Heavy-duty gas generator set cooling oil circulation digital twin system

Abstract

This invention relates to the field of gas-fired power generation equipment and digital twin control technology, specifically a digital twin system for cooling oil circulation in heavy-duty gas generator sets, comprising: an edge control device and a twin server; a high-frequency data acquisition module for acquiring characteristic parameters and attaching a unified global timestamp; a low-frequency data acquisition module for acquiring fluid temperature and attaching a unified global timestamp; a high-frequency state observation module for outputting a high-frequency estimated value of the fluid temperature in the target heating zone; a dynamic hysteresis alignment module for generating phase alignment residuals; and a feedforward feedback control module for generating feedforward adjustment commands, which drive the fluid power source. This invention solves the problems of difficult direct measurement and slow feedback response.

Description

Technical Field

[0001] This invention relates to the field of gas-fired power generation equipment and digital twin control technology, specifically a digital twin system for cooling oil circulation in heavy-duty gas generator sets. Background Technology

[0002] In the operating environment where heavy-duty gas generator sets participate in peak shaving, the units often experience continuous operating conditions such as rapid load increases and decreases, causing high-frequency and rapid changes in rotor thermal load and cooling oil system conditions. Since it is difficult to directly deploy high-frequency sensors in the high-temperature core area, the system can usually only obtain fluid temperature data with physical transmission lag in the downstream pipeline.

[0003] For cooling control, existing solutions generally adopt a passive feedback regulation mechanism based on downstream temperature measurement. Although this solution has a certain temperature regulation capability in steady-state scenarios, the time-varying transmission delay of the cooling oil flowing from the heat source to the downstream causes the measurement results to lag significantly behind the actual thermal state. This control method, which is highly dependent on hysteresis data, has a slow response and is difficult to proactively respond to thermal shocks in the heat-generating area within a time window smaller than the preset time threshold. This can easily lead to missed opportunities for preset intervention and thus operational risks. Therefore, how to achieve real-time high-frequency reconstruction of the thermal state of the target heat-generating area and precise feedforward control of the cooling power source under the premise of pure time delay of physical fluids has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a digital twin system for cooling oil circulation in heavy-duty gas generator sets. Specifically, the technical solution of this invention includes:

[0005] The edge control device and the twin server are connected via communication. The edge control device has a high-frequency data acquisition module, a high-frequency status observation module, and a feedforward feedback control module. The twin server has a low-frequency data acquisition module and a dynamic hysteresis alignment module.

[0006] The high-frequency data acquisition module is used to acquire the operating frequency of the fluid power source, the transient pressure data of the pipeline, and the characteristic parameters of the target heating zone of the fluid circulation system within a preset high-frequency period that is less than the minimum control response time of the fluid power source frequency change, and attach a unified global timestamp.

[0007] The low-frequency data acquisition module is used to acquire the downstream fluid temperature of the fluid circulation system within a preset low-frequency period that is greater than the thermal inertia delay time of a downstream fluid temperature sensor located in the fluid circulation system, and to attach a unified global timestamp.

[0008] The high-frequency state observation module is used to output a high-frequency estimated value of the fluid temperature in the target heating zone through a state observer based on the operating frequency of the fluid power source, the transient pressure data of the pipeline, and the characteristic parameters of the target heating zone.

[0009] The dynamic hysteresis alignment module is used to calculate the transmission delay time of the fluid based on the transient pressure data of the pipeline and the operating frequency of the fluid power source. It continuously stores the high-frequency prediction values ​​obtained in each cycle into a dynamic hysteresis alignment queue, extracts the historical high-frequency prediction values ​​after the transmission delay time of the global timestamp in the dynamic hysteresis alignment queue, compares the historical high-frequency prediction values ​​with the downstream fluid temperature of the current cycle, and generates phase alignment residuals.

[0010] The feedforward control module is used to inject the phase-aligned residual back into the high-frequency state observation module as an update weight, and generate feedforward adjustment commands based on the high-frequency prediction value, thereby driving the fluid power source through the feedforward adjustment commands.

[0011] Preferably, the fluid circulation system is a heavy-duty gas generator set cooling oil circulation system, the characteristic parameter of the target heating zone is the rotor speed, the fluid power source is the cooling oil pump, and the downstream fluid temperature is the return oil pipeline temperature.

[0012] Preferably, the process of the high-frequency data acquisition module acquiring pipeline transient pressure data includes: acquiring unfiltered pipeline transient pressure data as raw pipeline transient pressure data; filtering the raw pipeline transient pressure data using a Kalman filter to remove high-frequency mechanical vibration noise from the fluid power source, extracting the low-frequency envelope reflecting changes in flow resistance, and marking the low-frequency envelope as pipeline transient pressure data.

[0013] Preferably, the process of the high-frequency state observation module outputting the high-frequency predicted value of the fluid temperature in the target heating zone includes: constructing an extended Kalman filter model as a state observer;

[0014] The operating frequency of the fluid power source, the transient pressure data of the pipeline, and the characteristic parameters of the target heating zone are input into the extended Kalman filter model. The extended Kalman filter model is used to perform state prediction and covariance update, and the current covariance matrix is ​​generated during the update process. The high-frequency prediction value of the fluid temperature in the target heating zone is output.

[0015] Preferably, the process by which the dynamic hysteresis alignment module calculates the fluid transport delay time includes: obtaining the known pipe length and pipe cross-sectional area of ​​the fluid circulation system;

[0016] The transient flow velocity of the fluid is derived from the transient pressure data of the pipeline, the operating frequency of the fluid power source, and the cross-sectional area of ​​the pipeline.

[0017] Using the global timestamp corresponding to the current low-frequency cycle as a reference, we trace back along the historical time axis to obtain the displacement integral result by performing time integration on the transient flow velocity. We then compare the displacement integral result with the known pipe length to determine whether the displacement integral result reaches the known pipe length.

[0018] If the displacement integral result reaches the known pipe length, the corresponding integration time period is marked as the transmission delay time.

[0019] If the displacement integral result does not reach the known pipe length, continue time integration until the displacement integral result reaches the known pipe length.

[0020] Preferably, the process of extracting historical high-frequency estimates after the transmission delay time of the global timestamp in the dynamic hysteresis alignment queue, and comparing the historical high-frequency estimates with the downstream fluid temperature of the current cycle to generate phase alignment residuals includes: the dynamic hysteresis alignment queue is a first-in-first-out data structure;

[0021] Within each high-frequency cycle, the high-frequency estimate with a global timestamp is pushed into a dynamic hysteresis alignment queue;

[0022] Within the current low-frequency cycle, calculate the corresponding historical timestamp node based on the current global timestamp and transmission delay time, and extract the historical high-frequency estimate corresponding to the historical timestamp node from the dynamic hysteresis alignment queue.

[0023] Calculate the difference between the downstream fluid temperature of the current cycle and the historical high-frequency estimate, and label the difference as the phase-aligned residual.

[0024] Preferably, the process by which the feedforward control module injects the phase alignment residual into the high-frequency state observation module as an update weight includes: obtaining the current covariance matrix of the high-frequency state observation module; inputting the phase alignment residual as an observation update condition into the high-frequency state observation module; iteratively correcting the current covariance matrix based on the phase alignment residual to obtain the updated covariance matrix; and using the updated covariance matrix as the prediction weight for the next high-frequency cycle.

[0025] Preferably, the process by which the feedforward control module generates feedforward adjustment commands based on high-frequency prediction values ​​includes: setting a target safety threshold for the fluid temperature in the target heating zone; and comparing the high-frequency prediction value with the target safety threshold.

