Gated circulation unit coal-fired unit coordination modeling method fused with self-attention mechanism

Abstract

The invention discloses a self-attention mechanism fused gated cycle unit coal-fired unit coordination modeling method, which comprises the following steps of: 1, acquiring working condition data of a coal-fired unit, and preprocessing the data; 2, taking the preprocessed data as input, taking predicted values of coal-fired unit power and main steam pressure as output, and constructing a GRU network model fused with a self-attention mechanism; and 3, training and optimizing the GRU network model fused with the self-attention mechanism. According to the gated circulation unit neural network coordination modeling method fusing the self-attention mechanism, the modeling precision and real-time performance of a coal-fired unit under the complex working conditions of deep peak regulation, rapid load change and the like can be remarkably improved, and the method has great significance in guaranteeing safe, efficient and flexible operation of the unit.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of thermal process modeling of thermal power plants, and particularly relates to a coal-fired unit coordinated modeling method fusing a self-attention mechanism. BACKGROUND

[0002] The coal-fired unit has long been responsible for the dual tasks of basic load and peak shaving of the power grid. The characteristics of strong coupling, large inertia and nonlinearity of the furnace, the machine and the electricity subsystems of the coal-fired unit make it difficult for the traditional coordinated modeling method based on physical mechanism or linear simplification to maintain accuracy in the full operating condition range. With the normalization of deep peak shaving, the operating point of the unit frequently passes through the nonlinear region, the coupling relationship of the key variables presents high dimensionality, time variation and long-range dependence, and the traditional model is insufficient in terms of dynamic prediction error, parameter update rate and generalization ability. In recent years, the recurrent neural network and its variants have shown excellent performance in sequence modeling, but they still have problems such as insufficient capture of long-range dependence and unbalanced distribution of key feature weights in the face of multi-scale dynamics of the coal-fired unit with multi-dimensional input, second-level sampling and minute-level response. SUMMARY

[0003] The purpose of the present application is to provide a coal-fired unit coordinated modeling method fusing a self-attention mechanism, in order to solve the technical problem that the traditional mechanism model or the conventional data-driven model is difficult to meet the requirements of high accuracy, real-time performance and robustness at the same time.

[0004] The technical scheme of the present application is a coal-fired unit coordinated modeling method fusing a self-attention mechanism, comprising the following steps:

[0005] Step 1: Collecting the operating condition data of the coal-fired unit and preprocessing the data;

[0006] Step 2: Building a GRU network model fusing a self-attention mechanism with the preprocessed data as the input and the predicted values of the power of the coal-fired unit and the main steam pressure as the output;

[0007] Step 3: Training and optimizing the GRU network model fusing a self-attention mechanism;

[0008] Step 4: Predicting the power of the coal-fired unit and the main steam pressure based on the trained GRU network model fusing a self-attention mechanism.

[0009] Further, step 1 specifically comprises collecting the operating condition data of the coal-fired unit, including steady-state operation, load increase, load decrease and primary frequency modulation disturbance; the preprocessing adopts the maximum-minimum normalization method to linearly map all data to the [0, 1] interval, thereby eliminating the dimension effect; and the specific mathematical transformation formula is as follows:

[0010]

[0011] In the formula: x represents the original data point; x min and x max These represent the minimum and maximum values ​​of the variable recorded in the entire dataset, respectively. norm This is the normalized data obtained after calculation.

[0012] Furthermore, step 2 specifically includes the following steps:

[0013] