A siphon type axial flow pump flexible starting control method and system based on machine learning prediction

Abstract

The present application provides a kind of siphon type axial flow pump flexible starting control method and system based on machine learning prediction, belongs to axial flow pump control technical field, including S1: after the start of axial flow pump, the starting state feature vector of each sampling time in current control period is collected;S2: calculate the head and torque of each sampling time in current control period;S3: the starting state feature vector of all sampling times in control sampling period, head and torque are constructed sample;Sample is input into pre-trained machine learning prediction model, and the prediction result is output by model;S4: according to the prediction result of each prediction sampling time, calculate the starting risk of each prediction sampling time;S5: according to the prediction result in S3, combined with the starting risk in S4, differential control strategy is used.The technical problems that existing fixed starting rule and fixed air valve closing time are difficult to adapt to different starting conditions and are prone to starting impact in air valve closing or siphon forming stage are solved.

Description

Technical Field

[0001] This invention belongs to the field of axial flow pump control technology, specifically relating to a flexible start-up control method and system for a siphon axial flow pump based on machine learning prediction. Background Technology

[0002] Axial flow pumps, with their characteristics of large flow rate, low head, and high specific speed, are widely used in inter-basin water transfer, flood control and drainage, agricultural irrigation, and urban drainage projects. For axial flow pump stations equipped with siphon-type outlet channels, the unit relies on hydraulic action to gradually expel air from the channel during startup, causing the water flow to overcome the hump and form a stable siphon effect. This process involves impeller acceleration, flow rate establishment, head change, gas-liquid interface migration, and the opening and closing of the air valve, and is a typical unsteady transient process.

[0003] Existing pump station start-up controls often employ fixed start-up times, speed increase patterns, or air valve closing times. While these control methods are easy to implement, their parameters are typically set during the commissioning phase, making dynamic adjustments difficult during actual operation based on factors such as the inlet / outlet water level difference, residual air volume in the flow channel, air valve exhaust efficiency, and the unit's real-time vibration status. If the air valve closes too early, a large air pocket may remain in the flow channel, causing a sudden change in flow resistance, leading to an increase in peak head, torque, and axial force. If it closes too late, it prolongs the hydraulic air-carrying phase, increasing the unit's operating time in unstable flow conditions.

[0004] In practical engineering, signals such as pressure, flow rate, speed, motor current, vibration, and air valve opening during the startup process exhibit significant temporal correlations. Traditional threshold control can only respond based on current measurements, failing to predict potential hydraulic shock peaks caused by air valve closure or speed adjustments within the next few seconds. Furthermore, it struggles to implement differentiated control strategies for different stages, such as hydraulic air drive, hydraulic air carrying, and siphon stabilization.

[0005] Therefore, there is an urgent need to propose a flexible start-up control method. This method can comprehensively utilize real-time and historical data during the start-up process for prediction, and coordinately adjust the speed curve and air valve opening curve based on the prediction results, thereby achieving smoother and safer start-up control. Summary of the Invention

[0006] This invention proposes a flexible start-up control method and system for siphon-type axial flow pumps based on machine learning prediction. It solves the technical problems that existing fixed start-up rules and fixed air valve closing times are difficult to adapt to different start-up conditions and are prone to start-up shocks during the air valve closing or siphon formation stages.

[0007] To achieve the above objectives, the present invention proposes the following technical content:

[0008] A flexible start-up control method for a siphon-type axial flow pump based on machine learning prediction includes the following steps:

[0009] S1: After the axial flow pump starts, the monitoring data at each sampling moment within the sampling control cycle is sampled, and the start-up state feature vector is constructed;

[0010] S2: Based on the monitoring data, calculate the head and torque at each sampling moment within the current control cycle;

[0011] S3: Construct samples from the start-up state feature vectors, head, and torque at each sampling moment within the current control cycle; input the samples into a pre-trained machine learning prediction model, and have the model output prediction results; the prediction results include: head, torque, axial force, pressure pulsation amplitude, vibration amplitude, residual air volume, and start-up stage at multiple consecutive sampling moments in the future;

[0012] S4: Calculate the activation risk at each prediction sampling time based on the prediction results at each prediction sampling time;

[0013] S5: Based on the prediction results in S3 and combined with the start-up risk in S4, adopt a differentiated control strategy.

[0014] Furthermore, for the e-th control cycle, in steps S1 and S2, the start-up state feature vector at time k is:

[0015]

[0016] In the formula, This represents the feature vector of the starting state at time k; This represents the water level in the inlet pool at time k. This represents the water level in the outlet pool at time k. This represents the flow rate at time k. This represents the impeller speed at time k; This represents the motor current at time k; This represents the motor voltage at time k; This represents the pressure in the inlet channel at time k; This represents the pressure in the outlet channel at time k; This represents the air valve opening at time k; This represents the vibration amplitude at time k. This represents the amplitude of the pressure pulsation at time k; This represents the rate of increase in rotational speed at time k; This represents the flow rate growth rate at time k. This represents the rate of change of head at time k;

[0017] The corresponding head is:

[0018]

[0019] In the formula, This represents the head at time k; This represents the pressure in the inlet channel at time k; This represents the pressure in the outlet channel at time k; Indicates fluid density; Represents gravitational acceleration; This indicates the average flow velocity at the outlet cross-section; This indicates the average flow velocity at the inlet cross-section; Indicates the elevation of the pressure measuring point on the water outlet side; Indicates the elevation of the pressure measuring point on the inlet side;

[0020] The corresponding torque is:

[0021]

[0022] In the formula, This represents the torque at time k; Indicates the efficiency of the motor and transmission system; This represents the motor voltage at time k; This represents the motor current at time k; The angle represents the power factor at time k. This represents the impeller angular velocity at time k.