[0026] If the high-frequency forecast is greater than or equal to the target safety threshold, a feedforward adjustment command to increase the operating frequency of the fluid power source is generated; if the high-frequency forecast is less than the target safety threshold, a feedforward adjustment command to maintain or decrease the operating frequency of the fluid power source is generated.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. This invention collects high-frequency operating characteristics from the edge side and constructs a state observer to output a high-frequency predicted value of the fluid temperature in the heating zone in real time. This overcomes the limitation of traditional temperature measurement being restricted by the lag in downstream physical transmission, and solves the problems of difficulty in direct measurement and slow feedback response without increasing the number of sensors in the high-temperature core area.

[0029] 2. This invention dynamically calculates the fluid transport delay time and extracts the corresponding historical high-frequency estimate from the queue and compares it with the current downstream low-frequency temperature; this phase alignment mechanism based on real physical hysteresis effectively eliminates the temporal misalignment caused by flow velocity changes and generates a phase alignment residual with real physical meaning.

[0030] 3. This invention uses the phase alignment residual as an update weight and injects it back into the high-frequency state observer to iteratively correct the covariance matrix; this enables the system to dynamically adjust the confidence distribution between prediction and observation, giving the high-frequency thermal state reconstruction process self-correction capability and continuously adapting to operating condition drift, oil aging and equipment characteristic changes.

[0031] 4. Based on the comparison between the generated high-frequency prediction value and the target safety threshold, this invention directly generates feedforward adjustment commands to drive the cooling oil pump; this mechanism transforms thermal state perception into feedforward control capability, and implements cooling intervention in advance before the thermal shock is transmitted to the downstream, thus changing the lag of traditional passive feedback control.

[0032] 5. This invention utilizes a Kalman filter to process unfiltered pipeline transient pressure data, accurately filtering out high-frequency mechanical vibration noise from oil pumps and other sources. This data purification mechanism extracts a low-frequency envelope that truly reflects the slow changes in pipeline flow resistance, providing a highly reliable data foundation for subsequent transmission delay estimation and temperature reconstruction. Attached Figure Description

[0033] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0034] Figure 1 This is a schematic diagram of the module of the digital twin system for cooling oil circulation of heavy-duty gas generator sets provided in the embodiments of this application. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0036] A digital twin system for cooling oil circulation in heavy-duty gas generator sets is applied to fluid circulation systems. It includes: an edge control device and a twin server. The edge control device is connected to a high-frequency data acquisition module, a high-frequency status observation module, and a feedforward feedback control module. The twin server is connected to a low-frequency data acquisition module and a dynamic hysteresis alignment module.

[0037] The high-frequency data acquisition module is used to acquire the operating frequency of the fluid power source, the transient pressure data of the pipeline, and the characteristic parameters of the target heating zone of the fluid circulation system within a preset high-frequency period that is less than the minimum control response time of the fluid power source frequency change, and attach a unified global timestamp.

[0038] The low-frequency data acquisition module is used to acquire the downstream fluid temperature of the fluid circulation system within a preset low-frequency period that is greater than the thermal inertia delay time of a downstream fluid temperature sensor located in the fluid circulation system, and to attach a unified global timestamp.

[0039] The high-frequency state observation module is used to output a high-frequency estimated value of the fluid temperature in the target heating zone through a state observer based on the operating frequency of the fluid power source, the transient pressure data of the pipeline, and the characteristic parameters of the target heating zone.

[0040] The dynamic hysteresis alignment module is used to calculate the transmission delay time of the fluid based on the transient pressure data of the pipeline and the operating frequency of the fluid power source. It continuously stores the high-frequency prediction values ​​obtained in each cycle into a dynamic hysteresis alignment queue, extracts the historical high-frequency prediction values ​​after the transmission delay time of the global timestamp in the dynamic hysteresis alignment queue, compares the historical high-frequency prediction values ​​with the downstream fluid temperature of the current cycle, and generates phase alignment residuals.

[0041] The feedforward control module is used to inject the phase-aligned residual back into the high-frequency state observation module as an update weight, and generate feedforward adjustment commands based on the high-frequency prediction value, thereby driving the fluid power source through the feedforward adjustment commands.

[0042] This embodiment provides a digital twin control mechanism for cooling oil circulation in heavy-duty gas generator sets, such as... Figure 1 As shown; specifically, this mechanism is applied to the cooling oil circulation loop of a heavy-duty gas turbine generator set in a combined cycle power plant; when the unit participates in grid peak shaving, it usually experiences continuous operating conditions such as steady-state operation, rapid load increase, sudden load shedding, and re-stabilization; since there is a physical transmission time when the cooling oil flows from the high-temperature heating part to the downstream temperature measuring point, and the grid load change will cause changes in rotor heat load and oil pump operating conditions in a shorter time, the system adopts a collaborative working mode of edge control equipment and twin server. Among them, the edge control equipment undertakes high-frequency periodic data acquisition, status observation and feedforward control, while the twin server undertakes low-frequency periodic temperature acquisition and hysteresis alignment correction;

[0043] Specifically, the preset high-frequency period is determined based on the minimum control response time of the fluid power source frequency change, and the preset low-frequency period is determined based on the thermal inertia delay time of the downstream fluid temperature sensor, so that the sampling frequency matches the physical dynamic characteristics of the control execution side and the temperature measurement side respectively.

[0044] The details are as follows: The edge control device continuously receives three types of signals at a high frequency cycle. One type is the operating frequency of the fluid power source, which reflects the pumping capacity; another type is the transient pressure data of the pipeline, which reflects the flow resistance, local pressure drop, and fluid transport status; and the third type is the characteristic parameters of the target heat-generating area, which characterizes the strength of the heat source and its changing trend. The twin server collects the downstream fluid temperature at a low frequency cycle. Although this temperature lags behind the actual state of the high-temperature area in time, its measurement stability meets the preset standards, the installation environment meets the preset safety requirements, and it has long-term maintainability.

[0045] To enable comparison of fast and slow loop data on the same timeline, all collected data are appended with a unified global timestamp. The edge control device and the twin server are periodically synchronized via a precise time protocol or a network time protocol to ensure that the synchronization error of the global timestamp is controlled within a preset tolerance range.

[0046] Based on this, the state observation module in the edge control device continuously outputs a high-frequency predicted value of the fluid temperature in the target heating zone according to the pump operating frequency, pipeline transient pressure changes and heating zone operating condition changes. The high-frequency predicted value here is not the directly measured temperature, but the real-time reconstruction result of the oil thermal state in the high-temperature zone. Its physical meaning is to infer the temperature change of the part where a high-frequency sensor cannot be directly installed by using the available external process quantities.

[0047] The dynamic hysteresis alignment module in the twin server does not directly use the currently measured downstream temperature to correct the current prediction value. Instead, it first estimates the transmission delay time required for the cooling oil to travel from the heating zone to the downstream measuring point based on the current flow state, and then searches for the historical high-frequency prediction value corresponding to the delay in the dynamic hysteresis alignment queue. Only when the two belong to the same batch of oil in time are they compared and a phase alignment residual is generated.

[0048] To facilitate understanding of the data flow, a simplified time-series example can be used for illustration: Assume that a high-frequency prediction record is formed every 10 milliseconds on the edge side, denoted as T1, T2, T3, etc., and the server obtains a return oil temperature record every 1 second, denoted as R1, R2, R3, etc. If the transmission delay corresponding to the current flow state is approximately 4 seconds, then when the server receives R5, it does not compare it with the latest prediction value T500, but instead retrieves the set of historical prediction values ​​formed approximately 4 seconds ago from the queue and selects the record that matches the time phase of R5 for comparison.

[0049] The residual obtained in this way truly reflects the state difference of the same batch of oil at the upstream and downstream positions, rather than the misalignment difference between two objects at different times. The feedforward feedback control module sends the residual back to the state observation module to adjust the allocation of model and observation confidence in the next round of prediction, and at the same time directly generates pump adjustment commands based on the latest high-frequency forecast.

