Boiler three-dimensional temperature parameter prediction method based on deep reinforcement learning

Abstract

The invention is suitable for the technical field of boiler temperature prediction, and provides a boiler three-dimensional temperature parameter prediction method based on deep reinforcement learning, which comprises the following steps: acquiring boiler DCS historical operation data, and generating a boiler furnace three-dimensional temperature field reference data set under a corresponding typical working condition; fuzzy clustering type working condition classification is carried out on DCS historical operation data; performing adaptive resampling on the three-dimensional temperature field data generated by CFD to reduce the data scale; carrying out normalization processing on the data; and establishing a deep reinforcement learning prediction model taking the operation parameter characteristics and the space coordinates as input and taking the temperature as output. According to the method, real-time and high-precision prediction of the three-dimensional temperature field in the boiler can be realized, and effective support can be provided for boiler combustion optimization, safe operation and digital twinning construction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of boiler temperature prediction, and particularly relates to a boiler three-dimensional temperature parameter prediction method based on deep reinforcement learning. BACKGROUND

[0002] At present, the monitoring of the internal combustion temperature of a power station boiler mainly relies on two types of technologies, direct measurement and computational fluid dynamics (CFD) numerical simulation, but these methods have significant limitations: in the direct measurement technology, contact type temperature measurement such as thermocouples can only provide sparse point temperature data, cannot restore the three-dimensional temperature field panorama, and has problems of measurement delay, insufficient high-temperature resistance, and high maintenance cost; non-contact methods such as infrared temperature measurement are easily disturbed by the environment and can only obtain two-dimensional surface information, acoustic wave temperature measurement has low spatial resolution and poor noise resistance, and optical technology is complex and costly, which are difficult to realize large-scale engineering application; on the other hand, although CFD numerical simulation can theoretically reconstruct the three-dimensional temperature field, it has huge calculation time consumption and cannot meet the real-time monitoring demand, and its accuracy is highly dependent on the accuracy of complex physical models and boundary conditions, and the modeling process is highly professional and has weak generalization ability.

[0003] In summary, the existing technical means cannot balance real-time, globality, accuracy and economy, which restricts the fine management and intelligent control of the boiler combustion state. SUMMARY

[0004] The purpose of the embodiment of the application is to provide a boiler three-dimensional temperature parameter prediction method based on deep reinforcement learning, aiming to solve the problems proposed in the background technology.

[0005] The embodiment of the application is implemented in this way, the boiler three-dimensional temperature parameter prediction method based on deep reinforcement learning comprises the following steps:

[0006] Multi-source data acquisition and construction: synchronously acquire the DCS historical operation data of the boiler, and generate the three-dimensional temperature field benchmark data set of the boiler furnace under the corresponding typical working condition by using the computational fluid dynamics numerical simulation technology;

[0007] Data preprocessing: perform fuzzy clustering type working condition classification on the DCS historical operation data; perform adaptive resampling on the three-dimensional temperature field data generated by the CFD to reduce the data size; perform normalization processing on all input and output data;

[0008] Prediction model construction: establish a deep reinforcement learning prediction model with operating parameter features and spatial coordinates as inputs and temperature as output;

[0009] Real-time temperature prediction: after the real-time collected DCS operation data are preprocessed, the data are input into the prediction model together with the target spatial coordinates, so as to realize online real-time prediction of the temperature value at any position inside the boiler.

[0010] This invention comprehensively utilizes readily available DCS operating data and high-precision CFD simulation data from boilers. Through deep reinforcement learning algorithms, it constructs an efficient surrogate model, achieving real-time, high-precision, and low-cost online measurement of the three-dimensional temperature field inside the boiler. This provides crucial data support for boiler combustion optimization, safety early warning, and intelligent operation, overcoming the limitations of traditional measurement methods in terms of spatial coverage and real-time performance. By fusing actual operating data and CFD simulation data, and introducing fuzzy clustering-based condition classification and adaptive resampling strategies, it significantly improves data utility and model training efficiency. It has broad engineering application prospects and can be used for boiler combustion optimization, fault early warning, and digital twin construction, enhancing the safety and economy of unit operation. Attached Figure Description

[0011] Figure 1 A flowchart illustrating a method for predicting three-dimensional boiler temperature parameters based on deep reinforcement learning, provided in an embodiment of the present invention.

