A boiler water wall temperature field on-line prediction system and method

Abstract

The application discloses a boiler water cooling wall temperature field online prediction system and method, and belongs to the technical field of computer systems based on specific calculation models. The system comprises a data acquisition module, a physical priori encoder-decoder module, a phase state sensing heat transfer calculation graph network module, a physical residual driven hidden space online adaptive closed loop module and a result output warning module. The encoder structure maps the operation data into a low-dimensional hidden space vector, and orthogonally decouples it into three groups of sub-vectors and applies constraints; the decoder structure outputs the furnace heat flow density and the coking thermal resistance distribution. The calculation graph network module solidifies the heat transfer equation into fixed calculation nodes, forms a thermal load boundary through pipe section discrete sampling, and solves the full screen wall temperature in combination with phase state discrimination. The closed loop module constructs a residual, iteratively updates the hidden space vector and returns the iteration. The application can reconstruct the global temperature field with a small number of measuring points, and can realize water cooling wall temperature reconstruction, coking identification and over-temperature early warning.

Description

Technical Field

[0001] This invention belongs to the field of computer system technology based on specific calculation models, and specifically relates to an online prediction system and method for boiler water-cooled wall temperature field. Background Technology

[0002] Supercritical and ultra-supercritical once-through boilers frequently experience load fluctuations within a wide load range of 30%–100% BMCR, causing the working fluid inside the water-cooled wall tubes to undergo three distinct heat transfer modes: subcritical two-phase flow, critical phase transition, and supercritical single-phase flow. Overheating and thermal deviation of the fire-facing wall temperature are the primary causes of tube rupture failure, and accurate online prediction of the wall temperature is the foundation for preventative control.

[0003] Existing water-cooled wall temperature prediction / monitoring technologies mainly present four technical routes, but all of them have insurmountable engineering bottlenecks:

[0004] Category 1: Monitoring technologies based on hardware-based measurement point encryption and heat transfer inverse calculation

[0005] This type of technology adds a large number of thermocouple measuring points to the water-cooled wall tube screen outside the furnace (such as arranging measuring points at each throttling orifice ring corresponding to the outlet of the spiral water-cooled wall and the vertical water-cooled wall, or adding layered measuring points on the backfire side of the tube screen), and combines one-dimensional steady-state heat conduction to back calculate the fire-facing wall temperature, and then infers the coking location by calculating the enthalpy increase. Its inherent defects are: (1) Thermocouples can only be arranged on the tube section outside the furnace and cannot cover the fire-facing wall surface of the main combustion zone of the furnace; (2) The measuring point density is strictly limited by hardware cost and installation space. For a single boiler with thousands of water-cooled wall tubes, sparse measuring points cannot capture local thermal deviations; (3) The back calculation model assumes that the heat flux density is known or uniform along the tube length, but the actual heat load in the furnace is strongly non-uniformly distributed, resulting in large back calculation errors; (4) It cannot predict and can only achieve post-event monitoring.

[0006] The second category: Machine learning prediction techniques based on pure data-driven approaches.

[0007] Such technologies use historical operating data to train neural networks (such as LSTM, GRU, SVM, etc.), and use DCS measurable parameters (load, coal quantity, air volume, water supply, etc.) as input to predict wall temperature or heat flux density. Its inherent defects are: (1) The model is a black box, lacks physical interpretability, and cannot guarantee that the prediction results meet the basic physical constraints such as energy conservation and heat transfer law. The prediction reliability drops sharply under extrapolated operating conditions outside the training data distribution (such as extremely low loads or new coal types that have never been run before); (2) It cannot distinguish the physical root causes of wall temperature changes - the coupling effect between combustion heat load disturbance, coking thermal resistance increase, and heat transfer deterioration is submerged in the data association, causing the model to fail under new coal quality or different burner operating conditions; (3) It relies on a large amount of labeled data, but the "true value" of the full-screen temperature field of the water-cooled wall is not available in actual engineering, and cannot form an effective monitoring signal.

[0008] Category 3: Offline computing techniques based on high-fidelity numerical simulation (CFD)

[0009] This type of technology simulates combustion and heat transfer in the furnace through computational fluid dynamics to obtain the heat flux density distribution on the surface of the water-cooled wall, and then couples it with the hydrodynamic calculation inside the tube to obtain the wall temperature field. Its inherent defects are: (1) a single full furnace CFD calculation takes several hours to tens of hours, which cannot meet the real-time requirements of online prediction; (2) the CFD boundary conditions (such as coal element analysis, burner wear status, furnace coking morphology) change dynamically in actual operation and are difficult to obtain online, resulting in a systematic deviation between the offline calculation results and the actual operation; (3) the CFD model and the hydrodynamic model inside the tube are loosely coupled, which makes it difficult to reflect the dynamic coupling process of heat and mass transfer inside and outside the tube.

[0010] Category 4: Fusion modeling techniques based on Physics-Informed Neural Networks (PINNs)

[0011] PINNs is a technique that has emerged in recent years to embed physical equations into machine learning. It uses the residuals of partial differential equations describing physical laws as the regularization term of the loss function and optimizes them together with the data fitting loss. However, PINNs has fundamental defects when applied to boiler water-cooled wall temperature prediction: (1) Physical constraints are only soft constraints: The residuals of physical equations are balanced with the data loss in the loss function through weight coefficients, which means that physical constraints can only be approximately satisfied, not strictly satisfied; the adjustment of weight coefficients is extremely difficult, and the dimensions and numerical scales of different loss terms are very different, which easily leads to imbalance phenomena such as "physical terms being submerged by data terms" or "insufficient data fitting". (2) The training time is unacceptable: PINNs requires solving the residuals of high-order differential equations, and a single training takes several hours to tens of hours. Moreover, each change in operating conditions or update of boundary conditions requires retraining, which cannot meet the requirements of online real-time prediction at all. (3) Unable to handle hard switching of phase transitions: The heat transfer correlation in the water-cooled wall tube undergoes a fundamental switch at the subcritical-supercritical phase transition. The soft regularization of PINNs cannot characterize this hard switching behavior, resulting in a non-physically smooth transition in the critical region prediction. (4) Latent space vectors lack interpretability: The neural network weights and hidden layer features of PINNs are still black boxes without physical semantics, and cannot decouple coking and combustion disturbances.

[0012] In boiler design, engineers are well aware that furnace heat flux density is spatially non-uniformly distributed. They often employ principal component analysis (PCA), nonnegative matrix factorization (NMF), or orthogonal polynomial expansion to reduce the dimensionality of the heat flux density field under design conditions, extracting several spatial modal basis functions to facilitate design calculations and combustion optimization analysis. Furthermore, in the fields of furnace flame image processing and combustion diagnostics, there is also a common practice of extracting flame characteristic modes using dimensionality reduction techniques.

[0013] However, the aforementioned modal decomposition methods are currently only used as data processing tools in the design or post-analysis phases. The analysis results are presented to engineers in the form of reports, charts, or offline databases for design verification or operational diagnostics, but they do not participate in the real-time prediction calculation process, nor are they solidified into the structural parameters of the online prediction model topology. Existing technologies rarely consider that these modal basis functions can be reconstructed as weight parameters for the prediction system decoder, forming an end-to-end differentiable real-time prediction system together with the encoder, physical computation graph, and online adaptive closed loop. Transforming these "post-analysis tools" into "intrinsic components of the online prediction architecture" requires breaking through the traditional cognitive framework that separates the "analysis-operation" two-stage process; this is precisely one of the long-standing technological biases.

[0014] In summary, existing technologies suffer from limitations such as limited monitoring range, poor physical consistency, weak adaptability to operating conditions, inability to decouple multiple coupling interferences, and insufficient predictive stability, making it difficult to meet the precise monitoring requirements for the safe and stable operation of power units. Statistics show that unplanned outages caused by water-cooled wall overheating and tube rupture in supercritical and ultra-supercritical units can result in direct economic losses of several million to tens of millions of yuan per incident, while also creating a gap in the power grid's peak-shaving capacity, severely impacting the economic efficiency of thermal power unit operation and the stability of power grid supply. Therefore, addressing these technical pain points requires the development of an online prediction system and method for boiler water-cooled wall temperature field that balances physical constraints with data self-adaptation capabilities and enables precise reconstruction of the entire temperature field. This has urgent engineering application value, economic value, and power grid security value. Summary of the Invention

[0015] The purpose of this invention is to provide an online prediction system and method for boiler water-cooled wall temperature field, in order to solve the technical problem that traditional monitoring and prediction schemes cannot simultaneously meet the requirements of physical consistency, high prediction accuracy and long-term stable operation of the full-screen fire-facing wall temperature prediction of the water-cooled wall without increasing the density of measuring points in the furnace or adding in-furnace sensors.

[0016] The specific solution adopted in this invention is as follows:

[0017] On the one hand, the present invention provides an online prediction system for the temperature field of boiler water-cooled walls, the system comprising: a data acquisition module, a physical prior encoder-decoder module, a phase-sensing heat transfer calculation graph network module, a physical residual-driven latent space online adaptive closed-loop module, and a result output and early warning module;

[0018] The data acquisition module is used to collect time-series data of boiler sparse wall temperature measurement points, unit operating condition parameters, and soft measurement parameters of coal quality, and load prior knowledge of water-cooled wall structure to integrate and generate model input data.