[0050] Under abnormal operating conditions, when there is short-term packet loss of high-frequency data, low-frequency temperature data fails to arrive on time, global clock drift exceeds the set tolerance, or the current delay time exceeds the traceability range of the queue, the system will not perform misalignment correction, but will temporarily maintain the effective update weight of the previous round, and the edge side will continue to perform prediction and control according to the most recent reliable state; when communication is restored and the time base is recalibrated, phase alignment correction will be restored; if the downstream temperature sensor has distortions that exceed the preset tolerance range, such as sudden spikes or long-term constant values, the server can suspend residual back-injection to avoid erroneous measurements contaminating the high-frequency state observation process.

[0051] For example, during the evening peak shaving period of this combined cycle power plant, the unit load rapidly increases from 70% to 95%, the rotor heat load rises rapidly, and the oil pump frequency increases synchronously. However, the actual temperature of the return oil pipeline does not rise immediately because the fluid is still en route. At this time, the edge side provides an estimate of the oil temperature rise in the heating zone based on the high-frequency signal and increases the oil pump speed in advance. A few seconds later, the return oil temperature measurement point collects the temperature rise result, and the server uses this lagging but reliable data to correct the prediction result from a few seconds ago, making subsequent predictions closer to the actual thermal state. If the grid experiences load shedding, the system also uses the new delay time to reselect the alignment object, avoiding the misuse of the cold oil temperature after load shedding to correct the thermal prediction before load shedding.

[0052] To further clarify, to maintain consistency in terminology, in this embodiment and its dependent embodiments, whenever "edge side" appears, it refers to the edge control device; whenever "server" or "server side" appears, it refers to the twin server; whenever "pump frequency" or "oil pump frequency" appears, it refers to the operating frequency of the fluid power source; and whenever "return oil temperature" appears, it refers to the specific measured quantity of the downstream fluid temperature in this system. The aforementioned abbreviations are only used for simplification and do not indicate the addition of new equipment, modules, or technical objects different from those defined in the embodiments.

[0053] Correspondingly, the state observation module and the state observer used therein are related as a module and its internal algorithm carrier. When the feedforward feedback control module injects the phase alignment residual back into the high-frequency state observation module, it performs the update through the state observer inside the module, rather than setting up a separate second observation unit in the system.

[0054] The purpose of this step is to transform the physically lagging downstream measurements into effective observations that can be used to correct high-frequency predictions, and to enable the edge side to continuously reconstruct and feedforward adjust the fluid temperature in the heating zone without increasing the number of sensors in the high-temperature zone, thereby achieving an earlier and more stable control response for the cooling oil pump.

[0055] Furthermore, the fluid circulation system is a cooling oil circulation system for a heavy-duty gas generator set, the characteristic parameter of the target heating zone is the rotor speed, the fluid power source is the cooling oil pump, and the downstream fluid temperature is the return oil pipeline temperature.

[0056] This embodiment provides a parameter mapping mechanism for the cooling oil circulation loop of a heavy-duty gas generator set. Specifically, in the aforementioned peak-shaving operation scenario, the system specifically defines the fluid circulation system as the cooling oil circulation system of the gas generator set, specifically selects the rotor speed as the characteristic parameter of the heating zone, specifically defines the fluid power source as the cooling oil pump, and specifically defines the downstream fluid temperature as the return oil pipeline temperature.

[0057] The reasons for choosing rotor speed as a characteristic parameter of the heating zone are as follows: In heavy-duty gas generator sets, rotor speed is directly related to the shear state of the friction pair, the heating state of the bearing oil film, and the rhythm of load change of the unit. When the speed changes or is maintained at a high level under high load, the heat generation in the relevant parts usually appears before the return oil temperature. Therefore, rotor speed is suitable as an indicator of the heat intensity in high-frequency state observation.

[0058] The cooling oil pump is used as the fluid power source because the pump frequency directly determines the circulating oil volume and heat carrying capacity, which is the direct object of control execution. The return oil pipeline temperature is used as the downstream fluid temperature because the return oil location is far away from the high-temperature operating environment that exceeds the sensor's rated operating temperature. The sensor's operating environment is more stable and easier to maintain. At the same time, its temperature change still retains the result information after the upstream heat source state is propagated downstream.

[0059] If only arbitrary process parameters are used to characterize the heating zone, the coupling relationship between the parameters and the heat source may be lower than the preset correlation threshold, resulting in high-frequency predictions being insensitive to the actual thermal state. Therefore, this embodiment limits the parameter to the rotor speed to establish a more stable industrial correspondence with changes in the mechanical heat source. Similarly, if the downstream fluid temperature is set at a location beyond the effective monitoring distance, heat dissipation, oil mixing, and environmental disturbances along the way will weaken the measurement's indicative role in the upstream thermal state. If the distance between the measuring points is less than the preset safety threshold, the installation and maintenance conditions will deteriorate, and there may even be a problem of insufficient reliability. The return oil pipeline temperature takes into account both availability and representativeness in engineering.

[0060] As an anomaly handling mechanism, when the unit enters special operating conditions such as turning gear, start-up, shutdown, or extremely low load, the rotor speed can still be collected, but its sensitivity to the heat intensity may be different from that of the power generation operating condition. At this time, the system can switch to the state observation parameter set under the corresponding operating condition, or reduce the weight of the speed parameter in the observation. If the return oil temperature sensor maintenance is interrupted, the edge side prediction and pump frequency basic protection control are retained, and the hysteresis correction link is re-established after the low frequency temperature measurement is restored.

[0061] For example, during the rapid load-bearing phase of the evening peak of the same peak-shaving unit, the unit operates stably near the rated speed and undertakes a higher electrical load. The heat load of the rotor-related heat-generating parts increases, the frequency of the cooling oil pump increases, and the temperature of the return oil pipeline gradually rises after a few seconds. The system processes these quantities as representative quantities of heat source intensity, driving capability, and downstream response, respectively, thereby forming a digital twin control link suitable for the unit structure and maintenance conditions.

[0062] The purpose of this step is to explicitly define the abstract data collection objects onto the actual industrial measurement points and execution equipment that can be deployed, maintained, and operated long-term in heavy-duty gas generator sets, thereby achieving a consistent correspondence between the solution and the real power plant equipment.

[0063] Furthermore, the process of the high-frequency data acquisition module acquiring pipeline transient pressure data includes: acquiring unfiltered pipeline transient pressure data as raw pipeline transient pressure data; filtering the raw pipeline transient pressure data using a Kalman filter to remove high-frequency mechanical vibration noise from the fluid power source, extracting the low-frequency envelope reflecting changes in flow resistance, and marking the low-frequency envelope as pipeline transient pressure data.

[0064] This embodiment provides a pipeline transient pressure data purification mechanism. Specifically, in the aforementioned continuous scenario where the unit participates in grid peak shaving, the original pressure signal usually contains two types of components: one is high-frequency disturbances introduced by the mechanical rotation of the oil pump, the vibration of the coupling, and the micro-vibration of the mounting base; the other is low-frequency changes related to oil delivery capacity, pipeline resistance, and local flow state changes. If these two types of components are not distinguished, subsequent state observations will misjudge mechanical noise as flow state changes, thereby affecting the inference of the temperature of the heating zone. Therefore, the high-frequency data acquisition module first acquires the unfiltered original pressure data, and then extracts the low-frequency envelope that can represent the flow resistance change through a Kalman filter, which is used as the pipeline transient pressure data for subsequent use.

[0065] The specific process is as follows: In the Kalman filter model, the original pipeline transient pressure is set as the superposition of the baseline state with a rate of change less than a preset threshold and high-frequency mechanical vibration noise; by configuring the process noise covariance within a first preset range and the measurement noise covariance within a second preset range, and the value of the second preset range is greater than the value of the first preset range, the filter exhibits attenuation and suppression characteristics for transient high-frequency jitter, thereby smoothly fitting a low-frequency baseline, i.e., the low-frequency envelope, representing the slow change of pipeline flow resistance at the output end;

[0066] The details are as follows: At the physical level, the rotation of the oil pump will bring about periodic mechanical vibrations. These vibrations will be directly superimposed on the pressure transmitter, making the original pressure curve appear as a rapidly fluctuating waveform. However, what is really used to judge the flow hysteresis and heat exchange capacity is not these mechanical vibrations themselves, but the overall rise and fall trend of pressure with changes in load, oil temperature, valve position and oil viscosity. The core of the filtering process is not simply to make the curve smoother, but to retain the low-frequency signal trend related to the cooling oil delivery state and remove high-frequency interference components that are not directly related to the thermal state.

