: In the formula, This is the estimated vector of the entire network state at the current moment. This is the posterior state estimate. Here is the Kalman gain matrix. This represents the observation vector at the current moment; The relative deviation mean The calculation formula is: In the formula, M represents the total number of sensing nodes. The actual observation data collected by the m-th sensing node Let m be the prior predicted state of the m-th sensing node.

5. The method for dynamic hydraulic balance regulation of pipeline networks based on digital twin and model predictive control according to claim 1, characterized in that, S3 specifically includes: The predictive feature data is extracted and preprocessed by outlier removal and normalization; wherein, the predictive feature data includes current meteorological parameters, historical energy consumption data, and time characteristics; The spatiotemporal sequence network is constructed by fusing one-dimensional convolutional and long short-term memory networks, and the spatiotemporal sequence network is trained offline using the predicted feature data; The current meteorological parameters, current time characteristics, and historical energy consumption data within a recent sliding window are spliced ​​together online in real time to form an online inference feature vector, which is then input into the pre-trained spatiotemporal sequence network for forward propagation calculation. The traffic demand trajectories of each user terminal node are output from the spatiotemporal sequence network and combined and assembled in the spatial dimension to obtain a traffic demand matrix.

6. The method for dynamic hydraulic balance regulation of pipeline networks based on digital twin and model predictive control according to claim 1, characterized in that, S4 specifically includes: The system receives the overall network state estimation vector as the initial hydraulic state reference for the current time step, and receives the flow demand matrix as the disturbance boundary condition for changes in external loads within and outside the predicted step size in the future. Using the initial hydraulic state baseline as the starting point, and combining the demand forecast values ​​in the flow demand matrix, forward propagation calculations are performed in the updated reduced-order digital twin model of the pipeline network to simulate and calculate the pressure evolution values ​​and flow evolution trajectories of each pump, valve and node in the pipeline network at each predicted time in the future. Construct a global optimization objective function with the goal of minimizing the total energy consumption of the integrated operation of the pipeline network and maintaining the water pressure at the nodes; The global optimization objective function is iteratively optimized by combining a multidimensional constraint set to find the global optimal control quantity. The optimal frequency sequence to be allocated to each water pump and the pressure reference value to be issued to the main pipe valve are then extracted from the global optimal control quantity.

7. The method for dynamic hydraulic balance regulation of pipeline networks based on digital twin and model predictive control according to claim 6, characterized in that, The global optimization objective function, which aims to minimize the total energy consumption of the integrated pipeline network operation and maintain the water pressure at key nodes, is formulated as follows: In the formula, To optimize the objective function globally, Let k be the step size for future predictions, k be the step index of the future prediction control cycle, Pumps be the set of all pumps in the pipeline network, and Nodes be the set of all nodes in the network. Let i be the energy consumption function of the i-th pump in step k. To determine the predicted pressure value of the j-th node at step k, derived from the updated reduced-order digital twin model of the pipeline network. The minimum guaranteed pressure threshold required to maintain the most unfavorable node for hydraulic balance. The weighting coefficients for the energy consumption items are set. This represents the weighting coefficient for the penalty term for exceeding the pressure limit.

8. The method for dynamic hydraulic balance regulation of pipeline networks based on digital twin and model predictive control according to claim 1, characterized in that, S5 specifically includes: The pressure reference values ​​of each valve are received as the target trajectory for local tracking, and the rated flow-pressure drop physical characteristic curves of each valve are adaptively corrected using the dynamic compensation parameters to establish a local nonlinear predictive control model. Construct a local optimization cost function with the objectives of minimizing the control increment and minimizing the trajectory tracking error; Set physical action constraints for the equipment, perform rolling optimization on the local optimization cost function under the physical action constraints, calculate the control increment sequence, and accumulate the initial opening degree of the current cycle with the control increment sequence to obtain the optimal opening degree sequence of the valve. Redundant control quantities in subsequent time steps of the prediction sequence are discarded. The first-step frequency control command and the first-step opening control command corresponding to the current time step are extracted from the optimal frequency sequence and the optimal opening sequence, respectively, to form the first-step control command, which is then sent to the underlying controller for physical execution.

9. The method for dynamic hydraulic balance regulation of pipeline networks based on digital twin and model predictive control according to claim 8, characterized in that, The constructed local optimization cost function, which aims to minimize the control increment and the trajectory tracking error, is calculated as follows: In the formula, For local optimization cost function, For the prediction step size of local control, This is the predicted valve outlet pressure value calculated based on the local nonlinear predictive control model at the k-th step. This is the pressure reference value corresponding to step k. Let Q be the control increment at step k, Q be the semi-definite error weight matrix, and R be the positive definite control increment penalty weight matrix.

10. A dynamic hydraulic balance control device for a pipeline network based on digital twin and model predictive control, applied to a pipeline network including pumps, valves, and sensing nodes, non-sensing nodes, and user terminal nodes equipped with IoT sensors, based on the dynamic hydraulic balance control method for a pipeline network based on digital twin and model predictive control as described in any one of claims 1 to 9, characterized in that... The device includes: The construction module is used to acquire the physical topology data of the pipeline network to construct a graph network. The basic state reference value calculated based on the current operating parameters of the water pump and the current opening degree of the valve is input into the graph neural network to obtain the dynamic compensation parameters of the pipeline network, so as to construct a reduced-order digital twin model of the pipeline network based on the dynamic compensation parameters. The update module is used to input the real-time monitoring data collected by the sensing node into the state estimator, perform inversion calculation in combination with the reduced-order digital twin model of the pipeline network, output the current state estimation vector of the entire network and the predicted state of the pipeline network, and update the reduced-order digital twin model of the pipeline network according to the deviation between the predicted state of the pipeline network and the real-time monitoring data. The prediction module is used to obtain the current meteorological parameters of the area where the pipeline network is located and the historical energy consumption data of the user terminal node, input them into a pre-trained spatiotemporal sequence network, and obtain the flow demand matrix of the user terminal node within a set prediction step size in the future. The calculation module is used to call the updated reduced-order digital twin model of the pipeline network to perform state evolution calculation based on the overall network state estimation vector and flow demand matrix, solve the global optimization problem with the goal of minimizing the total energy consumption of the water pump, and calculate the optimal frequency sequence of the water pump and the pressure reference value of the valve. The execution module is used to solve a local optimization problem with the goal of minimizing the control increment based on the pressure reference value and dynamic compensation parameters, obtain the optimal opening sequence of the valve, and extract the first control command corresponding to the optimal frequency sequence and the optimal opening sequence at the current moment and send it to the water pump and valve for execution.

Images