Digital twin energy efficiency prediction control method for shell-and-tube heat exchanger and heat exchanger

Abstract

The invention relates to a digital twinborn energy efficiency prediction control method for a shell-and-tube heat exchanger and the heat exchanger, and the method comprises the steps: constructing a digital twinborn body containing a mechanism-data fusion channel, forming an energy efficiency prediction model through the combined training of a Transform network and a physical equation, and outputting an energy efficiency curve in the next 5-15 minutes; solving an optimal control sequence in a rolling manner by taking energy efficiency curve integral maximization as a target; an instruction is issued to an execution mechanism before the load changes; the heat exchanger comprises a variable frequency pump, a regulating valve, a data acquisition unit and an edge calculation unit, and the edge calculation unit is directly connected with an execution mechanism through an EtherCAT bus to achieve closed-loop control; according to the method, the problems of response lag, model misalignment and insufficient safety of a traditional control method are solved, and unification of energy efficiency optimization and equipment safety is realized.

Description

Technical Field

[0001] This invention belongs to the field of industrial process control and energy management technology, specifically relating to a digital twin energy efficiency prediction and control method for shell-and-tube heat exchangers and the heat exchanger itself. Background Technology

[0002] Shell-and-tube heat exchangers are common heat exchange equipment in industries such as petroleum, chemical, metallurgy, and central air conditioning. Their operating energy efficiency directly affects the energy consumption level of the entire system. Currently, energy-saving control methods for heat exchangers mainly fall into the following categories, all of which have obvious limitations:

[0003] PID-based feedback control is the mainstream method in current industrial applications. By measuring the outlet temperature and comparing it with the set value, the deviation is calculated by the PID controller and then the pump frequency or valve opening is adjusted. However, this method has a serious response lag. When the external load, such as the inlet temperature or flow rate, changes, the system must wait for the outlet temperature to deviate before it can start to act. It cannot achieve advance adjustment, resulting in poor control quality and high energy consumption.

[0004] Some advanced systems calculate a fixed or stepped optimal setpoint for pumps and valves based on empirical formulas or steady-state models. However, this method cannot cope with dynamically changing operating conditions, and its models usually do not take into account the time-varying nature of equipment performance, such as scaling and aging. As the operating time increases, the set optimal value has actually deviated from the true optimal point.

[0005] Existing research has attempted to apply Model Predictive Control (MPC) to heat exchanger control. Its core bottleneck lies in the accuracy of the prediction model. Pure mechanistic models are difficult to accurately describe complex nonlinear characteristics such as fouling and flow distribution, leading to inaccurate predictions. On the other hand, pure data-driven models, such as RNN and LSTM, lack physical constraints and are prone to making predictions that violate physical laws in areas not covered by training data, resulting in poor generalization ability and low extrapolation reliability.

[0006] Most existing control schemes separate performance optimization from equipment safety protection. Optimization algorithms pursue the highest energy efficiency, but may output dangerous commands that cause pump cavitation or system overpressure. This requires intervention and interlocking shutdown through an additional, independent Safety Instrumented System (SIS), making it impossible to achieve optimal control under the premise of safety. In addition, although there are rule-based or simple adaptive parameter adjustments, there is a lack of a learning mechanism that can simultaneously track scaling and equipment aging and intelligently balance historical data with the latest information.

[0007] In summary, existing technologies lack a closed-loop energy efficiency control solution that can accurately predict future energy efficiency trends, adapt to equipment performance degradation, strictly ensure operational safety, and deeply integrate algorithms with hardware. Summary of the Invention

[0008] In order to overcome the problems of the prior art, the present invention discloses a digital twin energy efficiency prediction and control method for a shell-and-tube heat exchanger and a heat exchanger thereof.

[0009] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0010] A digital twin energy efficiency prediction and control method for a shell-and-tube heat exchanger includes the following steps:

[0011] S1, Construct a digital twin of a shell-and-tube heat exchanger, the digital twin including a mechanism-data fusion channel and an energy efficiency time series prediction sub-model;

[0012] Among them, the mechanism channel calculates the baseline EER based on the real-time heat transfer coefficient and pump power, while the data channel outputs the fouling-aging deviation that is not captured by the mechanism through the residual network. The two channels are added together to form the final EER prediction value.

[0013] The energy efficiency time series prediction sub-model is jointly trained by the Transformer neural network model and the heat transfer-pump work physical equation. It is used to output the instantaneous energy efficiency ratio (EER) prediction curve for the next 5–15 minutes. The joint training adopts a physical loss regularization term to ensure that the output of the Transformer neural network model satisfies the basic laws of energy conservation and heat transfer.

[0014] S2, with the optimization objective of maximizing the integral of the EER prediction curve over time, the optimal control sequence of pump frequency and regulating valve opening is obtained by rolling solution;

[0015] S3, before the actual load change occurs, the optimal control sequence is sent to the actuator.

[0016] Preferably, the joint training of the Transformer neural network model and the heat transfer-pump work physical equation adopts an end-to-end training strategy. Its training loss function is calculated based on the error between the final output EER prediction value and the true value, and a physical loss regularization term is fused for backpropagation.

[0017] Preferably, when the absolute value of the deviation between the final predicted EER value and the measured EER value at N consecutive sampling points is greater than a set threshold δ, online correction is automatically triggered.

[0018] The fouling thermal resistance Rf and pump efficiency η are updated in real time using a sliding window recursive least squares method, and the updated parameters are fed into the mechanism channel immediately.

[0019] Where N is an integer from 3 to 10, and δ is from 0.05 to 0.15.

[0020] Preferably, the online correction step employs a lifelong learning strategy with a forgetting factor λ∈(0.95,0.99), and λ is adaptively adjusted with the dirt growth rate to further suppress the cumulative interference of historical dirt data on the current parameters.

[0021] Preferably, after obtaining the optimal control sequence, a safety-execution mapping layer is added to verify the pump frequency, valve opening command, net positive suction head, maximum head, and upper limit of shell-side-tube-side pressure difference in real time.

[0022] If any constraint exceeds the limit, a second planning process is immediately initiated to solve the problem again, in order to balance optimal energy efficiency and equipment safety.

[0023] Preferably, the time required to resolve the quadratic programming problem is ≤100ms;

[0024] After the solution is completed, the edge computing unit sends the pump frequency and valve opening setpoints to the frequency converter and electric regulating valve via OPC UA or Modbus TCP protocol;

[0025] After the signal is sent, the actual feedback signal is continuously monitored. If the deviation between the actual value and the set value exceeds the allowable hysteresis and continues for two sampling cycles, a recalculation is triggered to maintain control accuracy.

