Terminal pressure predictive pump group energy saving control system and method for fluid delivery

Abstract

The application provides an end pressure predictive pump group energy-saving control system for fluid delivery, which significantly reduces the energy consumption of the pump group, greatly improves the end pressure stability, ensures that the digital twin model maintains high precision for a long time, can realize real-time control and global optimization, and prolongs the service life of the pump group. It comprises: a perception layer comprising multiple types of sensors deployed in the pump house and the pipe network for real-time data acquisition; an edge computing layer arranged on site in the pump house, comprising a real-time data acquisition and preprocessing module, a local digital twin engine, an online model self-healing module, a predictive control module, and a real-time decision execution module; a cloud platform layer deployed in a public cloud or private cloud environment, comprising a big data storage and analysis module, a global digital twin model library, a deep prediction model training module, a global pump group scheduling optimization module, and a visual operation and maintenance module; and an execution layer.

Description

Technical Field

[0001] This invention relates to the technical field of fluid machinery control, specifically to an energy-saving control system for a predictive end-pressure pump set for fluid transport. The invention also provides a method for energy-saving control of a predictive end-pressure pump set for fluid transport. Background Technology

[0002] Pump systems consume enormous amounts of electricity in building water supply, industrial circulating water, and HVAC systems. According to statistics, pump systems in water supply and air conditioning systems account for 40% to 60% of total building energy consumption. Traditional pump control systems mainly suffer from the following technical problems: (1) Energy waste problem of traditional constant pressure control; The traditional constant pressure water supply method aims at constant pressure at the pump outlet, without considering the pressure change at the end of the pipeline. During the low water consumption period, the pump outlet maintains constant high pressure operation, and a large amount of energy is consumed in the friction resistance of the pipeline, resulting in serious waste. Existing technologies have improved this problem through end pressure monitoring, but most still use threshold trigger control, which has a delayed response and lacks predictive ability. (2) Existing terminal pressure control systems lack predictive capabilities; some existing technologies have introduced terminal pressure into pump control. For example, CN117662443A discloses an intelligent control system for pumps with terminal pressure monitoring function, which automatically adjusts the outlet pressure through terminal constant pressure control, thus improving the problem of insufficient terminal pressure during peak periods and energy waste during off-peak periods; CN121408195A discloses a variable frequency pump group coordinated adjustment method with terminal pressure feedback, which controls the terminal pressure fluctuation within ±0.02MPa by real-time monitoring of the pressure at the most unfavorable point of the terminal combined with rapid response adjustment within 50ms; however, the above technologies are essentially still reactive control, that is, they only adjust after detecting that the terminal pressure deviates from the target value, which cannot predict future load changes and is difficult to adapt to the dynamic drift of the hydraulic characteristics of the pipeline network; (3) The pump group scheduling strategy lacks global optimization capability; in the parallel operation of multiple pumps, how to select the optimal pump combination and speed ratio according to the real-time operating conditions is a complex multi-objective optimization problem. Existing technologies mostly adopt fixed priority rotation or simple threshold start-stop strategies, which are difficult to maximize energy efficiency. Some studies have attempted to use AI to regulate the constant pressure water supply strategy at the end, by collecting the total flow rate and the real-time health status data of the pumps to generate feasible pump group combination strategies and score them, and select the optimal solution from them. However, such methods are still limited to offline calculation in the cloud, and the response delay is large. (4) The separation of physical model and data-driven approach leads to model distortion. Digital twin-based pump station monitoring technology has been proposed, for example, CN120195983A discloses a digital twin-based pump station operating condition monitoring method, which generates predictive data of operating status through a hybrid digital twin model, improving the ability to identify abnormal operating conditions. However, most existing digital twin models are used for operating condition monitoring and early warning, failing to be deeply coupled with real-time control. As operating time increases, factors such as pipe wear and impeller scaling cause model parameter drift, and the deviation between model prediction and actual physical reality gradually accumulates, leading to model distortion. Existing technologies lack an effective online self-healing mechanism for models, making it impossible to dynamically correct model parameters during operation. (5) Insufficient collaboration between edge and cloud makes it difficult to balance real-time performance and intelligence. Existing technologies achieve millisecond-level real-time control through edge control units, and cloud AI computing power iteratively optimizes the basic control model, which can improve the energy utilization rate of pump rooms by 15% to 20% according to actual tests. However, its edge computing module only calls the pre-loaded basic control model (PID fusion model), and the model optimization in the cloud is an offline batch update, which fails to achieve online adaptive adjustment on the edge side and has limited adaptability to scenarios with rapidly changing operating conditions.

