Ultra-supercritical unit modeling method based on dynamic weight adaptive Transform
Through the dynamic weight adaptive Transformer model, combined with the channel self-attention mechanism and sliding window technology, the complexity problem of ultra-supercritical unit modeling is solved, and accurate prediction and efficient modeling of the unit are achieved.
Patent Information
- Application Number
- CN202510625481.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-15
- Publication Date
- 2025-09-16
AI Technical Summary
Existing technologies make it difficult to accurately model the multivariable, large delay, large inertia, strongly coupled time-varying and uncertain characteristics of ultra-supercritical units, resulting in insufficient model accuracy and applicability.
A dynamic weight adaptive Transformer model is adopted, combined with the channel self-attention mechanism and sliding window technology to separate global and local data features, and feature fusion is performed through a dynamic weight allocation module to build a parallel training framework and improve the model's predictive ability.
It has achieved accurate modeling of ultra-supercritical units, improved the model's generalization ability and prediction accuracy, and can effectively capture the unit's long-cycle operating conditions and minute-level dynamic fluctuations, thereby improving the prediction accuracy of the thermal system.
Smart Images
Figure CN120654533A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of power system modeling and intelligent control, and in particular relates to a modeling method of an ultra-supercritical unit based on a dynamic weight adaptive Transformer. Background Art
[0002] Thermal power generation primarily converts the heat released by combustible materials into electricity through power generation units. This includes coal-fired, gas-fired, and oil-fired power generation. Within coal-fired power generation, supercritical units have attracted significant attention due to their unique advantages. During operation, supercritical units maintain steam pressure and temperature exceeding critical values, typically exceeding 25 MPa and 580°C. They utilize variable-pressure operation, automatically adjusting operating parameters based on grid load fluctuations for efficient operation. Their high thermal efficiency, typically exceeding 40%, effectively reduces energy consumption. Supercritical units also excel in environmental protection. By utilizing advanced combustion technology and pollution control equipment, they significantly reduce emissions of pollutants such as sulfur dioxide, nitrogen oxides, and soot, meeting national environmental standards. Furthermore, supercritical units offer strong load adaptability, enabling stable operation across a wide load range, meeting the grid's flexibility requirements.
[0003] When an ultra-supercritical unit is in operation, water is heated above the critical point (374°C and 22.1 MPa) to produce high-temperature, high-pressure steam. In this state, there is no clear boundary between water and steam, and the density and thermal conductivity are high, thereby improving thermal efficiency. In the boiler, the high-temperature gas produced by the combustion of fossil fuels transfers heat to the water, converting it into supercritical steam. Some of the steam is reheated in the reheater to further increase the temperature and efficiency. Subsequently, the high-temperature, high-pressure steam drives the turbine to rotate, which in turn drives the generator to generate electricity. The used steam is condensed into water in the condenser, completing the cycle and being recycled. Due to its high efficiency and low coal consumption, this unit can significantly improve energy utilization efficiency and reduce environmental pollution compared to traditional subcritical units. Therefore, ultra-supercritical units are widely used in thermal power generation and have become the main units in the power industry.
[0004] However, the operation of supercritical units also faces several challenges. Their thermal systems are complex, involving the interconnectedness of multiple subsystems and highly variable operating conditions. These characteristics result in units with multiple variables, large delays, large inertia, strongly coupled time-varying behavior, and uncertainty, making accurate modeling of the units difficult. Current mechanistic modeling methods for thermal power units analyze the physical and chemical processes within the boiler-turbine system and construct mathematical models based on principles such as energy conservation, thermodynamics, and fluid mechanics. These methods reduce computational complexity by assuming uniform heat distribution, constant heat transfer efficiency, and simplified combustion models, but this also sacrifices model accuracy and applicability. Summary of the Invention
[0005] Purpose of the invention: The present invention provides a modeling method for ultra-supercritical units based on a dynamic weight adaptive Transformer. Taking into account the multivariable coupling, strong nonlinear characteristics, and global and local characteristics of the operating data of the unit, the channel self-attention mechanism is introduced into the Transformer model, a parallel training framework that captures global and local data features is proposed, and a dynamic weight allocation module is designed.
