A method and system for short-term heat load prediction control of a steam pipe network

Abstract

The application discloses a steam pipe network short-term heat load prediction control method and system, belonging to the technical field of industrial automation and thermal energy engineering. The method obtains real-time operation data of the steam pipe network, constructs a constraint operator based on fluid dynamics to quantify the physical safety threshold of the steam flow rate change rate, and generates an initial predicted heat load using a deep learning model based on physical perception. Subsequently, the initial prediction is corrected in real time using the fluid dynamics constraint operator to ensure that the prediction result meets the physical realizability. Finally, a feedforward control instruction is generated according to the corrected predicted heat load, and the feedforward control instruction and a feedback control instruction work together to accurately adjust the heat source output. The application combines physical laws and data-driven models to improve the physical credibility of the prediction and the response speed, stability and safety of the control system.

Description

Technical Field

[0001] This application belongs to the field of industrial automation and thermal energy engineering technology, specifically relating to a method and system for predicting and controlling short-term heat load in steam pipeline networks. Background Technology

[0002] With the continuous expansion of industrial parks and urban centralized heating systems, steam pipelines, as the core carrier of energy transmission, play a crucial role in improving overall energy utilization efficiency through stable and efficient operation. Steam pipeline systems involve complex fluid transport processes, where the physical properties of the steam medium directly affect the distribution and conversion efficiency of thermal energy. To ensure continuous industrial production and the quality of heating for residents, real-time monitoring of pipeline operating parameters and achieving dynamic supply-demand balance have become key research areas in modern thermal energy engineering.

[0003] Short-term heat load prediction and control of steam pipeline networks is a key component for achieving precise system regulation, aiming to adjust heat source output in advance by anticipating changes in user demand. Due to the significant compressibility of steam and the complex momentum and heat losses during long-distance transport, the load regulation process exhibits significant nonlinearity, time lag, and strong coupling characteristics. To address these challenges, related fields are dedicated to developing efficient automated control strategies to maintain dynamic stability of pipeline network pressure amidst complex operating condition fluctuations.

[0004] However, traditional feedback control methods in existing technologies often lag behind actual load changes, easily causing drastic fluctuations in pipeline pressure when facing large-scale load switching, making it difficult to meet the requirements of high-precision regulation. Meanwhile, conventional artificial intelligence prediction models rely excessively on historical data and lack constraints from fundamental physical laws of fluid mechanics. This leads to predictions that may violate basic physical principles during sudden changes in operating conditions, such as predicting flow rates exceeding the maximum flow capacity under current pressure differential conditions. Such predictions lacking physical reliability can not only cause control strategies to fail or generate severe control overshoot, but may even trigger major safety hazards such as water hammer under extreme conditions, seriously threatening the operational safety of the pipeline system. Therefore, a short-term heat load prediction and control scheme for steam pipeline networks is desired. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting and controlling short-term heat load in steam pipeline networks, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting and controlling short-term heat load in a steam pipeline network includes the following specific steps: Acquire real-time operating data of the steam pipeline network, including the pressure at the beginning end, the temperature at the beginning end, the pressure at the end end, and the heat load at the user end; A fluid dynamics constraint operator is constructed, which is used to quantify the physical safety threshold of the rate of change of steam flow under the current operating conditions based on the physical parameters of the steam pipeline network and real-time operating data. Real-time running data is input into a trained physical perception-based deep learning model to generate an initial predicted heat load; wherein the physical perception-based deep learning model is trained using a physical perception loss function. The initial predicted heat load is corrected in real time using fluid dynamics constraint operators. When the rate of change of the initial predicted heat load exceeds the physical safety threshold, it is corrected to be within the physical safety threshold to obtain the corrected predicted heat load. Based on the revised predicted heat load, a transfer function model describing the dynamic relationship between heat source output and pipeline end pressure response is established using a system identification method. The heat source outlet pressure adjustment is calculated based on the steady-state gain of the transfer function model and used as a feedforward control command. Pressure deviation is monitored in real time, and an incremental proportional-integral-derivative control algorithm is used to calculate the feedback control command. The feedforward control command and the feedback control command are weighted and superimposed to form a comprehensive control command to regulate the heat source output of the steam pipeline network.

[0007] Furthermore, the steps for constructing the fluid dynamics constraint operator further include: Obtain the physical parameters of the steam pipeline network, including the nominal diameter of the pipeline, the length of the pipeline, and the equivalent roughness of the inner wall of the pipeline; Based on the difference between the initial and final pressures in the real-time operating data, and combined with the physical parameters, the maximum limiting velocity that steam can reach in the pipeline under the current pressure difference is calculated. Based on the maximum limiting flow rate, the relationship between flow rate and time is differentiated to derive the maximum allowable flow rate change per unit time, which is then used as the physical safety threshold.

[0008] Furthermore, the steps for calculating the maximum limiting velocity include: Calculate the steam density and dynamic viscosity based on the head pressure and head temperature in the real-time operating data; Calculate the Reynolds number based on the steam density, dynamic viscosity, nominal pipe diameter, and steam velocity. Based on the Reynolds number and the relative roughness determined by the equivalent roughness of the pipe inner wall and the nominal diameter of the pipe, the friction factor is obtained by iteratively solving the Colbrook equation. Substituting the difference between the initial and final pressures, the friction factor, the pipe length, the nominal pipe diameter, and the steam density into Darcy's formula, the maximum limiting velocity can be calculated.

[0009] Furthermore, the physical perception loss function includes a data-driven loss term and a physical perception loss term; the data-driven loss term is used to measure the degree of deviation between the model prediction and the actual observation; the physical perception loss term is used to apply a positive penalty to the portion of the predicted flow rate change rate at adjacent time points that exceeds the physical safety threshold calculated by the fluid dynamics constraint operator.