[0023] Further, in step S3, the number of samples in the continuous sampling interval input to the machine learning prediction model is set to M, and the sampling times are 1, 2, ..., M; the machine learning prediction model outputs prediction results for N consecutive sampling times, with prediction times of M+1, M+2, ..., M+N; then the j-th prediction result is:

[0024]

[0025] In the formula, This represents the prediction result at the j-th prediction sampling time within the prediction window. And it is an integer; This represents the predicted head at the j-th sampling time. This represents the torque predicted at the j-th sampling time. This represents the axial force predicted at the j-th sampling time. This represents the predicted pressure pulsation amplitude at the j-th sampling time. This represents the predicted vibration amplitude at the j-th sampling time. The residual air volume is the predicted value of the residual air volume at the j-th prediction sampling time within the prediction window, and the residual air volume is used to characterize the state of the airbag. This indicates the start-up phase at the j-th sampling time of the prediction. The start-up phase includes: the hydraulic air-driving phase, the hydraulic air-carrying phase, or the siphon stabilization phase.

[0026] Furthermore, in step S4, the activation risk of the j-th prediction result is:

[0027]

[0028] In the formula, This represents the predicted startup risk at the j-th sampling time. Indicates the weighting coefficient. ; These represent the torque limit, axial force limit, head limit, pressure pulsation limit, vibration amplitude limit, and residual air volume limit, respectively.

[0029] Furthermore, in step S5, the strategy for differentiated control is as follows:

[0030] If the future continues L s If all predicted sampling times belong to the same startup phase, then that startup phase is taken as the phase confirmation result of the current control cycle; if the future consecutive L s If the start-up phases of the predicted sampling times are inconsistent, the current control cycle will be treated as a transitional state, the air valve will remain fully open or partially open, and the judgment will be re-evaluated based on the prediction results of the next control cycle.

[0031] 1. When the current control cycle stage confirmation result is the hydraulic aeration stage, the air valve remains fully open, and the speed gradually increases according to the unit's existing flexible start-up curve.

[0032] 2. When the current control cycle stage confirmation result is the hydraulic air-carrying stage, the air valve remains fully open or partially open to allow air in the flow channel to continue to be discharged; when the start-up risk and airbag status within the future prediction window meet the allowable valve closing conditions, the controller gradually closes the air valve according to the set air valve opening curve; when the start-up risk or airbag status within the future prediction window does not meet the allowable valve closing conditions, the air valve remains fully open or partially open.

[0033] 3. When the current control cycle stage confirmation result is the siphon stabilization stage, if the airbag state in the future prediction window is not higher than the set airbag state threshold, and the start-up risk in the future prediction window is not higher than the set start-up risk threshold, the controller controls the air valve to gradually close according to the set air valve opening curve, and smoothly transitions the speed to the target speed.

[0034] Furthermore, the opening and closing strategy of the air valve is as follows:

[0035]

[0036] In the formula, G c (e) represents the stage confirmation result of the e-th control cycle; Ω c This indicates the set of permitted valve-closing stages, including the hydraulic air-carrying stage and the siphon stabilization formation stage; This indicates the air valve closing permission flag for the e-th control cycle; when When, it indicates that the controller allows the air valve to enter the closing process; when When this time, it indicates that the air valve remains fully open or partially open; This indicates the risk of initiation at the j-th prediction sampling time. This indicates the set threshold for the risk of valve closure. This represents the normalized airbag state at the j-th prediction sampling time. This indicates the set threshold value for the airbag to be in a closed-valve state.

[0037] The normalized airbag state is as follows:

[0038]

[0039] In the formula, This represents the predicted residual air volume at the j-th prediction sampling time. This indicates the air volume at the initial startup time or under the set reference state.

[0040] when The controller will determine the moment when the valve closing condition is met as the start time of flexible valve closing. The air valve is controlled to gradually close according to the following air valve opening curve:

[0041]

[0042] in:

[0043]

[0044] In the formula, a(t) represents the relative opening of the air valve, a(t)=1 indicates that the air valve is fully open, and a(t)=0 indicates that the air valve is completely closed; Indicates the start time of flexible valve closure determined by the prediction results; T c ξ represents the duration from fully open to fully closed of the air valve; ξ represents the normalized valve closing time.

[0045] when At this time, the controller does not perform the valve closing action, the air valve remains fully open or partially open, and the prediction results continue to be updated based on subsequent sampling data.

