Hybrid modeling method for waste steam cascade heat supply system of thermal power plant

Abstract

The invention provides a hybrid modeling method for a steam exhaust cascade heat supply system of a thermal power plant, and the method comprises the following steps: collecting real-time and historical working condition data, and constructing a heat supply network part physical information sub-model and a heat source part physical information sub-model; the two sub-models adopt a deep learning network to dynamically identify and predict key time-varying parameters in a mechanism model from data, the key time-varying parameters are embedded into a mechanism differential equation set constructed based on a first principle, and data loss and physical loss are minimized through a collaborative optimizer. And the neural network weight is reversely propagated and updated. According to the hybrid modeling method, the limitation of a traditional method is effectively overcome, the established model has high precision and strong physical interpretability, a reliable basis can be provided for predictive control, real-time optimization and advanced control of the heat supply system, and the energy utilization efficiency is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind turbine unit fault monitoring, in particular to a hybrid modeling method for a thermal power plant exhausted steam cascade heat supply system. BACKGROUND

[0002] Energy efficient utilization and low carbon emission are major challenges faced by the current thermal power generation field. The exhausted steam cascade heat supply system of thermal power plants is one of the key technologies to improve the comprehensive energy utilization rate. It recovers the low-grade exhausted steam waste heat at the end of the steam turbine for municipal heat supply, significantly reduces the cold end loss, and realizes the cascade utilization of energy, which is of great significance for energy saving and consumption reduction of power plants.

[0003] The core of the exhausted steam cascade heat supply system of thermal power plants lies in the accurate modeling and optimal control of complex thermal processes. However, this system is a typical high-dimensional, nonlinear, and strongly coupled dynamic system, which contains complex heat transfer, mass transfer, and phase change processes inside, and the operating conditions frequently fluctuate, which makes it extremely difficult to establish a model that can accurately describe the dynamic characteristics of the whole operating condition.

[0004] Currently, the existing modeling methods for this system mainly fall into two categories, but both have significant limitations. The first category is the mechanism model based on differential equations. This method establishes a strict differential algebraic equation set according to the first principles such as mass, energy, and momentum conservation laws. Its advantage lies in its clear physical meaning. However, this system involves many parameters that are difficult to accurately calculate, and the model complexity increases exponentially with the system size, making it difficult to solve the model and the calculation cost is high, which makes it difficult to meet the needs of real-time monitoring and optimal control.

[0005] The second category is a data-driven model based on machine learning. This method uses system operation data to learn the mapping relationship between input and output through deep learning algorithms. Its advantage is that it does not require complex prior physical knowledge and has strong fitting ability for nonlinear systems. However, this method relies entirely on the quality and quantity of data, and requires a large amount of labeled data covering all operating conditions, which is scarce in actual power plant operation. More importantly, the training results lack physical consistency and may produce prediction results that violate physical laws, and the reliability drops sharply in operating conditions that the system has not experienced, which poses a safety risk.

[0006] In recent years, although some research has tried to combine mechanism and data-driven methods, most of them are simple series or parallel combinations. For example, a mechanism model is used to calculate the approximate output, and a data-driven model is used to compensate for the error. However, this simple hybrid does not achieve deep collaboration and does not embed physical laws into the learning process in a structured way, resulting in limited improvement in model performance in data-scarce areas and overall generalization ability.

[0007] In summary, the prior art cannot effectively balance the precision, efficiency and physical consistency of the modeling of the cogeneration plant exhausted steam cascade heat supply system. Developing a new modeling method that can deeply integrate the internal physical mechanism and data-driven learning ability has become a key technical problem to be solved in the field.

[0008] Therefore, the present application is proposed. SUMMARY

[0009] Therefore, the present application proposes a hybrid modeling method for a cogeneration plant exhausted steam cascade heat supply system. The method decomposes the hybrid modeling process for the cogeneration plant exhausted steam cascade heat supply system into a heat network part and a heat source part, and establishes core mechanism differential equation sets for each part. Then, the method uses a deep learning network to dynamically identify key time-varying parameters in the model. Finally, the method realizes the collaborative driving of mechanism and data by embedding the physical equation as a constraint condition into the neural network training loss function. The present application effectively overcomes the limitations of traditional methods, and the built model has high precision and strong physical interpretability, which can provide a reliable basis for predictive control, real-time optimization and advanced control of the heat supply system, and significantly improve energy utilization efficiency.

[0010] Specifically, the present application is realized by the following technical solutions:

[0011] The present application provides a hybrid modeling method for a cogeneration plant exhausted steam cascade heat supply system, comprising the following steps:

[0012] Collecting real-time and historical working condition data for constructing a heat network part physical information submodel and a heat source part physical information submodel;

[0013] Both submodels use a deep learning network to dynamically identify and predict key time-varying parameters in the mechanism model from data, embed the key time-varying parameters into a mechanism differential equation set constructed based on first principles, minimize data loss and physical loss by an optimizer, thereby back-propagating and updating neural network weights.