[0003] In summary, existing technologies suffer from problems such as reactive lag control, lack of load forecasting capabilities, lack of global optimization in pump scheduling, lack of online self-healing mechanisms in digital twin models, and insufficient edge-cloud collaboration. There is an urgent need for a predictive pump energy-saving control system that can deeply integrate end-point pressure big data, digital twin models, and cloud-edge collaborative architecture. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides an energy-saving control system for predictive end-point pressure pump sets for fluid transport. This system significantly reduces pump set energy consumption, greatly improves end-point pressure stability, ensures that the digital twin model maintains high accuracy over the long term, and can achieve a balance between real-time control and global optimization, while also extending the pump set lifespan in a balanced manner.

[0005] An energy-saving control system for predictive end-pressure pump sets used for fluid transport, characterized in that it comprises: The sensing layer includes various types of sensors deployed in pump stations and pipelines for real-time data acquisition. The edge computing layer, located at the pump station site, includes a real-time data acquisition and preprocessing module, a local digital twin engine, an online model self-healing module, a predictive control module, and a real-time decision execution module. The cloud platform layer, deployed in a public or private cloud environment, includes a big data storage and analysis module, a global digital twin model library, a deep prediction model training module, a global pump group scheduling optimization module, and a visual operation and maintenance module. The execution layer includes multiple variable frequency pumps and electric valves connected in parallel. The variable frequency pumps are equipped with variable frequency drivers. The execution layer receives control commands from the edge computing layer and executes actions.

[0006] Its further features are: The data collected by the sensing layer includes pump unit operating parameters, pipeline status parameters, and environmental and external data; The pump set operating parameters include the speed, current, power, vibration, and temperature of each pump. The pipeline status parameters include pump outlet pressure, pump outlet flow rate, pressure at each node of the main pipeline, and pressure at the most unfavorable point, wherein the pressure at the most unfavorable point is preferably the pressure at the end of the pipeline. Environmental and external data include pump room temperature and humidity, meteorological data, and smart water meter data at water terminals, with meteorological data obtained through meteorological API access; The real-time data acquisition and preprocessing module of the edge computing layer performs filtering, anomaly removal, and normalization on the sensor data. The local digital twin engine obtains a digital twin model by deploying a lightweight hydraulic network model and a pump characteristic model, and simulates the network pressure distribution in a rolling 250μs cycle. The online model self-healing module dynamically corrects the model parameters based on the deviation between the model prediction and the actual monitoring data using the recursive least squares (RLS) algorithm. The model parameters include, but are not limited to, pipe resistance coefficient and pump efficiency curve coefficient. The predictive control module loads an LSTM / Transformer time series prediction model trained in the cloud, and combines real-time data to predict the pressure change trend and water demand of each end node in the next 5 to 30 minutes to obtain the prediction results. The real-time decision execution module uses the model predictive control (MPC) algorithm to solve for the optimal pump group control command based on the prediction results and digital twin model, and sends the command to the electric valve and frequency converter via EtherCAT / Profinet bus. The big data storage and analysis module of the cloud platform layer is used to persistently store historical operation data, water usage pattern data, and fault logs; The cloud platform layer is configured to periodically synchronize running data from each edge node, retrain the deep prediction model based on global big data, and distribute the updated model to each edge node through containerized deployment. The cloud platform layer is also configured to solve long-term pump group scheduling strategies on a weekly or monthly basis, and distribute them to the edge computing layer as constraints and objective function weight references for model predictive control optimization. A global digital twin model library is used to build a high-precision hydraulic model of the entire pipeline network. The model parameters are periodically synchronized from the edge computing layer to achieve parameter alignment between the edge model and the cloud model. The deep prediction model training module trains LSTM and Transformer deep learning prediction models based on historical big data, and distributes them to each edge node after training is completed. The global pump scheduling optimization module solves for long-term optimization scheduling strategies for multiple pumps and distributes them to the edge computing layer as constraint references. The medium-term optimization scheduling strategies include, but are not limited to, pump rotation plans and optimal efficiency operating intervals. The visual operation and maintenance module provides functions such as system operation status monitoring, energy consumption analysis, and fault early warning.