[0067] A simplified timing example can illustrate the data flow: Suppose that five raw pressure points are continuously sampled within a short period of time, exhibiting high-frequency oscillations of different amplitudes, but the overall center value around which these points revolve is slowly shifting upwards; after filtering, the system no longer retains the aforementioned high-frequency fluctuation components, but instead outputs a slowly rising envelope trend line; the former mainly reflects the superposition effect of mechanical vibration in the measurement link, while the latter is closer to the actual changes in oil circuit resistance or flow velocity; subsequent modules use this envelope trend line, rather than the original jitter sequence;

[0068] Without this step, when the oil pump frequency is rapidly adjusted, the vibration of the unit base is enhanced, or local resonance occurs, the mechanical noise in the pressure signal will be significantly amplified, causing the transmission delay calculation to be off, and even causing the high-frequency state observation to mistake the vibration peak as an abnormal flow resistance; therefore, this embodiment optimizes the previous stage scheme by filtering and extracting the low-frequency envelope.

[0069] Under abnormal operating conditions, when the pressure sensor becomes saturated, disconnected, or drifts significantly, the system will no longer output a new effective envelope. Instead, it will mark the channel as faulty and temporarily use the representative value within the most recent stable time window for state observation. If the mechanical vibration noise increases abnormally, causing the filter to fail to form a stable envelope, the system will reduce the contribution of the pressure channel in the observation and maintain a conservative control mode based on speed and pump frequency to avoid incorrect feedforward regulation caused by pressure distortion.

[0070] For example, when the peak-shaving unit shifts from medium load to high load, the oil pump frequency converter increases its frequency, and the vibration of the pump body and pipeline flange increases for a short time, causing obvious high-frequency jitter in the original signal of the pressure transmitter. If the original pressure sequence is used directly, the system will mistakenly believe that the oil circuit resistance is fluctuating violently at high frequency. After filtering, only the upward shift trend of the envelope caused by changes in oil viscosity and actual flow rate is retained, and the subsequent delay estimation and temperature reconstruction are therefore more stable.

[0071] The purpose of this step is to extract effective information that truly corresponds to the flow state from the noisy field pressure signal, thereby enabling a more reliable judgment on fluid transport hysteresis and the thermal state of the heating zone.

[0072] Furthermore, the process of the high-frequency state observation module outputting the high-frequency predicted value of the fluid temperature in the target heating zone includes: constructing an extended Kalman filter model as a state observer; inputting the operating frequency of the fluid power source, the transient pressure data of the pipeline, and the characteristic parameters of the target heating zone into the extended Kalman filter model; using the extended Kalman filter model to perform state prediction and covariance update; generating the current covariance matrix during the update process; and outputting the high-frequency predicted value of the fluid temperature in the target heating zone.

[0073] This embodiment provides a high-frequency thermal state observation mechanism. Specifically, during the peak-shaving operation of the aforementioned unit, the oil temperature in the heating zone cannot be directly and continuously measured by a high-frequency sensor. Therefore, an extended Kalman filter model is constructed in the edge control device as a state observer. The oil pump operating frequency, filtered pipeline pressure data, and rotor speed are input together to continuously output a high-frequency estimated value of the fluid temperature in the target heating zone.

[0074] The reasons for using this type of state observer are as follows: the cooling oil circuit has obvious nonlinear characteristics. For example, the viscosity of the oil changes with temperature, and the viscosity change will in turn affect the flow rate, pressure drop and heat exchange effect. The unit speed change will affect the heat generation intensity, and the pump frequency change will affect the heat carrying capacity. This mutual coupling relationship cannot be fully described by a single linear proportion. The physical role of the state observer here is to organize the measurable driving and response quantities into a continuously evolving internal thermal state estimation process, so that the system can make high-frequency tracking of the oil temperature change in the heat generation zone even before the low-frequency temperature measurement has arrived.

[0075] To further explain, the extended Kalman filter model in this scheme operates on a high-frequency prediction and low-frequency correction rhythm. Within each high-frequency cycle, the model mainly extrapolates the state based on the previous moment's state, the pump operating frequency, pipeline transient pressure data, and rotor speed, while simultaneously propagating the current uncertainty. The covariance update in the high-frequency stage is specifically manifested as the covariance propagating recursively with the process model, rather than requiring each high-frequency cycle to rely on new temperature measurements to complete the observation correction.

[0076] When the twin server provides the phase alignment residual after dynamic hysteresis alignment, the residual is then used as an effective observation condition for low-frequency arrivals to participate in subsequent weight updates. Thus, the high-frequency state observation module is responsible for solving the problem of high-frequency continuous estimation, and the dynamic hysteresis alignment module is responsible for solving the problem of adaptive correction after the alignment residual arrives. The two are connected in time and cooperate with each other in function, rather than repeatedly performing the same observation action.

[0077] Furthermore, the state variables in the extended Kalman filter model include at least the core thermal state of the fluid temperature in the target heating zone, and can be combined with auxiliary internal variables that need to be included to characterize the heat accumulation or heat dissipation trend; the input variables are the oil pump operating frequency, the filtered pressure envelope, and the rotor speed; the output variable is the high-frequency prediction of the fluid temperature in the target heating zone.

[0078] In the state prediction process equation of the extended Kalman filter model, the rotor speed is specifically used as the input variable of the heat source generation term to calculate the heat input rate, and the transient flow velocity calculated from the oil pump operating frequency and pressure envelope is used as the input variable of the convective heat transfer term to calculate the heat transfer rate. Then, based on the dynamic energy balance relationship between the heat input rate and the heat transfer rate, the internal fluid temperature of the target heating zone is recursively predicted.

[0079] The purpose of this setup is to map changes in heat source, transport capacity, and flow resistance to the same thermal state evolution framework, so that the model retains the physical interpretability of the industrial object while avoiding simply treating the temperature of the heating zone as a static lookup result of a single input and single output.

[0080] A simplified timing example can be used to illustrate this: Assume that the input signal contains three sets of states at a certain time period. The first set is that the pump frequency is stable, the pressure is stable, and the speed is stable. The second set is that the pump frequency increases slightly, the pressure envelope shifts upward, and the speed increases. The third set is that the pump frequency increases further, the pressure change tends to stabilize, and the speed remains high. The state observer will not simply map these three sets of inputs into three isolated outputs. Instead, it will combine the continuity of the thermal state at the previous moment to form a high-frequency prediction trajectory in which the temperature first rises, is then suppressed by the flow-increasing cooling, and then enters a new thermal equilibrium. Here, the emphasis is on the continuous evolution of the thermal state, rather than a static lookup table at a certain instant.

[0081] If only the aforementioned acquisition and hysteresis alignment mechanism is relied upon, without such a state observer, the system can only perform correction after the low-frequency temperature is reached, but cannot make a forward-looking response to the thermal shock of the heating zone within a few seconds. Therefore, this embodiment makes up for the gap between high-frequency controllability and low-frequency measurability by establishing a state observer.

[0082] As an anomaly handling mechanism, when some inputs are temporarily missing, the observer can make predictions based on the continuation of the previous state within a short period of time. If the missing inputs continue for more than a set time, or if there are obvious physical contradictions between the inputs, such as a significant increase in pump frequency while the pressure remains unresponsive for a long time, or speed fluctuations that are inconsistent with the unit's operating condition records, the system will mark the prediction as low confidence, restricting it from directly triggering large-scale feedforward actions, and waiting for new reliable observations to participate in the correction. If the unit switches to abnormal power generation conditions such as start-up or shutdown, the parameter boundaries of the corresponding operating condition can be called to avoid interpreting the thermal state according to the conventional power generation conditions.