[0010] Furthermore, the physical perception loss term is defined as follows: for each time point to be predicted, when the absolute value of the predicted flow rate change is greater than the physical safety threshold, the square of the difference between the absolute value of the predicted flow rate change and the physical safety threshold is used as the penalty value; when the absolute value of the predicted flow rate change is less than or equal to the physical safety threshold, the penalty value is zero; the penalty values ​​of all time points in the prediction sequence are summed to obtain the physical perception loss term.

[0011] Furthermore, the steps for real-time correction of the initial predicted heat load using fluid dynamics constraint operators include: Based on the revised predicted heat load or the actual monitored heat load of the previous moment, calculate the predicted change of the initial predicted heat load at the current moment. Obtain the physical safety threshold calculated by the fluid dynamics constraint operator under the current operating condition; If the absolute value of the predicted change is greater than the physical safety threshold, the corrected predicted heat load at the current moment is set to the sum of the baseline value and the physical safety threshold, and the direction of change is consistent with the predicted change. If the absolute value of the predicted change is less than or equal to the physical safety threshold, then the corrected predicted heat load at the current moment is set as the initial predicted heat load.

[0012] Furthermore, it also includes the following steps: The acquired real-time operating data is preprocessed, including using a moving average filtering method to eliminate measurement noise and applying a minimum-maximum standardization method to map the characteristic parameters of different physical quantities to a unified numerical range. From the preprocessed historical load sequence, daily periodic features reflecting the day-night cycle and weekly periodic features reflecting the energy consumption difference between weekdays and weekends are extracted by fast Fourier transform. The periodic features are then combined with real-time operation data, environmental meteorological data, and current date attributes to construct a multi-dimensional input feature vector, which is used to train a physical perception-based deep learning model.

[0013] A short-term heat load forecasting and control system for a steam pipeline network, comprising: The data acquisition module is used to acquire real-time operating data of the steam pipeline network, including the head end pressure, head end temperature and heat load of the user terminal. The constraint operator construction module is used to construct fluid dynamics constraint operators, which are used to quantify the physical safety threshold of the rate of change of steam flow under the current operating conditions based on the physical parameters of the steam pipeline network and real-time operating data. The model building and training module is used to build a physical perception-based deep learning model. The model is based on a long short-term memory neural network and is trained using a physical perception loss function. The physical perception loss function is used to penalize the part of the model's predicted output that violates the physical laws quantified by the fluid dynamics constraint operator. The prediction module is used to input real-time running data into a trained physical perception-based deep learning model to generate an initial predicted heat load. The correction module is used to correct the initial predicted heat load in real time using fluid dynamics constraint operators. When the rate of change of the initial predicted heat load exceeds the physical safety threshold, it is corrected to be within the physical safety threshold to obtain the corrected predicted heat load. The collaborative control module is used to generate feedforward control commands based on the revised predicted heat load, and integrate the feedback control commands to regulate the heat source output of the steam pipeline network.

[0014] In summary, this application includes at least one of the following beneficial technical effects: 1. The feedforward control mechanism of this invention can adjust the heat source output in advance based on the physically corrected predicted value. This predictive adjustment effectively offsets the transmission delay, reducing the pressure fluctuation range at the user end by a predetermined amount. In extreme situations such as sudden start-up and shutdown of large users, the system can predict the pressure drop trend and increase the pressure in advance, ensuring the continuity of heating or production processes.

[0015] 2. By real-time correction of the predicted load change rate, this invention enables the system to automatically identify and intercept dangerous adjustment commands that may induce water hammer or severe vibration. This safety defense mechanism based on fluid dynamics constraint operators provides a solid guarantee for large-scale centralized heating in industrial parks, eliminating safety hazards caused by prediction distortion under complex operating conditions.

[0016] 3. This invention, through deep synergy between feedforward and feedback, enables the output at the heat source to precisely match the load demand, reducing exhaust steam losses caused by excessive steam supply and production fluctuations caused by insufficient steam supply. Overall energy utilization efficiency is improved by a predetermined percentage, providing a key technical means for achieving green and low-carbon operation of industrial heating systems. Attached Figure Description

[0017] Figure 1 This is an overall schematic diagram of the short-term heat load prediction and control method for steam pipe networks; Figure 2This is a diagram illustrating the core principles of a deep learning prediction model based on physical perception. Figure 3 It is a logic flowchart of feedforward and feedback coordinated control; Figure 4 It is a flowchart of real-time correction of the predicted output using fluid dynamics constraint operators. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1-4 The present invention will be further described in detail with reference to specific embodiments.

[0019] Firstly, a method for predicting and controlling short-term heat load in a steam pipe network is presented, which is implemented according to the following steps: First, for step S1, real-time acquisition and processing of steam pipeline network operation data is performed. This step aims to build a complete data acquisition and preprocessing mechanism to provide accurate, standardized, and physically information-rich input data for subsequent load forecasting models. The entire process covers all aspects from sensor deployment, data acquisition, signal processing to feature construction, and is implemented through the following sub-steps.

[0020] Step S101: Collect and preprocess real-time operating data. First, deploy a sensor network to acquire real-time operating data from the beginning of the steam pipeline network and each user terminal. High-precision pressure transmitters, platinum resistance temperature sensors, and ultrasonic flow meters are installed at the beginning of the pipeline network to continuously monitor the pressure, temperature, and flow rate of the steam. Integrated heat meters are installed at each user terminal to accurately measure the actual heat load consumption. The sampling frequency of the entire network is set between 1 Hz and 0.1 Hz to ensure that minute transient pressure fluctuations can be captured.