[0007] A method for energy-saving control of a predictive end-pressure pump set for fluid transport, employing an energy-saving control system for a predictive end-pressure pump set for fluid transport, characterized by comprising the following steps: S1, Data Acquisition: Real-time acquisition of pump unit operating parameters, pipeline status parameters, and environmental and external data through various types of sensors deployed in pump rooms and pipelines; S2, Edge Preprocessing: The edge computing layer preprocesses the collected data, uses real-time control data for millisecond-level control loops, and uploads non-real-time analysis data to the cloud platform. S3, local digital twin real-time simulation, the edge computing layer loads the local digital twin model, rolling simulation of the pressure distribution of each node in the pipeline network, and dynamically corrects the model parameters through an online model self-healing mechanism; S4, load forecasting and terminal pressure trend prediction, the edge computing layer calls the time-series prediction model, and combines real-time data to predict the pressure change trend of each terminal node of the pipeline network in the future time period. S5, Model Predictive Control (MPC) Rolling Optimization: Based on the digital twin model in step S2 and the prediction results in step S3, the Model Predictive Control (MP) algorithm is used to solve the composite optimization objective function and output pump group control commands. S6, command execution, sends control commands to electric valves and frequency converters to adjust the speed of each water pump and the valve opening; S7, closed-loop feedback, the sensor continuously collects the actual pipeline parameters after control and feeds them back to the edge computing layer to form a complete closed loop; S8, cloud-based global optimization and model iteration, the cloud platform layer periodically synchronizes running data from edge nodes, retrains the prediction model and solves the long-term pump group scheduling strategy, and distributes the updated model and strategy to edge nodes.

[0008] Its further features are: The online model self-healing mechanism described in step S4 is that when the deviation between the pressure prediction value of the digital twin model and the measured value of the terminal pressure continues to exceed a preset threshold, the key parameters of the hydraulic model are dynamically corrected using the recursive least squares algorithm so that the prediction deviation of the corrected model is restored to within the threshold. The composite optimization objective function mentioned in step S5 includes the sum of four terms: steady-state energy consumption, terminal pressure deviation penalty, pump switching cost, and pump life balancing.

[0009] A computer-readable storage medium having a computer program stored thereon for performing a method for predictive end-pressure energy-saving control of a pump set for fluid transport.

[0010] Compared with the prior art, the present invention has the following beneficial effects: (1) Significantly reduce pump energy consumption. Through predictive control of terminal pressure and MPC rolling optimization driven by digital twin, the pump unit always operates in the high-efficiency range. Theoretical calculations show that it can save 25%~40% energy compared with traditional constant pressure control, and further save 10%~15% energy compared with existing terminal pressure feedback control. (2) The stability of the terminal pressure is greatly improved. Predictive control eliminates the lag defect of reactive control. In scenarios with rapid load changes, the fluctuation of the terminal pressure can be controlled within ±0.01MPa, which is better than ±0.02MPa of the existing technology. (3) The digital twin model maintains high accuracy over a long period of time. Through the online model self-healing mechanism on the edge side, the model parameters are automatically corrected as the physical system changes, which solves the problem of "distortion" after long-term operation of traditional digital twin models and ensures the reliability of prediction and optimization decisions. (4) Real-time control and global optimization are coordinated and balanced. Edge-side MPC achieves millisecond-level real-time response, while cloud-based deep learning and long-term scheduling strategies achieve global optimization. The two are deeply coupled through the "model distribution-data upload-strategy constraint" mechanism. (5) The pump set life is extended in a balanced manner. The pump set life balance term is incorporated into the MPC objective function and combined with the cloud long-term scheduling strategy to avoid premature aging caused by long-term operation of a single pump, so as to extend the overall life of the pump set in a balanced manner. Attached Figure Description

[0011] Figure 1 This is an overall architecture diagram of a specific embodiment of the system of the present invention; Figure 2 This is a core flowchart of the method of the present invention; Figure 3 This is a schematic diagram of the internal module structure of the edge computing layer of the system of the present invention; Figure 4 The flowchart illustrates the online model self-healing mechanism in the method of this invention. Figure 5 A schematic diagram of the time domain for rolling optimization of the model predictive control (MPC) method of the invention; Figure 6 for Figure 1 A magnified view of a portion of the middle sensory layer; Figure 7 for Figure 1 A magnified view of a portion of the execution layer; Figure 8 for Figure 1 Enlarged view of the middle edge computing layer and cloud platform layer. Detailed Implementation

[0012] Energy-saving control system for predictive end-pressure pump sets for fluid transport, see [link / reference] Figures 1-8 It includes: The sensing layer includes various types of sensors deployed in pump stations and pipelines for real-time data acquisition. The edge computing layer, located at the pump station site, includes a real-time data acquisition and preprocessing module, a local digital twin engine, an online model self-healing module, a predictive control module, and a real-time decision execution module. The cloud platform layer, deployed in a public or private cloud environment, includes a big data storage and analysis module, a global digital twin model library, a deep prediction model training module, a global pump group scheduling optimization module, and a visual operation and maintenance module. The execution layer includes multiple parallel variable frequency pumps and electric valves. The variable frequency pumps are equipped with variable frequency drivers. The execution layer receives control commands from the edge computing layer and executes actions.