[0006] The present invention provides an ultra-supercritical unit modeling method based on a dynamic weight adaptive Transformer, comprising the following specific steps:
[0007] Collect unit data and divide the unit data into global input data D f and local input data D q Input into the improved transformer model;
[0008] The improved transformer model includes a transformer model and a dynamic weight adaptation module;
[0009] The Transformer model is used to process the global input data D f and local input data D q Extract the nonlinear features X F and nonlinear features X Q , and then the nonlinear feature X is transformed through the linear layer F Mapped into global prediction feature y f , the nonlinear feature X Q Mapped into predictions Mapped into local prediction features y q ;
[0010] The dynamic weight adaptive module is used to transform the global prediction feature y f and local prediction features y q A weighted combination is performed to obtain a composite feature y of the predicted value, and the composite feature y of the predicted value is mapped to the predicted value of the unit through a linear layer.
[0011] Furthermore, the unit data includes the coal flow instruction U B , total water flow D fw , throttle valve opening U t ; The predicted value of the unit is the unit load Ne, the main steam pressure value Pst or the separator enthalpy value hm.
[0012] Furthermore, the sliding window and slicing technology is used to divide the unit data into global input data D f and local input data D q, the specific division method is shown as follows:
[0013]
[0014] Among them, L is the unit data length, W is the window length, T is the step size, and n represents the number of windows. Represents the input data of the i-th group window length w, Represents the input data of the jth group with a window length of w.
[0015] Furthermore, for the global input data D f and local input data D q Extract the nonlinear features X F and nonlinear features X Q , the specific steps are:
[0016] The channel self-attention mechanism is used to convert the global input data D f Mapped into a weighted global composite feature X f , the local input data D q Mapped into weighted local composite feature X q ;
[0017] For the global composite feature X f Perform feature extraction to obtain nonlinear features X F ;
[0018] For the local composite feature X q Perform feature extraction to obtain nonlinear features X Q .
[0019] Furthermore, the channel self-attention mechanism is used to convert the global input data D f Mapped into a weighted global composite feature X f , the local input data D q Mapped into weighted local composite feature X q , the specific expression is as follows;
[0020]
[0021] Among them, X fq Represents the global composite feature X output after the channel self-attention mechanism f or local composite feature X q , H represents the feature height, W represents the feature width, D fq Represents the global input data D f Or local input data D q , W1 represents the weight of the FC1 fully connected layer, W2 represents the weight of the FC2 fully connected layer, σ and δ represent the ReLU and Sigmoid activation functions respectively.
[0022] Furthermore, the global composite feature X f Perform feature extraction to obtain nonlinear features X F , the specific steps are:
[0023] Global composite feature X f Feature extraction is performed through the self-attention mechanism, which is expressed as follows:
[0024] X′ F =LN(X f +Attention(Q f ,K f ,V f ))
[0025]
[0026] in, represents the output of the self-attention mechanism, n represents the number of samples, d model represents the input dimension, d q Indicates the dimension of the query, and represents the matrix parameters, d k Indicates the dimension of the key, d v represents the dimension of each header's value, Represents the corresponding query, key, value pair.
[0027] Next, the global data feature X′ F Input feedforward network module to extract nonlinear features X F , expressed as follows:
[0028]
[0029] in, Represents the final encoded output, represents the matrix parameters, d ff represents the hidden dimension, b1 and b2 represent the bias vectors.
[0030] For the local composite feature X q Perform feature extraction to obtain nonlinear features X Q The steps and global composite feature X f Perform feature extraction to obtain nonlinear features X F The steps are the same as
[0031] Furthermore, set the sliding window W to select the global prediction value y one by one f , evaluate the output fluctuation in each window, using the determination coefficient R2 As a measurement standard, the weights of global prediction features and local prediction features are assigned. The specific formula is as follows:
[0032]
[0033] Among them, L q represents the global loss, L f represents local loss, W q Global loss L q The weight, W f Represents the local loss L f Weight, R 2 represents the determination coefficient of the judgment mechanism, and α represents the threshold value;
[0034]
[0035] y q ,y f represents the local true value and the global true value, Represents the predicted value
[0036] The composite feature y of the predicted value is expressed as:
[0037] y=W f y f +W q y q .