[0026] Preferably, the digital twin uses a two-layer architecture consisting of a high-frequency closed-loop control layer and a low-frequency state evaluation layer.

[0027] The high-frequency closed-loop control layer refreshes every 1 minute to complete minute-level EER prediction and control sequence generation;

[0028] The low-frequency condition assessment layer is refreshed every 24 hours to assess the scaling and aging conditions and inject the aging condition data into the high-frequency closed-loop control layer to achieve multi-scale parameter cascade correction.

[0029] Preferably, the two-layer architecture is deployed in an edge-cloud collaborative environment, where the low-frequency state evaluation layer performs big data training in the cloud and the high-frequency closed-loop control layer performs real-time inference at the edge. The two layers communicate with each other through an event-driven interface to push state quantities in one direction.

[0030] Preferably, a shell-and-tube heat exchanger includes: a variable frequency circulating pump, an electric regulating valve, a data acquisition unit, and an edge computing unit;

[0031] The edge computing unit is connected to the frequency converter and electric regulating valve via a hard real-time EtherCAT bus, and carries executable code for a digital twin energy efficiency predictive control method for shell-and-tube heat exchangers, configured as follows:

[0032] It receives temperature, flow rate, and power signals uploaded by the data acquisition unit in real time.

[0033] The digital twin energy efficiency prediction control method for shell-and-tube heat exchangers generates pump frequency and valve opening control commands.

[0034] Commands are sent directly to the variable frequency circulating pump and electric regulating valve via the EtherCAT bus to achieve closed-loop control that maximizes the energy efficiency of the heat exchanger.

[0035] The beneficial effects of this invention are as follows:

[0036] By constructing a digital twin that integrates mechanism and data, and utilizing the powerful temporal feature capture capability of the Transformer network to learn complex nonlinear relationships, while constraining its output to conform to the basic laws of energy conservation and heat transfer through physical loss regularization terms, this hybrid modeling method combines the fitting accuracy of data-driven models with the extrapolation reliability of mechanistic models. It overcomes the shortcomings of inaccurate description by pure mechanistic models or poor generalization ability of pure data models, and achieves high-precision prediction of energy efficiency ratio (EER) for the next 5-15 minutes.

[0037] It possesses lifelong learning capabilities and maintains high fidelity over a long period. By introducing an online correction mechanism based on the sliding window recursive least squares method, it can automatically and in real time identify and update key time-varying parameters such as fouling thermal resistance and pump efficiency. Combined with an adaptive strategy with a forgetting factor, it can suppress the interference of historical dirty data and quickly track the trend of parameter changes. This enables the digital twin to adapt to the slow performance degradation of equipment such as fouling and aging, and maintain a state perception capability that is highly consistent with the physical entity throughout its entire life cycle.

[0038] Real-time online optimization under safety constraints is achieved by adding a safety-execution mapping layer, embedding key safety constraints such as net positive suction head, maximum head, and differential pressure into the optimization control loop, and using quadratic programming to resolve out-of-bounds commands at the millisecond level, ensuring that all issued commands are within the safe operating range of the equipment, resolving the conflict between optimization objectives and equipment safety, and achieving the unity of optimal energy efficiency and operational safety.

[0039] The control response is timely and the execution is accurate and reliable. It adopts a hard real-time communication link based on EtherCAT bus to send optimization instructions directly, quickly and definitively to the actuators, eliminating the delay and uncertainty caused by multi-layer network transmission in traditional control systems. It forms a complete intelligent control closed loop through the digital twin energy efficiency predictive control method, which effectively ensures the final execution effect of control instructions and improves the control accuracy and reliability of the system.

[0040] The system architecture is highly efficient, balancing real-time performance with global optimization. It adopts a two-layer architecture of cloud-edge collaboration, deploying minute-level real-time predictive control on the edge side to ensure low latency and high reliability of control, while deploying day-level deep state assessment and training on the cloud, making full use of cloud computing resources for big data analysis and model updates. The two layers collaborate through event-driven mechanisms to achieve decoupling and global optimization of tasks at different time scales, while meeting the needs of instantaneous response and long-term optimization.

[0041] In summary, this invention addresses the inherent shortcomings of traditional control methods in terms of prediction accuracy, adaptability, safety, and real-time performance through the deep integration of digital twin construction, online adaptive correction, safety constraint optimization, and cloud-edge collaborative architecture, forming a complete intelligent control closed loop. This solution not only improves the instantaneous operating energy efficiency of heat exchangers but also achieves continuous compensation and optimization for equipment performance degradation, providing a reliable and efficient technical path for energy saving and consumption reduction throughout the entire life cycle of industrial heat exchange equipment. Attached Figure Description

[0042] Figure 1 This is a flowchart of the steps of a digital twin energy efficiency prediction and control method for a shell-and-tube heat exchanger provided in Embodiment 1 of the present invention;

[0043] Figure 2 This is a flowchart illustrating the architecture of a shell-and-tube heat exchanger provided in Embodiment 2 of the present invention. Detailed Implementation

[0044] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description relating to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and systems consistent with some aspects of this application as detailed in the appended claims.

[0045] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0046] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0047] Example 1

[0048] Please see Figure 1 A digital twin energy efficiency prediction and control method for shell-and-tube heat exchangers includes the following steps:

[0049] S1, Construct a digital twin of a shell-and-tube heat exchanger, the digital twin including a mechanism-data fusion channel and an energy efficiency time series prediction sub-model;

[0050] Among them, the mechanism channel calculates the baseline EER based on the real-time heat transfer coefficient and pump power, while the data channel outputs the fouling-aging deviation that is not captured by the mechanism through the residual network. The two channels are added together to form the final EER prediction value.

[0051] The mechanism channel in the mechanism-data fusion channel receives real-time collected signals such as inlet temperature, outlet temperature, instantaneous flow rate, and actual power consumed by the pump from the pipe side and shell side as input;

[0052] Through the physical equation Q=K*A*ΔT m Calculate the heat exchange rate Q;

[0053] Among them, the overall heat transfer coefficient K is a key state variable characterizing the performance of the heat exchanger. Its value changes with the degree of fouling and equipment aging. During the initialization of the digital twin, K can use the design value or an empirical value under clean conditions. During operation, K will be adjusted according to the fouling thermal resistance R. f The online updated value is corrected in real time, and the calculation of K is related to R. f This is directly related, thus enabling the mechanism model to track the actual performance status of the device;

[0054] The total heat transfer area A is a fixed parameter determined by the design geometry of the heat exchanger, such as the number, diameter, and length of the heat exchange tubes. It is an inherent constant of the device that is input when the digital twin is constructed.