[0021] The analog electrical signals collected by the sensors are converted into digital signals by the signal acquisition module and then aggregated to the central processing unit in real time via industrial communication protocols.

[0022] To eliminate measurement noise introduced by factors such as electromagnetic interference, sensor vibration, or fluid pulsation, the system uses a moving average filtering method to smooth the data stream. This involves averaging the data point by point using a sliding window of a preset length to filter out high-frequency random noise.

[0023] The length of the sliding window is set to 5 to 10 sampling points, and the specific value is determined during the system debugging phase based on the actual noise characteristics measured on site.

[0024] Subsequently, a minimum-maximum standardization method is applied to linearly map all feature parameters to a numerical range of 0 to 1. This eliminates the differences in dimensionality between different physical quantities such as pressure, temperature, and flow rate, thereby improving the training convergence speed and weight allocation efficiency of the subsequent deep learning model. The standardization calculation formula is as follows: in These are the original eigenvalues. and These are the minimum and maximum values ​​of the feature in the training dataset, respectively. These are the standardized feature values.

[0025] Step S102: Construct a historical load sequence and extract periodic features. The system constructs a historical load sequence covering a predetermined time period. This sequence consists of instantaneous heat load values ​​obtained by converting second-level mass flow rate data. The instantaneous heat load is calculated based on the product of mass flow rate and the current steam enthalpy. The steam enthalpy is determined by real-time monitored pressure and temperature parameters using steam property formulas, such as the IAPWS-IF97 industrial standard, thus unifying mass flow rate as a heat load characterization quantity and ensuring consistency between the prediction target and the system control target. The predetermined time period is set to the past 30 to 90 days, with the specific duration determined based on the production stability and seasonal variation characteristics of the industrial park where the pipeline network is located, to ensure coverage of the complete energy consumption cycle. The time interval of the historical load sequence is set to 10 seconds, consistent with the update cycle of the control system.

[0026] To uncover the inherent patterns in load variations, the system performs a Fast Fourier Transform (FFT) on historical data, converting the time-domain signal into a frequency-domain signal, thereby accurately extracting the periodic characteristics of load changes. The formula for calculating the FFT is: in For the first The load value at each sampling point The total number of sampling points. For the first in the frequency domain The amplitude of each frequency component, The unit is imaginary. Peak detection is performed on the frequency domain amplitude spectrum to extract the periods corresponding to the frequency components with the largest amplitudes. The extracted features include daily periodic features reflecting the day-night cycle of production in the park, with a period length of 24 hours, and weekly periodic features reflecting the difference in energy consumption between weekdays and weekends, with a period length of 7 days.

[0027] Step S103: Construct a multi-dimensional input feature vector; combine the periodic feature terms extracted in step S102 with real-time environmental meteorological data, current date attribute and instantaneous operating parameters obtained in step S101 to form a multi-dimensional input feature vector.

[0028] The environmental meteorological data includes ambient temperature, humidity, and wind speed; the current date attribute is used to identify working days, statutory holidays, or special shutdown days, and is represented by a unique thermal encoding method, that is, each date type corresponds to an independent binary feature bit; the instantaneous operation data includes the head end pressure and head end temperature.

[0029] This input feature vector will serve as the training sample input for subsequent deep learning models. To construct a complete training dataset, the system divides historical data into several sample pairs in chronological order. Each sample pair consists of an input feature vector and a corresponding future load value. Specifically, the division method is as follows: using the current time as a baseline, and employing feature data from the current time and several previous times as input, the system predicts the load value at a predetermined future time point, such as 15 minutes, 30 minutes, or 60 minutes in the future.

[0030] In the training dataset, the input feature vectors and output labels are aligned in time to form the sample set required for supervised learning.

[0031] In summary, through step S1, the system completes the entire process from raw data acquisition to high-dimensional feature construction and training sample generation. Step S101 ensures the accuracy and standardization of the input data; step S102 uncovers the inherent periodicity of load changes; and step S103 fuses multi-source heterogeneous information into a unified structured feature representation and clarifies the data organization format required for model training. This series of operations provides the data foundation for the subsequent construction of a physical perception-based deep learning prediction model.

[0032] The next step, S2, involves constructing a fluid dynamics constraint operator. Step S2 involves building a mathematical constraint operator capable of quantifying the physical limits of the fluid, used for subsequent physical reliability verification and correction of the deep learning model's prediction output. This operator, based on classical fluid mechanics theory, calculates the physical safety boundary of the steam flow rate change under current operating conditions using the pipeline network's structural parameters and real-time operating data. This is specifically achieved through the following sub-steps.

[0033] Step S201: Obtain the physical parameters of the pipeline network and establish a parametric model. First, the physical structure of the steam pipeline network is parametrically modeled. Key parameters to be obtained include: the nominal diameter of each pipe segment, the actual length of the pipe, and the equivalent roughness of the pipe's inner wall. The nominal diameter is used to determine the flow cross-sectional area of ​​the pipe, the pipe length is used to calculate the friction resistance, and the equivalent roughness reflects the roughness of the pipe's inner wall, directly affecting the calculation of the friction factor. These parameters can be obtained from pipeline network design drawings or on-site surveys and stored in the system as structured data.

[0034] Step S202: Calculate the friction factor and pressure loss along the pipe. The friction factor is a core parameter reflecting the flow resistance of the pipe, and its value depends on the Reynolds number of the fluid and the relative roughness of the pipe wall. The formula for calculating the Reynolds number is: ,in The Reynolds number is dimensionless. This refers to the density of steam, expressed in kilograms per cubic meter. The steam velocity is expressed in meters per second. Pipe diameter, in meters; ρ represents the dynamic viscosity of steam, measured in Pascals per second. Relative roughness is the ratio of the equivalent roughness of the pipe's inner wall to its diameter.