[0046] Furthermore, in Strategy 3, the transition curve for rotational speed is as follows:

[0047]

[0048] In the formula, n(t) represents the rotational speed at time t; t1 represents the starting time of the rotational speed transition determined after satisfying the airbag status and start-up risk comparison conditions; t2 represents the time when the rotational speed reaches the target speed; n(t1) represents the rotational speed at the starting time of the rotational speed transition; n(t2) represents the target speed; γ represents the exponent of the drive curve determined by the controller's flexible start-up optimization module for the time period [t1, t2].

[0049] A flexible start-up control system for a siphon-type axial flow pump based on machine learning prediction includes the following modules:

[0050] Data acquisition module: Used to receive multi-source data of axial flow pump output from sensor group, and to complete data buffering, sampling frequency unification and time synchronization, so as to provide continuous time series data for subsequent prediction;

[0051] Feature construction module: used to calculate startup state features based on the raw data output by the data acquisition module;

[0052] Machine learning prediction module: used to predict the startup impact parameters and startup stage within a preset time window based on the startup state characteristics at several consecutive sampling times;

[0053] Start-up Risk Assessment Module: This module compares the start-up impact parameters output by the machine learning prediction module with preset limits and synthesizes the start-up risk assessment results.

[0054] Flexible start optimization module: used to generate speed setpoint curve and air valve opening curve based on start risk assessment results;

[0055] Speed ​​control module: Used to send the speed command curve to the frequency converter or speed control device, so that the impeller speed changes according to the flexible start curve;

[0056] Air valve control module: Used to send the air valve opening curve to the air valve actuator to realize the air valve opening, holding, partial closing, delayed closing or continuous closing;

[0057] Safety protection module: Used to perform speed reduction, keep the air valve open, delay valve closure, alarm or shutdown protection when the predicted or measured value exceeds the safety threshold, sensor data is abnormal, air valve operation is abnormal or the machine learning prediction module output is invalid.

[0058] The beneficial effects that can be achieved by adopting the above technologies are:

[0059] 1. Compared with existing fixed speed increase patterns and fixed air valve closing times, this invention can use machine learning prediction models to predict in advance the head, torque, axial force, pressure pulsation amplitude, vibration amplitude, airbag status and start-up stage within future time windows, and judge the start-up impact risk before the air valve closes or the speed transitions, avoiding lagging control based solely on current measured values ​​or fixed start-up times.

[0060] 2. This invention determines whether the air valve can be closed by combining the phase confirmation results, startup risk, and airbag status. When the unit has not yet entered the valve-closing permission stage, the startup risk exceeds the set threshold within the future prediction window, or the residual air status does not meet the valve-closing permission conditions, the controller keeps the air valve fully or partially open, thereby avoiding premature closure of the air valve when the residual air volume is large or the startup risk is high, and reducing the risk of sudden increases in peak head, peak torque, and peak axial force caused by fixed valve closure.

[0061] 3. This invention uses a smooth air valve opening curve to achieve continuous air valve closure, and coordinates the air valve opening curve with the speed transition curve to match the air valve closing process with the siphon stabilization process, thereby reducing the sudden changes in flow and pressure fluctuations caused by the step closure of the air valve, reducing the risk of sudden increases in pressure pulsation amplitude and vibration amplitude, and improving the safety and stability of the siphon axial flow pump startup process.

[0062] 4. The startup shock of this invention mainly refers to the phenomenon of a sudden increase in head, torque, axial force, pressure pulsation amplitude, or vibration amplitude in a short period of time during the startup process of a siphon-type axial flow pump. This is caused by residual air in the flow channel not being fully discharged, a mismatch between the closing time of the air valve and the siphon formation process, or a lack of coordination between the speed transition and the gas-liquid two-phase flow state. For startup methods using a fixed valve closing time, even if the unit is still in a hydraulic air-carrying or phase transition state, the air valve may still close at the preset time, causing abrupt changes in the gas-liquid two-phase flow boundary. This invention determines whether valve closure is permissible by confirming the phase within a future prediction window, judging startup risk, and judging the air bladder status. After meeting the permissible conditions, it gradually closes the air valve using a smooth air valve opening curve to reduce the risk of the aforementioned startup shock. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the identification and differentiated control during the start-up phase, showing the hydraulic air-driving stage, the hydraulic air-carrying stage, and the siphon stabilization stage, as well as their corresponding control strategies.

[0064] Figure 2This is a schematic diagram of the speed setting curve and the air valve opening curve, showing the relationship between the traditional fixed valve closing time, the flexible valve closing start time determined by the prediction results of this invention, the smooth opening curve of the air valve, and the smooth speed transition curve. It is used to illustrate that this invention reduces the risk of starting shock by avoiding high-risk valve closing periods and smoothing the air valve closing process.