[0012] Figure 2 The scatter plot and fitted line of the predicted temperature values ​​provided in the embodiments of the present invention;

[0013] Figure 3 This is a statistical histogram of predicted temperature value errors provided in an embodiment of the present invention.

[0014] Figure 4 This is a normal distribution diagram of the predicted residuals provided in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0017] like Figure 1 The flowchart shown is a method for predicting three-dimensional boiler temperature parameters based on deep reinforcement learning, provided in an embodiment of the present invention. The method includes the following steps:

[0018] Step 1: Multi-source data acquisition and construction: Simultaneously acquire historical operating data of the boiler DCS, and use computational fluid dynamics (CFD) numerical simulation technology to generate a three-dimensional temperature field benchmark dataset of the boiler furnace under corresponding typical operating conditions.

[0019] Step 2, Data Preprocessing: Perform fuzzy clustering-based operational condition classification on the historical DCS operating data; perform adaptive resampling on the 3D temperature field data generated by CFD to reduce the data size; and normalize all input and output data.

[0020] Step 3: Predictive Model Construction and Training: Establish a deep reinforcement learning model with operating parameter features and spatial coordinates as input and temperature as output, and train and optimize the model using preprocessed data;

[0021] Step 4: Real-time temperature prediction: After preprocessing, the real-time collected DCS operation data, along with the target spatial coordinates, is input into the trained model to achieve online real-time prediction of the temperature value at any location inside the boiler.

[0022] Taking a 350MW supercritical coal-fired boiler in a power plant as a specific implementation object, this paper elaborates on the above-mentioned method for predicting the three-dimensional temperature parameters of the boiler based on deep reinforcement learning. The boiler adopts a single furnace, balanced ventilation, solid ash discharge, and Π-shaped layout. The overall dimensions of the furnace are 58.30 meters high, 14.43 meters wide, and 14.43 meters deep. The boiler adopts a new tangential combustion method, and the main burner is fixed and arranged on the four walls of the water-cooled wall.

[0023] The specific implementation includes the following steps:

[0024] Step 1: Multi-source heterogeneous data acquisition and construction:

[0025] The foundation of this invention lies in constructing a high-quality dataset that can fully characterize the boiler combustion state. This step involves collecting data from both the actual operating environment and the high-precision digital simulation environment, specifically including:

[0026] 1. Real-time and historical operating data acquisition from the boiler's on-site DCS:

[0027] Through the data interface of the power plant's distributed control system (DCS), historical and real-time operating data of the boiler under various typical operating conditions such as 100%, 75%, and 50% load are acquired. This data directly and in real-time reflects the boiler's combustion status and serves as one of the key input features for the predictive model. The collected data mainly includes, but is not limited to:

[0028] Global parameters: unit load, total coal feed, total air volume, main steam flow, opening degree of each secondary damper, excess air coefficient, etc.

[0029] Burner-related parameters: coal feed rate, primary air volume and temperature, secondary air volume and temperature, etc. for each burner layer;

[0030] All data needs to undergo preliminary cleaning and screening to remove obviously invalid and abrupt data points, and be aligned according to a unified timestamp to form a time-series dataset that strictly corresponds to the boiler's operating status.