[0035] The friction factor is solved using the calculated Reynolds number and relative roughness, employing the Colbrook equation. The Colbrook equation takes the form: ,in The friction factor is dimensionless. This represents the equivalent roughness of the pipe's inner wall, expressed in meters. The equation is implicit and can be solved using an iterative method. The value.

[0036] After obtaining the friction factor, the pressure loss of steam along the pipeline is calculated using Darcy's formula: ,in This represents the pressure drop at the beginning and end of the pipeline, measured in Pascals. The friction factor is dimensionless. This refers to the length of the pipe, in meters. Pipe diameter, in meters; This refers to the density of steam, expressed in kilograms per cubic meter. The vapor velocity is expressed in meters per second. This formula describes the pressure difference required to maintain a specific flow velocity given pipe parameters and flow conditions.

[0037] Step S203: Derive the physical safety thresholds for the maximum limiting flow velocity and the rate of change of flow rate. Based on the pressure drop formula established in step S202, and combined with the real-time state equation of steam, the maximum limiting flow velocity that the pipeline network can support under the currently known pressure difference between the beginning and end of the pipeline can be derived. Specifically, the currently measured pressure difference between the beginning and end of the pipeline is taken as the known... Substituting into Darcy's formula, the corresponding flow velocity can be solved. This flow rate is the theoretically maximum value that can be achieved under the current pressure difference.

[0038] Furthermore, to obtain the maximum permissible change in flow rate per unit time, the derivative of the relationship between flow rate and time is calculated. The formula for calculating mass flow rate is: ;in Mass flow rate, measured in kilograms per second; This refers to the density of steam, expressed in kilograms per cubic meter. The cross-sectional area of ​​the pipe is expressed in square meters, and is determined by the pipe diameter. Calculated ; The steam velocity is expressed in meters per second.

[0039] Taking the derivative with respect to time yields the rate of change of flow: In engineering approximate calculations, if the effect of density changes over a short period of time is ignored, the calculation can be simplified to: ,in The acceleration of the flow velocity is expressed in meters per square second.

[0040] This acceleration is limited by the ability to change the driving pressure difference. According to the momentum equation (equation of motion) in fluid dynamics, for one-dimensional steady-state flow, the acceleration of the fluid in the pipe is determined by both the net pressure gradient and the inertial force acting on it. Neglecting the transient compressibility term of the steam, the following relationship can be established: in, This is the part used to overcome inertial forces in the driving pressure differential. For the pressure drop due to frictional resistance along the path, For the length of the pipe, The cross-sectional area of ​​the pipe. For vapor density, This represents the rate of change in mass flow rate.

[0041] Considering the physical response capability of the heat source regulating equipment, the maximum differential pressure change rate that the system can achieve per unit time is set to be... The maximum permissible rate of change of flow can then be derived from the following formula: When the frictional resistance pressure drop along the path Based on the current operating conditions and determined by Darcy's formula, combined with the maximum regulating capacity of the heat source... This allows for the calculation of the maximum permissible flow rate change per unit time. This value is the physical safety threshold, representing the acceleration limit that a fluid can achieve under finite power.

[0042] Step S204: The physical safety threshold is used as a fluid dynamics constraint operator for prediction verification. The maximum flow rate change obtained in step S203, i.e., the physical safety threshold, constitutes a fluid dynamics constraint operator. This operator is used to verify whether the predicted load change output by the deep learning model is physically feasible.

[0043] In practical applications, when the neural network model outputs the predicted load value at a certain future moment, the system will calculate the slope of change between the predicted value and the actual value at the previous moment or the predicted value in the previous prediction period, that is, the change in flow rate per unit time.

[0044] Subsequently, the change is compared with the physical safety threshold under the current operating conditions calculated in step S203. If the predicted change does not exceed the threshold, it is determined to be physically reliable; if it exceeds the threshold, the predicted value is determined to be a non-physical disturbance. For example, when the neural network predicts that the load will undergo an extremely large-scale jump far exceeding the current differential pressure driving capacity within the next minute, the constraint operator will determine that the prediction is unrealizable.

[0045] In summary, through step S2, the system establishes a complete fluid dynamics constraint operator. This operator combines the physical structural parameters of the pipeline network with real-time operating data to quantify the physical limits of steam flow variation under current operating conditions. This provides a clear mathematical basis and boundary conditions for the construction of the physical perception loss function in step S3 and the real-time correction of the predicted output in step S4. The introduction of this physical constraint mechanism effectively compensates for the shortcomings of purely data-driven models in terms of adherence to physical laws.

[0046] The next step, S3, involves constructing a deep learning model based on physical perception. This involves building and training a deep learning model that incorporates physical constraints to accurately predict the short-term heat load of the steam pipeline network. This model is based on a long short-term memory neural network architecture. By introducing a physical perception term into the loss function, it forces the model to learn and follow the fundamental laws of fluid mechanics, thus overcoming the shortcomings of purely data-driven models that may produce non-physical predictions during sudden changes in operating conditions. This is specifically achieved through the following sub-steps.

[0047] Step S301: Construct the model input and output data structure. The input data required for model training comes from the multi-dimensional input feature vector constructed in step S1. This feature vector is organized at 10-second intervals in the time dimension, consistent with the update cycle of the control system. The feature vector at each moment contains the following information: daily and weekly periodic feature terms extracted by fast Fourier transform, real-time environmental meteorological data including ambient temperature, humidity and wind speed, current date attribute represented by one-heat encoding, and instantaneous operating data including head-end pressure and head-end temperature.