[0013] Its further features are: The data collected by the sensing layer includes pump unit operating parameters, pipeline status parameters, and environmental and external data; Pump unit operating parameters include the speed, current, power, vibration, and temperature of each pump; The pipeline status parameters include pump outlet pressure, pump outlet flow rate, pressure at each node of the main pipeline, and pressure at the most unfavorable point, wherein the pressure at the most unfavorable point is preferably the pressure at the end of the pipeline. Environmental and external data include pump room temperature and humidity, meteorological data, and smart water meter data at water terminals, with meteorological data obtained through meteorological API access; The real-time data acquisition and preprocessing module of the edge computing layer performs filtering, anomaly removal, and normalization on sensor data. The local digital twin engine obtains a digital twin model by deploying a lightweight hydraulic network model and a pump characteristic model, and simulates the network pressure distribution in a rolling 250μs cycle. The online model self-healing module dynamically corrects the model parameters based on the deviation between the model prediction and the actual monitoring data using the recursive least squares (RLS) algorithm. The model parameters include, but are not limited to, pipe resistance coefficient and pump efficiency curve coefficient. The predictive control module loads an LSTM / Transformer time series prediction model trained in the cloud, and combines real-time data to predict the pressure change trend and water demand of each end node in the next 5 to 30 minutes to obtain the prediction results. The real-time decision execution module uses the model predictive control (MPC) algorithm to solve for the optimal pump group control command based on the prediction results and digital twin model, and sends the command to the electric valve and frequency converter via EtherCAT / Profinet bus. The big data storage and analysis module at the cloud platform layer is used for persistent storage of historical operation data, water usage pattern data, and fault logs; The cloud platform layer is configured to periodically synchronize running data from each edge node, retrain the deep prediction model based on global big data, and distribute the updated model to each edge node through containerized deployment. The cloud platform layer is also configured to solve long-term pump group scheduling strategies on a weekly or monthly basis, and distribute them to the edge computing layer as constraints and objective function weight references for model predictive control optimization. A global digital twin model library is used to build a high-precision hydraulic model of the entire pipeline network. The model parameters are periodically synchronized from the edge computing layer to achieve parameter alignment between the edge model and the cloud model. The deep prediction model training module trains LSTM and Transformer deep learning prediction models based on historical big data, and distributes them to each edge node after training is completed. The global pump scheduling optimization module solves for long-term optimization scheduling strategies for multiple pumps and distributes them to the edge computing layer as constraint references. The medium-term optimization scheduling strategies include, but are not limited to, pump rotation plans and optimal efficiency operating intervals. The visual operation and maintenance module provides functions such as system operation status monitoring, energy consumption analysis, and fault early warning.