[0038] Furthermore, the improved transformer model is trained using the loss function L DWAT Expressed as:
[0039]
[0040] Among them, W q Loss L q Weight, W f Indicates loss L q Weight, R 2 represents the determination coefficient of the judgment mechanism, and α represents the threshold value.
[0041] Beneficial effects: Through the channel self-attention mechanism (step S1), cross-dimensional interaction of input data is realized, breaking through the simplified assumption of variable coupling relationship in traditional mechanism models. Based on the improved Transformer architecture (step S2), a parallel processing channel is constructed: the global attention layer models the evolution law of the long-term operating conditions of the unit through a multi-head mechanism, while the local convolution layer focuses on minute-level dynamic fluctuations (. This dual-channel design enables the model to accurately capture local details such as micro-oscillations of the reheater tube wall temperature while retaining the macroscopic thermodynamic characteristics of the supercritical phase change process. The weight adaptive module is innovatively designed to achieve dynamic matching of global and local features through a differentiable gating network. The feature fusion weight can be automatically adjusted according to state parameters such as real-time steam pressure and main steam temperature. This patent invention effectively solves the problem of accurate modeling of strongly coupled time-varying systems of ultra-supercritical units, and provides key technical support for the construction of intelligent power generation systems under new power systems. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is a principle block diagram of the method of the present invention. DETAILED DESCRIPTION
[0043] The present invention will be further described below with reference to the accompanying drawings. It should be understood that the embodiments described herein are only a portion of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.
[0044] The ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer provided by the present invention has the following specific principles: Figure 1 As shown in the paper, first, a channel self-attention mechanism is introduced into the Transformer to efficiently identify the correlations between different input features and dynamically adjust the weights, thereby enhancing the model's ability to capture potential input features. Then, a parallel training framework that captures global and local data features is proposed. While utilizing global features, it fully considers local features, improving the network's ability to understand complex time series information and enhancing the model's generalization capabilities. Finally, a dynamic weight allocation module is designed to flexibly adjust the weights of global and local features based on the changing trends of the unit's output variables, thereby targetedly improving network performance. Experiments on real data from a 1000MW ultra-supercritical unit further verified the superiority of this method.
[0045] like Figure 1 As shown, the ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer of the present invention specifically includes the following steps:
[0046] Step 1: Data collection, i.e. collecting raw data x. Each set of raw data x includes coal flow instruction U B , total water flow D fw , throttle valve opening U t ;
[0047] In order to capture the historical trend of time series, the sliding window and slicing techniques are used to divide the collected data into global data D f Input and local data D q This method helps to extract representative features from time series to reflect the behavior of the system at different time scales. Global input data D f and local input data D q The division methods are shown as follows:
[0048]
[0049] Among them, L is the data length, W is the window length, T is the step size, and n represents the number of windows. Represents the input data of the i-th group of windows with a length of w. Represents the input data of the jth group with a window length of w.
[0050] Step 2: construct an improved transformer model (DWAT network), including a transformer model and a dynamic weight adaptation module; the transformer model includes a channel self-attention mechanism and a transformer network;
[0051] The divided data is globally input into data D f , local input data D q First, it is fed into the Transformer model. In the Transformer model, the channel self-attention mechanism extracts the global input data D f and local input data D q China Coal Flow Instruction U B , total water flow D fw , throttle valve opening U t Data features, assign weights according to the importance of each set of input features, and combine the global data features with the weights to map them into weighted global composite features X f , combining local data features with weights and mapping them into weighted local composite features X q The specific process is as follows:
[0052] Step 2.1, input data D fq Coal flow instruction U in B , total water flow D fw , throttle valve opening Ut The features are concatenated along the feature channel dimension to form a concatenated feature map:
[0053] D fq =Concat(U B ,D fw ,U t )
[0054] Among them, D fq Represents the global input data D f Or local input data D q , the concatenated feature map D fq The dimension is [H, W, C], where H represents the feature height, W represents the feature width, and C represents the number of original channels.
[0055] Then, through global average pooling, the two-dimensional features [H, W] of each channel are compressed into a one-dimensional real number, and the feature map dimension is compressed from [H, W, C] to [1, 1, C].
[0056] Next, a fully connected layer (FC1) reduces the channel dimension of the feature map vector to 1 / r of its original dimension, i.e., [1, 1, C × 1 / r]. After a Swish activation function, another fully connected layer (FC2) upscales the feature map back to its original dimension of [1, 1, C]. A sigmoid function is then used to convert the feature map vector into a normalized weight vector between 0 and 1.