[0055] Logarithmic mean temperature difference ΔT m This is a real-time calculated value, which is calculated based on the collected inlet and outlet temperatures of the tube side and shell side, according to the heat transfer formula. The calculation formula is different for counter-current, co-current or other flow modes, and needs to be preset in the system.

[0056] This leads to the output of the benchmark energy efficiency ratio (EER). mech =Q / W pump Among them, W pump The mechanism channel receives the actual power consumed by the pump in real time, and the mechanism channel ensures that the predicted output conforms to the basic laws of heat transfer and energy conservation.

[0057] The data channel in the mechanism-data fusion channel consists of a lightweight Transformer neural network model. In addition to real-time signals, its inputs include longer-term operational data and high-dimensional nonlinear features. The network is responsible for learning complex dynamic characteristics, local flow effects, and deviations that are difficult to measure in real time, such as slight leakage and sensor system errors, which are not included in the mechanism model, and outputs a deviation value ΔEER.

[0058] Finally, the outputs of the two channels are added together to form the final predicted instantaneous energy efficiency ratio (EER) = EER mech +ΔEER, this structure utilizes the generalization and interpretability of mechanistic models to capture unmodeled physical phenomena through data-driven models, thereby improving prediction accuracy;

[0059] The energy efficiency time series prediction sub-model is jointly trained by the Transformer neural network model and the heat transfer-pump work physical equation. It is used to output the instantaneous energy efficiency ratio (EER) prediction curve for the next 5–15 minutes. The joint training adopts a physical loss regularization term to ensure that the output of the Transformer neural network model satisfies the basic laws of energy conservation and heat transfer.

[0060] Defining the time window for the energy efficiency prediction curve as 5 to 15 minutes is an optimal engineering solution determined based on the dynamic characteristics and control requirements of the heat exchanger. The main basis for this is as follows:

[0061] From the adjustment of the actuator to the heat exchanger outlet temperature and system energy efficiency reaching a new steady state, the entire process of shell-and-tube heat exchangers exhibits significant pure time lag and capacity lag. Its time constant is usually on the order of several minutes to tens of minutes. A prediction window of 5-15 minutes is sufficient to cover the dynamic response process of a complete adjustment, enabling the predictive control strategy to play an effective role.

[0062] The main disturbances addressed by this invention, such as changes in ambient temperature and adjustments to production process loads, occur and have an impact at a speed on the order of minutes. Predictions shorter than 5 minutes cannot provide sufficient advance adjustment time; while predictions longer than 15 minutes will significantly reduce prediction accuracy due to greatly increased uncertainty, thus losing their optimization guidance significance.

[0063] Predictive control requires rolling optimization solutions. The prediction step size directly determines the scale of the optimization problem. A step size of 5-15 minutes, combined with a control cycle of 1 minute, means 5-15 prediction steps. This ensures that the optimization problem can be solved within a time much shorter than the sampling period under the limited computing resources of edge computing hardware, meeting the hard real-time requirements of industrial applications.

[0064] S2, with the optimization objective of maximizing the integral of the EER prediction curve over time, the optimal control sequence of pump frequency and regulating valve opening is obtained by rolling solution;

[0065] S3, before the actual load change occurs, sends the optimal control sequence to the actuator;

[0066] Specifically, in each control cycle, such as 1 minute, the digital twin executes the following process: First, it calls the energy efficiency time-series prediction sub-model to generate a series of EER prediction values ​​for a future prediction time domain, i.e., within 5-15 minutes, forming an EER prediction curve; then, using the integral of this curve over time, i.e., maximizing total energy efficiency, as the optimization objective, it solves for an optimal control sequence (pump frequency and valve opening); finally, based on the prediction of future energy efficiency, it sends the optimal control sequence to the actuator to take optimization actions in advance, rather than responding after the fact.

[0067] Furthermore, the joint training of the Transformer neural network model and the heat transfer-pump work physical equation adopts an end-to-end training strategy. Its training loss function is calculated based on the error between the final output EER prediction value and the true value, and the physical loss regularization term is fused for backpropagation.

[0068] The energy efficiency time-series prediction sub-model, comprising a Transformer network and subsequent physical computation layers, is constructed as a differentiable computational graph. During training, the calculation of the training loss function directly depends on the final output of the entire model, i.e., the predicted EER value. The training loss function consists of two parts:

[0069] Total loss = Data loss (EER predicted value, EER measured value) + λ * Physical loss regularization term

[0070] The total loss is the mean square error loss between the predicted EER value and the measured EER value. The measured EER value is directly calculated from the temperature, flow rate and power data measured by the field sensors.

[0071] The physical loss canonical phase is a physical consistency constraint term, which is constructed by calculating the degree of violation (residual) of the model predictions to fundamental physical laws, such as energy conservation. For example, it can be expressed as the difference norm between the heat predicted by the model and the heat calculated by independent measurements.

[0072] The hyperparameter λ is used to balance the weights of the two terms, and to balance the data fitting accuracy and physical consistency.

[0073] During training, optimization algorithms, such as Adam, calculate the gradient based on this total loss and use the backpropagation algorithm to make the gradient run through the entire differentiable computation graph, including the data-driven Transformer part and the physical equation part of the mechanism calculation, while updating all parameters of the Transformer neural network, such as weights and biases.

[0074] Specifically, the end-to-end training strategy avoids the error accumulation and suboptimal solution problems that may be caused by traditional staged training, enabling all components to work together to optimize the final EER prediction accuracy. Through backpropagation, the physical loss regularization term reflects the influence on the final output, directly guiding the feature extraction process of the front-end Transformer neural network, so that the feature vectors it learns themselves conform to physical laws, thereby enhancing the model's machine learning algorithm's ability to adapt to new samples.

[0075] Furthermore, when the absolute value of the deviation between the final predicted EER and the measured EER over N consecutive sampling points exceeds a set threshold δ, online correction is automatically triggered.

[0076] The fouling thermal resistance R was determined using the sliding window recursive least squares method. f The pump efficiency η is updated in real time, and the updated parameters are immediately fed into the mechanism channel.

[0077] The final EER prediction value output by the digital twin is continuously compared with the measured EER value calculated directly from the sensor data.