[0019] The physical prior encoder-decoder module is connected to the data acquisition module and includes an encoder structure and a decoder structure. The encoder structure extracts spatiotemporal features from the model input data, maps them to obtain a low-dimensional latent space vector, and explicitly orthogonally decouples the low-dimensional latent space vector into a combustion heat load basic coefficient vector, a coking thermal resistance disturbance coefficient vector, and a heat transfer deterioration indicator coefficient vector that are physically and semantically independent. Physical value constraints are applied to the three types of sub-vectors. The decoder structure decodes and outputs the furnace full-screen heat flux density distribution and the water-cooled wall coking thermal resistance distribution based on the decoupled combustion heat load basic coefficient vector and coking thermal resistance disturbance coefficient vector.

[0020] The phase-state sensing heat transfer calculation graph network module is communicatively connected to the physical prior encoder-decoder module, and has a built-in fixed physical calculation graph, pipe segment discrete unit, and phase state discriminator. The fixed physical calculation graph encodes the one-dimensional unsteady enthalpy recursive equation of the working fluid in the pipe, the radial heat conduction equation of the pipe wall, and the working fluid state equation into non-learnable fixed calculation graph nodes. The pipe segment discrete unit discretizes the water-cooled wall pipe along the pipe length into several independent pipe segment control bodies. Based on the spatial coordinate mapping relationship of the water-cooled wall pipe system, it samples and converts the decoded output furnace full-screen heat flux density distribution and water-cooled wall coking thermal resistance distribution to obtain the off-center heat flux density value and discrete coking thermal resistance value corresponding to each pipe segment control body, and uses them as the thermal load boundary input of each pipe segment control body. The phase state discriminator identifies the real-time phase state of the working fluid in the pipe, adaptively switches the heat transfer correlation and corrects the local heat transfer coefficient by combining the heat transfer deterioration indicator coefficient, and calculates the predicted value of the fire-facing side wall temperature of the water-cooled wall full-screen according to the law of conservation of energy.

[0021] The physical residual-driven latent space online adaptive closed-loop module is used to construct physical residuals based on measured wall temperatures and predicted wall temperatures. The latent space vector correction is obtained through constraint optimization. Under the premise that the network weights of the encoder and decoder structures are fixed, the low-dimensional latent space vector is updated online. The updated latent space vector is then fed back to the decoder structure and phase-sensing heat transfer calculation graph network module to achieve iterative closed-loop optimization. At the same time, the temperature field results after iterative optimization are pushed to the result output and early warning module.

[0022] Preferably, the result output and early warning module is used to output the full-screen temperature field, thermal deviation index and coking distribution area of ​​the water-cooled wall in real time, so as to realize dynamic online prediction and over-temperature early warning of the water-cooled wall temperature field.

[0023] Preferably, the data acquisition module establishes real-time communication with the power plant's DCS system via the OPC or MQTT protocol to complete the real-time acquisition of time-series data from sparse wall temperature measurement points, unit operating condition parameters, and soft coal quality measurement parameters.

[0024] Preferably, the encoder structure incorporates a spatiotemporal graph attention network, a latent space orthogonal decoupling module, and a physical constraint embedding layer. The spatiotemporal graph attention network is composed of a multi-layer graph attention network and a gated recurrent unit, used to extract the spatial correlation features of the water-cooled wall pipe system and the temporal evolution features of the operating data, and generate low-dimensional latent space vectors. The latent space orthogonal decoupling module is used to complete the orthogonal decoupling of the low-dimensional latent space vectors, obtaining three types of sub-vectors. The physical constraint embedding layer is used to apply physical value constraints to the three types of sub-vectors after decoupling.

[0025] Preferably, the physical value constraints applied by the physical constraint embedding layer include: applying non-negative constraints to the combustion heat load base coefficient vector and the coking thermal resistance disturbance coefficient vector, adding time-series smoothing constraints to the coking thermal resistance disturbance coefficient vector, and applying bounded constraints in the [0,1] interval to the heat transfer deterioration indicator coefficient vector. All physical value constraints are implemented through a differentiable soft constraint layer.

[0026] Preferably, each pipe section control body is equipped with an independent set of state variables, which includes the pipe wall temperature on the fire side, the pipe wall temperature on the back side, the working fluid temperature inside the pipe, the enthalpy of the working fluid inside the pipe, the pressure inside the pipe, the mass vapor content, or the working fluid density.

[0027] Preferably, the phase discriminator divides the working fluid into four regions based on the working fluid pressure and enthalpy: subcritical subcooled water region, subcritical two-phase region, critical transition region, and supercritical single-phase region. It uses the Gumbel-Softmax algorithm to achieve differentiable hard branch switching, and each phase region is matched with a dedicated heat transfer correlation. The heat transfer deterioration indicator coefficient vector is mapped to each pipe segment control volume through a learnable fully connected layer or a fixed partition mask matrix, including two mapping modes: pipe-by-pipe allocation and spatial clustering partitioning.

[0028] In the tube-by-tube allocation mode, the dimension of the heat transfer deterioration indicator coefficient vector is equal to the total number of water-cooled wall tubes; in the spatial clustering partitioning mode, the dimension of the heat transfer deterioration indicator coefficient vector is less than the total number of water-cooled wall tubes, and is allocated to each tube segment control body according to the furnace wall clustering partitioning results.

[0029] Preferably, the physical residual-driven latent space online adaptive closed-loop module constructs a physically constrained regularized least squares optimization model and configures a differentiated regularization strategy.

[0030] The differentiated regularization strategy is as follows: apply time-smoothing L2 regularization to the combustion heat load base coefficient vector, apply L1 sparse regularization to the coking thermal resistance disturbance coefficient vector, and apply conditionally activated L2 regularization to the heat transfer deterioration indicator coefficient vector.

[0031] The physical constraints include a base coefficient for combustion heat load, a non-negative coefficient for coking thermal resistance disturbance, and a heat transfer deterioration indicator coefficient with values ​​in the range [0,1].

[0032] Preferably, the physical residual-driven latent space online adaptive closed-loop module applies a first-order inertial time-series filter to the updated latent space vector.

[0033] On the other hand, the present invention also proposes an online prediction method for the temperature field of a boiler water-cooled wall, which is applied to the above-mentioned online prediction system for the temperature field of a boiler water-cooled wall, and includes the following steps:

[0034] S1. Data Acquisition: The data acquisition module collects real-time time-series data of boiler sparse wall temperature measurement points, unit operating condition parameters, and soft coal quality measurement parameters. Combined with the preset prior knowledge of water-cooled wall structure, the model input data is integrated and generated.

[0035] S2. Spatiotemporal feature extraction and latent space decoupling: Spatiotemporal features are extracted from the model input data through the encoder structure of the physical prior encoder-decoder module, and a low-dimensional latent space vector is obtained. The low-dimensional latent space vector is explicitly orthogonally decoupled into a combustion heat load base coefficient vector, a coking thermal resistance disturbance coefficient vector, and a heat transfer deterioration indicator coefficient vector that are physically and semantically independent, and physical value constraints are applied to the three types of sub-vectors respectively.

[0036] S3, Full-field physical field decoding and reconstruction: Through the decoder structure of the physical prior encoder-decoder module, combined with the pre-constructed combustion heat load mode basis function and coking thermal resistance disturbance basis function, the decoupled combustion heat load basis coefficient vector and coking thermal resistance disturbance coefficient vector are decoded and calculated respectively, and the corresponding output furnace full-screen heat flux density distribution and water-cooled wall coking thermal resistance distribution are output.

[0037] S4. Phase-aware heat transfer solution: The phase-aware heat transfer calculation graph network module calls the built-in fixed physical calculation graph to encode the one-dimensional unsteady enthalpy recursive equation of the working fluid in the tube, the radial heat conduction equation of the tube wall, and the working fluid state equation into unlearnable fixed calculation graph nodes; the water-cooled wall tube is discretized into several independent tube segment control bodies along the tube length through the tube segment discretization unit; based on the spatial coordinate mapping relationship of the water-cooled wall tube system, the furnace full-screen heat flux density distribution and water-cooled wall coking thermal resistance distribution output by the decoder structure are sampled and converted to obtain the boundary parameters of the heat dissipation load corresponding to each tube segment control body; the phase discriminator identifies the real-time phase state of the working fluid inside each tube segment control body, adaptively switches the appropriate heat transfer correlation, and uses the heat transfer deterioration indicator coefficient to correct the local heat transfer coefficient; the predicted value of the fire-facing side wall temperature of the water-cooled wall is calculated according to the law of conservation of energy.

[0038] S5. Physical Residual-Driven Adaptive Update of Latent Space: Through the physical residual-driven online adaptive closed-loop module of latent space, physical residuals are constructed by comparing the measured wall temperature and the predicted wall temperature at sparse measurement points. An optimization model is constructed by combining a preset differential regularization strategy and physical constraints. The latent space vector correction is solved by automatic differentiation across the entire link. Under the premise that the overall network weights of the encoder and decoder structures remain unchanged, the online update of the low-dimensional latent space vector is completed, and a first-order inertial time-series filter is applied to the updated latent space vector.

[0039] S6. Closed-loop iterative prediction and early warning: The filtered and updated low-dimensional latent space vector is fed back to the decoder structure and phase-sensing heat transfer calculation graph network module to iteratively update the furnace heat flux density, the full-field distribution of coking thermal resistance of water-cooled walls, and the local heat transfer coefficient, forming a closed-loop iterative calculation to continuously and dynamically update the full-screen temperature field of water-cooled walls; the full-screen temperature field of water-cooled walls, thermal deviation index, and coking distribution area are output in real time, and combined with the time sequence trend of unit load command, short-term early warning of over-temperature and visualization monitoring of furnace heating surface status are realized.

