A modeling method of a boiler intelligent combustion system based on incremental input and output
By adopting an incremental input-output modeling method for intelligent boiler combustion systems, and utilizing an LSTM encoder-decoder model and sliding window technology, the prediction lag and multi-objective interference problems of boiler combustion systems under non-stationary operating conditions are solved, achieving high-precision prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING GUITU TECH DEV
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-29
AI Technical Summary
Existing boiler combustion system modeling methods have insufficient generalization ability when facing non-stationary operating conditions such as load and coal quality fluctuations, and gradient interference is prone to occur when making multi-objective predictions, affecting the accuracy of predictions.
An incremental input-output modeling method for intelligent combustion systems in boilers is adopted. By calculating the incremental characteristics of control and target quantities, an independent LSTM encoder-decoder prediction model is constructed. The model is trained using a sliding window technique to avoid multi-target interference and improve the model's ability to track dynamic processes.
It significantly improves the prediction accuracy and robustness of boilers under complex operating conditions, effectively captures dynamic trends, eliminates interference between multiple objectives, and enhances the prediction accuracy of the model.
Smart Images
Figure CN122113603A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of thermal automation technology, and relates to boiler combustion system modeling technology, specifically to a boiler intelligent combustion system modeling method based on incremental input and output. Background Technology
[0002] The combustion system of a coal-fired boiler is a typical complex dynamic system characterized by multiple variables, strong coupling, large time lag, and nonlinearity. In actual operation, boiler load, coal quality, and other operating conditions fluctuate frequently, leading to significant nonstationarity in the combustion process. Existing modeling methods are mainly divided into mechanistic modeling and data-driven modeling. Mechanistic modeling is complex and struggles to address uncertainties in real-time. Traditional data-driven methods typically use the absolute values of process variables for training; however, this approach can cause the model to overemphasize the fitting of steady-state values while neglecting the dynamic incremental characteristics of system parameters over time, resulting in insufficient generalization ability under varying operating conditions. Furthermore, when the system contains multiple target variables with different physical meanings, using a single prediction model often leads to gradient interference between different targets, affecting the final prediction accuracy. Summary of the Invention
[0003] Purpose of the invention: In order to overcome the shortcomings of the existing technology, this invention provides a modeling method for intelligent combustion systems of boilers based on incremental input and output, which can effectively extract dynamic incremental features of the boiler combustion process and eliminate interference between multiple objectives, so as to improve the prediction accuracy and robustness of the model under complex boiler operating conditions.
[0004] Technical Solution: To achieve the above objectives, this invention provides a modeling method for a boiler intelligent combustion system based on incremental input-output, comprising the following steps:
[0005] S1: Read the historical operating data of the boiler combustion system to determine the control variables, status variables and target variables of the boiler combustion system;
[0006] S2: Normalize the historical operating data of the boiler combustion system and divide it into training set, validation set and test set according to the proportion;
[0007] S3: Calculate the increments of the control quantity and the target quantity, construct incremental time series data, and input the original data and incremental time series data into the prediction model together;
[0008] S4: Traverse all target quantities determined in step S1, construct an independent LSTM encoder-decoder prediction model for each target quantity, and use incremental time series data captured by the sliding window to iteratively train and update each model.
[0009] Furthermore, in step S1, the control quantities, state quantities, and target quantities of the boiler combustion system are determined based on the actual structure of the target boiler, and a one-to-one correspondence is established with the specific physical measurement points in the boiler system.
[0010] Furthermore, in step S2, the training set is randomly sampled to enhance the model's generalization ability, and the validation set is sampled sequentially over time to verify the accuracy of the prediction.
[0011] Furthermore, in step S3, the incremental feature, namely the difference between the current time and the previous time, is used to reflect the dynamic rate of change of the system. The formula for calculating the increment is as follows:
[0012]
[0013] In the formula, For the increment of input, Enter the value of the quantity at the current moment. The value of the input quantity from the previous moment;
[0014]
[0015] In the formula, For the output increment, The value of the output quantity at the current moment. This is the value of the output at the previous moment.
[0016] Furthermore, in step S4, an incremental feature input layer is constructed in the LSTM encoder-decoder prediction model, which forces the LSTM to focus on learning the dynamic rate of change of the sequence, thereby improving the model's ability to capture non-stationary boiler operating conditions.