[0078] When the condition that the predicted EER value - measured EER value > threshold δ is met for N consecutive sampling periods, i.e. N consecutive time points, the system determines that the current mechanism model (total heat transfer coefficient K, pump efficiency η) has significantly mismatched and automatically triggers the online correction process.

[0079] Once triggered, the system calls the sliding window recursive least squares method to identify the key state parameters that cause the deviation;

[0080] Recursive least squares does not require storing historical data. It can recursively calculate the parameter value at the current time using only the parameter estimate from the previous time step and the new observation data. It has a small computational load and is suitable for real-time operation.

[0081] To further enhance the algorithm's ability to track time-varying systems, this scheme employs a sliding window approach. It uses a recursive least squares method with a forgetting factor, or only uses the most recent fixed window size, such as data from the last 30 minutes, for calculation. This ensures the algorithm discards outdated data and closely tracks the system's current state, thus controlling the fouling thermal resistance R. f It is highly sensitive to changes in pump efficiency η;

[0082] Fouling thermal resistance R f A physical quantity characterizing the degree of fouling on the wall of a heat exchanger tube; fouling thermal resistance R. f Increasing the K value leads to a decrease in the overall heat transfer coefficient. This is a time-varying variable that cannot be directly measured and is a key point for identification.

[0083] Pump efficiency η is a physical quantity that characterizes the performance state of the pump itself. It decreases over time due to wear, cavitation and other reasons, and also needs to be identified online.

[0084] Implementation process:

[0085] The parameter to be estimated, the fouling thermal resistance R, will be included. f The physical model of pump efficiency η is linearized near the current operating point;

[0086] Fouling thermal resistance R f The physical model is as follows:

[0087] Basic heat transfer equation: Q = K * A * ΔT m ;

[0088] Overall heat transfer coefficient K and fouling thermal resistance R f Relationship:

[0089] The reciprocal of the overall heat transfer coefficient K (total thermal resistance) is equal to the sum of all individual thermal resistances. For a shell-and-tube heat exchanger, the functional relationship is K = f(R). f );

[0090]

[0091] Where: h shell This is the shell-side convective heat transfer coefficient, which is related to the shell-side flow rate and physical properties.

[0092] δ wall λ is the pipe wall thickness. wall h is the thermal conductivity of the pipe wall material. tube The tube-side convective heat transfer coefficient is related to the tube-side flow rate and physical properties; A is the total heat transfer area. tube This refers to the inner surface area of ​​the pipe.

[0093] Therefore, the complete functional relationship of the heat exchange Q can be expressed as:

[0094]

[0095] One of the goals of the online calibration algorithm, RLS algorithm, is to find the R value that best matches the measured Q with the model's calculated Q. f Value;

[0096] The physical model for pump efficiency η is as follows:

[0097] Pump power W pump The specific form of the functional relationship:

[0098] The basic power equation of a pump:

[0099]

[0100] Where ρ is the fluid density, which can be obtained from a table by looking up the temperature, g is the acceleration due to gravity, and Q is the fluid density. flow The volumetric flow rate is measured by the flow meter.

[0101] H is the pump head. For a given pump, at a specific speed (frequency), the head H and flow rate Q are related. flow There is a fixed relationship between them, namely, the pump characteristic curve H=H(Q) flow );

[0102] Pump power W pump The complete functional relationship can be expressed as:

[0103]

[0104] Another goal of online calibration algorithms is to find W that enables the model to calculate. pump Compared with the measured W pump The best-fit pump efficiency η value;

[0105] The recursive least squares method linearizes the above nonlinear equation at the current operating point, forming a linear form such as y=φ*θ;

[0106] y represents the observed value. The measured Q or W... pump Deviation from the model's calculated value;

[0107] φ is the regression vector, which consists of the known variables and coefficients;

[0108] θ is the parameter vector to be estimated, i.e., [R] f ,η] or their increments;

[0109] The algorithm continuously updates the estimated value of θ through recursive calculation, thereby achieving a solution for R. f Real-time tracking of η;

[0110] Updated R f The new estimates of η are immediately fed into the mechanistic channel of the digital twin to update the overall heat transfer coefficient K and pump work W. pump The calculation formula enables the mechanism model to adaptively approximate the actual current state of the heat exchanger, providing data for accurate prediction in the next round.

[0111] Where N is an integer from 3 to 10, and δ is 0.05 to 0.15;

[0112] Lower limit (N≥3): Setting it to at least 3 cycles is to prevent false triggering caused by noise or instantaneous disturbances. Deviation of a single or multiple sampling points may be due to measurement noise or instantaneous operating condition fluctuations, not model mismatch. Continuous deviation of more than 3 consecutive points can more reliably indicate the existence of systematic deviation, improving the anti-interference and reliability of triggering.

[0113] The upper limit (N≤10) is to ensure the system's response speed. If N is set too large, even if the model is mismatched, the system will need to wait a long time before it starts to correct, which will cause the control optimization effect to decrease during this period. N≤10 ensures that when the model has a problem, the system can respond quickly within a few minutes to a dozen minutes (assuming the sampling period is 1-2 minutes), maintaining the timeliness of the model.

[0114] Lower limit (δ≥0.05): Considering the measurement accuracy of the sensor itself and the unavoidable random error, a very small deviation (such as 0.01) may not be of practical significance. Setting the threshold to 0.05 relative error can filter out insignificant errors within the allowable range and avoid unnecessary parameter updates.

[0115] Upper limit (δ≤0.15): If the threshold is set too high, the system will be too insensitive to performance degradation. For example, if the EER has dropped from 3.0 to 2.6, with a deviation of more than 13%, and δ is set to 0.2, the system will still not start correction. This means that the system will run in an inefficient state for too long and lose energy. δ≤0.15 ensures that when energy efficiency degradation with practical engineering significance occurs, the system can capture and deal with it in time.

[0116] Specifically, the correction process is intelligently triggered by continuously monitoring the continuous deviation between the predicted and measured EER; a sliding window recursive least squares method is used to address the key internal state parameters (fouling thermal resistance R) that cause model mismatch. f The system tracks and updates the pump efficiency (η) in real time and feeds the updated parameters back to the mechanism model immediately, enabling the digital twin to overcome performance degradation caused by equipment aging and scaling, maintain high fidelity, and provide reliable assurance for upper-level predictive optimization control.