[0040] Compared with the prior art, the present invention has at least one of the following advantages or beneficial effects:

[0041] I. This invention, through a phase-aware heat transfer computational graph network module, solidifies the recursive law of the unsteady enthalpy value of the working fluid inside the water-cooled wall tube, the radial thermal conductivity characteristics of the tube wall, and the standard working fluid state equation into fixed, non-learnable computational graph nodes. This ensures that throughout the entire process of adaptive iterative updating of the model's latent space vector, regardless of how the low-dimensional latent space vector is adaptively updated, the solution calculation of the water-cooled wall temperature field always strictly follows the principle of energy conservation and the fundamental laws of heat transfer. Compared to the soft constraint scheme of traditional Physical Information Neural Networks (PINNs) that uses physical equations as regularization terms in the loss function, this invention, in the online operation phase, is a structured physical constraint that is unweighted, parameter-tuned, and precisely satisfied. It overcomes the industry pain points of PINNs models, such as approximate satisfaction of physical constraints, difficulty in adjusting loss weights, imbalance of multi-task losses, and susceptibility to non-physical prediction results, thus guaranteeing the physical authenticity of the prediction results from the architectural level.

[0042] Second, this invention decouples low-dimensional latent space vectors through structured orthogonal decoupling, resulting in three types of sub-vectors with independent and clear physical semantics: the combustion heat load basic coefficient vector, the coking thermal resistance disturbance coefficient vector, and the heat transfer deterioration indicator coefficient vector. This enables the model to possess intrinsic physical fault root cause identification capabilities, unlike traditional models which only output temperature results and lack attribution capabilities—a black box effect. When abnormal wall temperature rises locally in water-cooled walls, this invention can accurately trace the source through the spatial distribution characteristics of the corresponding sub-vectors. Specifically, if the coking thermal resistance sub-vector is abnormally high while the combustion sub-vector is normal, it is determined that the abnormal wall temperature is caused by coking ash accumulation, and soot blowing optimization and appropriate load reduction control strategies can be matched. If the combustion heat load sub-vector distribution is distorted, it is determined that the overheating is caused by furnace combustion deviation and uneven air distribution, which can guide the optimization of air distribution parameters. This identification capability is an inherent attribute of the model architecture, requiring no secondary attribution afterward, and can directly output actionable unit control suggestions, effectively avoiding control misjudgments and inaccuracies.

[0043] Third, this invention applies L1 sparse regularization constraints to the coking thermal resistance disturbance subvector. As can be seen from mathematical optimization principles, L1 regularization can force the sparsity of the solution, retaining only a small number of non-zero basis function coefficients. This corresponds to coking in the furnace occurring only in localized areas, which highly matches the true physical characteristics of boiler coking: "localized patchy distribution, not global coverage." Compared to traditional prediction schemes that use global uniform fitting and are prone to false coking areas, this invention ensures the accuracy and authenticity of coking location and range identification from both mathematical and physical perspectives.

[0044] IV. The overall architecture of this invention possesses millisecond / second-level high-efficiency computing characteristics, fully adapting to the real-time requirements of power plant online monitoring. The encoding and decoding network is a feedforward structure, with single spatiotemporal feature mapping and latent space vector decoding taking milliseconds. The phase-sensing heat transfer calculation graph is a one-dimensional chain-like fixed calculation structure, enabling second-level forward extrapolation for refined calculations involving thousands of water-cooled wall tubes and multiple control bodies per tube in boilers. The latent space correction employs regularized least squares solution. The total dimension of the low-dimensional latent space is typically less than 50, and the L1 regularization term can be equivalently transformed into linear inequality constraints, which can then be solved using the quadratic programming active set method or interior point method. Alternatively, the proximal gradient method or alternating direction multiplier method can be used for rapid iterative solution. The end-to-end closed-loop update cycle is in the minute range. Compared to traditional CFD offline simulations that take several hours and PINNs training iterations that take tens of hours, the real-time advantage is significant, fully meeting the engineering application requirements for online monitoring and dynamic early warning of the unit's water-cooled wall temperature field.

[0045] V. This invention overcomes the limitations of sparse measurement points, achieving precise reconstruction of the full-screen temperature field with "few measurements, many understandings," thus overcoming the inherent shortcomings of traditional monitoring technologies, which are "knowledgeable only at measurement points, unknown when no measurement points are available," and cannot cover the entire furnace blind zone. Based on the concept of physical prior mode decomposition, this invention reduces the complex furnace heat flux density distribution to a low-dimensional combustion basis coefficient vector. Only 5-10 sets of combustion mode basis functions are needed to cover more than 90% of the furnace heat flux density distribution variance. Combined with the temperature constraint equations of sparse measurement points in the field, the low-dimensional latent space vector is accurately solved through an optimized model, achieving full-field reconstruction under the constraint of a small number of measurement points.

[0046] VI. This invention reconstructs the causal reasoning chain of operating parameters, implicit physical states, and temperature fields, abandoning the simple data correlation fitting mode. The intermediate heat transfer calculation link is constrained by rigid physical laws and does not change with data distribution. For various unseen extreme operating conditions such as unit changes in coal quality, load, and new coal mill combinations, as long as the operating parameters can be mapped to a reasonable implicit space range, it can output physically reasonable and reliable temperature field results. This fundamentally solves the pain points of traditional AI models' weak operating condition transfer ability and failure in extreme operating condition prediction, and is suitable for the complex operating scenarios of thermal power units with wide loads, multiple coal types, and dynamic changing operating conditions. Attached Figure Description

[0047] The invention, its features, shape, and advantages will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. Like reference numerals denote like parts throughout the drawings. The drawings are not drawn to scale; the emphasis is on illustrating the gist of the invention.

[0048] Figure 1 This is a diagram of the online prediction system architecture for the boiler water-cooled wall temperature field in Embodiment 1 of the present invention.

[0049] Figure 2 This is a one-dimensional chain calculation diagram of a single water-cooled wall tube discretized into multiple control volumes along the tube length in Embodiment 1 of the present invention;

[0050] Figure 3 This is a schematic diagram of the physical residual-driven latent space online adaptive closed loop in Embodiment 1 of the present invention. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but these are not intended to limit the scope of the invention.

[0052] The core concept of this invention is to construct a "physical-data fusion" latent space representation framework, which encodes the spatiotemporal distribution of furnace-side heat flux density, coking thermal resistance distribution, and tube heat transfer coefficient distribution, which cannot be directly measured, into low-dimensional latent space vectors. The online adaptive update of the low-dimensional latent space vectors is driven by a physical residual closed loop, and then the updated low-dimensional latent space vectors are decoded into the full-screen water-cooled wall temperature field.

[0053] Based on this concept, the present invention proposes three interdependent and inseparable modules:

[0054] Module A: Physical-Prior Encoder-Decoder Module (PPED)

[0055] Module B: Phase-Aware Heat Transfer Computational Graph Network (PAHT-CGN)

[0056] Module C: Physics-Residual DrivenOnline Adaptive Closed-Loop Module (PRDOA)

[0057] The three modules are coupled through the unique interface of the latent space vector, forming a closed-loop prediction system of "encoding-computation-decoding-residual backpropagation-latent space correction". The latent space vector is the only bridge connecting the data-driven layer and the physical computing layer. Without this bridge, data and physics cannot achieve deep end-to-end integration.

[0058] Example 1:

[0059] Reference Figure 1 As shown, this embodiment provides an online prediction system for the temperature field of boiler water-cooled walls. The system specifically includes: a data acquisition module, a physical prior encoder-decoder module (module A), a phase-sensing heat transfer calculation graph network module (module B), a physical residual-driven latent space online adaptive closed-loop module (module C), and a result output and early warning module.

[0060] This data acquisition module is used to collect time-series data of boiler sparse wall temperature measurement points, unit operating condition parameters, and soft measurement parameters of coal quality, and load prior knowledge of water-cooled wall structure to integrate and generate model input data.

[0061] Specifically, the time-series data of the boiler sparse wall temperature measurement points are taken from the DCS external water-cooled wall outlet or intermediate header. Temperature sequence of thermocouple measuring points ,in For the measurement point number index, , This represents the total number of water-cooled wall tubes;

[0062] Unit operating parameters include unit load commands. Total coal supply Total air volume Primary air volume Secondary air volume Water supply flow rate Water supply temperature Furnace negative pressure , the sway angle of each burner wait.

[0063] Coal quality soft measurement parameters include the received basis lower heating value identified by an online coal quality analyzer or operational data. ash melting point characteristic temperature Ash content , sulfur content wait;

[0064] Prior knowledge of water-cooled wall structures includes the tube panel geometry matrix (tube length) Pipe diameter Wall thickness Thermal conductivity of materials Pitch (Circuit topology connection relationship) and furnace three-dimensional coordinate mapping matrix.

[0065] The physical prior encoder-decoder module connects to the data acquisition module and includes an encoder structure and a decoder structure. The encoder structure extracts spatiotemporal features from the model input data and maps them to obtain a low-dimensional latent space vector. It then explicitly orthogonally decouples the low-dimensional latent space vector into a combustion heat load basic coefficient vector, a coking thermal resistance disturbance coefficient vector, and a heat transfer deterioration indicator coefficient vector, which are physically and semantically independent. Physical value constraints are applied to the three types of sub-vectors. Based on the decoupled combustion heat load basic coefficient vector and coking thermal resistance disturbance coefficient vector, the decoder structure decodes and outputs the furnace full-screen heat flux density distribution and the water-cooled wall coking thermal resistance distribution.