[0014] A method for energy-saving control of end-pressure predictive pump sets for fluid transport, which employs an energy-saving control system for end-pressure predictive pump sets for fluid transport, see [link to relevant documentation]. Figures 1-8 It includes the following steps: S1, Data Acquisition: Real-time acquisition of pump unit operating parameters, pipeline status parameters, and environmental and external data through various types of sensors deployed in pump rooms and pipelines; S2, Edge Preprocessing: The edge computing layer preprocesses the collected data, uses real-time control data for millisecond-level control loops, and uploads non-real-time analysis data to the cloud platform. After receiving the raw data, the edge computing module performs cleaning, filtering, format standardization and compression processing to remove abnormal data. The preprocessed data is divided into real-time control data and non-real-time analysis data. The real-time control data enters the millisecond-level control loop, and the non-real-time analysis data is uploaded to the cloud platform through the standard API interface. S3, local digital twin real-time simulation, the edge computing layer loads the local digital twin model, rolling simulation of the pressure distribution of each node in the pipeline network, and dynamically corrects the model parameters through an online model self-healing mechanism; A lightweight hydraulic network model is loaded at the edge computing layer to simulate the pressure distribution of each node in the network with a rolling simulation period of 250μs. Unlike existing technologies, this invention deploys an online model self-healing module at the edge: this module compares the pressure prediction values ​​of the digital twin model with the measured values ​​of the end pressure sensors in real time and calculates the model prediction deviation. When the deviation continuously exceeds a preset threshold (e.g., 0.01MPa for more than 10 consecutive sampling periods), the recursive least squares (RLS) algorithm is used to dynamically correct the key parameters of the hydraulic model—including the pipe resistance coefficient and pump efficiency curve coefficient of each pipe segment—ensuring that the digital twin model always remains synchronized with the physical network. This mechanism solves the problem of model distortion caused by pipe wear, impeller scaling, and other factors after long-term operation of traditional digital twin models. S4, load forecasting and terminal pressure trend prediction, the edge computing layer calls the time-series prediction model, and combines real-time data to predict the pressure change trend of each terminal node of the pipeline network in the future time period. The edge computing layer calls the deployed time-series prediction model (which is trained on historical big data by the cloud and then distributed), and combines it with current real-time data (including real-time readings of smart water meters at water terminals, historical water usage patterns for the same period, and the current end-point pressure sequence, etc.) to predict the pressure change trend and water demand of each end node of the pipeline network within a time window of 5 to 30 minutes. Compared with the existing reactive end-point pressure control technology (which adjusts only after detecting pressure deviation from the target value), this invention achieves predictive control, that is, before the end-point pressure actually deviates from the target value, the system has already predicted the upcoming load change and adjusted the pump group operation status in advance. S5, Model Predictive Control (MPC) Rolling Optimization: Based on the digital twin model in step S2 and the prediction results in step S3, the Model Predictive Control (MP) algorithm is used to solve the composite optimization objective function and output pump group control commands. The objective function includes the following four items: Steady-state energy consumption term: Predicts the total power consumption of the pump unit within the time domain; End-point pressure deviation penalty: the weighted sum of squared deviations between the pressure at each end point and the target value; Pump set switching costs include the instantaneous energy consumption impact and equipment wear costs caused by pump start-up, shutdown, speed jumps, etc. Pump set life balance item: the balance index of the cumulative running time and cumulative start-stop count of each pump; By solving the above optimization problem in each control cycle (which can be set to 1~5 seconds), the MPC controller outputs the optimal speed sequence of each variable frequency pump and valve opening command in the next control time domain, and sends the first set of commands at the current moment to the execution layer. S6, command execution, sends control commands to electric valves and frequency converters to adjust the speed of each water pump and the valve opening; S7 features closed-loop feedback. Sensors continuously collect actual pipeline parameters after control and feed them back to the edge computing layer, forming a complete closed loop of "perception-twin-prediction-optimization-execution-feedback". At the same time, when the pump group's operating status (such as the cumulative running time of a certain pump exceeding the balance threshold) triggers the rotation condition, the system executes pump group rotation according to the long-term scheduling strategy issued by the cloud, avoiding the problem of neglecting the long-term equipment life balance due to model predictive control focusing on short-term energy consumption. S8 features cloud-based global optimization and model iteration. The cloud platform periodically (e.g., daily) synchronizes running data from each edge node, including actual energy consumption, end-point pressure control accuracy, and model prediction errors. The cloud retrains the deep prediction model based on global big data and distributes the updated model to each edge node through containerized deployment. In addition, the cloud solves long-term pump scheduling strategies (e.g., weekly pump rotation plans) and distributes them to the edge layer as constraints and objective function weights in MPC optimization, achieving synergy between short-term real-time control and long-term global optimization.

[0015] A computer-readable storage medium having a computer program stored thereon for performing a method of predictive end-pressure energy-saving control of a pump set for fluid transport.

[0016] In a specific embodiment, the system of this invention is deployed in the secondary water supply pump room of a 32-story high-rise residential community. The community consists of four buildings, each 32 stories high, with four households per floor. A parallel water supply scheme is adopted, with three variable frequency pumps operating in parallel (two in use and one as backup). The specific deployment is as follows: Sensing layer: Each water pump is equipped with a current transformer, vibration sensor and temperature sensor; a pressure transmitter and electromagnetic flow meter are installed on the pump outlet main pipe; end pressure transmitters are installed on the 30th floor (most unfavorable point), 20th floor and 10th floor of the 4 buildings respectively; data from the smart water meters already installed in the community are connected, and the sampling frequency is the cumulative water consumption once every 15 minutes.

[0017] Edge computing layer: Deploy an industrial edge computing gateway with a 4-core 2.0GHz CPU, 8GB of memory, 64GB of storage, running a Linux real-time operating system, and installing a digital twin engine, an online model self-healing module, an LSTM prediction model, and an MPC controller.