[0117] Furthermore, the online correction step employs a lifelong learning strategy with a forgetting factor λ∈(0.95,0.99), and λ is adaptively adjusted with the dirt growth rate to further suppress the cumulative interference of historical dirty data on the current parameters;

[0118] Based on the sliding window recursive least squares method, a forgetting factor λ is introduced to form a recursive least squares method with a forgetting factor FF-RLS, so as to realize lifelong learning.

[0119] The forgetting factor λ is used to exponentially reduce the weight of historical data in the current parameter estimation. Its value determines the storage length of the algorithm and its sensitivity to new data.

[0120] Lower limit (λ>0.95): The closer λ is to 1, the longer the algorithm stores historical data, and the smoother and more stable the estimation results. If λ is too low, historical data will be discarded too quickly, causing the algorithm to be too sensitive to the latest data, the estimation results to fluctuate wildly, and it will be easily interfered with by measurement noise and lose reliability. The lower limit is set to 0.95 to ensure that the system has sufficient memory depth to maintain the stability of parameter estimation.

[0121] Upper limit (λ<0.99): The closer λ is to 1, the slower the algorithm stores old data and the worse its ability to track parameter changes. For slow but continuous processes such as scaling and aging, an excessively high λ will make the algorithm too lazy and unable to track R in a timely manner. f The upper limit was set to 0.99 to reflect the actual change in η, ensuring that the algorithm has the necessary responsiveness to track time-varying parameters.

[0122] Engineering optimal trade-off: The range (0.95, 0.99) is an empirically optimal range in a large number of engineering practices. It strikes a balance between the stability of the estimate (noise resistance) and the speed of tracking (agility), and is very suitable for slowly changing industrial processes such as heat exchanger fouling and pump efficiency degradation.

[0123] Adaptive adjustment strategy for the forgetting factor λ:

[0124] The forgetting factor λ adaptively adjusts with the dirt growth rate, thereby achieving intelligent lifelong learning;

[0125] The system calculates the fouling thermal resistance R in real time. f The growth rate γ, used as an indicator to assess the drasticness of changes in the system, can be calculated using the following formula:

[0126]

[0127] That is, R within a unit time t f The relative change;

[0128] When the fouling growth rate γ is high, it indicates that the fouling process is accelerating, such as due to changes in fluid composition or adjustments in operating conditions. In this case, the value of the forgetting factor λ should be lowered so that the algorithm can discard old data from past clean states more quickly and give higher weight to new data, thereby speeding up the algorithm's tracking of the current rapid changing trend.

[0129] When the fouling growth rate γ is low, it indicates that the fouling process is slow and steady. The value of the forgetting factor λ should be increased to enhance the storage length of the algorithm and use longer historical data to smooth out the interference of random measurement noise, thereby obtaining more stable and reliable parameter estimates.

[0130] A mapping function or lookup table for λ=f(γ) is pre-defined. This function is a monotonically decreasing function. In each sampling period, the system dynamically queries or calculates a new value of λ based on the calculated γ, and uses it for the next step of recursive least squares FF-RLS calculation with forgetting factor.

[0131] Specifically, the online correction function is enhanced by introducing a forgetting factor λ that can be dynamically and adaptively adjusted within the range of (0.95, 0.99). Based on the rate of dirt growth, it automatically and smoothly switches between two modes: increasing storage length to resist interference and discarding data to accelerate change. This effectively suppresses the cumulative interference of historical dirty data or outliers on the current parameter estimation, improves the adaptability and robustness of the digital twin to different change rates throughout its life cycle, and ensures the accuracy and reliability of online parameter correction.

[0132] Furthermore, after obtaining the optimal control sequence, a safety-execution mapping layer is added to verify the pump frequency, valve opening command, net positive suction head, maximum head, and upper limit of shell-side-tube-side pressure difference in real time.

[0133] Before the optimal control sequence (pump frequency, regulating valve opening) obtained from the optimization calculation is sent to the actuator, it must be verified in real time by an independent functional module of the safety-execution mapping layer. The safety-execution mapping layer is the key to connecting the high-performance optimization algorithm with the actual physical equipment, ensuring that all control commands are strictly operated within the safe operating range allowed by the equipment, and preventing the pursuit of optimal energy efficiency at the expense of equipment safety.

[0134] The security constraints upon which the security-execution mapping layer verifies control commands are mainly divided into three categories:

[0135] The required net positive suction head (NPSH) constraint is an inherent property of the pump itself, provided by the pump manufacturer, and is usually given in the form of a curve (required NPSH vs. flow rate), which can be pre-stored in the control system.

[0136] The system calculates the required net positive suction head (NPSH) of the pump under the current operating conditions based on the pump frequency and predicted flow rate in the optimization instructions. At the same time, it calculates the effective NPSH of the device based on real-time measured parameters such as inlet pressure, temperature, and liquid level. The safety condition must be met: required NPSH > required NPSH + safety margin. The safety margin is a pre-set positive constant, usually taken as 0.5 to 1.5 meters of water column.

[0137] If the pump frequency command would cause this condition to be unmet, it would be considered an out-of-bounds error.

[0138] The maximum head constraint is a design upper limit determined by the pump's mechanical structure, such as casing strength, shaft power, and maximum motor torque, and is either a fixed constant or a curve related to the speed.

[0139] Based on the optimized pump frequency and flow rate, the head that the command will generate can be queried or calculated through the pump characteristic curve (HQ curve). The safety condition must be met: the generated head < the maximum head.

[0140] If the limits are exceeded, it means that the pump may be operating under overload, which poses a risk of damage.

[0141] The maximum permissible pressure difference between the shell and tube sides is a fixed upper limit determined by the mechanical design pressure of the heat exchanger body, the flange class, and the pressure bearing capacity of the piping system.

[0142] Based on the optimized valve opening and flow rate, the pressure difference that the command will generate between the shell side and the tube side of the heat exchanger is estimated by means of the valve flow characteristic curve or fluid dynamics formula. The safety condition must be met: the pressure difference generated between the shell side and the tube side of the heat exchanger < the maximum allowable pressure difference between the shell side and the tube side.

[0143] If the limits are exceeded, it may lead to internal leakage or structural damage to the heat exchanger;

[0144] If any constraint goes out of bounds, a second planning process is immediately initiated to solve the problem again, in order to balance optimal energy efficiency and equipment safety.

[0145] The safety-execution mapping layer performs parallel verification of the above three constraints. If any constraint condition goes out of bounds, the original control command is immediately blocked and the re-solution process is triggered.

[0146] Resolve Algorithm - Quadratic Programming:

[0147] The original optimization problem, maximizing the EER integral, is transformed into a constrained quadratic programming problem.