[0031] 2. High-precision CFD numerical simulation and reference temperature field generation of boiler combustion process:

[0032] To obtain the three-dimensional holographic temperature field distribution inside the boiler, which is difficult to measure directly with sensors, as a training label for the deep learning model, high-fidelity computational fluid dynamics (CFD) numerical simulations are performed on the typical operating conditions selected in step 1. This method specifically includes:

[0033] 2.1 Geometric Modeling and Mesh Generation: First, a precise three-dimensional geometric model is established based on the actual boiler design drawings. Then, the computational domain is meshed. To balance computational accuracy and efficiency, unstructured tetrahedral meshes with strong adaptability are used for local refinement in the complex burner nozzle area. For areas with regular structures such as the furnace body, hexahedral structured meshes with high computational efficiency and accuracy are used for mesh generation, thereby generating a high-quality computational mesh.

[0034] 2.2 Physical Model Selection and Setup: The simulation process was implemented based on the commercial CFD software ANSYS FLUENT platform, and a series of industrially validated advanced physical models suitable for boiler combustion processes were selected to ensure the reliability of the simulation results.

[0035] The boundary conditions are set strictly according to the actual DCS operating data (such as coal feed, air volume, air temperature, etc.) collected in step 1 under the corresponding operating conditions to ensure the consistency between the simulation environment and the real operating conditions. Through iterative solution, detailed data such as the convergent and stable three-dimensional spatial temperature field and component concentration field inside the boiler under the operating conditions are finally obtained. These CFD simulation results constitute a high-precision three-dimensional temperature field benchmark database.

[0036] Through the above two steps, a multi-source heterogeneous fusion dataset was constructed, including DCS running parameters and 2.77 million sets of three-dimensional temperature field data obtained by CFD, laying a solid data foundation for the subsequent training of deep learning models.

[0037] Step 2: Data Preprocessing and Feature Engineering

[0038] A series of refined preprocessing operations are performed on the multi-source heterogeneous fusion dataset to eliminate data redundancy, unify data scale, and extract key features, thereby significantly improving the training efficiency, prediction accuracy, and generalization ability of subsequent models. Specifically, these operations include:

[0039] 1. Operating condition classification and labeling:

[0040] Due to frequent changes in boiler operating load and strategies, its historical operating data covers a large number of different operating conditions. If all data are mixed for training, the model's learning objective will be unclear and its performance will degrade. Therefore, this embodiment of the invention first classifies the collected DCS historical operating dataset into operating conditions. Based on professional knowledge of boiler operation, core parameters such as unit load, total coal feed, total air volume, and air-coal ratio are selected as classification features. The Fuzzy C-Means (FCM) clustering algorithm is used to divide the massive historical operating data into several representative operating condition categories. Unlike traditional hard clustering, FCM allows each data sample to belong to multiple operating condition categories with different membership degrees, which is more in line with the transition and ambiguity between boiler operating conditions. After clustering, each data sample obtains an operating condition category label based on its maximum membership degree, laying the foundation for subsequent operating condition-specific modeling.

[0041] 2. Working condition classification method based on fuzzy clustering:

[0042] First, a fuzzy membership mechanism is introduced for each sample point. For the cluster center membership degree Values ​​in the range And satisfy the normalization constraint:

[0043] ;

[0044] in, As a historical working condition sample, This represents the total number of historical operating condition samples. Number of operating condition categories;

[0045] The objective function for fuzzy clustering is defined as:

[0046] ;

[0047] in, Indicates sample Belongs to the Membership degree of each working condition category; Represents the fuzziness index, controlling the degree of fuzziness in the membership distribution. The larger the sample size, the smaller the difference in membership degree between the samples and each category, and the more ambiguous the division of working conditions. Indicates the first Cluster centers for each working condition category; Indicates sample With cluster center The Euclidean distance between them;

[0048] Through iterative optimization, the cluster center and membership degree are updated alternately according to the following formula: The cluster center update formula is:

[0049] ;

[0050] Wherein, the numerator is the weighted sum of the samples, and the weight is the fuzzy membership degree of the sample to the cluster; the denominator is the normalization term of the weighted coefficient of the membership degree of the cluster.