[0014] Preferably, the construction method of the heat network part physical information submodel comprises:

[0015] Dynamically learning and predicting a key parameter in the primary pipe network, i.e., the comprehensive heat dissipation coefficient K;

[0016] Embedding the predicted K into a mechanism differential equation established based on the law of conservation of energy, solving the equation by a numerical integration method, and finally calculating the predicted water supply temperature;

[0017] The training and optimization of the model is realized by an optimizer containing data loss and physical loss minimizing data loss and physical loss, thereby back-propagating and updating neural network weights.

[0018] Preferably, the method for dynamically learning and predicting the key parameter in the pipe network, the comprehensive heat dissipation coefficient K, comprises:

[0019] According to the heat balance principle of water temperature change in the pipeline, the energy conservation equation is established:

[0020]

[0021] Wherein, p represents the water supply density (kg / m 3 ) ; Cp represents the specific heat capacity of water (J / (kg.℃) ) ; V represents the total volume of water in the pipeline (m 3 ) ; represents the rate of change of water supply temperature with time (℃ / s) ; Q in represents the input heat power, i.e. the heat supplied by the heat source to the pipe network (W) ; Q out represents the output heat power, i.e. the heat extracted by the user side (W) ; Q loss represents the heat dissipation of the pipeline to the environment (W) ;

[0022] Pipeline heat dissipation equation:

[0023] Q loss = K × A × (T supply -T amb ) (2) K is the comprehensive heat dissipation coefficient (W / (m 2 ·℃) ), which is a complex parameter related to flow rate, insulation material, environmental wind speed and other factors. The present application uses deep learning to dynamically predict this coefficient; A represents the pipeline heat dissipation surface area (m 2 ) ; T supply represents the water supply temperature; T amb represents the environmental temperature (℃).

[0024] Preferably, the method for calculating the predicted water supply temperature and model optimization comprises:

[0025] Substitute the calculated Q loss equation into (1) to obtain the differential equation of the control system:

[0026]

[0027] Then the comprehensive heat dissipation coefficient K is dynamically predicted by a long short-term memory neural network, the input of which is represented as X t At each time step t, the input features of the network include all measurable variables that affect K, the number of input features of the input layer of the network structure is equal to the number of nodes, the LSTM layer is 1-2 layers, each layer has 64-128 neurons, which is used to capture the time dependence; the fully connected layer is used to map the LSTM output to the final output; the output layer uses the ReLU activation function to ensure that the value of K is non-negative;

[0028] The loss function L is composed of three parts:

[0029]

[0030] wherein, L physics represents the constraint function value of the physical equation; represents the actual value of the change rate of the supply water temperature; represents the predicted value of the change rate of the supply water temperature;

[0031]

[0032] wherein, L data represents the constraint function value of the deep network; T supply-dnn represents the predicted value of the supply water temperature output by the neural network;

[0033] L total = α × L physics + β × L data (6)

[0034] wherein, L total represents the total loss function value of the physical information network model; α represents the physical loss weight; β represents the data loss weight; the loss function of formula (6) is the weighted sum of the physical loss and the data loss;

[0035] By the gradient descent algorithm, the total loss L total is minimized, and the weights of the LSTM network are updated at the same time, and finally the predicted supply water temperature is calculated.

[0036] Preferably, the construction method of the heat source part physical information sub-model comprises:

[0037] The exhaust steam heat exchanger efficiency η dc , the heat pump performance coefficient COP ahp and the extraction steam heat exchanger heat transfer coefficient K esh are dynamically predicted and output by the deep learning model.

[0038] The above three parameters are embedded in the core differential equation describing the three-stage heating process in real time, and the predicted values of the outlet temperatures of each stage are obtained by numerical solution. The data loss and the physical loss are minimized by the optimizer, so as to backward propagate and update the neural network weights.

[0039] Preferably, the representation method of the exhaust steam heat exchanger efficiency η dc , the heat pump performance coefficient COP ahp and the extraction steam heat exchanger heat transfer coefficient K esh comprises:

[0040] The return water absorbs the latent heat released by the exhaust steam condensation, and the temperature changes from Treturn Rise to T H1 :

[0041]

[0042] C dc = p x C p x V dc (8)

[0043] where C dc represents the heat capacity of the exhaust steam heat exchanger and related pipes (J / ℃); T H1 represents the temperature of the water at the outlet of the exhaust steam heat exchanger, i.e. the water temperature after the first heating (℃); t represents time (s); m dc represents the mass flow rate of the exhaust steam flowing into the exhaust steam heat exchanger (kg / s); h steam represents the specific enthalpy of the exhaust steam (J / kg); h water represents the specific enthalpy of the exhaust steam condensate (J / kg); η dc represents the thermal efficiency of the exhaust steam heat exchanger, which is a coefficient close to 1 but difficult to determine accurately; m W represents the mass flow rate of the network circulating water (kg / s); C p represents the specific heat capacity of water at constant pressure (J / (kg·℃)); T return represents the temperature of the heat network return water (℃);

[0044] The heat pump uses a driving heat source to pump the heat of a low-temperature heat source to a medium-temperature return water, so that the temperature of the return water rises from T H1 to T H2 :