[0048] The training samples are constructed using a sliding window approach. A time window length is set, for example, the feature data from the past 60 minutes constitutes an input sample used to predict the load value at a future time. Specifically, the current time is used as the baseline, and the feature data from the current time and several previous times are used as input. The corresponding output is the actual load value at the predetermined future time point.

[0049] The output prediction target can be set to the load value for the next 15, 30, or 60 minutes, depending on the application requirements. In this way, historical data is divided into several input-output sample pairs in chronological order, forming the training dataset required for supervised learning.

[0050] Step S302: Design the deep learning network architecture; the model adopts a long short-term memory neural network as the core prediction architecture. The long short-term memory neural network effectively captures long-distance dependencies in the load sequence through three gating mechanisms: input gate, forget gate, and output gate, solving the gradient vanishing problem that exists in traditional recurrent neural networks when processing long-term sequences.

[0051] The network structure is as follows: the number of nodes in the input layer corresponds to the dimension of the input feature vector in step S301; the hidden layer is set to a multi-layer structure, with each layer containing a preset number of neurons, which can be selected from 64, 128, or 256 depending on the model complexity; each neuron uses a linear rectified function as its activation function, the expression of which is: This function can introduce nonlinearity and alleviate the gradient vanishing problem; the output layer contains one neuron, which outputs the predicted future load value.

[0052] Step S303: Design the physical perception loss function. The goal of model training is to make the prediction results both fit the patterns of historical data and conform to the basic physical principles of fluid mechanics. Therefore, the total loss function is designed as a weighted sum of the data-driven loss and the physical perception loss, and its expression is: in The data-driven loss measures the deviation between model predictions and actual observations. It is calculated using the mean squared error and is expressed as follows: In the formula The total number of samples, For the model to the first The predicted load value for each sample, For the first The actual observation load value of each sample.

[0053] This is the physical perception loss term, used to penalize the portion of the prediction result that violates physical laws. Specifically, it is defined as the sum of squares of the portion of the predicted flow rate change that exceeds the physical limit calculated in step S2. Let... This represents the predicted change in flow rate between adjacent time points. Given the maximum allowable flow rate change under the current operating conditions calculated in step S2, the physical sensing loss is calculated as follows: In the formula To predict the number of time points in the sequence where the rate of change needs to be verified, For the first Predict the rate of change in traffic at each time point This represents the physical safety threshold at the corresponding time point. When the predicted rate of change does not exceed the physical limit, this value is 0; once it exceeds the limit, this value generates a positive penalty value, with the penalty increasing the greater the exceedance.

[0054] Weighting coefficient This is used to balance the relative importance of data-driven loss and physical perception loss in the total loss. The preset weights are set within the range of 0.01 to 0.1. The specific values ​​are dynamically fine-tuned based on the convergence of the model on the validation set. The optimization is usually based on the combined performance of the mean square error and the physical violation rate on the validation set.

[0055] Step S304: Perform model training and optimization; train the model using the training dataset constructed in step S301. During training, the input feature vector is fed into the network for forward propagation to calculate the predicted load value, and then the total loss value is calculated according to the loss function in step S303.

[0056] Subsequently, the gradient of the loss function with respect to the weights of each layer of the network is calculated using the backpropagation algorithm, and the network weights are iteratively updated using an adaptive moment estimation optimizer. The training process continues until the loss function converges on the validation set or the preset maximum number of training epochs is reached.

[0057] Through the above training process, the penalty generated by the physical perception loss term will guide the optimization direction of the network weights, forcing the model to learn and follow the physical laws implied by the fluid continuity equation and momentum equation.

[0058] When the predicted trend indicates that the future load will undergo a drastic change, if the rate of change of the flow corresponding to the change exceeds the physical limit, the physical perception loss term will generate a significant positive penalty, thereby enabling the model to gradually adjust the weights during training, suppress such non-physical prediction spikes, and make the prediction curve smoother and more physically reliable.

[0059] In summary, step S3 completes the construction and training of a deep learning prediction model that integrates physical laws. This model is endowed with physical perception capabilities during the training phase, enabling it to output prediction results that conform to both historical data statistical patterns and fundamental principles of fluid mechanics. This model provides high-quality initial prediction values ​​for real-time correction of the prediction output in subsequent step S4, and lays the foundation for the precise implementation of feedforward control in step S5.

[0060] For step S4, fluid dynamics constraint operators are used to correct the predicted output in real time. During the model inference stage, even if the model has undergone physical perception training in step S3, deviations may still occur due to extreme data anomalies or unmodeled perturbations at the input end. To ensure the physical reliability of the prediction results, step S4 introduces a dynamic limiter logic to perform real-time physical verification and correction of the model's original predicted output. This is specifically implemented through the following sub-steps.

[0061] Step S401: Set the prediction period and correction trigger conditions; the system executes the correction logic according to the preset prediction period. The prediction period is consistent with the update period of the control system, set to be executed once every 10 seconds. Within each prediction period, the central processing unit first obtains the original predicted heat load value output by the neural network model in step S3, denoted as... ,in This is the current predicted time.

[0062] To determine whether the predicted value is physically achievable, the heat load needs to be converted into mass flow rate, and then the rate of change of flow rate needs to be physically constrained for verification. The specific conversion process is as follows: Based on the currently monitored steam pressure and temperature, the steam property formula, such as the IAPWS-IF97 industrial standard, is used to calculate the enthalpy of the current steam. The unit is kJ / kg. Dividing the predicted heat load by the enthalpy value converts it into the predicted mass flow rate. The conversion formula is: Similarly, the baseline heat load value from the previous moment... Converted to baseline mass flow rate using the same enthalpy value. The selection principle for the benchmark value is as follows: if there is an output value corrected by step S4 at the previous time, then the corrected value is used as the benchmark; if no correction has been made at the previous time, then the actual monitoring value at the previous time is used as the benchmark.