[0057] Finally, the normalized weights are multiplied by the original input feature map channel by channel to generate the weighted composite feature X fq .
[0058]
[0059] Among them, X fq Represents the global composite feature X output after the channel self-attention mechanism f or local composite feature X q , W1 represents the weight of the FC1 fully connected layer, W2 represents the weight of the FC2 fully connected layer, σ and δ represent the ReLU and Sigmoid activation functions.
[0060] Step 2.2: The global composite feature X f or local composite feature X q Input to the encoder in the Transformer network to obtain the nonlinear feature X F and nonlinear features X Q ;
[0061] The encoder is composed of multiple identical stacked layers. Each layer includes two main sub-modules: the self-attention mechanism and the feedforward neural network. After the output of the self-attention and feedforward networks, residual connections are introduced to prevent input information from being lost. At the same time, layer normalization is used to ensure the stability of data distribution.
[0062] Taking the global composite feature X f For example, for the global composite feature X f For global data feature extraction, the process can be expressed as follows:
[0063] X′ F =LN(X f +Attention(Q f ,K f ,V f ))
[0064]
[0065] in, represents the output of SA (self-attention mechanism), and represents the matrix parameters, Represents the corresponding query and key-value pairs.
[0066] Next, the global data feature X′ F Input feedforward network module to extract nonlinear features X F , expressed as follows:
[0067]
[0068] in, Represents the final encoded output, represents the matrix parameters, d ff represents the hidden dimension, b1 and b2 represent the bias vectors.
[0069] For the local composite feature X q Perform feature extraction to obtain nonlinear features X Q The steps and global composite feature X f Perform feature extraction to obtain nonlinear features X F The steps are the same as
[0070] Step 2.3, the nonlinear feature X is transformed through the linear layer F Mapped into the global prediction value y f , local data feature X′ q Mapped into predictions Mapped into local prediction values y q ;
[0071] Similarly, the decoder structure is also composed of multiple identical modules stacked together, but it differs from the encoder in that a lower triangular mask matrix is added to ensure that information at future times cannot be accessed in advance during the decoding process to avoid information leakage.
[0072] Step 2.4: Set the sliding window W to select the global prediction value y one by one f , evaluate the output fluctuation in each window, using the determination coefficient R 2 As a measure, based on R 2 The analysis results of the value are used to dynamically assign feature weights. Set a threshold α and compare α with R 2 The size of the local and global weight distribution is determined by the size of the local and global weight distribution. 2 When R is greater than α, it indicates that the model prediction performance in this window is good and the unit output state is stable. At this time, the global and local feature weights are allocated according to the global and local training losses. 2 When ≤α, it indicates that the model prediction effect in this window is poor and the unit output state fluctuates greatly. At this time, it is necessary to analyze the fluctuation trend in the window more carefully and calculate the value of the model according to R. 2 The value of is used to assign weights to global prediction features and local prediction features. The specific formula is as follows:
[0073]
[0074] The global and local feature dynamic weight adaptive module adjusts the global prediction feature y according to the global change trend of the unit output variable f and local prediction features y q The weight ratio between them is used to obtain the predicted composite feature y, which is expressed as:
[0075] y=W f y f +W q y q
[0076] The composite feature y of the predicted value is then input into the linear mapping layer to obtain the final unit prediction value, which is the unit's unit load Ne, main steam pressure Pst, or separator enthalpy hm. This process helps identify long-term trends and local changes in unit performance, improving the accuracy of the prediction results.
[0077] Finally, to ensure the accuracy of the improved transformer model, the following loss function L is designed: DWAT , according to the global and local feature training loss allocation weights, effectively integrating global and local data features to meet the requirements of model training. Loss function L DWAT Expressed as:
[0078]
[0079] Among them, W q 、W f Indicates loss L q Weight, L f Weight, R 2 represents the determination coefficient of the judgment mechanism, and α represents the threshold value.