[0045]

[0046] where C ahp represents the heat capacity of the heat pump and related pipes (J / ℃); T H2 represents the temperature of the water at the outlet of the heat pump, i.e. the water temperature after the second heating (℃); COP ahp represents the performance coefficient of the heat pump; Q drive represents the heat power of the heat pump driving heat source (W):

[0047] Q drive = m drive x (h drive - h cond ) (10)

[0048] where m drive is the driving steam flow rate;

[0049] The water is finally heated to the set supply water temperature T supply using exhaust steam of a higher grade:

[0050]

[0051] where the logarithmic mean temperature difference ΔT lm is calculated as:

[0052]

[0053] where C esh represents the heat capacity of the heat pump and related pipes (J / ℃); T H2 represents the outlet water temperature of the heat pump, i.e. the water temperature after the second stage heating (℃); T supply represents the primary network supply water temperature (℃); K esh represents the comprehensive heat transfer coefficient of the extraction steam heat exchanger (W / (m 2 ·℃)); A esh represents the heat exchange area of the extraction steam heat exchanger (m 2 ); T steam represents the saturation temperature of the extraction steam (℃).

[0054] Preferably, the solving method of the predicted values of the outlet temperatures of each stage and the method optimized by the optimizer include the following steps:

[0055] The η dc , COP ahp , K esh are predicted by the long short-term memory neural network; the input features X {t-n:t} at the current and historical time are fed into the LSTM network to obtain the predicted values of the key parameters η {dc,pred} , COP {ahp,pred} , K {esh,pred} ; these predicted parameters are substituted into the aforementioned mechanism differential equations; the equation set is solved using a numerical integration method to calculate the predicted values of the node temperatures T H1_pred , T H2_pred , T supply_pred :

[0056] A hybrid loss function L total is designed to perform backpropagation to optimize the neural network weights:

[0057] L total = α × L data + β × L physics (13)

[0058] L data = MSE(T H1_true , T H1_pred ) + MSE(T H2_true , T H2_pred ) + MSE(T supply_true , T supply_pred)(14)

[0059] L physics represents the physical constraint error: the residual of the left and right sides of the differential equation is calculated and minimized; for the steam extraction heat exchanger equation, the physical residual is:

[0060]

[0061] The L is obtained by weighted sum of the residual of each equation physics The total loss L is minimized by gradient descent algorithm total , and the predicted value of the outlet temperature of each stage is solved.

[0062] The application also provides a hybrid modeling system of a hybrid modeling method for a cogeneration plant steam cascade heat supply system, comprising:

[0063] The construction module is used for collecting real-time and historical working condition data, and is used for constructing the physical information sub-model of the heat supply network part and the physical information sub-model of the heat source part.

[0064] The optimization module is used for dynamically identifying and predicting the key time-varying parameters in the mechanism model from the data by using the deep learning network for the two sub-models, embedding the key time-varying parameters into the mechanism differential equation group constructed based on the first principle, minimizing the data loss and the physical loss by using the optimizer, and then reversely propagating and updating the neural network weight.

[0065] In summary, the application constructs a physical equation and deep learning network fusion modeling framework aiming at the problems of insufficient accuracy of the traditional mechanism model and poor physical consistency of the pure data driven model. The method divides the system modeling process into a heat supply network part and a heat source part, and establishes the core mechanism differential equation group for each part. Then, the deep learning network is used to dynamically identify the key time-varying parameters in the model. Finally, the physical equation is embedded into the neural network training loss function as a constraint condition to realize the collaborative driving of the mechanism and the data. The application effectively overcomes the limitations of the traditional method, and the built model has high precision and strong physical interpretability, which can provide a reliable basis for the prediction control, real-time optimization and advanced control of the heat supply system, and significantly improve the energy utilization efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0066] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings, which serve as best modes for carrying out the application and are not intended to limit the same, and in which like reference numerals designate like parts throughout the several views, wherein:

[0067] Figure 1 is the overall architecture and implementation process of the hybrid modeling method of the application;

[0068] Figure 2 is a structural schematic diagram of the heat network part of the method of the present application;

[0069] Figure 3 is a structural schematic diagram of the heat source part of the method of the present application;

[0070] Figure 4 is a comparison diagram of the effect of the method of the present application and the traditional mechanism modeling method - differential equation;

[0071] Figure 5 is a flow schematic diagram of a computer device provided for an embodiment of the present application. DETAILED DESCRIPTION

[0072] The exemplary embodiments will be described in detail herein below, examples of which are shown in the drawings, and the following description refers to the accompanying drawings that show, by way of illustration, specific embodiments in which the application can be placed. In this regard, the description taken with the drawings makes apparent to those skilled in the art how the application can be practiced. The following exemplary embodiments described hereinbelow are illustrative of various embodiments consistent with the present disclosure and as such do not represent all the only embodiments consistent with the present disclosure, rather they are merely examples of apparatus and methods in conformance with aspects of the present disclosure as detailed in the appended claims,