[0017] Furthermore, the method of using a sliding window to extract incremental time-series data in step S4 includes: setting the time steps of the encoder and decoder respectively, extracting historical time-series incremental data based on the encoder time step as model input, and extracting future time-series incremental data based on the decoder time step as model prediction target.
[0018] Furthermore, in step S4, the independent training algorithm that iterates through each target and updates each model includes the following methods:
[0019] Suppose the system has K target variables to be predicted, and the training process is as follows:
[0020] A1: For each target variable Construct an incremental sequence subset targeting this objective from the total training set. :
[0021]
[0022] In the formula, For encoder input, For decoder input, For the first The true incremental label of each target The total number of samples;
[0023] A2: Constructing the first A separate LSTM network , in These are parameters unique to this network; during forward propagation, the prediction increment is calculated. ;
[0024] A3: Based on the predicted increment Construct the loss function ;
[0025] A4: For each sub-model Gradient descent updates are performed independently; the Adam optimization algorithm is used, based on the loss function. For parameters Perform iterations:
[0026]
[0027] Through the above arrive Through the traversal loop, the system completes the construction and convergence of prediction models for all target variables, ensuring that the prediction error gradient of one target will not backpropagate and affect the feature extraction parameters of other targets, thereby achieving physical decoupling.
[0028] Furthermore, in step A3, the L1 norm is used as the loss function. To enhance the model's robustness to abnormal fluctuations:
[0029]
[0030] This invention aims to address the shortcomings of existing data-driven models, which focus solely on static absolute values and exhibit weak dynamic trend capture and prediction lag under non-stationary conditions such as load fluctuations. The core of this method lies in enhancing the learning of dynamic processes by introducing incremental features. Specifically, historical data is first acquired from the DCS (Distributed Control System) to determine control variables, state variables, and target variables. Then, the increments of the control and target variables are calculated and used, along with the original data, as input features to reflect the system's rate of change. Subsequently, the technical solution employs an independent modeling strategy, constructing a dedicated LSTM encoder-decoder model for each target variable and applying a sliding window technique for sampling and training. By combining incremental features with independent LSTM models, this invention effectively avoids multi-target interference and significantly improves the model's ability to track and predict the complex dynamic characteristics of boilers.
[0031] Beneficial Effects: Compared with existing technologies, this invention solves the problems of weak dynamic trend capture and multi-objective interference in existing models by introducing incremental data features and an independent modeling strategy. Specifically, the primary technical problem this invention addresses is that traditional data-driven models often only focus on the static absolute values of process quantities, making it difficult for them to capture the dynamic trends of system parameter changes under non-stationary conditions such as load fluctuations or coal quality fluctuations, resulting in significant prediction lag. This invention, by introducing incremental features, enables the model to learn parameter changes, thereby improving its ability to track dynamic processes. It can effectively extract dynamic incremental features of the boiler combustion process, eliminate interference between multiple objectives, and improve the prediction accuracy and robustness of the model under complex boiler operating conditions. Attached Figure Description
[0032] Figure 1 This is a flowchart of the method of the present invention;
[0033] Figure 2 A graph showing the multi-step prediction results of the separator outlet steam temperature;
[0034] Figure 3 Graph showing the results of multi-step prediction of reheat steam temperature;
[0035] Figure 4 A graph showing the results of multi-step prediction of flue gas temperature;
[0036] Figure 5 The figure shows the results of the multi-step prediction of nitrogen oxide emissions. Detailed Implementation
[0037] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.
[0038] Example 1:
[0039] like Figure 1 As shown in the figure, this embodiment provides a modeling method for a boiler intelligent combustion system based on incremental input and output, including the following steps:
[0040] S1: Read the historical operating data of the boiler combustion system to determine the control variables, status variables and target variables of the boiler combustion system;
[0041] The control quantities, state quantities, and target quantities of the boiler combustion system are determined based on the actual structure of the target boiler, and a one-to-one correspondence is established with the specific physical measurement points in the boiler system.
[0042] The control variables, state variables, and target variables obtained in this embodiment are shown in Table 1.
[0043] Table 1 Boiler Modeling Variables
[0044]
[0045] S2: Normalize the historical operating data of the boiler combustion system and divide it into training set, validation set and test set according to the proportion;
[0046] In this embodiment, the training set is randomly sampled to shuffle the sample time sequence and enhance the model's generalization ability. The validation set is sampled in time sequence to verify the accuracy of the prediction and ensure the validity of the validation results.