[0083] For example, during the evening peak period of the power plant, after the unit receives the load increase command, the rotor-related heat sources are enhanced first, and the oil pump frequency is increased to increase the circulation of cooling oil. According to the physical rhythm of heat source enhancement first and cooling enhancement later, the condition observer first outputs a high-frequency prediction that the oil temperature in the heating zone has an upward trend, and then gradually corrects the temperature rise rate as the pumping capacity increases. Even if the temperature at the return oil measuring point has not yet risen significantly, the edge control equipment can already take adjustments in advance based on the high-frequency prediction.

[0084] The purpose of this step is to continuously reconstruct the target's thermal state using real-time operating information available at the edge when sensors cannot directly cover the high-temperature core area, thereby enabling early identification and rapid control of thermal shock.

[0085] Furthermore, the process by which the dynamic hysteresis alignment module calculates the fluid transmission delay time includes: obtaining the known pipe length and cross-sectional area of ​​the fluid circulation system; deriving the transient flow velocity of the fluid based on the transient pressure data of the pipe, the operating frequency of the fluid power source, and the cross-sectional area of ​​the pipe; using the global timestamp corresponding to the current low-frequency cycle as a reference, tracing back along the historical time axis, performing time integration on the transient flow velocity to obtain the displacement integration result, comparing the displacement integration result with the known pipe length, and determining whether the displacement integration result reaches the known pipe length; if the displacement integration result reaches the known pipe length, then the corresponding integration time period is marked as the transmission delay time; if the displacement integration result does not reach the known pipe length, then time integration continues until the displacement integration result reaches the known pipe length.

[0086] This embodiment provides a time-varying transmission delay determination mechanism. Specifically, in the aforementioned continuous operation scenario where the unit transitions from a steady state to rapid load increase and load shedding, the time required for the cooling oil to flow from the heating zone to the return oil temperature measurement point is not constant. As the oil temperature, viscosity, pump frequency, and pipeline resistance change, the actual flow rate of the oil will drift. If a fixed hysteresis time is still used, alignment errors will occur. Therefore, the dynamic hysteresis alignment module estimates the current fluid transmission delay time based on the known pipe length, pipeline cross-sectional area, and real-time flow status.

[0087] The steps are described below: The engineering meaning of this step can be understood as determining how long it takes for a certain batch of oil to reach the downstream measuring point from upstream; the system reads the pipe length and cross-sectional area obtained from the pre-modeling, which are structural parameters of the cooling oil system; combined with the current pressure envelope and pump operating frequency, the transient flow velocity is estimated; here, the goal is not to establish a complex three-dimensional flow field, but to extract the main engineering features, namely the transient flow velocity change trend of the overall oil delivery at the current moment; then, the system gradually accumulates the flow displacement in different time slices, and when the accumulated displacement reaches the actual pipe length, it is determined that the accumulated time corresponds to the transmission delay of the current batch of oil;

[0088] To further explain, the transient flow velocity of the fluid is derived from the transient pressure data of the pipeline, the operating frequency of the fluid power source, and the cross-sectional area of ​​the pipeline. It is preferable to use the engineering mapping relationship pre-calibrated for the cooling oil system of the unit, rather than applying ideal fluid formulas without considering the actual mechanical properties of the equipment. Specifically, a correspondence table, piecewise function, or lookup table model between the oil pump frequency, pressure envelope characteristics, and volumetric flow rate can be established during the unit commissioning phase.

[0089] During operation, the server first determines the corresponding volumetric flow rate online based on the current oil pump frequency and pressure envelope, and then calculates the transient flow velocity by combining the pipeline cross-sectional area. In other words, the pressure data in this scheme mainly reflects the current flow resistance and delivery status, while the oil pump frequency mainly reflects the pumping capacity. Together, they determine the flow rate estimation result that is closer to the actual situation on site.

[0090] For example, if the estimated volumetric flow rate for the current period is The cross-sectional area of ​​the pipeline is The transient velocity of that period can be calculated as follows: Conversion; Accumulate the flow rate by time slice, if the first... The duration of each time slice is The result of the displacement integral can be calculated as follows: Gradually accumulate; among them, Indicates the first The transient flow rate used within each time slice, Indicates the first The duration of a time slice Indicates the cumulative total up to the [number]. The displacement integral results over each time slice Indicates the time slice number that participated in the cumulative process. Indicates a known pipe length; when Reaching or exceeding the known pipe length for the first time When the cumulative time is equal to the transmission delay time, the corresponding cumulative time is considered as the transmission delay time; if a time slice causes the cumulative displacement to just cross the threshold... Then, a fraction of a time can be compensated linearly within that time slice to refine the hysteresis time boundary, thereby avoiding significant jumps caused by high-frequency sampling discretization.

[0091] The specific calculation method for this fractional time is as follows: when the known pipe length is reached or exceeded for the first time... Cumulative displacement at time With known pipe length The difference, divided by the current number. Transient flow rate per time slice This yields an extra amount of time; the first The fractional time is the sum of the duration of each time slice minus the extra time. This fractional time is then added to the previous time slice. The precise transmission delay time is obtained over each cumulative time slice;

[0092] Furthermore, the above integration process preferably uses the global timestamp corresponding to the current low-frequency temperature record as the endpoint and traces backward along the time axis where the historical high-frequency records are located. In other words, the system does not predict when the oil will arrive at the measuring point in the future, but rather, when the current downstream temperature has already been reached, it infers approximately how long ago this batch of oil left the heating zone upstream. The delay time obtained in this way is consistent with the subsequent queue backtracking to extract the historical high-frequency estimate, and can directly serve the generation of phase alignment residuals.

[0093] A simplified timing example can be used to illustrate this: Assume the system divides the continuous flow process into four time slices: S1, S2, S3, and S4. In each time slice, the oil advances a certain distance. When the oil pump is running at high frequency, the oil temperature is high, and the viscosity is low, the advance distance in the first two time slices may be large, and by S3, the entire distance from the upstream to the return oil measuring point has been reached. However, when the oil temperature is low, the oil is more viscous, and the pump frequency is low, the advance distance in the aforementioned four time slices is small, and it may be necessary to accumulate to S4 or even later time slices to reach the same pipe length. Based on this, the system uses the time required to reach the pipe length as the current hysteresis time, rather than using a fixed constant to cover all operating conditions.

[0094] Without this dynamic determination process, although the previous scheme has the ability to compare historical queues, it will encounter the problem of matching the wrong historical object when the flow rate changes significantly. Especially after the load is shed, the drop in oil temperature will change the oil viscosity and the actual flow rate, and the previously applicable fixed delay will no longer be reliable. As a result, the residual does not reflect the true phase difference, but is mixed with timing misalignment error. Therefore, this embodiment further improves the original scheme by calculating the hysteresis time in real time.

[0095] As an anomaly handling mechanism, when pressure data is abnormal or pump frequency signal is interrupted briefly in a certain cycle, resulting in the inability to form a reliable flow velocity estimate, the system can temporarily use the hysteresis time of the previous effective cycle and mark this value as the transition hysteresis value; if an effective estimate cannot be recovered for several consecutive cycles, the generation of residuals based on hysteresis alignment is suspended, and only high-frequency prediction and basic protection control are retained; if the integral process cannot reach the known pipe length for a long time, it indicates that the current flow velocity estimate is too low, there are signs of blockage or input abnormality, and the system can issue a flow abnormality prompt and switch the control strategy to conservative flow increase mode;

[0096] For example, after the peak-shaving unit has been operating at high load for a period of time, the return oil temperature rises, causing the oil viscosity to decrease. The cooling oil is then transported faster at the same pump frequency. Based on this, the system determines that the actual transmission time from the heat-generating area to the return oil measuring point has shortened. When the power grid experiences load shedding, the heat source weakens, the oil gradually cools down, and the local flow resistance changes again. The system then automatically lengthens the lag time. Throughout the entire process, the server always searches for historical prediction values ​​based on the actual transportation time required under the current operating conditions, rather than mechanically using a fixed delay.