[0063] Subsequently, the predicted change in mass flow rate is calculated. : This predicted change is the predicted change in mass flow rate per unit time, which is used for subsequent comparison with the physical safety threshold calculated by the fluid dynamics constraint operator.

[0064] After correcting for the flow rate dimension, the corrected mass flow rate needs to be converted back to a heat load value. Let the corrected mass flow rate be... The corrected heat load for: Among them, enthalpy value The calculation was performed using the same current operating parameters as during the conversion.

[0065] Step S402: Obtain the real-time physical safety threshold; the system obtains the maximum allowable flow rate change under the current operating condition from the fluid dynamics constraint operator constructed in step S2, denoted as... This threshold is calculated based on real-time operating parameters at the current moment, including the current pressure difference between the inlet and outlet, steam density, and pipeline geometry parameters, reflecting the maximum acceleration limit that the fluid can achieve under the current driving conditions.

[0066] Step S403: Execute the dynamic limiter correction logic; calculate the predicted change amount obtained in step S401. The physical security threshold obtained in step S402 Compare them.

[0067] like If the predicted change does not exceed the physical safety threshold, the predicted value is deemed physically reliable, and the system directly uses the original predicted value output in step S3 as the corrected predicted output for this period, denoted as . .like If the predicted change exceeds the physical safety threshold, the predicted value is determined to be a non-physical disturbance, and the dynamic limiter intervenes to correct it.

[0068] The correction method is as follows: the current prediction output is locked at the baseline value plus the physical safety threshold. The specific correction formula is as follows: ,in This is a sign function used to keep the direction of change constant; that is, if the predicted value is too high, it is positively corrected, and if the predicted value is too low, it is negatively corrected. Through this correction, predicted values ​​that exceed physical limits are brought back to physically feasible boundary values.

[0069] To facilitate understanding, consider the following example: Assume the baseline flow rate at the previous moment was 100 tons per hour, and the physical safety threshold under the current operating conditions is a flow rate change of 5 tons per hour. Tons per hour. If the model predicts the flow rate for the next time step to be 108 tons per hour, the predicted change is 8 tons per hour, exceeding the physical limit of 5 tons per hour. At this point, the dynamic limiter intervenes, locking the corrected flow rate output at 105 tons per hour, which is the baseline value plus the physical safety threshold.

[0070] Step S404: Output the correction result and update the baseline value; after the correction is completed, the system will... This final output value for the current prediction cycle is used for the feedforward control calculation in subsequent step S5. Simultaneously, this correction value is stored as the baseline value for the next prediction cycle. It is used to calculate the changes in subsequent periods.

[0071] In summary, step S4 establishes a real-time correction mechanism based on fluid dynamics constraint operators, which can intervene in a timely manner when non-physical fluctuations occur in the model prediction, pulling the prediction output back to the physically feasible boundary, and completely eliminating the risk of control command overshoot caused by sensor instantaneous failure, data noise, or load pseudo-jumps from the algorithm level.

[0072] Finally, step S5, the implementation of feedforward and feedback coordinated control, is executed by the coordinated control module. This module achieves deep integration and collaborative operation of feedforward and feedback control. Feedforward control, based on the physically corrected predicted load value output in step S4, adjusts the heat source output in advance to offset predicted disturbances; feedback control, based on real-time monitored pressure deviations, compensates for unmodeled disturbances and predicted residuals. The two are weighted and superimposed to form a comprehensive control signal, driving the actuator to achieve precise adjustment. This is specifically implemented through the following sub-steps.

[0073] Step S501: Establish the feedforward control transfer function model. Feedforward control requires the prior establishment of a mathematical model describing the dynamic relationship between the heat source output and the pressure response at the end of the pipeline network. The system uses a system identification method to construct this model. The specific operation is as follows: First, an excitation signal is applied to the heat source actuator, such as a step signal or a pseudo-random binary sequence signal applied to the boiler steam pressure setpoint or the opening of the main steam regulating valve.

[0074] Simultaneously, pressure response data is collected at pressure monitoring points at the end of the pipeline network. The collected input and output data constitute the identification sample set.

[0075] Subsequently, the parameters of the transfer function model are estimated using either the least squares method or the prediction error method. The established transfer function model accurately describes the dynamic characteristics from the application of control commands to the end-pressure response, including gain, time constant, and pure time delay.

[0076] Step S502: Calculate the feedforward compensation amount. Based on the transfer function model established in step S501, the feedforward controller calculates the compensation amount in advance according to the corrected predicted load value output in step S4. Specifically, the system first maps the predicted load change to the heat source outlet pressure adjustment required to maintain stable terminal pressure. The mapping relationship is determined based on the hydraulic balance equation of the pipe network and the steady-state gain of the transfer function model. This mapping adopts an analytical model based on steady-state hydraulic characteristics, as follows: For steam pipeline networks, under steady-state conditions, the initial pressure is... With end pressure The following relationship exists between them: in This refers to the steam mass flow rate. The comprehensive resistance coefficient can be obtained through hydraulic calculations based on the geometric parameters of the pipeline network, friction factors, and steam properties, or through on-site identification.

[0077] Let the predicted change in flow rate at the current moment be... The current actual traffic is The predicted future flow is To maintain the terminal pressure stable at the set target value. Required head pressure It should meet the following requirements: The required adjustment amount of heat source outlet pressure for: in This represents the current measured pressure at the front end.