[0066] Specifically, the encoder structure incorporates a Spatio-Temporal Graph Attention Network (ST-GAT), a latent space orthogonal decoupling module, and a physical constraint embedding layer. The ST-GAT consists of a multi-layer graph attention network and a gated recurrent unit, used to extract the spatial correlation features of the water-cooled wall pipe system and the temporal evolution features of the operating data, respectively, and obtain a low-dimensional latent space vector through dimensionality reduction mapping. The latent space orthogonal decoupling module is used to explicitly orthogonally decouple the low-dimensional latent space vector into a combustion heat load base coefficient vector, a coking thermal resistance disturbance coefficient vector, and a heat transfer deterioration indicator coefficient vector that are physically and semantically independent. The physical constraint embedding layer is used to apply differentiated physical value constraints to the three types of sub-vectors after decoupling. The physical value constraints applied by the physical constraint embedding layer include: applying non-negative constraints to the combustion heat load base coefficient vector and the coking thermal resistance disturbance coefficient vector, adding temporal smoothing constraints to the coking thermal resistance disturbance coefficient vector, and applying bounded constraints in the [0,1] interval to the heat transfer deterioration indicator coefficient vector. All physical value constraints are implemented through a differentiable soft constraint layer.

[0067] Encoder structure A spatiotemporal graph attention network is used to model the water-cooled wall pipe system as a graph structure. ,in, (Vertex / Vertices): Represents a set of vertices (nodes); (Edge / Edges): Represents a set of edges; nodes Corresponding to the root water-cooled wall tube, side Corresponding to the loop connection relationship or spatial adjacency relationship.

[0068] The encoder input is a temporal feature matrix. ,in Dimensional features include:

[0069] node Temperature at the measuring point (if a measuring point exists) or estimated temperature obtained through spatial interpolation;

[0070] node The structural parameters of the corresponding pipe section include pipe length, pipe diameter, wall thickness, material thermal conductivity, pitch, and loop topology.

[0071] Global operating condition parameters (embedded to each node via a broadcast mechanism).

[0072] in For the current moment The node feature matrix; the encoder uses a sliding window to display the historical node feature matrix sequence from the past T steps { ( -T+1),…, The time sequence is recursively processed by the gated loop unit, where T is the preset time window length (T=60 in this embodiment).

[0073] The encoder extracts spatial correlation features through multi-layer graph attention convolution, then extracts temporal evolution features through gated recurrent units (GRU), and finally reduces the dimensionality to a low-dimensional latent space vector. .

[0074] The core innovation of the physical prior encoder-decoder module lies in the latent space orthogonal decoupling module, which explicitly orthogonally decouples the low-dimensional latent space vector into a combustion heat load basic coefficient vector, a coking thermal resistance disturbance coefficient vector, and a heat transfer deterioration indicator coefficient vector that are physically and semantically independent (also known as explicit decoupling). Traditional autoencoders use black-box vectors with no physical semantics, while the low-dimensional latent space vectors of this invention... (in The total dimension of the latent space is explicitly decomposed into three orthogonal subvectors:

[0075]

[0076] in: It characterizes the spatial distribution of radiative heat flux density determined by the combustion organization within the furnace. Dimensionality reduction is performed on the heat flux density sample library obtained from historical CFD calculations or furnace flame image inversion using principal component analysis (PCA) or nonnegative matrix factorization (NMF) to extract... Orthogonal heat load spatial modal basis functions . The components are the linear combination coefficients of these basis functions, and their physical meaning is "the first". Activation intensity of each combustion heat load mode.

[0077] It is important to emphasize that the modal basis functions here are not post-analysis tools, but rather are embedded as structured weight parameters in the decoder, forming an end-to-end differentiable prediction system together with the encoder, physical computation graph, and residual loop. This paradigm shift from "analysis basis functions → architecture parameters" makes the prior knowledge of the spatial distribution of heat load an intrinsic property of the model topology. This is fundamentally different from the well-known practice of using modal decomposition only for offline data analysis.

[0078] It characterizes the spatial distribution of additional thermal resistance caused by the coking layer on the surface of the water-cooled wall. Similarly, the basis functions for the coking thermal resistance perturbation are constructed through mode decomposition. , Characterizes the severity of coking in different areas of the furnace. Coking thermal resistance. Represented as:

[0079]

[0080] in The baseline thermal resistance for cleaning the wall surface (negligible or taken as an empirical value). Basis functions. A radial basis function (RBF) with local support is employed to take advantage of the sparsity prior of the spatial distribution of the coke.

[0081] It characterizes the correction of the local heat transfer coefficient caused by heat transfer deterioration phenomena such as deviation from nucleus boiling (DNB) or dryout due to excessively high local heat flux density or excessively low mass flow rate. This sub-vector can be implemented in two equivalent ways: (1) per-tube allocation: assign a scalar deterioration indicator coefficient to each tube (the dimension of the heat transfer deterioration indicator coefficient vector is equal to the total number of tubes in the water-cooled wall), at this time (2) Spatial clustering and partitioning: The dimension of the heat transfer deterioration indicator coefficient vector is smaller than the total number of water-cooled wall tubes. Based on the furnace wall clustering and partitioning results, it is assigned to each tube control volume, that is, the tubes are spatially clustered (e.g., partitioned by furnace wall), and a coefficient is assigned to each partition. The specific implementation depends on the boiler size and computational resource constraints. The heat transfer deterioration indicator coefficient vector is mapped to the control volume of each pipe section through a learnable fully connected layer or a fixed partition mask matrix. It should be noted that the learnable fully connected layer is determined during offline training along with the encoder and decoder structures; the fixed partition mask matrix is ​​predefined and determined during offline training based on the spatial clustering results; both remain fixed during online operation and do not participate in the online update of the latent space vector.

[0082] The physical constraint embedding layer is also the core layer of the physical prior encoder-decoder module, located in the last layer of the encoder. Apply a non-negative constraint (heat flux density is positive) to Apply non-negativity constraints and add temporal smoothing constraints (non-negative coking thermal resistance, and limited amplitude of change between adjacent time points, to characterize the temporal continuity of coking accumulation and soot blowing processes), for Apply bounded constraints (heat transfer degradation correction factor in) These constraints are implemented through differentiable soft constraint layers (such as ReLU and Sigmoid) to ensure that the latent space vectors are always within the physically feasible region, while also guaranteeing the feasibility of end-to-end gradient propagation. This embedding layer makes the physical semantics of the latent space vectors not a post-training interpretation, but a structural property locked in during the architecture design phase.

[0083] It should be noted that the differentiable soft constraint layer system of the present invention uses a differentiable activation function to restrict the value domain of the subvector to the physical feasible domain, which is a structural constraint that is fundamentally different from the soft constraint of PINNs in the background art, which uses the physical equation residual as the regularization term of the loss function.

[0084] decoder Latent space vectors The decoupled combustion heat load basis coefficient vector and coking thermal resistance disturbance coefficient vector are mapped (decoded output) to the full-screen heat flux density distribution of the water-cooled wall. With coking thermal resistance distribution :

[0085]

[0086] in The baseline heat flux density distribution under design conditions (obtained through design calculations or cold-state commissioning). This is a residual term, learned by the deconvolutional network in the decoder, used to capture local details that the modal basis functions cannot describe. The introduction of this residual term ensures a balance between the simplicity of modal decomposition and the completeness of the network's expressive power.

[0087] The core innovation of the Phase-Aware Heat Transfer Computational Graph Network (PAHT-CGN) lies in reconstructing the one-dimensional unsteady-state heat transfer differential equations and working fluid thermophysical property calculations into non-learnable fixed layers of a deep learning computational graph, rather than using loss function regularization terms as soft constraints, or simply "implementing physical formulas with a computer program." Specifically, the discretized conservation equations are encoded into non-learnable fixed computational nodes in the computational graph. The inputs to these nodes are the full-screen heat flux density distribution and coking thermal resistance distribution of the water-cooled wall obtained from latent space vector decoding, and the outputs are the pipe wall temperature and working fluid state. The connections between nodes are determined by the physical causal chain of the heat transfer process. In this architecture, any input must sequentially undergo deterministic transformations through these physical layers; there are no alternative paths to bypass physical constraints.

[0088] It is important to note that while deep learning frameworks contain fixed, non-learnable computational nodes such as pooling layers and Dropout layers, this "non-learnability" is a design choice for engineering convenience (dimensionality reduction, regularization), and their computational rules can be replaced or modified. However, the "non-learnability" of the physical equation nodes in the phase-aware heat transfer computation graph network module is not a design choice, but a physical inevitability—the energy conservation equation and the working fluid state equation are objective natural laws describing the heat transfer process of the boiler water-cooled wall, and cannot be modified by the neural network weights. Once these equations are learned and modified, the model will lose its physical consistency and no longer describe the true heat transfer process.