[0097] The purpose of this step is to establish a true one-to-one correspondence between historical predicted values ​​and currently measured temperatures, thereby compensating for pure hysteresis drift under time-varying viscosity and time-varying flow velocity conditions.

[0098] Furthermore, the process of extracting historical high-frequency estimates from the dynamic hysteresis alignment queue after the transmission delay time of the global timestamp, and comparing the historical high-frequency estimates with the downstream fluid temperature of the current cycle to generate phase alignment residuals includes: the dynamic hysteresis alignment queue is a first-in-first-out data structure; in each high-frequency cycle, the high-frequency estimates with global timestamps are pushed into the dynamic hysteresis alignment queue; in the current low-frequency cycle, the corresponding historical timestamp node is calculated based on the current global timestamp and the transmission delay time, and the historical high-frequency estimates corresponding to the historical timestamp node are extracted from the dynamic hysteresis alignment queue; the difference between the downstream fluid temperature of the current cycle and the historical high-frequency estimates is calculated, and the difference is marked as the phase alignment residual.

[0099] This embodiment provides a phase alignment residual generation mechanism. Specifically, in the aforementioned continuous peak shaving scenario of the unit, the edge side continuously outputs high-frequency prediction values, while the server side obtains the return oil temperature once at relatively long intervals. In order to use these two types of data with different cycles for comparison of the same physical object, the system constructs a dynamic hysteresis alignment queue with a first-in-first-out data structure, and pushes the high-frequency prediction value with a timestamp into the queue in each high-frequency cycle. When the low-frequency temperature arrives, the corresponding historical prediction value is extracted back based on the current hysteresis time, thereby generating the phase alignment residual.

[0100] The following is a detailed explanation: First-in-first-out (FIFO) queues are suitable for storing high-frequency predictions that arrive sequentially over time, because the cooling oil transport process itself has a first-to-first-to-arrive time sequence. The head of the queue usually stores the earliest historical predictions, while the tail stores the latest predictions. When a low-frequency temperature is reached, the server does not take the latest value, but instead searches the queue for the prediction results of the same batch of oil at the upstream time based on the physical transport time corresponding to that temperature. The difference between the two is the phase alignment residual, which is not simply a prediction error, but rather the model's interpretation bias of the thermal state of the same batch of oil after eliminating transport lag.

[0101] A simplified timing example can be used to illustrate this: Assume the queue contains P1, P2, P3, P4, and P5, representing the upstream temperature estimates formed by five high-frequency cycles. At this time, the server receives a low-frequency return oil temperature M5. If the current lag time corresponds to reviewing three high-frequency cycles, M5 is not compared with P5, but with P2. If M5 is higher than P2, it indicates that the previous estimate of the thermal state of this batch of oil was too low; if M5 is lower than P2, it indicates that the previous prediction was too high. The difference between the two is the residual. This residual only has corrective significance when the time phase is consistent.

[0102] If this queuing mechanism is not used, and the current low-frequency temperature is used to directly correct the current high-frequency forecast, significant misalignment will occur in high-load, rapidly changing scenarios. For example, the current return oil temperature reflects the thermal state several seconds ago, while the current high-frequency forecast reflects the thermal state at this moment. Directly subtracting the two will mistake normal physical hysteresis for prediction deviation, resulting in incorrect correction direction. Therefore, this embodiment establishes a traceable historical prediction cache through a first-in-first-out queue to solve the asynchronous alignment problem of multi-rate data.

[0103] To further explain, pushing the high-frequency prediction value with a global timestamp into the dynamic hysteresis alignment queue within each high-frequency cycle corresponds to the implementation of continuously caching historical prediction values. Storing the high-frequency prediction value before the transmission delay time into the dynamic hysteresis alignment queue means that within any current low-frequency cycle, the historical high-frequency prediction values ​​that actually participate in the phase alignment candidate range are those located before the current time and whose time position falls within the transmission delay time backtracking range. In other words, the queue continuously receives all high-frequency prediction records in its implementation, but only calls the historical record that meets the hysteresis backtracking conditions in its usage. This maintains continuous updates to the data structure while remaining consistent with the object selection logic before the transmission delay time in the embodiment.

[0104] Furthermore, the queue length is preferably designed to cover at least the maximum allowable transmission delay time window of the system, and to reserve additional time margin to absorb clock jitter and sampling discrepancy errors. In this way, when the current global timestamp is reduced by the transmission delay time to form a historical timestamp node, the system can complete the backtracking search in the same queue without having to reassemble historical data across modules. If the historical timestamp node falls between two adjacent high-frequency records, it is preferable to select the effective node with the smaller time difference first. When the time difference is still within the set tolerance, the estimated values ​​of two adjacent historical high frequencies can also be interpolated according to the time ratio to reduce the phase quantization error caused by the non-divisibility of the high and low frequency sampling periods. The aforementioned interpolation is only used to refine the determined historical timestamp nodes and does not change the basic process of using a first-in-first-out queue for hysteresis backtracking and residual generation.

[0105] As an anomaly handling mechanism, when a completely consistent historical node cannot be found based on the timestamp, the system can select the nearest valid node as the comparison object; if the nearest deviation exceeds the set tolerance, the residual of this round will not participate in model correction; if the queue length is insufficient to cover the current lag time, it indicates that the system has just started, the historical cache has not yet been fully accumulated, or the lag time has increased abnormally. In this case, the server only performs data caching and does not generate residuals; if low-frequency temperature suddenly disappears, the queue is continuously updated, and phase comparison is resumed after the new temperature arrives.

[0106] For example, after the unit completes a rapid load increase, the edge side generates a large number of high-frequency predictions within a few seconds and stores them in the queue; when the return oil measuring point first shows a significant temperature rise, the server finds the prediction record from the queue when the heat was just released from the heating zone a few seconds ago, based on the lag time calculated at that time, compares it with the current return oil temperature, and obtains a set of residuals; these residuals are used to correct prediction deviations under subsequent similar operating conditions.

[0107] The purpose of this step is to reconstruct data collected at different sampling frequencies and physical locations onto the same time phase, thereby achieving true interpretability of the residuals and effectiveness of subsequent corrections.

[0108] Furthermore, the process by which the feedforward control module injects the phase alignment residual back into the high-frequency state observation module as the update weight includes: obtaining the current covariance matrix of the high-frequency state observation module; inputting the phase alignment residual as the observation update condition into the high-frequency state observation module; iteratively correcting the current covariance matrix based on the phase alignment residual to obtain the updated covariance matrix; and using the updated covariance matrix as the prediction weight for the next high-frequency cycle.

[0109] This embodiment provides an adaptive update mechanism for prediction weights based on phase-aligned residuals. Specifically, based on the aforementioned high-frequency state observation mechanism, the system does not simply add the residuals formed each time as temperature corrections, but uses them as observation update conditions to reverse the prediction weight allocation within the high-frequency state observation module, so that the confidence level of the model trend and the actual observation in the next high-frequency cycle is adaptively adjusted.

[0110] The following is a detailed explanation: In industrial settings, thermal condition prediction deviations are not usually a simple case of being fixedly large or small, but rather vary with load levels, oil viscosity, ambient temperature, and sensor status. If the same correction method is used every time, it is easy to undercorrect under some operating conditions and overcorrect under others.