[0018] Cloud platform layer: Use public cloud servers (4 vCPUs, 16GB RAM, 500GB storage) to deploy big data storage and analysis modules, a global digital twin model library, a deep prediction model training module, and a visual operation and maintenance platform.

[0019] Execution layer: The three variable frequency pumps are driven by three variable frequency drives (VFDs) and connected to the edge computing gateway via an EtherCAT bus. The valves are controlled by electric regulating valves.

[0020] Model pre-training, or system initialization phase, involves collecting historical operational data from the past six months of the community (including pump outlet pressure, flow rate, terminal pressure, and water consumption at different times) before the system officially goes into operation. An LSTM time series prediction model is then trained in the cloud. The model input consists of the terminal pressure sequence, flow rate sequence, and external features (time, day of the week, and holiday markers) from the past 24 hours. The output is the terminal pressure prediction sequence for the next 30 minutes. After training, the model is converted to TensorFlow Lite format and deployed to the edge computing gateway.

[0021] During the online operation phase, it operates according to the control method, including the following steps: S1, Data Acquisition: After the system starts up, the edge computing gateway collects all sensor data at 100ms intervals; S2, Edge Data Preprocessing: The lightweight hydraulic network model is loaded with initial parameters (pipe diameter, pipe length, elevation difference, initial pipe resistance coefficient), and the pressure distribution of each node in the network is calculated in a 1-second cycle. S3, Online Model Self-Healing: Suppose that in the third month of system operation, due to slight scaling on the inner wall of the pipe, the actual pipe resistance coefficient increases by about 8% compared to the initial value. At this time, the pressure at the end of the 30th floor predicted by the edge-side digital twin model (assumed to be 0.28 MPa) and the actual sensor measurement (0.30 MPa) consistently deviates by about 0.02 MPa. The online model self-healing module detects that this deviation exceeds the preset threshold of 0.01 MPa for 20 consecutive sampling cycles (i.e., 20 seconds), triggering parameter correction: the pipe resistance coefficient is updated recursively in 20-second windows using the RLS algorithm. After about 5 minutes (about 15 recursions), it converges to the new pipe resistance coefficient value. After this, the model prediction deviation recovers to within ±0.005 MPa, and the digital twin model is realigned with the physical pipe network.

[0022] S4, Load Forecasting and Terminal Pressure Prediction: At 7:20 a.m. on a weekday, the LSTM prediction model, based on the input data from the past 24 hours, predicted that the water consumption in the community would rise rapidly within the next 30 minutes (7:20 to 7:50 a.m.) – this period coincides with the peak time for residents to get up and wash – and the terminal pressure would drop from the current 0.30 MPa to about 0.24 MPa at 7:40 a.m., which is lower than the target value of 0.28 MPa. The system predicted this trend at 7:22 a.m., about 5 minutes before the actual terminal pressure began to drop.

[0023] S5, MPC scrolling optimization: The edge computing gateway performs MPC optimization every 3 seconds. In the 7:22 control cycle, the MPC controller, based on a digital twin model and the predicted end-point pressure trajectory for the next 30 minutes, solves the composite optimization objective function and outputs the speed commands for the three variable frequency pumps in the next control time domain. The optimization result suggests switching from the current scheme of operating only one pump (pump 2, frequency 45Hz) to operating two pumps (pump 2 frequency 38Hz + pump 3 frequency 32Hz), adjusting the pump outlet pressure from 0.48MPa to 0.44MPa. This scheme, while ensuring the end-point pressure is not lower than 0.28MPa, predicts energy consumption to be approximately 12% lower than the "one pump operating at high frequency" scheme.

[0024] S6 and S7 command execution and closed-loop: The first set of commands output by the MPC (38Hz for pump 2, 32Hz for pump 3) is sent to the frequency converter via the EtherCAT bus, and the two pumps complete the speed adjustment within approximately 200ms. Sensors continuously monitor the actual terminal pressure change trajectory. During the peak period of 7:30-7:50, the actual terminal pressure on the 30th floor remains between 0.27 and 0.29 MPa, with fluctuations controlled within ±0.01 MPa. In contrast, if traditional terminal pressure feedback control (frequency increase only after pressure drops below 0.28 MPa) is used, the terminal pressure will drop as low as 0.22 MPa, with maximum fluctuations exceeding ±0.04 MPa.