[0051] The membership update formula is:

[0052] ;

[0053] Where the denominator is the sample The cumulative term of the distance ratio to all cluster centers ensures that the sum of the membership degrees of each sample in all clusters is 1;

[0054] The iterative process continues until the membership matrix is ​​reached. When the change in value is less than a preset threshold or the maximum number of iterations is reached, the operating condition classification is completed. Finally, each sample can obtain an operating condition category label based on the cluster center corresponding to its maximum membership degree, while retaining the complete membership degree distribution to reflect the transition and ambiguity of boiler operating conditions.

[0055] 3. Resampling of three-dimensional temperature field data:

[0056] The three-dimensional temperature field data generated by CFD simulation is usually defined on millions or even tens of millions of grid nodes. The amount of data is extremely large. Directly using all grid data as training labels will lead to an explosion of model parameters, low training efficiency, and easy introduction of noise.

[0057] To address this issue, this invention provides a rational dimensionality reduction method for high-resolution three-dimensional temperature field data obtained through CFD simulation. While preserving key temperature distribution features, the method significantly reduces the data size, thereby improving model training efficiency. This method adaptively determines the sampling density of each region based on the temperature field gradient magnitude.

[0058] First, calculate the CFD temperature field in three directions at each grid node. The partial derivatives on the x-axis yield the gradient vector. , It is the temperature field in three-dimensional space, and then its gradient magnitude is calculated. This value reflects the degree of drastic temperature change;

[0059] For each grid node Assign an importance weight to the sampled item. :

[0060] ;

[0061] in, It is the total number of grid nodes in the entire CFD computation domain. It is an adjustment constant used to ensure that a basic probability is sampled even in flat regions where the gradient is zero;

[0062] Regions with large gradient magnitudes correspond to areas of drastic temperature changes and are assigned higher sampling probabilities, while regions with small gradient magnitudes correspond to areas of stable temperature and are assigned lower sampling probabilities, based on the calculated probability distribution. By using random sampling or systematic sampling methods, a preset number of subsets are extracted from the original grid nodes. This approach helps to effectively reduce the computational complexity and time cost of subsequent model training while maintaining the main physical characteristics of the temperature field, making efficient learning of large-scale three-dimensional temperature field data possible.

[0063] 4. Data normalization processing:

[0064] Input data includes DCS operating parameters (such as air volume and coal volume) and spatial coordinates. These features have different dimensions and orders of magnitude. The huge difference in numerical range can lead to instability in the model training process. Features with large values ​​will dominate the model weights, while features with small values ​​may be ignored.

[0065] To ensure training stability and convergence speed, this embodiment of the invention normalizes all input features (including DCS parameter values ​​and spatial coordinates) and output labels (temperature values). Using methods such as Min-Max normalization, various types of data are linearly mapped to the interval [0,1] or [-1,1] to eliminate the influence of dimensions. After normalization, all features are on the same order of magnitude, and the model can learn the complex relationship between all input features and output equally, thereby effectively improving the model's prediction accuracy and training efficiency.

[0066] Thus, after the above preprocessing steps, a high-quality dataset of moderate size, clear annotation, and uniform scale is obtained, which is fully prepared for the next step of building an efficient and accurate deep learning prediction model.

[0067] Step 3: Construction and training of the deep reinforcement learning prediction model:

[0068] A deep reinforcement learning (DRL) model is constructed that can accurately learn the characteristics of boiler operating parameters and the complex nonlinear mapping relationship between three-dimensional spatial coordinates and temperature values. This model introduces an interaction mechanism between the agent and the environment, aiming to maximize cumulative rewards, to replace high-fidelity CFD simulation and achieve real-time, high-precision full-field temperature prediction.

[0069] 1. Network model architecture design:

[0070] The deep reinforcement learning model adopts an Actor-Critic architecture, which consists of a policy network and a value network.