[0111] Therefore, this embodiment does not regard the residual as a one-time static error compensation term, but rather as a feedback signal of the current model's credibility. The current covariance matrix can be understood as the representation of uncertainty within the state observer. The larger the residual, the more obvious the difference between the previous prediction and the actual aligned observation, and the more the influence of the observation information should be increased in the future. The smaller the residual, the more reliable the current prediction structure, and the more important it is to maintain the continuity of the model in the future.

[0112] To further explain, the core of iteratively correcting the current covariance matrix based on phase-aligned residuals is not to directly write the residual values ​​into the covariance matrix in equal amounts, but to adjust the relative weights of observation uncertainty and prediction uncertainty according to the residual amplitude, residual duration, and sensor health status.

[0113] Specifically, when the residuals remain large within a reasonable range, the system increases the confidence level of the aligned observation information, enabling the prediction results of the next high-frequency cycle to converge to the new observation values ​​more quickly; when the residuals remain small, the system maintains or moderately restores its confidence in the continuous evolution of the model, so as to avoid unnecessary jitter in high-frequency predictions caused by accidental fluctuations in low-frequency observations; thus, the covariance matrix in this scheme plays the role of adjusting whether the model or the observation is more dominant, rather than being regarded as a container for simply storing error values.

[0114] Furthermore, to ensure that the updated covariance matrix can continue to be used as the prediction weight for the next high-frequency cycle, the iterative correction preferably satisfies three constraints: First, the updated result maintains the same physical dimension and state correspondence as the original matrix; second, the updated main diagonal elements are not less than a preset lower limit to avoid misjudging the system as completely without uncertainty; third, the update magnitude is limited by a set upper bound to prevent a single abnormal residual from causing abrupt changes in subsequent weights; if necessary, the matrix can also be symmetricized and its boundaries pruned after the update to ensure the stability and continuity of the subsequent recursive process.

[0115] The specific iterative correction logic is as follows: calculate the ratio of the absolute value of the phase alignment residual to the preset residual tolerance threshold, and use this ratio as the adaptive adjustment coefficient; when the adaptive adjustment coefficient is greater than 1, use the adaptive adjustment coefficient to linearly amplify the process noise covariance matrix of the extended Kalman filter model, and then combine it with the standard Kalman gain formula to complete the update calculation of the current covariance matrix.

[0116] To further clarify, the current covariance matrix, the updated covariance matrix, and the prediction weights for the next high-frequency cycle all correspond to the same set of state variables within the extended Kalman filter model, and a second set of covariance systems parallel to this extended Kalman filter model is not introduced separately.

[0117] In other words, the state prediction and covariance recursion in the high-frequency stage, as well as the weight correction after the low-frequency phase alignment residual arrives, all operate on the same observer structure within the same high-frequency state observation module. They only perform two functions, continuous extrapolation and hysteresis correction, respectively, in terms of time tick. This explanation ensures that the process of covariance update and iterative correction based on phase alignment residual are consistent in terms of object and connected in terms of time sequence, avoiding the misunderstanding that they are repetitive processing for different matrices or different observers.

[0118] A simplified time series example can be used to illustrate this: assuming the phase alignment residuals obtained from three consecutive low-frequency cycles are small, medium, and large, respectively; in the first case, the system assumes that the current high-frequency observation model is basically consistent with the actual thermal state, and the original prediction weights are maintained in the next stage; in the second case, the system appropriately increases the participation of observation correction; in the third case, it indicates that the unit may have experienced new thermal conditions or changes in flow conditions, and the system further enhances the update weights, so that subsequent high-frequency predictions can more quickly approach the actual operating conditions; the key point here is not how a certain matrix element changes, but the dynamic adjustment of the model's own credibility.

[0119] If the high-frequency predictions output by the aforementioned state observer are relied upon without injecting the phase-aligned residuals back into the weight update, the model may gradually develop systematic drift after long-term operation. This drift will continue to accumulate, especially when the flow resistance characteristics change due to seasonal changes, oil aging, or unit maintenance. Therefore, this embodiment utilizes the aligned residual information to continuously correct the uncertainty assessment within the observer and further optimize the previous stage scheme.

[0120] As an anomaly handling mechanism, when a residual in a certain round is determined to originate from sensor failure, abnormal communication latency, or queue alignment failure, that residual will not be used for weight updates, and the system will continue to use the effective weights from the previous round. If the residual increases abnormally for several consecutive cycles, accompanied by physical inconsistencies between pressure, pump frequency, and speed, the system can determine that the current model's basic assumptions deviate significantly from the actual field conditions, triggering re-initialization or switching to conservative control mode. If the residual remains close to zero but the sensor maintains a constant value for a long period, false stability must also be prevented. In this case, the system can decide whether to postpone updates based on the sensor's health status.

[0121] For example, after the unit experiences a sudden load shedding, the return oil temperature measurement point returns a set of residuals that deviate significantly from historical high-frequency predictions after the hysteresis time arrives. This indicates that the unit's heat release rhythm is no longer consistent with the observer settings under the original high-load steady state. Based on this, the system adjusts the prediction weights for the next stage, making the observers more sensitive to the new cooling oil flow and heat dissipation status. When the unit re-enters a stable load, the residuals gradually decrease, and the system gradually restores a more stable prediction weight allocation.

[0122] The purpose of this step is to enable the high-frequency condition observation process to have long-term self-correction capabilities, thereby achieving continuous adaptation to the effects of operating condition drift, changes in oil properties, and equipment aging.

[0123] Furthermore, the process by which the feedforward control module generates feedforward adjustment commands based on high-frequency prediction values ​​includes: setting a target safety threshold for the fluid temperature in the target heating zone; comparing the high-frequency prediction value with the target safety threshold; if the high-frequency prediction value is greater than or equal to the target safety threshold, generating a feedforward adjustment command to increase the operating frequency of the fluid power source; if the high-frequency prediction value is less than the target safety threshold, generating a feedforward adjustment command to maintain or decrease the operating frequency of the fluid power source.

[0124] This embodiment provides a feedforward adjustment mechanism based on high-frequency temperature prediction. Specifically, in the continuous control link constituted by the aforementioned embodiments, the feedforward feedback control module pre-sets a safety threshold for the fluid temperature in the target heating zone and compares the high-frequency prediction value with the threshold. When the high-frequency prediction value reaches or exceeds the threshold, the system generates an adjustment command to increase the operating frequency of the oil pump before the return oil temperature fully reflects the thermal shock. When the high-frequency prediction value is lower than the threshold, an adjustment command to maintain or reduce the oil pump frequency is generated.

[0125] The details are as follows: This safety threshold is not an arbitrarily set number, but a control boundary determined based on bearing lubrication safety, oil film stability, and the allowable temperature rise range of the equipment. When the oil temperature in the heating zone approaches this boundary, the oil viscosity may decrease, and the oil film's load-bearing capacity and heat-carrying capacity will be affected. If we wait for feedback from the downstream temperature sensor at this time, we often miss the best time for intervention. Therefore, this embodiment uses the high-frequency prediction value directly for feedforward regulation, so that pump frequency regulation occurs during the risk formation stage rather than after the risk has been transmitted downstream.

[0126] A simplified state sequence example can be used to illustrate this: Assume the system has a safe temperature boundary, and the edge side continuously outputs three high-frequency predicted states, respectively representing below the boundary, close to the boundary, and reaching the boundary. In the first state, the system can maintain the current pump frequency, or moderately reduce it while meeting cooling requirements, to reduce unnecessary continuous high-frequency operation. In the second state, the system can enter an early warning monitoring state, preparing to increase flow. In the third state, the system directly issues a feedforward command to increase the pump frequency. This process reflects control based on thermal risk advance, rather than passive remediation after waiting for lagging temperature feedback.

[0127] If only the aforementioned prediction, alignment, and correction are performed, without directly converting the high-frequency prediction results into execution instructions, the system can only monitor the thermal state and cannot fully utilize the role of digital twins in feedforward control. Therefore, this embodiment further implements the high-frequency temperature reconstruction results into the oil pump frequency conversion control, forming a complete closed loop.