[0065] Figure 3 This is a block diagram of the control system, showing the signal relationships between the sensor group, data acquisition module, feature construction module, machine learning prediction module, start-up risk assessment module, flexible start-up optimization module, speed control module, air valve control module, frequency converter or speed control device, air valve actuator, and safety protection module. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Example 1: As Figure 1 and Figure 2 As shown, a flexible start-up control method for a siphon-type axial flow pump based on machine learning prediction includes the following steps:

[0068] S1: After the axial flow pump starts, the monitoring data at each sampling moment within the sampling control cycle is sampled, and the start-up state feature vector is constructed.

[0069] For the e-th control cycle, taking the k-th time as an example, the corresponding start-up state feature vector is:

[0070]

[0071] In the formula, This represents the feature vector of the starting state at time k; This represents the water level in the inlet pool at time k. This represents the water level in the outlet pool at time k. This represents the flow rate at time k. This represents the impeller speed at time k; This represents the motor current at time k; This represents the motor voltage at time k; This represents the pressure in the inlet channel at time k; This represents the pressure in the outlet channel at time k; This represents the air valve opening at time k; This represents the vibration amplitude at time k. This represents the amplitude of the pressure pulsation at time k; This represents the rate of increase in rotational speed at time k; This represents the flow rate growth rate at time k. This represents the rate of change of head at time k.

[0072] S2: Calculate the head and torque at each moment within the control cycle.

[0073] Continuing with the example at time k, the corresponding head is:

[0074]

[0075] In the formula, This represents the head at time k; This represents the pressure in the inlet channel at time k; This represents the pressure in the outlet channel at time k; Indicates fluid density; Represents gravitational acceleration; The average flow velocity at the outlet cross section is a known quantity obtained from monitoring. This represents the average flow velocity at the inlet cross-section, which is a known quantity obtained from monitoring. Indicates the elevation of the pressure measuring point on the water outlet side; This indicates the elevation of the pressure measurement point on the inlet side.

[0076] The corresponding torque is:

[0077]

[0078] In the formula, This represents the torque at time k; Indicates the efficiency of the motor and transmission system; This represents the motor voltage at time k; This represents the motor current at time k; The power factor angle at time k is a known quantity obtained from monitoring. Let represent the impeller angular velocity at time k, which is a known quantity obtained from monitoring.

[0079] S3: Construct samples from the start-up state feature vectors, head, and torque at all sampling moments within the control cycle; input the samples into a pre-trained machine learning prediction model, and the model outputs the prediction results.

[0080] The machine learning prediction model employs long short-term memory networks, gated recurrent unit networks, temporal convolutional networks, random forests, gradient boosting trees, support vector machines, or combinations thereof. Among these, head, torque, axial force, pressure pulsation amplitude, vibration amplitude, and residual air volume are predicted using regression methods, while the startup phase is predicted using classification methods.

[0081] For each control cycle, the start-up state feature vector, head, and torque of M consecutive sampling times are selected to form the input sequence of the machine learning prediction model. The model outputs the head, torque, axial force, pressure pulsation amplitude, vibration amplitude, residual air volume, and start-up stage of N consecutive prediction sampling times after the input sequence.

[0082] During the model training phase, the output labels are the measured data or labeled data corresponding to the next N sampling times in the historical startup process; during the online prediction phase, the data of the most recent M consecutive sampling times are input into the pre-trained machine learning prediction model, and the model outputs the prediction results for the next N consecutive sampling times.

[0083] The input sequence of the machine learning prediction model is set to have a length of M sampling times, with the input sampling times being 1, 2, ..., M; the model outputs the prediction results for N consecutive prediction sampling times, with the corresponding prediction sampling times being M+1, M+2, ..., M+N.

[0084] Set the j-th prediction ( The prediction result is (and is an integer). ,but:

[0085]

[0086] In the formula, This represents the prediction result at the j-th prediction sampling time within the prediction window; This represents the predicted head at the j-th sampling time. This represents the torque predicted at the j-th sampling time. This represents the axial force predicted at the j-th sampling time. This represents the predicted pressure pulsation amplitude at the j-th sampling time. This represents the predicted vibration amplitude at the j-th sampling time. This represents the predicted value of the residual air volume at the j-th prediction sampling time within the prediction window. The residual air volume is used to characterize the state of the air bladder. This indicates the start-up phase at the j-th sampling time of the prediction. The start-up phase includes: the hydraulic air-driving phase, the hydraulic air-carrying phase, or the siphon stabilization phase.

[0087] S4: Calculate the activation risk for each prediction sampling time based on the prediction results at each prediction sampling time.

[0088]

[0089] In the formula, This represents the predicted startup risk at the j-th sampling time. Indicates the weighting coefficient. ; These represent the torque limit, axial force limit, head limit, pressure pulsation limit, vibration amplitude limit, and residual air volume limit, respectively.

[0090] S5: Based on the predicted start-up phase in S3 and combined with the start-up risks in S4, adopt a differentiated control strategy.

[0091] The prediction initiation phase refers to the phase results output by the machine learning prediction model for each prediction sampling time within the future prediction window, rather than making a single-phase judgment for the entire continuous sampling interval. Let the initiation phase for the j-th prediction sampling time be:

[0092]

[0093] In the formula, This indicates the hydrodynamic air-driving stage. This indicates the stage of water-driven air transport. This indicates the stable formation stage of the siphon.