[0089] Explanation of "unsteady state" and "steady state": The Phase-Aware Heat Transfer Computational Graph Network (PAHT-CGN) is an overall unsteady state model, reflected in the recursive evolution of the working fluid enthalpy along the pipe length (driven by time-varying thermal boundary conditions) and the time-varying evolution of the latent space vector. Radial heat conduction through the pipe wall, however, adopts a quasi-steady state assumption because the heat capacity of the pipe wall metal (approximately several hundred joules per kilogram per Kelvin) is much smaller than the heat capacity of the working fluid inside the pipe (approximately several thousand joules per kilogram per Kelvin for water / steam), and the metal wall thickness is only a few millimeters. The pipe wall temperature can reach thermal equilibrium within seconds, thus it can be considered an instantaneous response relative to the working fluid dynamics (minutes). This assumption is a standard engineering simplification in boiler hydrodynamic calculations and does not affect the model's accurate characterization of the unsteady process on the working fluid side.

[0090] This creates a fundamental difference from PINNs and conventional numerical solvers: in PINNs, the physical equations are "external rules" (indirectly constraining network behavior through the loss function), while conventional numerical solvers are "independent tools" (executed offline outside the deep learning framework, not supporting end-to-end automatic differentiation with the neural network); whereas in PAHT-CGN, the physical equations are "internal structure" (directly constituting the network's forward propagation path). The resulting technical advantages are: regardless of changes in the latent space vectors, forward propagation always strictly satisfies energy conservation; physical constraints do not require weight tuning, eliminating the problem of multi-task loss imbalance; forward propagation consists only of deterministic algebraic operations, resulting in extremely fast computation speed; and more importantly, because the physical computation graph and the neural network share the same computation graph, end-to-end automatic differentiation is supported—a crucial capability lacking in traditional independent numerical solvers.

[0091] The phase-aware heat transfer calculation graph network module communicates with the physical prior encoder-decoder module, and has a built-in fixed physical calculation graph, pipe segment discrete unit, and phase discriminator. The fixed physical calculation graph encodes the one-dimensional unsteady enthalpy recursive equation of the working fluid in the pipe, the radial heat conduction equation of the pipe wall, and the working fluid state equation into non-learnable fixed calculation graph nodes. The pipe segment discrete unit is used to discretize the water-cooled wall pipe along the pipe length into several independent pipe segment control bodies. Based on the spatial coordinate mapping relationship of the water-cooled wall pipe system, it samples and converts the decoded output furnace full-screen heat flux density distribution and water-cooled wall coking thermal resistance distribution to obtain the off-center heat flux density value and discrete coking thermal resistance value corresponding to each pipe segment control body, and uses them as the thermal load boundary input of each pipe segment control body. The phase discriminator identifies the real-time phase state of the working fluid in the pipe, adaptively switches the heat transfer correlation and corrects the local heat transfer coefficient by combining the heat transfer deterioration indicator coefficient, and calculates the predicted value of the fire-facing side wall temperature of the water-cooled wall full-screen according to the law of conservation of energy.

[0092] Specifically, such as Figure 2 As shown, the discrete element of the pipe section will divide the water-cooled wall into sections. The root canals are discretized along their length as follows: Each control entity (node) is used to construct a one-dimensional chained computation graph. Each computation graph node contains the following state variables:

[0093] Pipe wall temperature on the fire side

[0094] tube wall backfire side temperature

[0095] Temperature of working fluid inside the pipe

[0096] Enthalpy of working fluid inside the pipe

[0097] Pipe pressure

[0098] Vapor content by mass (Subcritical region) or working fluid density (Supercritical region)

[0099] Each control node embeds a phase discriminator. Based on the current working fluid pressure and enthalpy within the control volume, the current phase region is determined (i.e., divided into the following four major regions). Each phase region is matched with a specific heat transfer correlation equation, as follows:

[0100] Region I (Subcritical Subcooled Water Region): The single-phase liquid forced convection heat transfer correlation is adopted (such as the Dittus-Boelter equation).

[0101] Region II (Subcritical Two-Phase Region): Two-phase flow boiling heat transfer correlations (such as the Chen formula or the Liu-Winterton formula) are adopted, and a drying point discrimination is introduced;

[0102] Region III (Critical Transition Zone): and A corrected correlation based on the transition zone of dimensionless pressure and temperature is adopted;

[0103] Region IV (Supercritical Single-Phase Region): And all operating conditions that do not meet the criteria for Region III, covering the liquid-like region under supercritical pressure ( ) and the vapor-like region ( All of these employ supercritical fluid heat transfer correlations (such as the Bishop or Jackson equations), and the physical properties in the correlations are determined by the working fluid's equation of state based on local pressure and enthalpy; when At that time, saturation enthalpy , If this condition does not exist, the criteria for regions I and II will automatically become invalid. Explanation: pressure The enthalpy of the saturated liquid phase and the enthalpy of the saturated vapor phase; This is the critical pressure of water (approximately 22.064 MPa). The critical transition region pressure bandwidth; Preset enthalpy bandwidth; It is the quasi-critical enthalpy (the enthalpy corresponding to the temperature with the maximum specific heat capacity under supercritical pressure).

[0104] The output of the phase discriminator is a four-dimensional One-Hot vector or soft attention weights. The Gumbel-Softmax technique is used to achieve differentiable hard branch switching (approximately soft weighting in backpropagation and hard selection in forward propagation).

[0105] Local heat transfer coefficient The calculation is as follows:

[0106]

[0107] in For the region The corresponding local convective heat transfer coefficient, Hidden space Mapped to the The heat transfer deterioration indicator coefficient of the root canal (a value of 0 indicates normal heat transfer, and a value close to 1 indicates severe deterioration). Depend on dimensional vector Obtained through mapping with a fully connected layer or partition mask matrix: in per-pipe allocation mode = The mapping relationship is a one-to-one correspondence between tubes; in the spatial clustering partitioning mode... < Partition mask matrix Each zone coefficient is assigned to all pipes covered by that zone; the heat transfer deterioration indicator coefficient of the same pipe takes the same value across all control volumes of that pipe to characterize the overall heat transfer deterioration of that pipe or its zone.

[0108] The temperature and enthalpy updates for each control volume follow discrete heat transfer and flow conservation equations, which are directly encoded as node operations of the computation graph (non-learnable fixed operations):

[0109] Radial heat conduction through the pipe wall (steady-state assumption, since the heat capacity of the pipe wall is much smaller than that of the working fluid):

[0110]

[0111] in For the thermal resistance of the pipe wall, The outer diameter of the pipe. This refers to the inner diameter of the pipe.

[0112] Pipe wall-working fluid convection heat transfer:

[0113]

[0114] Note: The two formulas above calculate the same heat flux from the outer side of the pipe wall (radial heat conduction through the coking layer and the metal pipe wall) and the inner side of the pipe wall (convective heat transfer to the working fluid inside the pipe), respectively. Under the assumption of radial steady-state heat conduction through the pipe wall (the heat capacity of the pipe wall metal is much smaller than that of the working fluid inside the pipe and can be ignored), the heat transferred into the pipe wall is equal to the heat transferred out of the pipe wall, therefore the two equations must be equal. Solving these two equations simultaneously allows us to determine the temperatures on both sides of the pipe wall and the working fluid temperature, which is a classic method for handling coupled boundary conditions of the pipe wall.

[0115] The enthalpy of the working fluid is derived along the pipe length (considering radially uniform and axially one-dimensional flow):

[0116]

[0117] in To control the body length (distance from the walk). For the first The reference flow rate of the root canal loop is pre-calculated by the hydrodynamic calculation module (based on the loop pressure drop balance and throttling coil resistance characteristics). During operation, it is regarded as a slow time-varying parameter and is updated by the hydrodynamic calculation module at a preset period.

[0118] Equation of state for working fluid: ,in The water vapor temperature state function defined by the IAPWS-IF97 standard is calculated using the industrial water vapor property formula. The calculation graph is embedded using a differentiable neural network fitting method (3-layer MLP, 64 neurons) or a differentiable lookup table method to ensure end-to-end differentiability.

[0119] Determination of pipe pressure along pipe length: Pipe pressure of each control volume The pressure drop is calculated by the hydrodynamic pre-calculation module based on the loop pressure drop balance equation, or determined recursively along the pipe length using a frictional pressure drop and local resistance pressure drop model (a homogeneous flow pressure drop model can be used in the two-phase region). Given that the pressure drop inside the water-cooled wall tubes is small relative to the working fluid system pressure, a linear pressure drop approximation based on the inlet header pressure can also be used in engineering. The pressure inside the tubes, as the input for phase discrimination and working fluid property calculation, is pre-given in the above manner and does not participate in the online correction of the latent space vector. It should be noted that the continuous heat flux density distribution output by the decoder... With coking thermal resistance distribution By mapping the spatial coordinates of the water-cooled wall piping system, discrete values ​​at the center of each control volume are sampled. and This serves as the boundary condition input for the PAHT-CGN computation graph. The aforementioned update rule propagates forward segment by segment from the pipe inlet (with known feedwater temperature / enthalpy) to the outlet, forming a complete temperature field prediction. Since all operations are differentiable algebraic operations and logical switching, the entire PAHT-CGN can calculate the gradient of the output with respect to the input (latent space vector) using automatic differentiation techniques. This is the mathematical basis for PRDOA's operation.

[0120] The Physical Residue-Driven Latent Space Online Adaptive Closed-Loop (PRDOA) module is used to construct physical residuals based on measured wall temperatures and predicted wall temperatures. The latent space vector correction is obtained through constraint optimization. Under the premise that the network weights of the encoder and decoder structures are fixed, the low-dimensional latent space vector is updated online. The updated latent space vector is then fed back to the decoder structure and phase-sensing heat transfer calculation graph network module to achieve iterative closed-loop optimization. At the same time, the temperature field results after iterative optimization are pushed to the result output and early warning module.