[0047] The dataset is divided into 70% training set, 15% validation set, and 15% test set.
[0048] S3: Calculate the increments of the control quantity and the target quantity, construct incremental time series data, and input the original data and incremental time series data into the prediction model;
[0049] When constructing time-series samples, the system employs an incremental approach for feature extraction. Specifically, the system not only extracts the values of control and state variables at the current moment but also simultaneously calculates the differences between the control and state variables at the current moment and the previous moment, obtaining the control variable increment and state variable increment, respectively. The final input consists of both the original data and the incremental data. This hybrid input method explicitly introduces the first-order rate of change information of the parameters, enhancing the model's sensitivity to the dynamic trends of the combustion process.
[0050] The increment characteristic is the difference between the current time and the previous time, used to reflect the dynamic rate of change of the system. The formula for calculating the increment is as follows:
[0051]
[0052] In the formula, For the increment of input, Enter the value of the quantity at the current moment. The value of the input quantity from the previous moment;
[0053]
[0054] In the formula, For the output increment, The value of the output quantity at the current moment. This is the value of the output at the previous moment.
[0055] S4: Traverse all target quantities determined in step S1, construct an independent LSTM encoder-decoder prediction model for each target quantity, and use incremental time series data captured by the sliding window to iteratively train and update each model.
[0056] The method of using a sliding window to extract incremental time series data includes: setting the time steps of the encoder and decoder respectively, extracting historical time series incremental data based on the encoder time step as the model input, and extracting future time series incremental data based on the decoder time step as the model prediction target.
[0057] The LSTM model consists of an encoder and a decoder. The encoder receives historical data, and the decoder outputs the future prediction sequence. The specific configuration of the LSTM encoder-decoder prediction model in this embodiment is shown in Table 2.
[0058] Table 2 Feature space and hyperparameters of LSTM models
[0059]
[0060] This invention designs an incremental prediction architecture based on independent channels. Existing technologies often use a single model to predict multiple target variables simultaneously, which easily leads to gradient interference between different physical quantities. The improved architecture of this invention is characterized by "decoupling":
[0061] The system iterates through each objective quantity and constructs an independent LSTM prediction model for it. The parameters of each model are not shared, thereby eliminating the mutual constraints in multi-objective optimization.
[0062] In the input layer, this invention changes the traditional structure of directly inputting absolute values and instead constructs an incremental feature input layer, forcing the LSTM to focus on learning the dynamic rate of change of the sequence, thereby improving the model's ability to capture non-stationary boiler operating conditions.
[0063] To achieve multi-objective decoupling, this invention abandons the traditional joint training mode and instead adopts an independent training algorithm that iterates through each object. Assume the system has K target variables to be predicted, and the training process is as follows:
[0064] 1) For each target variable Construct an incremental sequence subset targeting this objective from the total training set. :
[0065]
[0066] In the formula, For encoder input, For decoder input, For the first The true incremental label of each target The total number of samples.
[0067] 2) Construct the first A separate LSTM network , in These are parameters unique to this network. During forward propagation, the prediction increment is calculated. This invention uses the L1 norm as the loss function. To enhance the model's robustness to abnormal fluctuations:
[0068]
[0069] 3) For each sub-model Gradient descent updates are performed independently. The Adam optimization algorithm is used, based on the loss function. For parameters Perform iterations:
[0070]
[0071] Through the above arrive Through the traversal loop, the system completes the construction and convergence of prediction models for all target variables, ensuring that the prediction error gradient of one target will not backpropagate and affect the feature extraction parameters of other targets, thereby achieving physical decoupling.
[0072] Example 2:
[0073] To verify the effectiveness and impact of the method of the present invention, this embodiment applies the method of the present invention to a real boiler combustion system, as follows:
[0074] 1) An incremental multi-step prediction model was established based on historical data from the power plant's DCS system. The data sampling interval was 10 seconds, with a total of 100,000 data points.
[0075] 2) To build a high-performance LSTM prediction model, its key structural parameters and training parameters need to be systematically configured and optimized. A grid search method is used to optimize the network structure and training parameters. Specific parameters and their value ranges are as follows: the number of hidden layer units is set to {64, 128, 256}, the number of hidden layers is {1, 2, 3, 4}, the regularization parameter Dropout probability is {0.1, 0.2, 0.3, 0.4}, the learning rate during model training is {0.01, 0.001, 0.0001}, the encoder stride is {10, 20, 40}, and the decoder stride is {40, 60, 80}. These parameter combinations collectively determine the structural complexity and training stability of the LSTM encoder-decoder network.