[0094] For the e-th control cycle, if its future consecutive L s If all predicted sampling times belong to the same startup phase, then that startup phase is taken as the phase confirmation result for the current control cycle:

[0095]

[0096] In the formula, This indicates the stage confirmation result for the e-th control cycle; Indicates any of the startup phases G1, G2, or G3; L s This represents the number of consecutive prediction sampling points required for phase stability confirmation, and 1 <L s ≤N.

[0097] If the future continues L s If the phase results at each predicted sampling time are inconsistent, the current control cycle is considered to be in a phase transition state. The controller does not execute the air valve closing action, the air valve remains fully open or partially open, and the judgment is re-evaluated based on the predicted results of the next control cycle.

[0098] Specifically, this includes the following situations:

[0099] I. The stage confirmation result of the e-th control cycle At this time, the unit is considered to be in the hydraulic air-driving stage. At this time, the air in the flow channel is still in the initial process of being driven out by the water flow. The air valve is kept fully open to allow the air in the flow channel to be continuously discharged. The speed gradually increases according to the unit's existing flexible start-up curve, limiting the speed growth rate to avoid the flow rate and head increasing too quickly in the initial start-up stage.

[0100] II. When the stage confirmation result of the e-th control cycle At this time, the unit is considered to be in the hydraulic air-carrying stage. The air valve remains fully or partially open to allow air to continue to be discharged from the flow channel.

[0101] When the valve is closed, the permission flag C is enabled. b When (e)=1, the controller determines the moment when the valve closing condition is met as the start time of flexible valve closing. And gradually close the air valve according to the set air valve opening curve;

[0102] When C b When (e)=0, the air valve remains fully open or partially open, and is re-evaluated based on the prediction results of the next control cycle.

[0103] The opening and closing strategy of the air valve is as follows:

[0104]

[0105] In the formula, This indicates the stage confirmation result for the e-th control cycle; This indicates the set of permitted valve-closing stages, including the hydraulic air-carrying stage and the siphon stabilization formation stage; This indicates the air valve closing permission flag for the e-th control cycle; when When, it indicates that the controller allows the air valve to enter the closing process; when When this time, it indicates that the air valve remains fully open or partially open; This indicates the risk of initiation at the j-th prediction sampling time. This indicates the set threshold for the risk of valve closure. This represents the normalized airbag state at the j-th prediction sampling time. This indicates the set threshold for the airbag to be in a closed state.

[0106] The normalized airbag state is as follows:

[0107]

[0108] In the formula, This represents the predicted residual air volume at the j-th prediction sampling time. This indicates the air volume at the initial startup time or under the set reference state.

[0109] when The controller will determine the moment when the valve closing condition is met as the start time of flexible valve closing. The air valve is controlled to gradually close according to the following air valve opening curve:

[0110]

[0111] in:

[0112]

[0113] In the formula, a(t) represents the relative opening of the air valve, a(t)=1 indicates that the air valve is fully open, and a(t)=0 indicates that the air valve is completely closed; Indicates the start time of flexible valve closure determined by the prediction results; T c ξ represents the duration from fully open to fully closed of the air valve; ξ represents the normalized valve closing time.

[0114] when At this time, the controller does not perform the valve closing action, the air valve remains fully open or partially open, and the prediction results continue to be updated based on subsequent sampling data.

[0115] The above air valve opening curve is in and With a zero rate of change, it can avoid sudden changes in flow and pressure fluctuations caused by abrupt changes in the opening of the air valve.

[0116] III. The stage confirmation result of the e-th control cycle At this point, the unit is considered to be in the siphon stabilization formation stage. The controller compares the airbag status and startup risk at each predicted sampling time within the future prediction window. If the maximum airbag status within the future prediction window is not higher than the set airbag status threshold, and the maximum startup risk within the future prediction window is not higher than the set startup risk threshold, then it is considered that the residual air in the flow channel has been reduced to the allowable range, and there is no risk of over-limit startup impact during the subsequent transition process. At this time, the controller controls the air valve to gradually close according to the set air valve opening curve, and determines the current time as the speed transition start time t1, so that the speed smoothly transitions to the target speed according to the speed transition curve.

[0117] The comparison conditions are as follows:

[0118]

[0119] and:

[0120]

[0121] In the formula, This represents the normalized airbag state at the j-th prediction sampling time. This indicates the set threshold for the airbag status; R represents the activation risk at j prediction sampling times; th This represents the set activation risk threshold; N represents the number of prediction sampling points for the future prediction window.

[0122] If any of the above conditions are not met, the controller will not enter the speed transition phase, the air valve will remain fully open or partially open, the speed will continue to operate according to the current flexible start-up pattern, and a new judgment will be made based on the new prediction results in the next control cycle.

[0123] The speed transition curve is as follows:

[0124]

[0125] In the formula, n(t) represents the rotational speed at time t; t1 represents the starting time of the rotational speed transition determined after satisfying the airbag status and start-up risk comparison conditions; t2 represents the time when the rotational speed reaches the target speed; n(t1) represents the rotational speed at the starting time of the rotational speed transition; n(t2) represents the target speed; γ represents the exponent of the drive curve determined by the controller's flexible start-up optimization module for the time period [t1, t2].