[0073] The terminology used in the present disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the present disclosure. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof,

[0074] It should be understood that although the terms first, second, third, etc. can be employed in this disclosure to describe various information, these information should not be limited to these terms solely. These terms are only used to distinguish one information from another information. Thus, a first information could be termed a second information or a second information could be termed a first information without departing from the scope of the present disclosure, as the context can dictate, the word "if" can be interpreted to mean "when" or "upon" or "in response to determining" or at least "if" in some embodiments,

[0075] EMBODIMENTS

[0076] The operation data of the unit is obtained through the plant-level monitoring information system, which includes the parameter variables required for modeling. The modeling process of the present application is divided into two parts: the heat network part and the heat source part. First, the modeling method of the heat network part is introduced, and the model of the heat network part is established by using differential equation and physical information network, and the modeling process is as follows:

[0077] (1) Establishing the mathematical model of the heat network part

[0078] The mechanism differential equation of the primary network thermal process is embedded into the training process of the deep learning neural network as a constraint condition, the neural network learns and compensates for the part that is difficult to accurately calculate or measure in the mechanism model, and finally outputs the calculation result of the primary heat supply water temperature.

[0079] According to the heat balance principle of water temperature change in the pipeline, the energy conservation equation is established:

[0080]

[0081] Where, ρ represents the density of the water supply (kg / m 3 ) ; Cp represents the specific heat capacity of water (J / (kg.℃) ) ; V represents the total volume of water in the pipeline (m 3 ) ; represents the change rate of the water supply temperature with time (℃ / s) ; Q in represents the input heat power, that is, the heat supplied by the heat source to the pipe network (W) ; Q out represents the output heat power, that is, the heat extracted by the user side (W) ; Q loss represents the heat dissipation of the pipeline to the environment (W).

[0082] Pipeline heat dissipation equation:

[0083] Q loss = K * A * (T supply -T amb ) (2) K is a comprehensive heat dissipation coefficient (W / (m 2 ·℃) ), which is a complex parameter related to flow rate, thermal insulation material, environmental wind speed and other factors, and the present application uses deep learning to dynamically predict the coefficient; A represents the pipe heat dissipation surface area (m 2 ) ; T supply represents the water supply temperature; T amb represents the environmental temperature (℃), which is an input of the model.

[0084] The above equation is substituted and arranged to obtain the differential equation of the control system:

[0085]

[0086] The equation will be a physical constraint for neural network training. The following is a long short-term memory neural network to dynamically predict the comprehensive heat dissipation coefficient K.

[0087] The network input is represented as X t At each time step t, the input features of the network include all the measurable variables that affect K: the current time primary network flow; the current time environmental temperature; the current time environmental wind speed; the last time water supply temperature; time information; the network output is represented as Y tis the predicted overall heat dissipation coefficient at the current time.

[0088] Network structure. Input layer: the number of nodes is equal to the number of input features; LSTM layer: 1-2 layers, 64-128 neurons in each layer, used to capture time dependence; fully connected layer: maps the LSTM output to the final output; output layer: 1 node, using ReLU activation function to ensure that the K value is non-negative.

[0089] The loss function L consists of three parts, which are the core of collaborative driving:

[0090]

[0091] where L physics represents the constraint function value of the physical equation; represents the actual value of the change rate of the supply water temperature; represents the predicted value of the change rate of the supply water temperature; the loss function of formula (4) is used to measure the difference between the predicted physical process and the real physical process.

[0092]

[0093] where L data represents the constraint function value of the deep network; T supply-dnn represents the predicted value of the supply water temperature output by the neural network; the loss function of formula (5) is used to measure the difference between the predicted supply water temperature and the real measured value.

[0094] L total = α × L physics + β × L data (6)

[0095] where L total represents the total loss function value of the physical information network model; α represents the physical loss weight; β represents the data loss weight; the loss function of formula (6) is the weighted sum of the physical loss and the data loss.

[0096] Through the gradient descent algorithm, the total loss L total is minimized, and the weights of the LSTM network are updated at the same time; this process forces the network to learn a K value that can both fit the data and strictly comply with the law of conservation of energy; through the operation of the above model, the supply water temperature of the first network is finally calculated.

[0097] (2) Establish the mathematical model of the heat source part

[0098] First, the mechanism differential equation model of the heat source part is constructed.

[0099] The application aims to provide a differential equation and deep learning collaborative driving modeling method for a heat source part of a cascade heat supply system.

[0100] The application constructs a mechanism differential equation model of the heat source part, which includes three series heating units, and the heat balance process thereof is described by the following differential algebraic equation set:

[0101] The exhaust steam-water heat exchanger model:

[0102] The return water absorbs the latent heat released by the exhaust steam condensation, and the temperature increases from T return to T H1 .