[0148] To ensure energy efficiency is as close as possible to the original optimal control sequence;

[0149] The above three safety constraints and the physical upper and lower limits of the actuator are all transformed into linear or quadratic inequality constraints of decision variables (pump frequency, valve opening), forming a new and strict feasible region.

[0150] The problem can be solved in real time by calling efficient QP solvers, such as OSQP and ActiveSet methods. Since there are only two decision variables, the solution can be completed in milliseconds.

[0151] Specifically, by adding a safety-execution mapping layer, the three safety defenses are checked in real time before the control command is executed. Once a risk of exceeding the limit is detected, a re-solution mechanism based on quadratic programming is immediately activated. The safety problem is treated as a hard constraint to recalculate a new control command that is safe and closest to the original optimal solution, thereby resolving the potential contradiction between energy efficiency optimization and equipment safety and achieving optimal energy efficiency under the premise of safety.

[0152] Furthermore, the time required to resolve the quadratic programming problem is ≤100ms;

[0153] The time required for resolving quadratic programming is determined by the sampling period and system response time of industrial process control.

[0154] The rolling optimization cycle of predictive control is usually 1 minute (60 seconds) or less. As part of the optimization process, the time taken by the re-solution of the quadratic programming must be much less than that of the main control cycle. 100ms accounts for 0.17% of the 1-minute cycle, which means that the re-solution process hardly occupies the computing resources of the main cycle, leaving enough time for other tasks such as data communication and model prediction, and ensuring the stability and reliability of the real-time operation of the entire edge computing unit.

[0155] There is an inherent inertial delay between the issuance of control commands and the physical system (such as flow rate and temperature) starting to respond. A solution speed of 100ms means that the system can complete safety checks and command corrections before the physical process changes significantly, thus suppressing potential safety hazards in time and ensuring the responsiveness of the control system.

[0156] Current mainstream edge computing hardware, such as high-performance ARM processors, industrial-grade GPUs, and efficient QP solver libraries, such as OSQP, are fully capable of solving optimization problems with few variables and constraints within milliseconds. 100ms is a reasonable and achievable balance between high performance requirements and current hardware capabilities.

[0157] After the solution is completed, the edge computing unit sends the pump frequency and valve opening setpoints to the frequency converter and electric regulating valve via OPC UA or Modbus TCP protocol;

[0158] Edge computing units, acting as the execution entities, meet real-time requirements and avoid control lag caused by network latency.

[0159] OPC UA is the preferred protocol for modern industrial communication, with advantages such as cross-platform compatibility, high security, and information modeling, making it suitable for complex and highly reliable application scenarios.

[0160] Modbus TCP is the most widely used and simple protocol in the industrial field. It is highly compatible and suitable for connecting most existing frequency converters and electric valves.

[0161] Supporting both protocols simultaneously demonstrates a thorough consideration of compatibility with different customer field equipment;

[0162] The instructions are sent directly to the frequency converter (to control the pump frequency) and the electric regulating valve (to control the opening degree) in the form of set values.

[0163] After the signal is sent, the actual feedback signal is continuously monitored. If the deviation between the actual value and the set value exceeds the allowable hysteresis and lasts for two sampling cycles, a recalculation is triggered to maintain control accuracy.

[0164] After the command is issued, the edge computing unit does not consider the task to be over, but continues to monitor the actual status signals returned from the frequency converter and regulating valve, such as the actual pump frequency and the actual valve position.

[0165] If the absolute value of the deviation between the actual returned value and the issued set value exceeds the allowable hysteresis, which is a threshold preset based on the accuracy of the actuator, such as pump frequency deviation > 0.5Hz or valve position deviation > 2%;

[0166] The above deviation state lasts for 2 sampling periods. The purpose of adding the duration judgment is to filter out instantaneous jitter or communication interference, prevent unnecessary recalculation, and improve the anti-interference capability of the system.

[0167] Once the triggering conditions are met, the system determines that the actuator may be malfunctioning or there may be unknown interference, causing the control objective to fail to be achieved. At this time, a new round of predictive control recalculation will be automatically triggered to regenerate the optimal control sequence based on the latest and real system state, so as to ultimately ensure control accuracy.

[0168] Specifically, a highly reliable, hard real-time closed-loop control link is constructed from three dimensions: computing performance, industrial communication, and execution feedback. By limiting the re-solution time, the real-time performance of safe control is ensured; by adopting standard industrial protocols, the compatibility of system integration is guaranteed; and finally, by implementing deviation monitoring and recalculation triggering mechanisms, the last line of defense for control effectiveness is achieved, giving the system the ability to detect anomalies and self-correct.

[0169] Furthermore, the digital twin adopts a two-layer architecture consisting of a high-frequency closed-loop control layer and a low-frequency state evaluation layer:

[0170] The high-frequency closed-loop control layer refreshes every 1 minute to complete minute-level EER prediction and control sequence generation;

[0171] This cycle matches the typical advanced control cycle of industrial process control systems (DCS / PLC), ensuring real-time control.

[0172] Receive real-time sensor data, run the "Energy Efficiency Time Series Prediction Sub-model", and output the instantaneous energy efficiency ratio (EER) prediction curve for the next 5-15 minutes in a rolling manner.

[0173] Using the predicted curve as input, the model predictive control optimization algorithm is executed to solve for the optimal pump frequency and valve opening control sequence in a rolling manner.

[0174] It performs safety checks, re-solution, and command issuance functions to complete high-frequency closed-loop control;

[0175] The high-frequency closed-loop control layer focuses on dynamic process response and high-frequency optimization at the second / minute level;

[0176] The low-frequency condition assessment layer is refreshed every 24 hours to assess the scaling and aging conditions and inject the aging condition quantity into the high-frequency closed-loop control layer to achieve multi-scale parameter cascade correction.

[0177] The low-frequency condition assessment layer selects one day as the cycle because equipment performance degradation, such as scaling and aging, is a slow and long-term process, and its changes cannot be significantly observed on a minute scale.

[0178] Collect and aggregate all system operating data (temperature, pressure, flow, power, control commands, etc.) from the past 24 hours to form a rich dataset. Utilize this sufficiently long historical data for in-depth offline analysis and data mining.

[0179] Through trend analysis and advanced parameter identification algorithms, the fouling thermal resistance R can be evaluated more accurately. f The long-term trend and current level;

[0180] Analyze slowly varying parameters related to equipment lifespan, such as assessing the long-term decay curve of pump efficiency η, or identifying the degree of slow degradation of the overall heat transfer performance of the heat exchanger.