[0128] Under abnormal operating conditions, if the high-frequency prediction value exceeds the safety threshold, but the sensor health status is abnormal, the prediction reliability is low, or the oil pump has reached the frequency limit, the system can enter the restricted feedforward mode, only executing the maximum allowable flow increase protected by the equipment constraints, and simultaneously issuing a high-temperature risk alarm; if the high-frequency prediction value is lower than the threshold, but the unit is in start-up warm-up, post-maintenance trial operation, or other special operating conditions, the system can prohibit active frequency reduction to avoid new risks caused by insufficient cooling; if the pressure or flow response does not meet expectations after the pump frequency adjustment is executed, it indicates that the actuator may be stuck or the oil circuit is abnormal. In this case, feedback protection takes priority, and the feedforward action can be limited or canceled.

[0129] For example, after the peak-shaving unit experiences rapid load, the oil temperature in the estimated heating zone on the edge side is close to the safety boundary, while the temperature in the return oil pipeline has not yet fully risen. Based on this, the system first increases the frequency of the cooling oil pump, allowing more cold oil to enter the high-temperature zone before the thermal shock expands further. When subsequent return oil measurement points confirm that the temperature rise has been suppressed and the high-frequency prediction drops back below the safety boundary, the system then maintains or gradually reduces the pump frequency to avoid long-term high-frequency operation causing additional operating load.

[0130] The purpose of this step is to transform the high-frequency thermal state sensing capability into a direct feedforward control capability for the cooling oil pump, thereby enabling early intervention in the thermal risks of the heat-generating area while taking into account the equipment safety boundary and operational stability.

[0131] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A digital twin system for cooling oil circulation in heavy-duty gas generator sets, applied to fluid circulation systems, characterized in that: include: An edge control device and a twin server are provided. The edge control device is communicatively connected to a high-frequency data acquisition module, a high-frequency status observation module, and a feedforward feedback control module. The twin server is communicatively connected to a low-frequency data acquisition module and a dynamic hysteresis alignment module. The high-frequency data acquisition module is used to acquire the fluid power source operating frequency, pipeline transient pressure data and target heating zone characteristic parameters of the fluid circulation system within a preset high-frequency period that is less than the minimum control response time of the fluid power source frequency change, and attach a unified global timestamp. The low-frequency data acquisition module is used to acquire the downstream fluid temperature of the fluid circulation system within a preset low-frequency period that is greater than the thermal inertia delay time of a downstream fluid temperature sensor located in the fluid circulation system, and to attach a unified global timestamp. The high-frequency state observation module is used to output a high-frequency estimated value of the fluid temperature in the target heating zone through a state observer based on the operating frequency of the fluid power source, the transient pressure data of the pipeline, and the characteristic parameters of the target heating zone. The dynamic hysteresis alignment module is used to calculate the transmission delay time of the fluid based on the transient pressure data of the pipeline and the operating frequency of the fluid power source, continuously store the high-frequency prediction values ​​obtained in each cycle into the dynamic hysteresis alignment queue, extract the historical high-frequency prediction values ​​after the transmission delay time from the global timestamp in the dynamic hysteresis alignment queue, compare the historical high-frequency prediction values ​​with the downstream fluid temperature of the current cycle, and generate phase alignment residuals. The feedforward control module is used to inject the phase alignment residual back into the high-frequency state observation module as an update weight, and generate a feedforward adjustment command based on the high-frequency prediction value, and drive the fluid power source through the feedforward adjustment command.

2. The heavy-duty gas generator set cooling oil circulation digital twin system according to claim 1, characterized in that, The fluid circulation system is a cooling oil circulation system for a heavy-duty gas generator set, the characteristic parameter of the target heating zone is the rotor speed, the fluid power source is a cooling oil pump, and the downstream fluid temperature is the temperature of the return oil pipeline.

3. The heavy-duty gas generator set cooling oil circulation digital twin system according to claim 1, characterized in that, The process of the high-frequency data acquisition module acquiring pipeline transient pressure data includes: acquiring unfiltered pipeline transient pressure data as raw pipeline transient pressure data; filtering the raw pipeline transient pressure data using a Kalman filter to remove high-frequency mechanical vibration noise from the fluid power source, extracting a low-frequency envelope reflecting changes in flow resistance, and marking the low-frequency envelope as the pipeline transient pressure data.

4. The heavy-duty gas generator set cooling oil circulation digital twin system according to claim 1, characterized in that, The process by which the high-frequency state observation module outputs the high-frequency predicted value of the fluid temperature in the target heating zone includes: constructing an extended Kalman filter model as the state observer; The operating frequency of the fluid power source, the transient pressure data of the pipeline, and the characteristic parameters of the target heating zone are input into the extended Kalman filter model. The extended Kalman filter model is used to perform state prediction and covariance update, and the current covariance matrix is ​​generated during the update process. The high-frequency predicted value of the fluid temperature of the target heating zone is output.

5. The heavy-duty gas generator set cooling oil circulation digital twin system according to claim 1, characterized in that, The process by which the dynamic hysteresis alignment module calculates the fluid transport delay time includes: obtaining the known pipe length and pipe cross-sectional area of ​​the fluid circulation system; The transient flow velocity of the fluid is derived based on the transient pressure data of the pipeline, the operating frequency of the fluid power source, and the cross-sectional area of ​​the pipeline. Using the global timestamp corresponding to the current low-frequency cycle as a reference, trace back along the historical time axis, perform time integration on the transient flow velocity to obtain the displacement integration result, compare the displacement integration result with the known pipe length, and determine whether the displacement integration result reaches the known pipe length; If the displacement integral result reaches the known pipe length, then the corresponding integration time period is marked as the transmission delay time; If the displacement integral result does not reach the known pipe length, then time integration continues until the displacement integral result reaches the known pipe length.

6. The heavy-duty gas generator set cooling oil circulation digital twin system according to claim 1, characterized in that, The process of extracting the historical high-frequency estimate of the global timestamp in the dynamic hysteresis alignment queue after the transmission delay time, comparing the historical high-frequency estimate with the downstream fluid temperature of the current period, and generating the phase alignment residual includes: the dynamic hysteresis alignment queue is a first-in-first-out data structure; Within each high-frequency cycle, the high-frequency estimate with a global timestamp is pushed into the dynamic hysteresis alignment queue; Within the current low-frequency cycle, the corresponding historical timestamp node is calculated based on the current global timestamp and the transmission delay time, and the historical high-frequency estimate corresponding to the historical timestamp node is extracted from the dynamic hysteresis alignment queue. Calculate the difference between the downstream fluid temperature in the current cycle and the historical high-frequency estimate, and label the difference as the phase alignment residual.

7. The heavy-duty gas generator set cooling oil circulation digital twin system according to claim 4, characterized in that, The process by which the feedforward control module injects the phase alignment residual back into the high-frequency state observation module as an update weight includes: obtaining the current covariance matrix of the high-frequency state observation module; inputting the phase alignment residual as an observation update condition into the high-frequency state observation module; iteratively correcting the current covariance matrix based on the phase alignment residual to obtain the updated covariance matrix; and using the updated covariance matrix as the prediction weight for the next high-frequency cycle.

8. The heavy-duty gas generator set cooling oil circulation digital twin system according to claim 1, characterized in that, The process by which the feedforward control module generates a feedforward adjustment command based on the high-frequency prediction value includes: preset a target safety threshold for the fluid temperature in the target heating zone; and comparing the high-frequency prediction value with the target safety threshold. If the high-frequency estimated value is greater than or equal to the target safety threshold, a feedforward adjustment command to increase the operating frequency of the fluid power source is generated; if the high-frequency estimated value is less than the target safety threshold, a feedforward adjustment command to maintain or decrease the operating frequency of the fluid power source is generated.

Images