[0103]

[0104] C dc = ρ × C p × V dc (8)

[0105] Wherein, C dc represents the heat capacity of the exhaust steam heat exchanger and the related pipeline water (J / ℃); T H1 represents the exhaust steam heat exchanger outlet water temperature, i.e. the first stage heating water temperature (℃); t represents time (s); m dc represents the exhaust steam mass flow rate flowing into the exhaust steam heat exchanger (kg / s); h steam represents the specific enthalpy of the exhaust steam (J / kg); h water represents the specific enthalpy of the exhaust steam condensate water (J / kg); η dc represents the heat efficiency of the exhaust steam heat exchanger, which is a coefficient close to 1 but difficult to accurately determine; m W represents the network circulating water mass flow rate (kg / s); C p represents the constant-pressure specific heat capacity of water (J / (kg·℃)); T return represents the heat network return water temperature (℃).

[0106] The absorption heat pump model:

[0107] The heat pump uses a driving heat source (usually steam extraction) to pump the heat of the low-temperature heat source (exhaust steam condensate water) to the medium-temperature return water, so that the temperature increases from T H1 to T H2 .

[0108]

[0109] Wherein, C ahprepresents the heat capacity of water in the heat pump and related pipelines (J / ℃); T H2 represents the water temperature at the outlet of the heat pump, i.e., the water temperature after the second-stage heating (℃); COP ahp represents the performance coefficient of the heat pump. This is a key parameter, and its value changes nonlinearly with the driving heat source temperature, source-side temperature, load rate, and other factors, and it is difficult for a traditional model to accurately describe; Q drive represents the heat power of the driving heat source of the heat pump (W).

[0110] Q drive = m drive × (h drive -h cond ) (10)

[0111] wherein m drive is the driving steam flow rate.

[0112] Steam-water heat exchanger model:

[0113] The water is finally heated to the set water supply temperature T supply

[0114]

[0115] wherein the logarithmic mean temperature difference ΔT lm is calculated as:

[0116]

[0117] wherein C esh represents the heat capacity of water in the heat pump and related pipelines (J / ℃); T H2 represents the water temperature at the outlet of the heat pump, i.e., the water temperature after the second-stage heating (℃); T supply represents the primary network water supply temperature (℃); K esh represents the comprehensive heat transfer coefficient of the steam heat exchanger (W / (m 2 ·℃)), which is another key parameter and changes dramatically with the flow rate, dirt condition, and steam load, and is a main source of model error; A esh represents the heat exchange area of the steam heat exchanger (m 2 ); T steam represents the saturation temperature of the steam (℃);

[0118] Second, a deep learning model is introduced to dynamically identify and compensate for key parameters.

[0119] The key uncertain parameters η dc , COP ahp , and K esh in the mechanism model are dynamically predicted and output by the deep learning model.

[0120] Deep learning model input Xt: At each time step t, the input features include the system measurable operating variables, as well as their historical sequences to capture the dynamic characteristics.

[0121] Deep learning model output yt: η dc , COP ahp , K esh .

[0122] Network structure: A recurrent neural network structure containing one or more layers of LSTM is adopted, followed by a fully connected layer, and the final output layer uses activation functions such as Sigmoid (for efficiency η), ReLU (for heat transfer coefficient K and COP) to constrain the output range.

[0123] Third, a collaborative driving algorithm of differential equations and deep learning is realized.

[0124] The calculation process of this collaborative driving modeling algorithm is as follows:

[0125] Data preparation: Collect system historical operation data, including all input features Xtand corresponding real measured values.

[0126] Forward propagation: input the current and historical input features X {t-n:t} into the LSTM network to obtain the predicted values of key parameters η {dc,pred} , COP {ahp,pred} , K {esh,pred} ; Substitute these predicted parameters into the aforementioned mechanism differential equations; use numerical integration method to solve the equation set to calculate the predicted values of node temperatures T H1_pred , T H2_pred , T supply_pred .

[0127] Loss function calculation and optimization: design a hybrid loss function L total , perform backpropagation, and optimize the neural network weights.

[0128] L total = α × L data + β × L physics (13)

[0129] L data = MSE(T H1_true , T H1_pred ) + MSE(T H2_true , T H2_pred ) + MSE(T supply_true , T supply_pred ) (14)

[0130] L physicsResidual error of physical constraint: calculate the residual error of the left and right sides of the differential equation and minimize it; for the steam extraction heat exchanger equation, the physical residual error is:

[0131]

[0132] The L is obtained by weighted sum of the residual error of each equation physics Minimize L by gradient descent algorithm total Force the parameters learned by the neural network to make the model output fit the real data and satisfy the physical law of energy conservation.

[0133] The application also provides a hybrid modeling system, comprising:

[0134] A construction module: used for collecting real-time and historical working condition data, for constructing the physical information sub-model of the heat network part and the physical information sub-model of the heat source part;

[0135] An optimization module: used for dynamically identifying and predicting the key time-varying parameters in the mechanism model by the deep learning network from the data for both the above two sub-models, embedding the key time-varying parameters into the mechanism differential equation group constructed based on the first principle, minimizing the data loss and the physical loss by the optimizer, thereby back-propagating and updating the neural network weights.

[0136] In specific implementation, the above modules can be implemented as independent entities, or can be combined as the same or several entities, and the specific implementation of the above units can be referred to the method embodiments above, which will not be described here.