[0025] S8 Cloud Global Optimization and Model Iteration: Every day at 2:00 AM, the edge computing gateway packages and uploads the day's operational data to the cloud. The cloud platform retrains the LSTM prediction model weekly and calculates pump rotation recommendations—for example, based on the cumulative runtime of each pump, it recommends replacing this week's configuration of "Pump 2 as the main pump, Pump 2 as the auxiliary pump, and Pump 3 as the backup" with "Pump 1 as the main pump, Pump 2 as the auxiliary pump, and Pump 3 as the backup" next week. The updated prediction model and rotation plan are deployed to the edge computing gateway via containerization and automatically take effect the following morning.

[0026] Energy-saving effect: The following table shows the comparison of energy-saving effects measured over three consecutive months on the secondary water supply system of a 32-story high-rise residential community: The actual test results show that: Compared to traditional constant pressure control at the outlet, this invention saves approximately 40% of energy; Compared to existing end-point pressure feedback variable frequency control, this invention achieves an additional energy saving of approximately 18%. The maximum fluctuation range of the terminal pressure was reduced from ±0.02MPa to ±0.01MPa, doubling the control accuracy; After three months of operation, the average prediction error of the digital twin model remained within 3%, verifying the effectiveness of the online self-healing mechanism.

[0027] Compared with the prior art, the present invention has the following beneficial effects: (1) Significantly reduce pump energy consumption. Through predictive control of terminal pressure and MPC rolling optimization driven by digital twin, the pump unit always operates in the high-efficiency range. Theoretical calculations show that it can save 25%~40% energy compared with traditional constant pressure control, and further save 10%~15% energy compared with existing terminal pressure feedback control. (2) The stability of the terminal pressure is greatly improved. Predictive control eliminates the lag defect of reactive control. In scenarios with rapid load changes, the fluctuation of the terminal pressure can be controlled within ±0.01MPa, which is better than ±0.02MPa of the existing technology. (3) The digital twin model maintains high accuracy over a long period of time. Through the online model self-healing mechanism on the edge side, the model parameters are automatically corrected as the physical system changes, which solves the problem of "distortion" after long-term operation of traditional digital twin models and ensures the reliability of prediction and optimization decisions. (4) Real-time control and global optimization are coordinated and balanced. Edge-side MPC achieves millisecond-level real-time response, while cloud-based deep learning and long-term scheduling strategies achieve global optimization. The two are deeply coupled through the "model distribution-data upload-strategy constraint" mechanism. (5) The pump set life is extended in a balanced manner. The pump set life balance term is incorporated into the MPC objective function and combined with the cloud long-term scheduling strategy to avoid premature aging caused by long-term operation of a single pump, so as to extend the overall life of the pump set in a balanced manner.

[0028] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0029] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An energy-saving control system for predictive end-pressure pump sets used for fluid transport, characterized in that, It includes: The sensing layer includes various types of sensors deployed in pump stations and pipelines for real-time data acquisition. The edge computing layer, located at the pump station site, includes a real-time data acquisition and preprocessing module, a local digital twin engine, an online model self-healing module, a predictive control module, and a real-time decision execution module. The cloud platform layer, deployed in a public or private cloud environment, includes a big data storage and analysis module, a global digital twin model library, a deep prediction model training module, a global pump group scheduling optimization module, and a visual operation and maintenance module. The execution layer includes multiple variable frequency pumps and electric valves connected in parallel. The variable frequency pumps are equipped with variable frequency drivers. The execution layer receives control commands from the edge computing layer and executes actions.

2. The energy-saving control system for predictive end-pressure pump sets for fluid transport according to claim 1, characterized in that: The data collected by the sensing layer includes pump unit operating parameters, pipeline status parameters, and environmental and external data.

3. The energy-saving control system for predictive end-pressure pump sets for fluid transport according to claim 1, characterized in that: The pump set operating parameters include the speed, current, power, vibration, and temperature of each pump. The pipeline status parameters include pump outlet pressure, pump outlet flow rate, pressure at each node of the main pipeline, and pressure at the most unfavorable point, wherein the pressure at the most unfavorable point is preferably the pressure at the end of the pipeline. Environmental and external data include pump room temperature and humidity, meteorological data, and smart water meter data at water terminals, with meteorological data obtained through meteorological API access.