[0121] Specifically, such as Figure 3 As shown, the key innovation of the physical residual-driven latent space online adaptive closed-loop module lies in constructing an end-to-end sensitivity propagation chain of physical residuals to latent space vectors. Specifically, it involves transforming the model prediction values ​​at sparse measurement points in the DCS into... Compared with measured values The deviation is defined as the physical residual:

[0122]

[0123] Since the forward propagation of PAHT-CGN is a deterministic physical computation (fully differentiable), the Jacobian matrix can be obtained by performing end-to-end differentiation on the entire PPED-PAHT-CGN computation link using automatic differentiation techniques.

[0124]

[0125] The Jacobian matrix quantizes the latent space. The component is related to the first... The degree of influence of temperature at each measuring point has a clear physical meaning: if A larger absolute value indicates that the first... The physical processes corresponding to each latent space vector (such as the activation of combustion heat load modes or coking thermal resistance disturbances in a certain region) have a significant impact on the temperature at that measuring point. This sensitivity matrix is ​​not obtained through numerical difference approximation, but rather through automatic differentiation of the computational graph, with a computational cost on the same order of magnitude as forward propagation.

[0126] The update of low-dimensional latent space vectors is not achieved by minimizing the black-box loss function through gradient descent, but rather by solving a physically constrained regularized least squares problem:

[0127]

[0128] Constraints:

[0129]

[0130] in:

[0131] The time-smoothing regularization parameter (L2 regularization) for the combustion heat load coefficient is used to prevent severe physical oscillations in the combustion mode;

[0132] The L1 norm (L1 sparsity regularization) is adopted, and the spatial sparsity prior of coking thermal resistance (coking usually occurs in local areas rather than being uniformly distributed throughout the furnace) is used to promote the spatial sparsity of the solution.

[0133] The regularization parameter (L2 regularization) for the heat transfer deterioration indicator coefficient is activated only when the temperature at the measuring point deviates significantly from the predicted normal heat transfer value.

[0134] The updated latent space vector is:

[0135]

[0136] in This is the optimal correction for the aforementioned least squares problem with physical constraints; To achieve an adaptive learning rate, it can be dynamically adjusted based on the magnitude of the residual norm—when the residual is large (such as after dust removal or at the initial stage of variable load), a larger value is used. Accelerate convergence; when the residual is small, take the smaller value. Suppress noise disturbances.

[0137] The evolution of the updated latent space vectors is further enhanced by a first-order inertial filter:

[0138]

[0139] in, This is the time smoothing coefficient. This prior is based on engineering physics: furnace coking is a slow, cumulative process (timescale from hours to days), while the combustion heat load distribution is regulated by the combination of coal mills and the air distribution method (timescale on the order of minutes), exhibiting significantly separated timescale characteristics. This filter effectively prevents spurious fluctuations in the latent space caused by measurement point noise, ensuring that the evolution of the latent space conforms to the timescale characteristics of the physical process.

[0140] Latent space vectors updated by the PRDOA module The decoder structure of the re-entered PPED and the PAHT-CGN module (where, in the hidden space) Without going through a decoder structure, the heat transfer coefficient is directly corrected by the PAHT-CGN module, and the temperature field of the fire-facing side wall of the full-screen water-cooled wall is calculated. .

[0141] The results output and early warning module outputs the following prediction and early warning information:

[0142] Real-time full-screen temperature field visualization: The predicted temperature field is mapped onto the three-dimensional geometric model of the furnace, and the temperature distribution is displayed through a color temperature map;

[0143] Thermal Deviation Index: Calculates the standard deviation and maximum deviation of the temperature on the fire-facing side wall of each tube outlet in the entire furnace (or can be calculated separately for each tube panel group):

[0144]

[0145] superscript Indicates the first The exit control body for the root canal. This represents the cross-sectional average temperature of the fire-facing sidewall of the entire furnace.

[0146] Identification of coking location and degree: based on The spatial distribution of the decoded data identifies high-risk areas for coking and the estimated equivalent thermal resistance of the coking layer.

[0147] Advanced forecasting: Introducing unit load commands As a feedforward input, it is combined with the trend extrapolation of the latent space vector (e.g., using a first-order inertial model for prediction). Predicting the future The wall temperature field after 5 to 15 minutes (e.g.) enables over-temperature early warning.

[0148] The three core modules of this invention (PPED, PAHT-CGN, and PRDOA) are tightly coupled and indispensable. Without explicit decoupling of the latent space in PPED, the model degenerates into a conventional black-box neural network or a pure physical model: black-box neural networks cannot obtain interpretability of coking-combustion decoupling, and operators cannot formulate targeted control strategies based on prediction results; pure physical models require known full-screen heat flux density boundary conditions, which are not available in engineering. Explicit decoupling of the latent space is the only bridge to transform "unobservable full-screen heat flux density" into "low-dimensional identifiable coefficients"; without explicit decoupling, the residual backpropagation of PRDOA will be unable to determine whether to correct the combustion mode or the coking thermal resistance, resulting in ambiguity in the correction direction and failure of the closed loop to converge. Without the phase-aware hard-coding of PAHT-CGN, the following issues arise: In the subcritical-to-supercritical transition region, the step change in the heat transfer coefficient cannot be accurately captured by the black-box network, leading to energy non-conservation or non-physical jumps in temperature prediction; Under deep peak-shaving conditions, the phase distribution within the pipe changes dynamically over time, and the training weights of soft-constrained PINNs cannot be adaptively adjusted. However, the hard-coded phase discriminator of this invention automatically switches the heat transfer correlation, ensuring physical consistency across all operating conditions; Without a hard-coded deterministic physical computation graph, PRDOA cannot obtain the precise sensitivity matrix of the residual to the latent space vector through automatic differentiation, and the online adaptive closed loop will lose its mathematical foundation. Without the physical residual online adaptive closed loop of PRDOA, the latent space vector will be entirely determined by the encoder feedforward output. The encoder weights are fixed after training, making it unable to track time-varying factors such as the slow accumulation of coking, burner wear, and coal drift; as time progresses, the model prediction error will continuously accumulate, ultimately losing its engineering usability. PRDOA corrects the hidden state in real time through the physical residuals of sparse measurement points, which is a necessary mechanism to ensure the long-term online availability of the model. Without PRDOA, the hard-coded physical equations of PAHT-CGN would lack online correction means for boundary conditions, resulting in the inability to eliminate the systematic error of "correct structure but biased input".

[0149] In particular, traditional state observers or Kalman filters are typically designed for low-dimensional centralized state systems. Even when using dimensionality reduction filtering techniques such as ensemble Kalman filtering to process high-dimensional distributed systems, accurate state-space models and noise statistics are still required, making engineering implementation difficult. PRDOA, through dimensionality reduction backpropagation of "high-dimensional residual → low-dimensional implicit vector correction," is adapted to the differentiability of the PAHT-CGN hard-coded computation graph, forming an online adaptive mechanism suitable for full-screen distributed temperature fields.

[0150] In summary, the three technical features are not simply a combination of functions, but rather constitute a complete technical closed loop of "decoupled representation → physical calculation → closed-loop correction," and none of the three can be omitted. The absence of any module will cause the entire closed loop to break, and the technical effect of this invention cannot be achieved.

[0151] Example 2:

[0152] This embodiment discloses an online prediction method for the temperature field of a boiler water-cooled wall. This method is applied to the aforementioned online prediction system for the temperature field of a boiler water-cooled wall, and includes the following steps:

[0153] S1. Data Acquisition: The data acquisition module collects real-time time-series data of boiler sparse wall temperature measurement points, unit operating condition parameters, and soft coal quality measurement parameters. Combined with the preset prior knowledge of water-cooled wall structure, the model input data is integrated and generated.

[0154] S2. Spatiotemporal feature extraction and latent space decoupling: Spatiotemporal features are extracted from the model input data through the encoder structure of the physical prior encoder-decoder module, and a low-dimensional latent space vector is obtained. The low-dimensional latent space vector is explicitly orthogonally decoupled into a combustion heat load base coefficient vector, a coking thermal resistance disturbance coefficient vector, and a heat transfer deterioration indicator coefficient vector that are physically and semantically independent, and physical value constraints are applied to the three types of sub-vectors respectively.

[0155] S3, Full-field physical field decoding and reconstruction: Through the decoder structure of the physical prior encoder-decoder module, combined with the pre-constructed combustion heat load mode basis function and coking thermal resistance disturbance basis function, the decoupled combustion heat load basis coefficient vector and coking thermal resistance disturbance coefficient vector are decoded and calculated respectively, and the corresponding output furnace full-screen heat flux density distribution and water-cooled wall coking thermal resistance distribution are output.

[0156] S4. Phase-aware heat transfer solution: The phase-aware heat transfer calculation graph network module calls the built-in fixed physical calculation graph to encode the one-dimensional unsteady enthalpy recursive equation of the working fluid in the tube, the radial heat conduction equation of the tube wall, and the working fluid state equation into unlearnable fixed calculation graph nodes; the water-cooled wall tube is discretized into several independent tube segment control bodies along the tube length through the tube segment discretization unit; based on the spatial coordinate mapping relationship of the water-cooled wall tube system, the furnace full-screen heat flux density distribution and water-cooled wall coking thermal resistance distribution output by the decoder structure are sampled and converted to obtain the boundary parameters of the heat dissipation load corresponding to each tube segment control body; the phase discriminator identifies the real-time phase state of the working fluid inside each tube segment control body, adaptively switches the appropriate heat transfer correlation, and uses the heat transfer deterioration indicator coefficient to correct the local heat transfer coefficient; the predicted value of the fire-facing side wall temperature of the water-cooled wall is calculated according to the law of conservation of energy.