Claims
1. The ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer is characterized by: The following steps are involved: Collect unit data and divide the unit data into global input data D f and local input data D q Input into the improved transformer model; The improved transformer model includes a transformer model and a dynamic weight adaptation module; The Transformer model is used to process the global input data D f and local input data D q Extract the nonlinear features X F and nonlinear features X Q ; After the linear layer, the nonlinear feature X F Mapped into global prediction feature y f , the nonlinear feature X Q Mapped into predictions Mapped into local prediction features y q ; The dynamic weight adaptive module is used to transform the global prediction feature y f and local prediction features y q A weighted combination is performed to obtain a composite feature y of the predicted value, and the composite feature y of the predicted value is mapped to the predicted value of the unit through a linear layer.
2. The ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer according to claim 1 is characterized in that: The unit data includes coal flow instruction U B , total water flow D fw , throttle valve opening U t .
3. The ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer according to claim 1 is characterized in that: The unit data is divided into global input data D using sliding window and slicing technology. f and local input data D q , the specific division method is shown as follows: Among them, L is the unit data length, W is the window length, T is the step size, and n represents the number of windows. Represents the input data of the i-th group window length w, Represents the input data of the jth group with a window length of w.
4. The ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer according to claim 1 is characterized in that: For global input data D f and local input data D q Extract the nonlinear features X F and nonlinear features X Q , the specific steps are: The channel self-attention mechanism is used to convert the global input data D f Mapped into a weighted global composite feature X f , the local input data D q Mapped into weighted local composite feature X q ; For the global composite feature X f Perform feature extraction to obtain nonlinear features X F ; For the local composite feature X q Perform feature extraction to obtain nonlinear features X Q .
5. The ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer according to claim 4 is characterized in that: The channel self-attention mechanism is used to convert the global input data D f Mapped into a weighted global composite feature X f , the local input data D q Mapped into weighted local composite feature X q , the specific expression is as follows; Among them, X fq Represents the global composite feature X output after the channel self-attention mechanism f or local composite feature X q , H represents the feature height, W represents the feature width, D fq Represents the global input data D f Or local input data D q , W1 represents the weight of the FC1 fully connected layer, W2 represents the weight of the FC2 fully connected layer, σ and δ represent the ReLU and Sigmoid activation functions respectively.
6. The ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer according to claim 4 is characterized in that: The global composite feature X f Perform feature extraction to obtain nonlinear features X F , the specific steps are: Global composite feature X f Feature extraction is performed through the self-attention mechanism, which is expressed as follows: X′ F =LN(X f +Attention(Q f ,K f ,V f )) in, represents the output of the self-attention mechanism, n represents the number of samples, d model represents the input dimension, d q Indicates the dimension of the query, and represents the matrix parameters, d k Indicates the dimension of the key, d v represents the dimension of each header's value, Represents the corresponding query, key, and value pairs; Next, the global data feature X′ F Input feedforward network module to extract nonlinear features X F , expressed as follows: X F =LN(X f +(max(0,X f W1 T +b1)W2 T +b2)) in, Represents the final encoded output, represents the matrix parameters, d ff represents the hidden dimension, b1 and b2 represent the bias vectors. For the local composite feature X q Perform feature extraction to obtain nonlinear features X Q The steps and global composite feature X f Perform feature extraction to obtain nonlinear features X F The steps are the same as 7. The ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer according to claim 1 is characterized in that: Set the sliding window W to select the global prediction value y one by one f , evaluate the output fluctuation in each window, using the determination coefficient R 2 As a measurement standard, the weights of global prediction features and local prediction features are assigned. The specific formula is as follows: Among them, L q represents the global loss, L f represents local loss, W q Global loss L q The weight, W f Represents the local loss L f Weight, R 2 represents the determination coefficient of the judgment mechanism, and α represents the threshold value; y q ,y f represents the local true value and the global true value, Represents the predicted value The composite feature y of the predicted value is expressed as: y=W f y f +W q y q 。 8. The ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer according to claim 1 is characterized in that: The improved transformer model is trained using the loss function L DWAT Expressed as: Among them, W q Loss L q Weight, W f Indicates loss L q Weight, R 2 represents the determination coefficient of the judgment mechanism, and α represents the threshold value.
9. The ultra-supercritical unit modeling method based on dynamic weight adaptive Transformer according to claim 1 is characterized in that: The predicted value of the unit is the unit load Ne, the main steam pressure value Pst or the separator enthalpy value hm.