[0071] The input is a 10-dimensional fused feature vector, where 7 dimensions are preprocessed DCS operating parameter features representing the current boiler operating status, and the other 3 dimensions are spatial coordinates. Specify the temperature location to be predicted; the input vector serves as the current state. This is the foundation for the interaction between intelligent agents and the environment;

[0072] Both the policy network and the value network employ a 5-layer fully connected neural network structure as feature extractors. The layer output is:

[0073] ;

[0074] in, and The first Layer weight matrix and bias vector;

[0075] Activation function ReLU (Rectified Linear Unit) is used to enhance the model's nonlinear fitting ability, thereby better capturing complex thermophysical processes:

[0076] ;

[0077] Policy Network Output Probability Distribution , indicating the state Predicting the action distribution of temperature at a specific spatial point; Value network output scalar That is, the value estimation of the state, which is used to evaluate the merits of the current prediction strategy; it enables the model to unify boiler operating conditions, spatial coordinates, and temperature predictions under the reinforcement learning framework, and realize dynamic learning of complex spatial fields.

[0078] 2. Model Training and Optimization:

[0079] The preprocessed data is used as the environmental feedback source. Each sample consists of a state, an action, and a reward, where the state is the working condition and spatial coordinates, the action is the predicted temperature value, and the reward is... Defined as the negative mean square error between the predicted value and the CFD simulation result:

[0080] ;

[0081] in, For predicted values, The larger the reward is for CFD simulation results, the closer the prediction result is to the CFD simulation value; if the prediction error is large, the reward will be negative.

[0082] The goal of the training process is no longer to directly minimize the error, but to maximize the long-term cumulative reward:

[0083] ;

[0084] in, This is a discount factor used to balance the accuracy of short-term and long-term forecasts. Indicates a time step. This represents the expected value of all possible future states and actions given a policy π.

[0085] We employ a gradient-based reinforcement learning optimization method (Proximal Policy Optimization, PPO) to update the parameters of the policy network and the value network. To avoid overfitting and improve generalization ability, we introduce methods such as Dropout and L2 regularization.

[0086] The agent continuously updates its policy and value networks by repeatedly interacting with the environment, i.e. predicting temperature and receiving CFD feedback rewards. It monitors performance on the validation set and uses early stopping when the rewards converge or there are signs of overfitting. The resulting agent can infer the temperature at any location in real time based on the boiler's operating status and spatial coordinates, achieving high-precision prediction.

[0087] To quantitatively evaluate the overall performance advantages of the technical solutions in the embodiments of the present invention, comprehensive performance verification was conducted on an independent test set after model training was completed. Multiple statistical indicators were used for comprehensive evaluation, including the coefficient of determination (R²). 2 ), mean squared error (MSE), mean absolute error (MAE), etc., to comprehensively measure prediction accuracy;

[0088] To verify the effectiveness of the fuzzy C-means algorithm-based method for classifying historical boiler operating conditions proposed in this invention, it was compared with several classic clustering methods. All clustering methods were processed using the same deep reinforcement learning model for performance evaluation. The results are shown in Table 1.

[0089] Table 1. Comparison of model performance under different data processing methods

[0090] Data processing method [R 2 ]] MSE MAE MAPE k-means 0.749 42794 148.67 18.08 GMM 0.715 47513 154.96 19.68 k-medoids 0.786 36550 134.96 16.85 Embodiments of the invention 0.827 34369 129.998 17.291

[0091] As shown in Table 1, the working condition classification method based on Fuzzy C-Means used in this embodiment of the invention significantly outperforms traditional clustering methods in all key indicators. Among them, the coefficient of determination R0...2 The highest value (0.827) indicates the best consistency between the model's predictions and the actual values; the lowest mean squared error (MSE) and mean absolute error (MAE) are 34369 and 129.99 respectively, proving the optimal prediction accuracy; compared with traditional hard clustering methods such as K-Means and K-Medoids, this embodiment of the invention, by introducing a fuzzy membership mechanism, more fully characterizes the transition and fuzziness between working conditions, making R... 2 The results show that the FCM-based operating condition classification can more accurately reflect the intrinsic relationship between boiler operating states, providing higher-quality input features for subsequent prediction models, thereby significantly improving prediction performance and generalization ability.