[0181] These parameters, which represent the long-term state, can be regarded as aging state quantities and injected unidirectionally and from top to bottom into the mechanism channel of the digital twin of the high-frequency closed-loop control layer.

[0182] The low-frequency state assessment layer evaluates the more accurate R. f The latest values ​​of parameters such as η cover the corresponding parameter values ​​originally in the high-frequency closed-loop control layer model;

[0183] Subsequently, when the high-frequency closed-loop control layer performs the next prediction and optimization, its mechanism model will be calculated based on these parameters provided by the low-frequency state assessment layer, which are closer to the long-term real state of the equipment.

[0184] To achieve cascaded correction of parameters in the high-frequency closed-loop control layer model and make control decisions that better align with the optimal goals throughout the equipment's lifecycle;

[0185] Specifically, by innovatively designing a two-layer architecture, the contradiction of computing resources between high-frequency real-time control and low-frequency fine-grained state evaluation in industrial digital twins is cleverly resolved. The high-frequency closed-loop control layer ensures the immediacy of control, while the low-frequency state evaluation layer ensures the accuracy of the state. Through a 24-hour cycle evaluation-injection mechanism, the low-frequency state evaluation layer continuously calibrates the high-frequency closed-loop control layer, enabling the system to not only see the current operating condition fluctuations but also perceive the performance degradation trend on a daily basis, ultimately achieving cascaded optimized control.

[0186] Furthermore, the two-layer architecture is deployed in an edge-cloud collaborative environment, where the low-frequency state evaluation layer is trained with big data in the cloud and the high-frequency closed-loop control layer performs real-time inference at the edge. The two layers communicate with each other through an event-driven interface to push state quantities in one direction.

[0187] Define the physical deployment method of the two-tier architecture, adopt the edge-cloud collaborative computing paradigm, and allocate tasks with different computing loads to appropriate computing layers;

[0188] The high-frequency closed-loop control layer is deployed at the edge, specifically at the edge computing unit. The edge computing unit is typically located in the industrial field, close to heat exchangers, sensors, and actuators (frequency converters, electric valves).

[0189] Responsible for running all high real-time tasks, including:

[0190] Real-time data acquisition and preprocessing;

[0191] Minute-level EER prediction (energy efficiency time series prediction sub-model inference);

[0192] Model predictive control optimization solution;

[0193] Security verification and quadratic programming re-solution;

[0194] Control command issuance and execution status monitoring;

[0195] Deploying the control closed loop at the edge ensures low control latency, meets real-time requirements of ≤100ms, high reliability, independent operation during network interruption, and data security, ensuring that raw data does not leave the factory area.

[0196] The low-frequency state assessment layer is deployed in the cloud, specifically in a cloud server cluster, which has scalable computing and storage resources.

[0197] Responsible for running low-frequency, computationally intensive deep analysis tasks, including:

[0198] Aggregate and store large amounts of heat exchanger operation data from multiple edge devices, covering the entire plant or even multiple factories;

[0199] Offline, batch processing-based deep mining and training of massive historical data is performed on a 24-hour cycle.

[0200] Execute more complex and accurate long-term condition assessment algorithms, such as advanced time series analysis and deep learning models, to accurately assess scaling and aging trends;

[0201] Make full use of cloud computing power to train complex models and obtain more accurate state assessment results than edge analysis of a single device based on a larger dataset;

[0202] The interaction between the two layers is not a periodic request-response, but rather adopts an efficient event-driven pattern;

[0203] The low-frequency state assessment layer completes the 24-hour cycle analysis task and generates new, usable aging state quantities, such as the updated R. f The event is triggered when the long-term trend value of η is obtained.

[0204] After the event is triggered, the cloud-based low-frequency state assessment layer proactively sends new state data packets to the designated edge high-frequency closed-loop control layer through one-way state quantity push.

[0205] A lightweight message queue protocol is typically used, which is suitable for asynchronous, one-to-many message pushing in IoT scenarios.

[0206] The pushed data packets only contain the evaluated result parameters, rather than the original massive amount of runtime data, thus reducing the communication bandwidth requirements.

[0207] The communication direction is strictly unidirectional, from the cloud to the edge. The high-frequency closed-loop control layer at the edge only receives instructions and data from the cloud, without needing to actively request data from the cloud, which simplifies the logic at the edge and reduces resource consumption.

[0208] Specifically, by deploying the low-frequency state assessment layer in the cloud and the high-frequency closed-loop control layer at the edge, computing resources are rationally allocated, balancing millisecond-level real-time control response with day-level big data in-depth analysis. The two layers collaborate through an event-driven interface and a lightweight one-way push mechanism, enabling the macro-level decision-making of the low-frequency state assessment layer to efficiently and securely empower the micro-level control of the high-frequency closed-loop control layer.

[0209] Example 2

[0210] Please see Figure 2 A shell-and-tube heat exchanger includes: a variable frequency circulating pump, an electric regulating valve, a data acquisition unit, and an edge computing unit;

[0211] The variable frequency circulating pump drive motor is connected to the frequency converter to receive speed commands and change the fluid flow rate. It is the main regulating actuator.

[0212] Electric regulating valves are installed in pipelines to regulate flow resistance or flow distribution; they are actuators.

[0213] The edge computing unit is a dedicated computer (such as an industrial PC or high-performance embedded controller) deployed in an industrial site. The edge computing unit is connected to the frequency converter and electric regulating valve via a hard real-time EtherCAT bus and carries executable code for a digital twin energy efficiency predictive control method for shell-and-tube heat exchangers. The configuration is as follows:

[0214] The data acquisition unit typically consists of a distributed control system (DCS) or a programmable logic controller (PLC) and various high-precision sensors (temperature, pressure, flow, power signals), and is responsible for real-time acquisition and uploading of all operating status signals of the system.

[0215] The digital twin energy efficiency prediction control method for shell-and-tube heat exchangers generates pump frequency and valve opening control commands.

[0216] Commands are sent directly to the variable frequency circulating pump and electric regulating valve via the EtherCAT bus to achieve closed-loop control that maximizes the energy efficiency of the heat exchanger.

[0217] The EtherCAT bus uses a master-slave architecture and data processing mechanism, which has very low communication latency, such as 1ms, and high time determinism. It can meet the stringent real-time requirements of ≤100ms for resolving quadratic programming, ensuring the timely and accurate issuance of control commands.