[0078] Furthermore, this pressure adjustment amount needs to be combined with the steady-state gain of the transfer function model established in step S501. Corrections are made to reflect the steady-state proportional relationship between the pressure setpoint adjustment and the actual pressure response. The final feedforward compensation control command is then executed. Calculate using the following formula: This instruction will act on the heat source actuator before the load fluctuation actually reaches the pipeline network, thereby offsetting the impact of the delivery delay.

[0079] Step S503 implements feedback regulation control; the feedback regulation system, as a supplement to feedforward control, is used to eliminate prediction bias and unmodeled disturbances. The feedback loop adopts an incremental proportional-integral-derivative control algorithm to monitor the pressure deviation at the user end in real time, that is, the difference between the end pressure setpoint and the current actual monitored pressure value.

[0080] The incremental proportional-integral-derivative controller calculates the proportional term contribution based on the change in the current pressure deviation from the previous time step, the integral term contribution based on the cumulative amount of the current deviation, and the derivative term contribution based on the trend of deviation change. The sum of these three factors is then added to the output of the previous time step to form the feedback control command for the current time step.

[0081] The proportional, integral, and derivative coefficients are tuned according to the dynamic characteristics of the pipeline network and the control quality requirements. When the terminal pressure deviates from the set value due to prediction errors or unmodeled disturbances, the feedback controller quickly calculates the incremental compensation signal to bring the terminal pressure back to the set value.

[0082] Step S504: Perform weighted superposition and limiting of feedforward and feedback signals; the feedforward control command calculated in step S502 and the feedback control command calculated in step S503 are weighted and superimposed to form a comprehensive control command. The weight coefficient of the feedback command is set according to the feedforward prediction accuracy and feedback response requirements. When the prediction accuracy is higher, the weight can be appropriately reduced to reduce the feedback adjustment burden.

[0083] Before being sent to the actuator, the integrated control command must be limited to the safe travel range of the actuator. For example, the opening of the regulating valve is limited to 0% to 100%, and the speed of the boiler feedwater pump is limited to the minimum allowable speed to the maximum allowable speed.

[0084] In step S505, the actuator responds and performs a periodic update; the limited integrated control signal is sent to the corresponding actuator. The actuator includes a boiler feedwater pump driven by a frequency converter and an electro-hydraulic linkage main steam regulating valve. After receiving the digital control signal, the actuator uses a built-in high-precision positioner to precisely adjust the valve opening or pump speed, with an adjustment accuracy better than 0.5%.

[0085] The system executes the entire control process according to a predetermined update cycle, which is set to 10 seconds. Within each update cycle, the system repeatedly executes steps S401 to S404 to obtain the corrected predicted load value, and then executes steps S501 to S505 to calculate the new control command.

[0086] This periodic update mechanism ensures that the control system can respond quickly to load changes and maintain dynamic stability of pipeline pressure.

[0087] Through steps S501 to S505, the system achieves deep synergy between feedforward control and feedback control. Feedforward control, based on physically corrected predicted loads, acts in advance to effectively offset the impact of delivery delays; feedback control compensates for prediction deviations and unmodeled disturbances in real time, eliminating steady-state errors. The combined control signal formed by the weighted superposition of the two is both predictive and stable, ensuring the accurate and stable operation of the steam pipeline network under complex operating conditions.

[0088] Additionally, the steam network short-term heat load prediction and control system disclosed in this application includes: The data acquisition module is used to acquire real-time operating data of the steam pipeline network, including inlet pressure, inlet temperature, outlet pressure, and heat load at the user end. This module acquires data through a sensor network deployed at the inlet and outlet of the pipeline network and aggregates the data to the central processing unit at predetermined intervals.

[0089] The constraint operator construction module is used to construct fluid dynamics constraint operators. These operators quantify the physical safety threshold of the steam flow rate change under current operating conditions based on the physical parameters of the steam pipeline network and real-time operating data. This module calculates the maximum allowable flow rate change per unit time based on physical parameters such as the nominal pipe diameter, pipe length, and equivalent roughness of the pipe inner wall, combined with Darcy's formula and the momentum equation.

[0090] The model building and training module is used to construct a physics-aware deep learning model. This model is based on a long short-term memory neural network architecture and trained using a physics-aware loss function. This loss function penalizes parts of the model's predicted output that violate the physical laws quantified by the fluid dynamics constraint operators. During training, a weighted sum of the data-driven loss and the physics-aware loss ensures that the model adheres to fundamental fluid dynamics principles while fitting historical data.

[0091] The prediction module takes real-time operating data and feeds it into a trained, physically-aware deep learning model to generate an initial predicted heat load. This module then outputs predicted heat load values ​​for future times according to a set prediction period.

[0092] The correction module is used to correct the initial predicted heat load in real time using fluid dynamics constraint operators. When the rate of change of the initial predicted heat load exceeds the physical safety threshold, the correction module limits it to within the physical safety threshold to obtain the corrected predicted heat load, thereby ensuring that the prediction result is physically achievable.

[0093] The collaborative control module generates feedforward control commands based on the revised predicted heat load and integrates these commands with feedback control commands to regulate the heat source output of the steam pipeline network. This module employs a system identification method to establish a transfer function model, calculates the feedforward compensation, and combines incremental proportional-integral-derivative feedback control to form a weighted superposition of these commands to generate a comprehensive control signal that drives the actuators to achieve precise regulation.

[0094] The above modules work together to form a complete "prediction-correction-control" closed-loop system, which can effectively improve the physical reliability of short-term heat load prediction of steam pipeline networks and the stability and safety of the control process.