[0126] Calculation example:

[0127] like Figure 2 As shown, traditional control methods typically follow a preset fixed valve closing time t. c Perform the air valve closing action. Since the fixed valve closing time does not adjust with changes in the residual air state in the flow channel, the startup phase, or startup risks, if at t... c At any given time, the unit is still in the hydraulic air-carrying stage or the transition stage, and the residual air in the flow channel has not been fully discharged. Closing the air valve will change the gas-liquid two-phase flow boundary in the flow channel, causing local flow resistance and pressure distribution to change in a short period of time, which may lead to a sudden increase in head, torque, axial force, pressure pulsation amplitude or vibration amplitude.

[0128] This invention does not directly close the air valve at a fixed time, but rather confirms the result G within a future prediction window. c (e) Risk of initiation j and normalized airbag state Determine whether valve closure is permitted. If the current control cycle does not belong to the set Ω of permitted valve closure phases. c Or, if there is an excessive risk of activation or an excessive airbag status within the future prediction window, the valve closure permission flag C will be activated. b (e)=0, the controller keeps the air valve fully or partially open and continues to re-determine based on the prediction results of the next control cycle.

[0129] When the phase confirmation results, activation risks, and airbag status all meet the conditions for permissible valve closure, the valve closure permission flag C is activated.b (e)=1, the controller determines the moment when the condition is met as the start time of flexible valve closure. The air valve is gradually closed according to the air valve opening curve a(t). Since the air valve opening curve changes smoothly at both the start and end of valve closure, it avoids sudden changes in flow and pressure fluctuations caused by abrupt changes in the air valve opening. At the same time, the speed smoothly transitions to the target speed according to the speed transition curve, so that the air valve closing process is coordinated with the siphon stabilization process, thereby reducing the risk of start-up shock.

[0130] To further illustrate the process by which this invention suppresses the risk of startup shock, the following implementation example is provided:

[0131] In one implementation example, the traditional control method uses a fixed air valve closing time t. c When the control time reaches t c In traditional control methods, the startup stage, startup risk, and airbag status within the future prediction window are not considered. Instead, the air valve is directly closed according to the preset startup procedure. If the residual air in the flow channel has not been fully discharged at this time, or if the unit is still in the hydraulic air-carrying stage or the transition stage, the closure of the air valve will cause a sudden change in the gas-liquid two-phase flow boundary in the flow channel, resulting in a short-term increase in head, torque, axial force, pressure pulsation amplitude, or vibration amplitude.

[0132] When using the method of the present invention, at a fixed valve closing time t c The corresponding control period k c Internally, the controller first calculates the valve closing permission flag based on the future prediction window results output by the machine learning prediction model. This is done if any of the following conditions are met:

[0133]

[0134] or:

[0135]

[0136] or:

[0137]

[0138] Then we have:

[0139]

[0140] At this time, the controller does not close the air valve, but keeps it fully or partially open, allowing air to continue to be discharged from the flow path, and then reassesses based on the prediction results of the next control cycle. This avoids closing the air valve before the unit has entered the permissible valve-closing stage, when the predicted start-up risk is high, or when the residual air condition does not meet the permissible valve-closing conditions.

[0141] As the startup process continues, when a certain subsequent control cycle k * satisfy:

[0142]

[0143] and:

[0144]

[0145] and:

[0146]

[0147] Then we have:

[0148]

[0149] At this time, the controller will control cycle k * The corresponding time is determined as the start time of flexible valve closure. The air valve is controlled to gradually close according to the air valve opening curve a(t), while the speed is smoothly transitioned to the target speed according to the speed transition curve.

[0150] As can be seen from the above control process, the traditional fixed-time valve closing method may perform valve closing action when the residual air state and start-up risk are unknown. However, this invention comprehensively judges the start-up stage, start-up risk, and air bladder state within the future prediction window before closing the valve. When any condition is not met, the air valve does not close; only when all conditions are met does the air valve enter the smooth closing process. Therefore, this invention can avoid the air valve from suddenly closing when the residual air volume is large or the start-up risk exceeds the limit, reducing the risk of sudden increases in head peak, torque peak, axial force peak, pressure pulsation amplitude, and vibration amplitude caused by abrupt changes in the gas-liquid two-phase flow boundary.

[0151] Example 2: As Figure 3 As shown, a flexible start-up control system for a siphon-type axial flow pump based on machine learning prediction includes the following modules:

[0152] Data acquisition module: Used to receive multi-source data of axial flow pump output from sensor group, and to complete data buffering, sampling frequency unification and time synchronization, so as to provide continuous time series data for subsequent prediction;

[0153] Feature construction module: used to calculate startup state features based on the raw data output by the data acquisition module; startup state features include at least one of the following: head, speed growth rate, flow rate growth rate, head change rate, pressure pulsation amplitude, vibration root mean square value, air valve opening change rate, and torque estimate.