4. The energy-saving control system for predictive end-pressure pump sets for fluid transport according to claim 1, characterized in that: The real-time data acquisition and preprocessing module of the edge computing layer performs filtering, anomaly removal, and normalization on the sensor data. The local digital twin engine obtains a digital twin model by deploying a lightweight hydraulic network model and a pump characteristic model, and simulates the network pressure distribution in a rolling 250μs cycle. The online model self-healing module dynamically corrects the model parameters based on the deviation between the model prediction and the actual monitoring data using the recursive least squares (RLS) algorithm. The model parameters include, but are not limited to, pipe resistance coefficient and pump efficiency curve coefficient. The predictive control module loads an LSTM / Transformer time series prediction model trained in the cloud, and combines real-time data to predict the pressure change trend and water demand of each end node in the next 5 to 30 minutes to obtain the prediction results. The real-time decision execution module uses the model predictive control (MPC) algorithm based on the prediction results and digital twin model to solve for the optimal pump group control command, and sends it to the electric valve and frequency converter via EtherCAT / Profinet bus.

5. The energy-saving control system for predictive end-pressure pump sets for fluid transport according to claim 1, characterized in that: The big data storage and analysis module of the cloud platform layer is used to persistently store historical operation data, water usage pattern data, and fault logs; The cloud platform layer is configured to periodically synchronize running data from each edge node, retrain the deep prediction model based on global big data, and distribute the updated model to each edge node through containerized deployment.

6. The energy-saving control system for predictive end-pressure pump sets for fluid transport according to claim 5, characterized in that: The cloud platform layer is also configured to solve long-term pump group scheduling strategies on a weekly or monthly basis, and distribute them to the edge computing layer as constraints and objective function weight references for model predictive control optimization.

7. The energy-saving control system for a predictive end-pressure pump set for fluid transport according to claim 1, characterized in that: A global digital twin model library is used to build a high-precision hydraulic model of the entire pipeline network. The model parameters are periodically synchronized from the edge computing layer to achieve parameter alignment between the edge model and the cloud model. The deep prediction model training module trains LSTM and Transformer deep learning prediction models based on historical big data, and distributes them to each edge node after training is completed. The global pump scheduling optimization module solves for long-term optimization scheduling strategies for multiple pumps and distributes them to the edge computing layer as constraint references. The medium-term optimization scheduling strategies include, but are not limited to, pump rotation plans and optimal efficiency operating intervals. The visual operation and maintenance module provides functions such as system operation status monitoring, energy consumption analysis, and fault early warning.

8. A method for energy-saving control of a predictive end-pressure pump set for fluid transport, comprising the energy-saving control system for a predictive end-pressure pump set for fluid transport as described in any one of claims 1-7, characterized in that, It includes the following steps: S1, Data Acquisition, uses multiple types of sensors deployed in pump rooms and pipelines to collect pump unit operating parameters, pipeline status parameters, and environmental and external data in real time; S2, Edge Preprocessing: The edge computing layer preprocesses the collected data, uses real-time control data for millisecond-level control loops, and uploads non-real-time analysis data to the cloud platform. S3, local digital twin real-time simulation, the edge computing layer loads the local digital twin model, rolling simulation of the pressure distribution of each node in the pipeline network, and dynamically corrects the model parameters through an online model self-healing mechanism; S4, load forecasting and terminal pressure trend prediction, the edge computing layer calls the time series prediction model, and combines real-time data to predict the pressure change trend of each terminal node of the pipeline network in the future time period. S5, Model Predictive Control (MPC) Rolling Optimization: Based on the digital twin model in step S2 and the prediction results in step S3, the Model Predictive Control (MP) algorithm is used to solve the composite optimization objective function and output pump group control commands. S6, command execution, sends control commands to electric valves and frequency converters to adjust the speed of each water pump and the valve opening; S7, closed-loop feedback, the sensor continuously collects the actual pipeline parameters after control and feeds them back to the edge computing layer to form a complete closed loop; S8, cloud-based global optimization and model iteration, the cloud platform layer periodically synchronizes running data from edge nodes, retrains the prediction model and solves the long-term pump group scheduling strategy, and distributes the updated model and strategy to edge nodes.

9. The method for energy-saving control of a predictive end-pressure pump set for fluid transport according to claim 8, characterized in that: The online model self-healing mechanism described in step S4 is that when the deviation between the pressure prediction value of the digital twin model and the measured value of the terminal pressure continues to exceed a preset threshold, the key parameters of the hydraulic model are dynamically corrected using the recursive least squares algorithm so that the prediction deviation of the corrected model is restored to within the threshold. The composite optimization objective function mentioned in step S5 includes the sum of four terms: steady-state energy consumption, terminal pressure deviation penalty, pump switching cost, and pump life balancing.

10. A computer-readable storage medium, characterized in that: It stores a computer program, which is used to perform the method for energy-saving control of end-pressure predictive pump sets for fluid transport as described in claim 8 or 9.

Images