[0137] The flowchart Figure 1 is the overall architecture and implementation process of the method of the application. The flowchart is divided into two stages of offline training and online application, embodying the deep integration of physical mechanism and data-driven technology.

[0138] The upper part of the flow chart is the offline training phase, which is the core of the model construction. This phase starts with the collection and preprocessing of historical operation data, forming a high-quality data set. Subsequently, the process enters two key sub-model construction links in parallel: the heat network part physical information sub-model and the heat source part physical information sub-model. Both of these sub-models use a unified "physical information neural network" framework: first, use a deep learning network to dynamically identify and predict key time-varying parameters in the mechanism model (such as the heat dissipation coefficient K of the primary network, the heat pump COP of the heat source part, and the heat transfer coefficient K_esh of the heat exchanger, etc.); Then, embed these predicted parameters into the mechanism differential equation set based on first principles; Finally, through a collaborative optimizer, minimize the data loss and physical loss simultaneously, thereby backpropagating and updating the neural network weights. This training mechanism ensures that the final trained model not only accurately fits the historical data, but also strictly follows the physical conservation laws.

[0139] The lower part of the flow chart is the online application phase. The trained hybrid model is deployed to the actual system, receives real-time collected operating condition data, and performs online calculation. The model finally outputs high-precision system state variables, thereby constructing a digital model of the system. This digital model provides a reliable and physically interpretable core calculation engine for subsequent real-time optimization setup, model predictive control, fault diagnosis and warning, and other advanced applications.

[0140] In summary, this flow chart fully reveals the complete technical path of the invention from data preparation, hybrid model collaborative training to final deployment application, highlighting its core advantage of balancing model accuracy and physical consistency.

[0141] Figure 2 For the heat network part model structure diagram in the invention, it can be seen from the diagram that:

[0142] When the model is working, first, real-time and historical operating condition data are input into an LSTM deep learning network. The core task of this network is to dynamically learn and predict the key parameter - the comprehensive heat dissipation coefficient K in the primary pipe network, which is difficult to accurately model.

[0143] Subsequently, the predicted K value is embedded into the mechanism differential equation based on the law of conservation of energy. This equation describes the physical relationship between the change of supply water temperature over time and the input, output, and dissipation of heat. By solving this equation through numerical integration, the predicted supply water temperature is finally calculated.

[0144] The training optimization target of the model is guided by a unique hybrid loss function. It contains both data loss (error between predicted temperature and actual measured value) and physical loss (residual of the mechanism equation itself). By minimizing the total loss through the optimizer, the neural network weights are updated through backpropagation, forcing the model to accurately fit the data while strictly adhering to physical laws, thus becoming a digital model with high precision, strong generalization ability, and physical interpretability.

[0145] Figure 3 is the core structure of the heat source part modeling algorithm. The model is composed of a deep learning parameter identifier and a mechanism differential equation solver coupled in parallel. As can be seen from the figure, the workflow is as follows: the real-time collected system operating condition data is first input to the LSTM network, which dynamically identifies and outputs three key physical parameters that are difficult to accurately model: the exhaust gas heat exchanger efficiency η dc , the heat pump performance coefficient COP ahp , and the steam extraction heat exchanger heat transfer coefficient K esh . These parameters are embedded in real time into the core differential equations that describe the three-stage heating process, and the predicted values of the outlet temperatures of each stage are obtained by numerical solving.

[0146] The synergistic driving mechanism is reflected in the unique loss function. The total loss of the model is composed of data loss (error between predicted temperature and actual measured temperature) and physical loss (mathematical residual of each differential equation). By minimizing these two losses through backpropagation, the weights of the LSTM network are constantly optimized, ensuring that the model's prediction results not only closely match the actual measured data but also strictly adhere to the physical laws of energy conservation, ultimately achieving the unity of accuracy and physical consistency.

[0147] Figure 4 Through 100 sets of time series data, the performance differences between the physical information neural network modeling method proposed in the invention and the traditional mechanism modeling method - differential equation modeling method in predicting the primary network water supply temperature are intuitively compared.

[0148] Figure 4 The above figure shows the comparison of temperature prediction curves. The black solid line represents the true value of the system, and its fluctuations reflect the complex changes of the actual heating load. The gray dotted line represents the prediction results of the differential equation model, and it can be seen that there is a significant phase lag and amplitude error, especially in the peak and valley regions of rapid load changes, the prediction deviation is larger. The gray dashed line represents the prediction results of the physical information neural network model of the invention, and its curve is significantly better fitted to the true value than the differential equation modeling method.

[0149] Figure 4The following figure further quantifies the absolute error of the two methods. It can be clearly seen that the gray error column representing the differential equation model is generally high and fluctuates greatly; while the black error column representing the model of the application always maintains a low level close to zero, and has good stability.

[0150] The performance indicators marked in the figure show that the RMSE of the application is much lower than that of the traditional model-differential equation model, and the accuracy is improved by more than 90%. This comparison proves that the application effectively overcomes the defects of parameter solidification and poor adaptability of traditional mechanism models by fusing physical equations and deep learning, and realizes high-precision and high-robustness modeling of the dynamic behavior of the system.