[0218] All connected slave devices (frequency converters, valves) can use a distributed clock to achieve high-precision synchronization at the sub-microsecond level;

[0219] Based on Ethernet, it supports linear, tree and other topologies, simplifies cabling, and eliminates the need for expensive dedicated switches;

[0220] The workflow of an edge computing unit:

[0221] The edge computing unit is configured to execute a continuous, automated closed loop;

[0222] Through its communication interfaces, such as Modbus TCP and OPC UA, it receives all sensor signals, such as temperature, flow rate, and power, uploaded from the data acquisition unit (DCS / PLC) in real time;

[0223] The system implements a complete set of digital twin energy efficiency prediction and control methods, which uses digital twins to predict energy efficiency, perform rolling optimization solutions, and conduct safety verification, ultimately generating pump frequency setpoints and valve opening setpoints.

[0224] Instead of using the traditional DCS system for roundabout control, the generated control commands are directly sent to the inverter and electric regulating valve drivers in milliseconds and deterministically in the form of process data objects via the hard real-time EtherCAT bus.

[0225] After receiving the command, the frequency converter and electric valve immediately execute it, changing the flow rate and thus affecting the operating efficiency of the heat exchanger. The new state is then collected by the sensor, and so on, ultimately achieving closed-loop control to maximize the energy efficiency of the heat exchanger.

[0226] Specifically, by integrating innovative intelligent algorithms (digital twins, predictive control, online correction) with industrial hardware such as edge computing units, EtherCAT buses, and variable frequency drives, the problem of traditional optimization control algorithms being significantly less effective due to execution link delays and uncertainties is solved.

Claims

1. A digital twin energy efficiency prediction and control method for a shell-and-tube heat exchanger, characterized in that, The method comprises the following steps: S1, constructing a digital twin of a shell-and-tube heat exchanger, the digital twin comprising a mechanism-data fusion channel and an energy efficiency time sequence prediction submodel; wherein the mechanism channel calculates a benchmark EER according to a real-time heat transfer coefficient and pump power, the data channel outputs a fouling-aging deviation not captured by the mechanism through a residual network, and the two channels are added to form a final EER prediction value; the energy efficiency time sequence prediction submodel is jointly trained by a Transformer neural network model and a heat-pump power physical equation, and is used to output an instantaneous energy efficiency ratio EER prediction curve in the next 5-15 minutes, and the joint training adopts a physical loss regularization term, so that the output of the Transformer neural network model satisfies the law of conservation of energy and the basic law of heat transfer; S2, taking the integral of the EER prediction curve with respect to time as an optimization objective, and rolling to obtain an optimal control sequence of pump frequency and regulating valve opening degree; S3, before the actual load change occurs, the optimal control sequence is sent to the actuator.

2. The digital twin energy efficiency predictive control method of a shell-and-tube heat exchanger of claim 1, wherein, The joint training of the Transformer neural network model and the heat-pump power physical equation adopts an end-to-end training strategy, the training loss function is calculated based on the error between the final output EER prediction value and the true value, and the physical loss regularization term is used for back propagation.

3. The digital twin energy efficiency predictive control method of a shell-and-tube heat exchanger of claim 2, wherein, When the absolute value of the deviation of the final EER prediction value and the EER measured value at N consecutive sampling points is greater than a set threshold δ, online correction is automatically triggered: The fouling thermal resistance Rf and pump efficiency η are updated in real time by using a sliding window recursive least squares method, and the updated parameters are fed into the mechanism channel in real time; wherein N is an integer between 3 and 10, and δ is between 0.05 and 0.

15.

4. The digital twin energy efficiency predictive control method of a shell-and-tube heat exchanger of claim 3, wherein, The online correction step adopts a lifelong learning strategy with a forgetting factor, the forgetting factor λ is between 0.95 and 0.99, and the λ is adaptively adjusted according to the fouling growth rate to further suppress the cumulative interference of historical dirty data on the current parameters.

5. The digital twin energy efficiency predictive control method of a shell-and-tube heat exchanger of claim 1, wherein, After obtaining the optimal control sequence, a safety-execution mapping layer is added to real-time check the pump frequency, valve opening degree instruction, and the upper limit of the net positive suction head, maximum head and shell-side-tube-side pressure difference; If any constraint is out of range, a quadratic programming re-solution is immediately started to balance the energy efficiency optimization and equipment safety.

6. The digital twin energy efficiency predictive control method of a shell-and-tube heat exchanger of claim 5, wherein, The quadratic programming re-solution takes less than 100 ms; After the solution is completed, the edge computing unit sends the pump frequency and valve opening degree set values to the frequency converter and electric regulating valve through OPC UA or Modbus TCP protocol; After sending, the actual return signal is continuously monitored, and if the deviation between the actual value and the set value exceeds the allowed return difference and lasts for 2 sampling periods, a re-calculation is triggered to maintain control accuracy.

7. The digital twin energy efficiency predictive control method of a shell-and-tube heat exchanger of claim 6, wherein, The digital twin uses a high-frequency closed-loop control layer and a low-frequency state evaluation layer double-layer architecture: The high-frequency closed-loop control layer is refreshed at a 1-minute cycle to complete the minute-level EER prediction and control sequence generation; The low-frequency state evaluation layer is refreshed at a 24-hour cycle to evaluate the fouling and aging states and inject the aging state quantity into the high-frequency closed-loop control layer to realize multi-scale parameter cascade correction.

8. The digital twin energy efficiency predictive control method of a shell-and-tube heat exchanger of claim 7, wherein, The double-layer architecture is deployed in an edge-cloud collaborative environment, wherein the low-frequency state evaluation layer performs big data training in the cloud, the high-frequency closed-loop control layer performs real-time inference at the edge, and one-way state quantity pushing is completed between the two layers through an event-driven interface.

9. A shell and tube heat exchanger comprising: A variable frequency circulating pump, an electric regulating valve, a data acquisition unit, and an edge computing unit; The edge computing unit is connected with the frequency converter and the electric regulating valve through a hard real-time EtherCAT bus, and carries an executable code of the digital twin energy efficiency prediction control method according to any one of claims 1 to 8, characterized in that The executable code specifically receives temperature, flow rate, and power signals uploaded by the data acquisition unit in real time; The digital twin energy efficiency prediction control method of the shell-and-tube heat exchanger generates pump frequency and valve opening control instructions; The instructions are directly issued to the variable frequency circulating pump and the electric regulating valve through the EtherCAT bus to achieve maximum closed-loop control of the energy efficiency of the heat exchanger.

Images