[0092] To verify the superiority of the deep reinforcement learning model constructed in this embodiment of the invention, it is compared with a variety of commonly used machine learning models. All models are compared fairly using the same training and test sets generated by the data processing method proposed in this embodiment of the invention. The comparison results are shown in Table 2:

[0093] Table 2 Performance Comparison of Different Prediction Models

[0094] Network model [R 2 ]]> MSE MAE MAPE ELM 0.741 43097 147.47 19.05 MLP 0.711 49880.6 158.14 19.75 XGboost 0.75 43109 148.47 19 Embodiments of the invention 0.788 36156 135.15 17.68

[0095] As shown in Table 2, the deep reinforcement learning model provided in this embodiment of the invention outperforms other comparative models in all aspects, with its R... 2 The performance was improved by 5.1% compared to the second-best XGBoost model, and the MSE error was reduced by 16.1%. This demonstrates that the multi-hidden-layer structure, ReLU activation function, and Adam optimizer designed in this embodiment of the invention can more effectively learn the complex nonlinear mapping relationship between boiler operating parameters, spatial coordinates, and temperature field, exhibiting stronger expressive power and higher prediction accuracy.

[0096] The scatter plot and fitted line of the final predicted boiler temperature values ​​in this embodiment of the invention are shown below. Figure 2 As shown, the predicted temperature has good overall consistency with the actual temperature, and the model can effectively capture the trend of boiler temperature change, achieving relatively accurate temperature prediction under most operating conditions. Figure 3 The histogram of errors in the predicted temperature values ​​shows that the peak values ​​of the error distribution are high and concentrated, indicating that most of the prediction errors are small deviations and the probability of large errors is low. Figure 4 To predict the normal distribution of the residuals, the residuals generally showed good normality, further validating the statistical rationality of the error distribution.

[0097] In summary, the data processing and prediction model have fully demonstrated that the technical solution proposed in this invention has significant and outstanding advantages over existing known methods in terms of accuracy, stability, and reliability in predicting the three-dimensional temperature field of boilers. It successfully solves the industry pain point of insufficient accuracy and difficulty in achieving both computational efficiency in traditional methods, and provides an effective technical means for intelligent monitoring and optimized operation of boilers.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting three-dimensional temperature parameters of a boiler based on deep reinforcement learning, characterized in that, Includes the following steps: Multi-source data acquisition and construction: synchronously acquire historical operating data of boiler DCS, and use computational fluid dynamics numerical simulation technology to generate a three-dimensional temperature field benchmark dataset of boiler furnace under corresponding typical operating conditions. Data preprocessing: Fuzzy clustering is used to classify the operating conditions of historical DCS data; adaptive resampling is performed on the three-dimensional temperature field data generated by CFD to reduce the data size; Normalize all input and output data; Prediction model construction: Establish a deep reinforcement learning prediction model with operating parameter features and spatial coordinates as input and temperature as output; Real-time temperature prediction: After preprocessing, the real-time collected DCS operation data is input into the prediction model along with the target spatial coordinates to achieve online real-time prediction of the temperature value at any location inside the boiler.

2. The method for predicting three-dimensional boiler temperature parameters based on deep reinforcement learning according to claim 1, characterized in that, The steps for generating a reference dataset of the three-dimensional temperature field of the boiler furnace under typical operating conditions using computational fluid dynamics numerical simulation technology are as follows: Geometric modeling and mesh generation: Based on the actual boiler design drawings, an accurate three-dimensional geometric model is established, and the computational domain is meshed. Physical model selection and setting: A physical model is selected, and the boundary conditions are set based on the actual DCS operation data collected under the corresponding operating conditions to ensure the consistency between the simulation environment and the real operating conditions. Through iterative solution, detailed data of the convergent and stable three-dimensional spatial temperature field and component concentration field inside the boiler under the operating conditions are obtained, forming a three-dimensional temperature field benchmark dataset.