[0154] Machine learning prediction module: used to predict the startup impact parameters and startup stage within a preset time window based on the startup state characteristics at several consecutive sampling times. The startup impact parameters include at least one of the following: peak head, peak torque, peak axial force, pressure pulsation amplitude, vibration amplitude, and airbag state; the startup stage includes: hydraulic air-driving stage, hydraulic air-carrying stage, or siphon stabilization formation stage.

[0155] Start-up Risk Assessment Module: This module compares the start-up impact parameters output by the machine learning prediction module with preset limits and synthesizes the start-up risk assessment results.

[0156] Flexible start optimization module: used to generate speed setpoint curve and air valve opening curve based on start risk assessment results;

[0157] Speed ​​control module: Used to send the speed command curve to the frequency converter or speed control device, so that the impeller speed changes according to the flexible start curve;

[0158] Air valve control module: Used to send the air valve opening curve to the air valve actuator to realize the air valve opening, holding, partial closing, delayed closing or continuous closing;

[0159] Safety protection module: Used to perform speed reduction, keep the air valve open, delay valve closure, alarm or shutdown protection when the predicted or measured value exceeds the safety threshold, sensor data is abnormal, air valve operation is abnormal or the machine learning prediction module output is invalid.

[0160] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A flexible start-up control method for a siphon-type axial flow pump based on machine learning prediction, characterized in that, Includes the following steps: S1: After the axial flow pump starts, the monitoring data at each sampling moment within the sampling control cycle is sampled, and the start-up state feature vector is constructed; S2: Based on the monitoring data, calculate the head and torque at each sampling moment within the current control cycle; S3: Construct samples from the start-up state feature vectors, head, and torque at each sampling moment within the current control cycle; input the samples into a pre-trained machine learning prediction model, and have the model output prediction results; the prediction results include: head, torque, axial force, pressure pulsation amplitude, vibration amplitude, residual air volume, and start-up stage at multiple consecutive sampling moments in the future; S4: Calculate the activation risk at each prediction sampling time based on the prediction results at each prediction sampling time; S5: Based on the prediction results in S3 and combined with the start-up risk in S4, adopt a differentiated control strategy.

2. The flexible start-up control method for a siphon-type axial flow pump based on machine learning prediction according to claim 1, characterized in that, In steps S1 and S2, for the e-th control cycle, the start-up state feature vector at time k is: In the formula, This represents the feature vector of the starting state at time k; This represents the water level in the inlet pool at time k. This represents the water level in the outlet pool at time k. This represents the flow rate at time k. This represents the impeller speed at time k; This represents the motor current at time k; This represents the motor voltage at time k; This represents the pressure in the inlet channel at time k; This represents the pressure in the outlet channel at time k; This represents the air valve opening at time k; This represents the vibration amplitude at time k. This represents the amplitude of the pressure pulsation at time k; This represents the rate of increase in rotational speed at time k; This represents the flow rate growth rate at time k. This represents the rate of change of head at time k; The corresponding head is: In the formula, This represents the head at time k; This represents the pressure in the inlet channel at time k; This represents the pressure in the outlet channel at time k; Indicates fluid density; Represents gravitational acceleration; This indicates the average flow velocity at the outlet cross-section; This indicates the average flow velocity at the inlet cross-section; Indicates the elevation of the pressure measuring point on the water outlet side; Indicates the elevation of the pressure measuring point on the inlet side; The corresponding torque is: In the formula, This represents the torque at time k; Indicates the efficiency of the motor and transmission system; This represents the motor voltage at time k; This represents the motor current at time k; The angle represents the power factor at time k. This represents the impeller angular velocity at time k.

3. The flexible start-up control method for a siphon-type axial flow pump based on machine learning prediction according to claim 2, characterized in that, In step S3, the number of samples input to the machine learning prediction model is set to M, and the sampling times are 1, 2, ..., M; the machine learning prediction model outputs prediction results for N consecutive sampling times, with prediction times of M+1, M+2, ..., M+N; then the j-th prediction result is: In the formula, This represents the prediction result at the j-th prediction sampling time within the prediction window. And it is an integer; This represents the predicted head at the j-th sampling time. This represents the torque predicted at the j-th sampling time. This represents the axial force predicted at the j-th sampling time. This represents the predicted pressure pulsation amplitude at the j-th sampling time. This represents the predicted vibration amplitude at the j-th sampling time. The residual air volume is the predicted value of the residual air volume at the j-th prediction sampling time within the prediction window, and the residual air volume is used to characterize the state of the airbag. This indicates the start-up phase at the j-th sampling time of the prediction. The start-up phase includes: the hydraulic air-driving phase, the hydraulic air-carrying phase, or the siphon stabilization phase.

4. The flexible start-up control method for a siphon-type axial flow pump based on machine learning prediction according to claim 3, characterized in that, In step S4, the activation risk of the j-th prediction result is: In the formula, This represents the predicted startup risk at the j-th sampling time. Indicates the weighting coefficient. ; These represent the torque limit, axial force limit, head limit, pressure pulsation limit, vibration amplitude limit, and residual air volume limit, respectively.