[0095] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0096] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for predicting and controlling short-term heat load in a steam pipeline network, characterized in that, Includes the following steps: Acquire real-time operating data of the steam pipeline network, including the pressure at the beginning end, the temperature at the beginning end, the pressure at the end end, and the heat load at the user end; A fluid dynamics constraint operator is constructed, which is used to quantify the physical safety threshold of the rate of change of steam flow under the current operating conditions based on the physical parameters of the steam pipeline network and real-time operating data. Real-time running data is input into a trained physical perception-based deep learning model to generate an initial predicted heat load; wherein the physical perception-based deep learning model is trained using a physical perception loss function. The initial predicted heat load is corrected in real time using fluid dynamics constraint operators. When the rate of change of the initial predicted heat load exceeds the physical safety threshold, it is corrected to be within the physical safety threshold to obtain the corrected predicted heat load. Based on the revised predicted heat load, a transfer function model describing the dynamic relationship between heat source output and pipeline end pressure response is established using a system identification method. The heat source outlet pressure adjustment is calculated based on the steady-state gain of the transfer function model and used as a feedforward control command. Pressure deviation is monitored in real time, and an incremental proportional-integral-derivative control algorithm is used to calculate the feedback control command. The feedforward control command and the feedback control command are weighted and superimposed to form a comprehensive control command to regulate the heat source output of the steam pipeline network.

2. The method according to claim 1, characterized in that, The steps for constructing fluid dynamics constraint operators further include: Obtain the physical parameters of the steam pipeline network, including the nominal diameter of the pipeline, the length of the pipeline, and the equivalent roughness of the inner wall of the pipeline; Based on the difference between the initial and final pressures in the real-time operating data, and combined with the physical parameters, the maximum limiting velocity that steam can reach in the pipeline under the current pressure difference is calculated. Based on the maximum limiting flow rate, the relationship between flow rate and time is differentiated to derive the maximum allowable flow rate change per unit time, which is then used as the physical safety threshold.

3. The method according to claim 2, characterized in that, The steps for calculating the maximum limiting velocity include: Calculate the steam density and dynamic viscosity based on the head pressure and head temperature in the real-time operating data; Calculate the Reynolds number based on the steam density, dynamic viscosity, nominal pipe diameter, and steam velocity. Based on the Reynolds number and the relative roughness determined by the equivalent roughness of the pipe inner wall and the nominal diameter of the pipe, the friction factor is obtained by iteratively solving the Colbrook equation. Substituting the difference between the initial and final pressures, the friction factor, the pipe length, the nominal pipe diameter, and the steam density into Darcy's formula, the maximum limiting velocity can be calculated.

4. The method according to claim 1, characterized in that, The physical perception loss function includes a data-driven loss term and a physical perception loss term; the data-driven loss term is used to measure the degree of deviation between the model prediction and the actual observation; the physical perception loss term is used to apply a positive penalty to the portion of the predicted flow rate change rate at adjacent time points that exceeds the physical safety threshold calculated by the fluid dynamics constraint operator.

5. The method according to claim 4, characterized in that, The physical perception loss term is defined as follows: for each time point to be predicted, when the absolute value of the predicted flow rate change is greater than the physical safety threshold, the square of the difference between the absolute value of the predicted flow rate change and the physical safety threshold is used as the penalty value; when the absolute value of the predicted flow rate change is less than or equal to the physical safety threshold, the penalty value is zero; the penalty values ​​of all time points in the prediction sequence are summed to obtain the physical perception loss term.

6. The method according to claim 1, characterized in that, The steps for real-time correction of the initial predicted heat load using fluid dynamics constraint operators include: Based on the revised predicted heat load or the actual monitored heat load of the previous moment, calculate the predicted change of the initial predicted heat load at the current moment. Obtain the physical safety threshold calculated by the fluid dynamics constraint operator under the current operating condition; If the absolute value of the predicted change is greater than the physical safety threshold, the corrected predicted heat load at the current moment is set to the sum of the baseline value and the physical safety threshold, and the direction of change is consistent with the predicted change. If the absolute value of the predicted change is less than or equal to the physical safety threshold, then the corrected predicted heat load at the current moment is set as the initial predicted heat load.

7. The method according to claim 1, characterized in that, It also includes the following steps: The acquired real-time operating data is preprocessed, including using a moving average filtering method to eliminate measurement noise and applying a minimum-maximum standardization method to map the characteristic parameters of different physical quantities to a unified numerical range. From the preprocessed historical load sequence, daily periodic features reflecting the day-night cycle and weekly periodic features reflecting the energy consumption difference between weekdays and weekends are extracted by fast Fourier transform. The periodic features are then combined with real-time operation data, environmental meteorological data, and current date attributes to construct a multi-dimensional input feature vector, which is used to train a physical perception-based deep learning model.

8. A short-term heat load prediction and control system for a steam pipeline network, applied to the method described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire real-time operating data of the steam pipeline network, including the pressure at the beginning end, the temperature at the beginning end, and the heat load at the user end. The constraint operator construction module is used to construct fluid dynamics constraint operators, which are used to quantify the physical safety threshold of the rate of change of steam flow under the current operating conditions based on the physical parameters of the steam pipeline network and real-time operating data. The model building and training module is used to build a physical perception-based deep learning model. The model is based on a long short-term memory neural network and is trained using a physical perception loss function. The physical perception loss function is used to penalize the part of the model's predicted output that violates the physical laws quantified by the fluid dynamics constraint operator. The prediction module is used to input real-time running data into a trained physical perception-based deep learning model to generate an initial predicted heat load. The correction module is used to correct the initial predicted heat load in real time using fluid dynamics constraint operators. When the rate of change of the initial predicted heat load exceeds the physical safety threshold, it is corrected to be within the physical safety threshold to obtain the corrected predicted heat load. The collaborative control module is used to generate feedforward control commands based on the revised predicted heat load, and integrate the feedback control commands to regulate the heat source output of the steam pipeline network.

Images