[0157] S5. Physical Residual-Driven Adaptive Update of Latent Space: Through the physical residual-driven online adaptive closed-loop module of latent space, physical residuals are constructed by comparing the measured wall temperature and the predicted wall temperature at sparse measurement points. An optimization model is constructed by combining a preset differential regularization strategy and physical constraints. The latent space vector correction is solved by automatic differentiation across the entire link. Under the premise that the overall network weights of the encoder and decoder structures remain unchanged, the online update of the low-dimensional latent space vector is completed, and a first-order inertial time-series filter is applied to the updated latent space vector.

[0158] S6. Closed-loop iterative prediction and early warning: The filtered and updated low-dimensional latent space vector is fed back to the decoder structure and phase-sensing heat transfer calculation graph network module to iteratively update the furnace heat flux density, the full-field distribution of coking thermal resistance of water-cooled walls, and the local heat transfer coefficient, forming a closed-loop iterative calculation to continuously and dynamically update the full-screen temperature field of water-cooled walls; the full-screen temperature field of water-cooled walls, thermal deviation index, and coking distribution area are output in real time, and combined with the time sequence trend of unit load command, short-term early warning of over-temperature and visualization monitoring of furnace heating surface status are realized.

[0159] Specifically, the method of this invention can be deployed on the computing server of a power plant's SIS / MIS system. It retrieves operational data from the DCS real-time database via OPC or MQTT protocols, and the prediction results are provided to operators for reference via a web visualization interface or by writing them back to a custom logic area of ​​the DCS. The recommended system hardware configuration is: a multi-core CPU server (≥16 cores), ≥64GB of memory, and a GPU accelerator card (optional, used for forward inference acceleration of the PPED encoder).

[0160] Taking a 600MW supercritical once-through boiler as an example, its water-cooled wall adopts a spiral tube coil + vertical tube panel structure, with a total arrangement of There are several furnace outer wall temperature measurement points (distributed at the spiral section outlet, intermediate mixing header, and vertical section outlet), and the total number of water-cooled wall pipes. root.

[0161] By analyzing boiler design CFD calculations and processing historical flame images, 200 samples of water-cooled wall heat flux density distributions under different loads (50%, 75%, 100% BMCR), different mill combinations (A / B / C / D / E mill combinations), and different air distribution methods were obtained. After spatial standardization (mapping to a unified furnace coordinate system) of each sample, nonnegative matrix factorization (NMF) was used to extract... thermal load mode basis functions .in, Characterizing the basic mode of high heat load in the center of the furnace, , Characterize the thermal deviation modes of the left and right walls respectively. Characterizing the thermal deviation modes of the front and rear walls, - Higher-order modes characterizing inter-layer differences in the burner. The non-negativity of the basis functions ensures the physical rationality of the heat flux density prediction; furthermore, a variant of orthogonal non-negative matrix decomposition can be used to make the basis functions of each mode approximately orthogonal, thereby reducing redundant coupling between modes.

[0162] Based on the analysis of the coking mechanism, coking is mainly concentrated in the burner region (elevation 20~35m) along the height of the furnace, and exhibits a patchy distribution along the width. A Gaussian radial basis function (RBF) was used to construct... The basis function of the coke thermal resistance disturbance Each basis function corresponds to a local region in the furnace (such as the left front wall region, the middle front wall region, the right rear wall region, etc.), possessing spatial local support. This design utilizes the prior knowledge of the sparsity of coking space, enabling the L1 regularization of PRDOA to effectively locate the coking region.

[0163] The encoder employs a 3-layer spatiotemporal graph attention network (8 attention heads per layer, 64 hidden dimensions) + 2-layer GRU (32 hidden dimensions) + a fully connected mapping layer. The input is the node feature sequence of all N=1152 nodes over the past 10 minutes (sampling period 10s, 60 time steps): for each node's temperature features, nodes with measuring points directly use the measured temperature from the corresponding thermocouple; nodes without measuring points use a distance-weighted interpolation estimate based on the spatial location of the measuring points. The node temperature features, node structural parameters, and global operating condition parameters embedded via a broadcast mechanism together constitute the node feature matrix. The output is a latent space vector. ( ).

[0164] A semi-supervised strategy is adopted during the training phase: the pipe segment temperature covered by the measuring points in historical operational data is used as the supervision signal; for pipe segments without measuring points, the physical calculation output of PAHT-CGN is used as the self-supervised signal. The loss function is:

[0165]

[0166] in The mean square error at the measuring point. Hinge loss is defined by non-negative constraints on heat flux density and coking thermal resistance. L2 regularization for latent space vectors; and These are the loss weight coefficients for the physical constraint term and the regularization term, respectively.

[0167] Each pipe is discrete along its length as follows: The control system comprises 30 spiral sections and 20 vertical sections. The thermophysical properties of the working fluid inside the pipe are calculated using the IAPWS-IF97 formula. A differentiable lookup table method (based on a pre-calculated fine property grid using IAPWS-IF97, with lookup achieved through bilinear or trilinear interpolation, the interpolation operation itself being fully differentiable) is preferred and embedded in the calculation graph to ensure industrial-grade accuracy and strict energy conservation. Neural network fitting (e.g., a 3-layer MLP) is only used as an alternative for extrapolation under extreme conditions. The phase discriminator employs a hard discriminant based on a pressure-enthalpy threshold combined with a Gumbel-Softmax soft approximation to ensure backpropagation feasibility. The heat transfer correlation automatically switches according to the phase state.

[0168] Supercooled zone: Dittus-Boelter

[0169] Two-phase region: Chen's formula,

[0170] Supercritical region: Jackson's formula

[0171] illustrate: For Nusselt numbers, Let Reynolds number be 1. For Prandtl numbers, The overall heat transfer coefficient in the two-phase region is... Contributes to micro-liquid layer evaporation. Contributes to macroscopic convection; The Reynolds number is based on the mainstream physical properties. The average Prandtl number, The density of the working fluid at the wall surface. The density of the working fluid at the center of the flow channel. This is the average isobaric specific heat capacity. The specific heat capacity at constant pressure at the center of the flow channel. This is the empirical constant of Jackson's formula.

[0172] After the system is put into operation, a closed-loop update will be performed every minute.

[0173] Collect current DCS data and obtain the initial latent space estimate through PPED encoding. ;

[0174] The full-screen temperature field is calculated using PAHT-CGN forward propagation, and the predicted values ​​at the measuring points are extracted. ;

[0175] Calculate physical residuals ;

[0176] Calculate the Jacobian matrix using automatic differentiation. (dimension) );

[0177] Solving the constrained least squares problem yields the latent space correction. ;

[0178] Update hidden space: ;

[0179] Time smoothing: ;

[0180] The updated latent space is used to re-decode and output the full-screen temperature field, thermal deviation index, and coking distribution map.

[0181] Receive grid AGC load command Constructing predictive encoder inputs for future moments: load command and load condition parameter components. The setpoints for coal feed, air volume, and water flow are calculated based on the coordinated control feedforward logic. The temperature component at the measuring point maintains the current measured sequence (approximately assuming that the measurement response lags behind the changes in the operating condition command). Based on this, the encoder outputs an estimate of the future hidden space trend. Meanwhile, based on the coking thermal stagnation space First-order inertial extrapolation (time constant taken as 4 hours, corresponding to the typical cumulative time scale of coking) is used to predict the future wall temperature field. When the predicted fire-facing wall temperature at a pipe outlet exceeds 95% of the material's allowable temperature, a graded over-temperature warning (such as a yellow warning or a red warning) is issued 5-15 minutes in advance; when the predicted thermal deviation index is... When the set threshold is exceeded, combined with and The decoding distribution automatically generates targeted combustion adjustment or soot blowing suggestions.

[0182] Those skilled in the art should understand that variations can be implemented by combining existing technology with the above embodiments. Such variations do not affect the essence of the present invention and will not be elaborated upon here.

[0183] The preferred embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and the devices and structures not described in detail should be understood as being implemented in a conventional manner in the art. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the present invention. This does not affect the essential content of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the present invention's technical solutions still fall within the protection scope of the present invention.