[0151] Figure 5 A structural schematic diagram of a computer device disclosed by the application is shown in the figure. Figure 5 The computer device 400 at least includes a memory 402 and a processor 401; the memory 402 is connected with the processor through a communication bus 403, and is used for storing computer instructions executable by the processor 401; and the processor 401 is used for reading the computer instructions from the memory 402 to realize the steps of the method described in any of the above embodiments.

[0152] For the above-mentioned device embodiments, since they basically correspond to the method embodiments, the related parts are described in the part of the method embodiments. The above-described device embodiments are only schematic, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. According to actual needs, part or all of the modules can be selected to achieve the purpose of the present application. Those skilled in the art can understand and implement it without creative labor.

[0153] The computer readable medium suitable for storing computer program instructions and data includes all forms of non-volatile memory, media and memory devices, such as semiconductor memory devices (such as EPROM, EEPROM and flash memory devices), magnetic disks (such as internal magnetic disks or removable disks), magneto-optical disks and CD ROM and DVD-ROM disks. The processor and the memory can be supplemented by or incorporated into special logic circuits.

[0154] Finally, it should be noted that although this specification contains many specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily used to describe the features of specific embodiments of a particular invention. Certain features described in the various embodiments of this specification may also be implemented in combination in a single embodiment. On the other hand, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation of a sub-combination.

[0155] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0156] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0157] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A hybrid modeling method for a cogeneration cascade heat supply system of a thermal power plant, characterized in that, The method comprises the following steps: Collecting real-time and historical working condition data for constructing a physical information sub-model of a heat supply network and a physical information sub-model of a heat source; Both of the two sub-models dynamically identify and predict key time-varying parameters in a mechanism model from data by using a deep learning network, embed the key time-varying parameters into a mechanism differential equation group constructed based on a first principle, minimize data loss and physical loss by using an optimizer, thereby back-propagating and updating neural network weights.

2. The hybrid modeling method of claim 1, wherein, The method for constructing the physical information sub-model of the heat supply network comprises: Dynamically learning and predicting a key parameter, i.e., a comprehensive heat dissipation coefficient K, in a primary pipe network; Embedding the predicted K into a mechanism differential equation established based on the law of conservation of energy, solving the equation by using a numerical integration method, and finally calculating a predicted supply water temperature; The training and optimization of the model are achieved by minimizing data loss and physical loss by using an optimizer containing data loss and physical loss, thereby back-propagating and updating neural network weights.

3. The hybrid modeling method of claim 2, wherein, The method for dynamically learning and predicting the key parameter, i.e., the comprehensive heat dissipation coefficient K, in the primary pipe network comprises: According to the heat balance principle of water temperature change in a pipeline, an energy conservation equation is established: wherein p represents the water supply density (kg / m 3 ) ; Cp represents the specific heat capacity of water (J / (kg.°C) ) ; V represents the total volume of water in the pipe (m 3 ) ; represents the rate of change of the water supply temperature with time (°C / s) ; Q in represents the input heat power, i.e. the amount of heat delivered by the heat source into the pipe network (W) ; Q out represents the output heat power, i.e. the amount of heat extracted by the user side (W) ; Q loss represents the amount of heat dissipated by the pipe into the environment (W) ; A pipeline heat dissipation equation is established: Q loss = K x A x (T supply -T amb ) (2) K is the comprehensive heat dissipation coefficient (W / (m 2 ), a complex parameter related to multiple factors such as flow rate, heat preservation material, and environmental wind speed, and the present application uses deep learning to dynamically predict the coefficient; A represents the pipe heat dissipation surface area (m 2 ); T supply represents the water supply temperature; and T amb represents the environmental temperature (℃).

4. The hybrid modeling method of claim 3, wherein, The method for calculating the predicted supply water temperature and model optimization comprises: The calculation Q loss Substituting equation into (1) and rearranging, the differential equation of the control system is obtained: Then the synthetic heat dissipation coefficient K is dynamically predicted by a long short-term memory neural network, the input of which is represented as X t At each time step t, the input features of the network contain all the measurable variables that affect K, the number of input features of the input layer of the network structure is equal to the number of nodes, the LSTM layer is 1-2 layers, and there are 64-128 neurons in each layer to capture the time dependence; the fully connected layer is used to map the LSTM output to the final output; the output layer uses the ReLU activation function to ensure that the value of K is non-negative; A loss function L is composed of three parts: wherein L physics represents a constraint function value of a physical equation; represents an actual value of a supply water temperature change rate; represents a predicted value of a supply water temperature change rate; wherein L data represents a constraint function value of the deep network; T supply-dnn represents a predicted value of the water supply temperature output by the neural network; L total = a x L physics + b x L data (6) wherein L total represents the total loss function value of the physical information network model; a represents the physical loss weight; β represents the data loss weight; the loss function of formula (6) is the weighted sum of the physical loss and the data loss; By gradient descent algorithm, the total loss L is minimized total The weights of the LSTM network are updated simultaneously, and the predicted supply water temperature is finally calculated.