3. The method for predicting three-dimensional boiler temperature parameters based on deep reinforcement learning according to claim 1, characterized in that, The steps for performing fuzzy clustering-based operational condition classification on historical DCS operating data are as follows: Operating condition classification and labeling: Core parameters are selected as classification features, and fuzzy C-means clustering algorithm is used to analyze historical operating data, dividing the operating data into several representative operating condition categories.

4. The method for predicting three-dimensional boiler temperature parameters based on deep reinforcement learning according to claim 3, characterized in that, The step of using the fuzzy C-means clustering algorithm to analyze historical operating data and divide the operating data into several representative operating condition categories is as follows: Introducing a fuzzy membership mechanism, for each sample point For the cluster center membership degree Values ​​in the range And satisfy the normalization constraint: ； in, As a historical working condition sample, This represents the total number of historical operating condition samples. Number of operating condition categories; The objective function for fuzzy clustering is defined as: ； in, Indicates sample Belongs to the Membership degree of each working condition category; Represents the fuzziness index, controlling the degree of fuzziness in the membership distribution. The larger the sample size, the smaller the difference in membership degree between the samples and each category, and the more ambiguous the division of working conditions. Indicates the first Cluster centers for each working condition category; Indicates sample With cluster center The Euclidean distance between them; Through iterative optimization, the cluster center and membership degree are updated alternately according to the following formula: The cluster center update formula is: ； Wherein, the numerator is the weighted sum of the samples, and the weight is the fuzzy membership degree of the sample to the cluster; the denominator is the normalization term of the weighted coefficient of the membership degree of the cluster. The membership update formula is: ； Where the denominator is the sample The cumulative term of the distance ratio to all cluster centers ensures that the sum of the membership degrees of each sample in all clusters is 1; The iterative process continues until the membership matrix is ​​reached. When the change is less than the preset threshold or the maximum number of iterations is reached, the working condition classification is completed.

5. The method for predicting three-dimensional boiler temperature parameters based on deep reinforcement learning according to claim 1, characterized in that, The step of adaptively resampling the three-dimensional temperature field data generated by CFD to reduce the data size is as follows: Calculate the CFD temperature field in three directions at each grid node The partial derivatives on the x-axis yield the gradient vector. , It is the temperature field in three-dimensional space, and then its gradient magnitude is calculated. ; For each grid node Assign an importance weight to the sampled item. : ； in, It is the total number of grid nodes in the entire CFD computation domain. It is an adjustment constant used to ensure that a basic probability is sampled even in flat regions where the gradient is zero; Regions with large gradient magnitudes correspond to areas of drastic temperature changes and are assigned higher sampling probabilities, while regions with small gradient magnitudes correspond to areas of stable temperature and are assigned lower sampling probabilities, based on the calculated probability distribution. A preset number of subsets are drawn from the original grid nodes using random sampling or systematic sampling methods.

6. The method for predicting three-dimensional boiler temperature parameters based on deep reinforcement learning according to claim 1, characterized in that, The steps for establishing a deep reinforcement learning prediction model that takes operating parameter features and spatial coordinates as input and temperature as output are as follows: The deep reinforcement learning model adopts an Actor-Critic architecture, which consists of a policy network and a value network. The input is a 10-dimensional fused feature vector, where 7 dimensions are preprocessed DCS operating parameter features representing the current boiler operating status, and the other 3 dimensions are spatial coordinates. Specify the temperature location to be predicted; the input vector serves as the current state. ; Both the policy network and the value network employ a 5-layer fully connected neural network structure as feature extractors. The layer output is: ； in, and The first Layer weight matrix and bias vector; Activation function ReLU is used to enhance the model's nonlinear fitting capability: ； Policy Network Output Probability Distribution , indicating the state Predict the motion distribution of temperature at a specific spatial point; Value network output scalar That is, the value estimate of the state, which is used to evaluate the merits of the current prediction strategy.

Images