5. The flexible start-up control method for a siphon-type axial flow pump based on machine learning prediction according to claim 4, characterized in that, In step S5, the strategy for differentiation control is as follows: If the future continues L s If all predicted sampling times belong to the same startup phase, then that startup phase is taken as the phase confirmation result of the current control cycle; if the future consecutive L s If the start-up phases of the predicted sampling times are inconsistent, the current control cycle will be treated as a transitional state, the air valve will remain fully open or partially open, and the judgment will be re-evaluated based on the prediction results of the next control cycle.

1. When the current control cycle stage confirmation result is the hydraulic aeration stage, the air valve remains fully open, and the speed gradually increases according to the unit's existing flexible start-up curve.

2. When the current control cycle stage confirmation result is the hydraulic air-carrying stage, the air valve remains fully open or partially open to allow air in the flow channel to continue to be discharged; when the start-up risk and airbag status within the future prediction window meet the allowable valve closing conditions, the controller gradually closes the air valve according to the set air valve opening curve; when the start-up risk or airbag status within the future prediction window does not meet the allowable valve closing conditions, the air valve remains fully open or partially open.

3. When the current control cycle stage confirmation result is the siphon stabilization stage, if the airbag state in the future prediction window is not higher than the set airbag state threshold, and the start-up risk in the future prediction window is not higher than the set start-up risk threshold, the controller controls the air valve to gradually close according to the set air valve opening curve, and smoothly transitions the speed to the target speed.

6. The flexible start-up control method for a siphon-type axial flow pump based on machine learning prediction according to claim 5, characterized in that, The opening and closing strategy of the air valve is as follows: In the formula, G c (e) represents the stage confirmation result of the e-th control cycle; Ω c This indicates the set of permitted valve-closing stages, including the hydraulic air-carrying stage and the siphon stabilization formation stage; This indicates the air valve closing permission flag for the e-th control cycle; when When, it indicates that the controller allows the air valve to enter the closing process; when When this time, it indicates that the air valve remains fully open or partially open; This indicates the risk of initiation at the j-th prediction sampling time. This indicates the set threshold for the risk of valve closure. This represents the normalized airbag state at the j-th prediction sampling time. This indicates the set threshold value for the airbag to be in a closed-valve state. The normalized airbag state is as follows: In the formula, This represents the predicted residual air volume at the j-th prediction sampling time. Indicates the air volume at the initial startup moment or under the set reference state; when The controller will determine the moment when the valve closing condition is met as the start time of flexible valve closing. The air valve is controlled to gradually close according to the following air valve opening curve: in: In the formula, a(t) represents the relative opening of the air valve, a(t)=1 indicates that the air valve is fully open, and a(t)=0 indicates that the air valve is completely closed; Indicates the start time of flexible valve closure determined by the prediction results; T c ξ represents the duration from fully open to fully closed of the air valve; ξ represents the normalized valve closing time. when At this time, the controller does not perform the valve closing action, the air valve remains fully open or partially open, and the prediction results continue to be updated based on subsequent sampling data.

7. The flexible start-up control method for a siphon-type axial flow pump based on machine learning prediction according to claim 6, characterized in that, The speed transition curve is as follows: In the formula, n(t) represents the rotational speed at time t; t1 represents the starting time of the rotational speed transition determined after satisfying the airbag status and start-up risk comparison conditions; t2 represents the time when the rotational speed reaches the target speed; n(t1) represents the rotational speed at the starting time of the rotational speed transition; n(t2) represents the target speed; γ represents the exponent of the drive curve determined by the controller's flexible start-up optimization module for the time period [t1, t2].

8. A flexible start-up control system for a siphon axial flow pump based on machine learning prediction, based on the flexible start-up control method for a siphon axial flow pump based on machine learning prediction as described in any one of claims 1-7, characterized in that, Includes the following modules: Data acquisition module: Used to receive multi-source data of axial flow pump output from sensor group, and to complete data buffering, sampling frequency unification and time synchronization, so as to provide continuous time series data for subsequent prediction; Feature construction module: used to calculate startup state features based on the raw data output by the data acquisition module; Machine learning prediction module: used to predict the startup impact parameters and startup stage within a preset time window based on the startup state characteristics at several consecutive sampling times; Start-up Risk Assessment Module: This module compares the start-up impact parameters output by the machine learning prediction module with preset limits and synthesizes the start-up risk assessment results. Flexible start optimization module: used to generate speed setpoint curve and air valve opening curve based on start risk assessment results; Speed ​​control module: Used to send the speed command curve to the frequency converter or speed control device, so that the impeller speed changes according to the flexible start curve; Air valve control module: Used to send the air valve opening curve to the air valve actuator to realize the air valve opening, holding, partial closing, delayed closing or continuous closing; Safety protection module: Used to perform speed reduction, keep the air valve open, delay valve closure, alarm or shutdown protection when the predicted or measured value exceeds the safety threshold, sensor data is abnormal, air valve operation is abnormal or the machine learning prediction module output is invalid.

Images