[0076] Based on the above settings to verify the effectiveness of the method, a test set data sample was used for experimentation, and MAPE was used as the main evaluation indicator during the testing phase.
[0077] Prediction results are as follows Figures 2-5 As shown:
[0078] Figures 2-5 The results shown all demonstrate that the incremental multi-step prediction model of this invention has excellent predictive performance. The 80-step prediction MAPE for separator outlet steam temperature is 1.239%, for reheat steam temperature it is 0.730%, for flue gas temperature it is 0.253%, and for nitrogen oxide emissions it is 11.079%. The accuracy of each prediction target has reached a high level, providing solid data support and scientific basis for power plant combustion strategy optimization and energy efficiency improvement.
Claims
1. A modeling method for a boiler intelligent combustion system based on incremental input-output, characterized in that, Includes the following steps: S1: Read the historical operating data of the boiler combustion system to determine the control variables, status variables and target variables of the boiler combustion system; S2: Normalize the historical operating data of the boiler combustion system and divide it into training set, validation set and test set according to the proportion; S3: Calculate the increments of the control quantity and the target quantity, construct incremental time series data, and input the original data and incremental time series data into the prediction model; S4: Traverse all target quantities determined in step S1, construct an independent LSTM encoder-decoder prediction model for each target quantity, and use incremental time series data captured by the sliding window to iteratively train and update each model.
2. The modeling method for a boiler intelligent combustion system based on incremental input / output according to claim 1, characterized in that, In step S1, the control quantities, state quantities, and target quantities of the boiler combustion system are determined based on the actual structure of the target boiler, and a one-to-one correspondence is established with the specific physical measurement points in the boiler system.
3. The modeling method for a boiler intelligent combustion system based on incremental input / output according to claim 1, characterized in that, In step S2, the training set is randomly sampled to enhance the model's generalization ability, and the validation set is sampled sequentially over time to verify the accuracy of the prediction.
4. The modeling method for a boiler intelligent combustion system based on incremental input / output according to claim 1, characterized in that, In step S3, the incremental feature, which is the difference between the current time and the previous time, is used to reflect the dynamic rate of change of the system. The formula for calculating the increment is as follows: ； In the formula, For the increment of input quantity, Enter the value of the quantity at the current moment. The value of the input quantity from the previous moment; ； In the formula, For the output increment, The value of the output quantity at the current moment. This is the value of the output at the previous moment.
5. The modeling method for a boiler intelligent combustion system based on incremental input / output according to claim 1, characterized in that, In step S4, an incremental feature input layer is constructed in the LSTM encoder-decoder prediction model, forcing the LSTM to focus on learning the dynamic rate of change of the sequence.
6. The modeling method for a boiler intelligent combustion system based on incremental input / output according to claim 1, characterized in that, The method of using a sliding window to extract incremental time series data in step S4 includes: setting the time steps of the encoder and decoder respectively, extracting historical time series incremental data based on the encoder time step as model input, and extracting future time series incremental data based on the decoder time step as model prediction target.
7. The modeling method for a boiler intelligent combustion system based on incremental input / output according to claim 1, characterized in that, The step S4 employs an independent training algorithm that iterates through each target, and the methods for iteratively training and updating each model include: Suppose the system has K target variables to be predicted, and the training process is as follows: A1: For each target variable Construct an incremental sequence subset targeting this objective from the total training set. : ； In the formula, For encoder input, For decoder input, For the first The true incremental label of each target The total number of samples; A2: Constructing the first A separate LSTM network , in These are parameters unique to this network; during forward propagation, the prediction increment is calculated. ; A3: Based on the predicted increment Construct the loss function ; A4: For each sub-model Gradient descent updates are performed independently; the Adam optimization algorithm is used, based on the loss function. For parameters Perform iterations: ； Through the above arrive Through iterative loops, the system completes the construction and convergence of prediction models for all target variables.
8. The modeling method for a boiler intelligent combustion system based on incremental input / output according to claim 7, characterized in that, In step A3, the L1 norm is used as the loss function. To enhance the model's robustness to abnormal fluctuations: 。