Claims

1. An online prediction system for the temperature field of a boiler water-cooled wall, characterized in that, The system includes: a data acquisition module, a physical prior encoder-decoder module, a phase-sensing heat transfer calculation graph network module, a physical residual-driven latent space online adaptive closed-loop module, and a result output and early warning module. The data acquisition module is used to collect time-series data of boiler sparse wall temperature measurement points, unit operating condition parameters, and soft measurement parameters of coal quality, and load prior knowledge of water-cooled wall structure to integrate and generate model input data. The physical prior encoder-decoder module is connected to the data acquisition module and includes an encoder structure and a decoder structure. The encoder structure extracts spatiotemporal features from the model input data, maps them to obtain a low-dimensional latent space vector, and explicitly orthogonally decouples the low-dimensional latent space vector into a combustion heat load basic coefficient vector, a coking thermal resistance disturbance coefficient vector, and a heat transfer deterioration indicator coefficient vector that are physically and semantically independent. Physical value constraints are applied to the three types of sub-vectors. The decoder structure decodes and outputs the furnace full-screen heat flux density distribution and the water-cooled wall coking thermal resistance distribution based on the decoupled combustion heat load basic coefficient vector and coking thermal resistance disturbance coefficient vector. The phase-state sensing heat transfer calculation graph network module is communicatively connected to the physical prior encoder-decoder module, and has a built-in fixed physical calculation graph, pipe segment discrete unit, and phase state discriminator. The fixed physical calculation graph encodes the one-dimensional unsteady enthalpy recursive equation of the working fluid in the pipe, the radial heat conduction equation of the pipe wall, and the working fluid state equation into non-learnable fixed calculation graph nodes. The pipe segment discrete unit discretizes the water-cooled wall pipe along the pipe length into several independent pipe segment control bodies. Based on the spatial coordinate mapping relationship of the water-cooled wall pipe system, it samples and converts the decoded output furnace full-screen heat flux density distribution and water-cooled wall coking thermal resistance distribution to obtain the off-center heat flux density value and discrete coking thermal resistance value corresponding to each pipe segment control body, and uses them as the thermal load boundary input of each pipe segment control body. The phase state discriminator identifies the real-time phase state of the working fluid in the pipe, adaptively switches the heat transfer correlation and corrects the local heat transfer coefficient by combining the heat transfer deterioration indicator coefficient, and calculates the predicted value of the fire-facing side wall temperature of the water-cooled wall full-screen according to the law of conservation of energy. The physical residual-driven latent space online adaptive closed-loop module is used to construct physical residuals based on measured wall temperatures and predicted wall temperatures. The latent space vector correction is obtained through constraint optimization. Under the premise that the network weights of the encoder and decoder structures are fixed, the low-dimensional latent space vector is updated online. The updated latent space vector is then fed back to the decoder structure and phase-sensing heat transfer calculation graph network module to achieve iterative closed-loop optimization. At the same time, the temperature field results after iterative optimization are pushed to the result output and early warning module.

2. The online prediction system for boiler water-cooled wall temperature field according to claim 1, characterized in that, The result output and early warning module is used to output the full-screen temperature field, thermal deviation index and coking distribution area of ​​the water-cooled wall in real time, so as to realize dynamic online prediction and over-temperature early warning of the water-cooled wall temperature field.

3. The online prediction system for boiler water-cooled wall temperature field according to claim 1, characterized in that, The data acquisition module establishes real-time communication with the power plant's DCS system via the OPC or MQTT protocol to complete the real-time acquisition of time-series data from sparse wall temperature measurement points, unit operating condition parameters, and soft measurement parameters of coal quality.

4. The online prediction system for boiler water-cooled wall temperature field according to claim 1, characterized in that, The encoder structure incorporates a spatiotemporal graph attention network, a latent space orthogonal decoupling module, and a physical constraint embedding layer. The spatiotemporal graph attention network is composed of a multi-layer graph attention network and a gated recurrent unit, used to extract spatial correlation features of the water-cooled wall pipe system and temporal evolution features of operating data, and generate low-dimensional latent space vectors. The latent space orthogonal decoupling module is used to complete the orthogonal decoupling of the low-dimensional latent space vectors, obtaining three types of sub-vectors. The physical constraint embedding layer is used to apply physical value constraints to the three types of sub-vectors after decoupling.

5. The online prediction system for boiler water-cooled wall temperature field according to claim 4, characterized in that, The physical value constraints applied by the physical constraint embedding layer include: applying non-negative constraints to the combustion heat load base coefficient vector and the coking thermal resistance disturbance coefficient vector; adding time-series smoothing constraints to the coking thermal resistance disturbance coefficient vector; and applying bounded constraints in the [0,1] interval to the heat transfer deterioration indicator coefficient vector. All physical value constraints are implemented through a differentiable soft constraint layer.

6. The online prediction system for boiler water-cooled wall temperature field according to claim 1, characterized in that, Each pipe section control body is equipped with an independent set of state variables, which includes pipe wall temperature on the fire side, pipe wall temperature on the back side, working fluid temperature inside the pipe, working fluid enthalpy inside the pipe, pipe pressure, and mass vapor content or working fluid density.

7. The online prediction system for boiler water-cooled wall temperature field according to claim 1, characterized in that, The phase discriminator divides the working fluid into four regions based on the working fluid pressure and enthalpy: subcritical subcooled water region, subcritical two-phase region, critical transition region, and supercritical single-phase region. It uses the Gumbel-Softmax algorithm to achieve differentiable hard branch switching, and each phase region is matched with a specific heat transfer correlation. The heat transfer deterioration indicator coefficient vector is mapped to each pipe segment control volume through a learnable fully connected layer or a fixed partition mask matrix, including two mapping modes: pipe-by-pipe allocation and spatial clustering partitioning. In the tube-by-tube allocation mode, the dimension of the heat transfer deterioration indicator coefficient vector is equal to the total number of water-cooled wall tubes; in the spatial clustering partitioning mode, the dimension of the heat transfer deterioration indicator coefficient vector is less than the total number of water-cooled wall tubes, and is allocated to each tube segment control body according to the furnace wall clustering partitioning results.

8. The online prediction system for boiler water-cooled wall temperature field according to claim 1, characterized in that, The physical residual-driven latent space online adaptive closed-loop module constructs a physically constrained regularized least squares optimization model and configures a differentiated regularization strategy. The differentiated regularization strategy is as follows: apply time-smoothing L2 regularization to the combustion heat load base coefficient vector, apply L1 sparse regularization to the coking thermal resistance disturbance coefficient vector, and apply conditionally activated L2 regularization to the heat transfer deterioration indicator coefficient vector. The physical constraints include a base coefficient for combustion heat load, a non-negative coefficient for coking thermal resistance disturbance, and a heat transfer deterioration indicator coefficient with values ​​in the range [0,1].

9. The online prediction system for boiler water-cooled wall temperature field according to claim 1, characterized in that, The physical residual-driven latent space online adaptive closed-loop module applies a first-order inertial timing filter to the updated latent space vector.

10. A method for online prediction of boiler water-cooled wall temperature field, characterized in that, The method, applied to the online prediction system for boiler water-cooled wall temperature field according to any one of claims 1 to 9, includes the following steps: S1. Data Acquisition: The data acquisition module collects real-time time-series data of boiler sparse wall temperature measurement points, unit operating condition parameters, and soft coal quality measurement parameters. Combined with the preset prior knowledge of water-cooled wall structure, the model input data is integrated and generated. S2. Spatiotemporal feature extraction and latent space decoupling: Spatiotemporal features are extracted from the model input data through the encoder structure of the physical prior encoder-decoder module, and a low-dimensional latent space vector is obtained. The low-dimensional latent space vector is explicitly orthogonally decoupled into a combustion heat load base coefficient vector, a coking thermal resistance disturbance coefficient vector, and a heat transfer deterioration indicator coefficient vector that are physically and semantically independent, and physical value constraints are applied to the three types of sub-vectors respectively. S3, Full-field physical field decoding and reconstruction: Through the decoder structure of the physical prior encoder-decoder module, combined with the pre-constructed combustion heat load mode basis function and coking thermal resistance disturbance basis function, the decoupled combustion heat load basis coefficient vector and coking thermal resistance disturbance coefficient vector are decoded and calculated respectively, and the corresponding output furnace full-screen heat flux density distribution and water-cooled wall coking thermal resistance distribution are output. S4. Phase-aware heat transfer solution: The phase-aware heat transfer calculation graph network module calls the built-in fixed physical calculation graph to encode the one-dimensional unsteady enthalpy recursive equation of the working fluid in the tube, the radial heat conduction equation of the tube wall, and the working fluid state equation into unlearnable fixed calculation graph nodes; the water-cooled wall tube is discretized into several independent tube segment control bodies along the tube length through the tube segment discretization unit; based on the spatial coordinate mapping relationship of the water-cooled wall tube system, the furnace full-screen heat flux density distribution and water-cooled wall coking thermal resistance distribution output by the decoder structure are sampled and converted to obtain the boundary parameters of the heat dissipation load corresponding to each tube segment control body; the phase discriminator identifies the real-time phase state of the working fluid inside each tube segment control body, adaptively switches the appropriate heat transfer correlation, and uses the heat transfer deterioration indicator coefficient to correct the local heat transfer coefficient; the predicted value of the fire-facing side wall temperature of the water-cooled wall is calculated according to the law of conservation of energy. S5. Physical Residual-Driven Adaptive Update of Latent Space: Through the physical residual-driven online adaptive closed-loop module of latent space, physical residuals are constructed by comparing the measured wall temperature and the predicted wall temperature at sparse measurement points. An optimization model is constructed by combining a preset differential regularization strategy and physical constraints. The latent space vector correction is solved by automatic differentiation across the entire link. Under the premise that the overall network weights of the encoder and decoder structures remain unchanged, the online update of the low-dimensional latent space vector is completed, and a first-order inertial time-series filter is applied to the updated latent space vector. S6. Closed-loop iterative prediction and early warning: The filtered and updated low-dimensional latent space vector is fed back to the decoder structure and phase-sensing heat transfer calculation graph network module to iteratively update the furnace heat flux density, the full-field distribution of coking thermal resistance of water-cooled walls, and the local heat transfer coefficient, forming a closed-loop iterative calculation to continuously and dynamically update the full-screen temperature field of water-cooled walls; the full-screen temperature field of water-cooled walls, thermal deviation index, and coking distribution area are output in real time, and combined with the time sequence trend of unit load command, short-term early warning of over-temperature and visualization monitoring of furnace heating surface status are realized.

Images