5. The hybrid modeling method of claim 1, wherein, The method for constructing the physical information sub-model of the heat source comprises: dynamically predicting output deaerator heat exchanger efficiency η by a deep learning model dc , heat pump coefficient of performance COP ahp and extraction heat exchanger heat transfer coefficient K esh ; Embedding the three parameters in real time into core differential equations describing a three-stage heating process, obtaining predicted values of outlet temperatures of each stage by numerical solution, minimizing data loss and physical loss by using an optimizer, thereby back-propagating and updating neural network weights.

6. The mixed modeling method of claim 5, wherein, An efficiency η of a steam exhaust heat exchanger dc A coefficient of performance COP of a heat pump ahp A heat transfer coefficient K of a steam extraction heat exchanger esh An expression method includes: The return water absorbs the latent heat released by the condensation of the steam, and the temperature increases from T return to T H1 . C dc = p x C p x V dc (8) where C dc represents the heat capacity of the deaerator heat exchanger and related piping water (J / °C); T H1 represents the deaerator heat exchanger outlet water temperature, i.e. the first stage heating after water temperature (°C); t represents time (s); m dc represents the deaerator heat exchanger inlet water mass flow rate (kg / s); h steam represents the specific enthalpy of the deaerator heat exchanger (J / kg); h water represents the specific enthalpy of the deaerator condensate (J / kg); η dc represents the deaerator heat exchanger thermal efficiency, which is a factor close to 1 but difficult to determine accurately; m W represents the network circulating water mass flow rate (kg / s); C p represents the specific heat capacity of water at constant pressure (J / (kg·°C)); T return represents the heat network return water temperature (°C); The heat pump uses a driving heat source to pump the heat of a low-temperature heat source to a medium-temperature return water, so that the temperature of the return water is raised from T H1 to T H2 . where C ahp represents the heat capacity of the heat pump and related piping water (J / °C); T H2 represents the heat pump outlet water temperature, i.e. the water temperature after the second stage heating (°C); COP ahp represents the performance coefficient of the heat pump; Q drive represents the heat power of the heat pump driving heat source (W): Q drive = m drive × (h drive - h cond ) (10) wherein m drive is the driving steam flow rate; With the higher grade extraction steam the water is finally heated to the set feed water temperature T supply : wherein the logarithmic mean temperature difference ΔT lm calculated as: where C esh represents the heat capacity of the heat pump and related pipes (J / ℃); T H2 represents the outlet water temperature of the heat pump, i.e. the water temperature after the second stage heating (℃); T supply represents the primary network water supply temperature (℃); K esh represents the comprehensive heat transfer coefficient of the extraction steam heat exchanger (W / (m 2 ·℃)); A esh represents the heat exchange area of the extraction steam heat exchanger (m 2 ); T steam represents the saturation temperature of the extraction steam (℃).

7. The mixed modeling method of claim 6, wherein, The method for solving the predicted values of the outlet temperatures of each stage and the method for optimization by using the optimizer comprise the following steps: Predicting η by long short-term memory neural network dc , COP ahp , K esh , the input features X of the current and historical moments {t-n:t} are sent into the LSTM network to obtain the predicted values η of the key parameters {dc,pred} , COP {ahp,pred} , K {esh,pred} ; these predicted parameters are substituted into the aforementioned mechanism differential equation set; the equation set is solved by using a numerical integration method to calculate the predicted values T of the node temperatures H1_pred , T H2_pred , T supply_pred : Design a hybrid loss function L total Perform backpropagation, optimize neural network weights: L total = a x L data + b x L physics (13) L data = MSE(T H1_true , T H1_pred ) + MSE(T H2_true , T H2_pred ) + MSE(T supply_true , T supply_pred )(14) L physics represents the physical constraint error: the residual of the left and right sides of the differential equation is calculated and minimized; for the steam extraction heat exchanger equation, the physical residual is: The equation residuals are weighted and summed to obtain L physics The total loss L is minimized by a gradient descent algorithm total The solution for the predicted outlet temperature at each level is obtained.

8. Hybrid modeling system for a hybrid modeling method for a cogeneration cascade heat supply system of a thermal power plant according to any one of claims 1 to 7, characterized in that The method comprises: A construction module is configured to collect real-time and historical working condition data for constructing a physical information sub-model of a heat supply network and a physical information sub-model of a heat source; An optimization module is configured to dynamically identify and predict key time-varying parameters in a mechanism model from data by using a deep learning network for both of the two sub-models, embed the key time-varying parameters into a mechanism differential equation group constructed based on a first principle, minimize data loss and physical loss by using an optimizer, thereby back-propagating and updating neural network weights.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is run by a processor to execute the steps of the hybrid modeling method for a cascade heat supply system of a cogeneration power plant according to any one of claims 1-7.

10. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the hybrid modeling method for a cascade heat supply system of a cogeneration power plant according to any one of claims 1-7.

Images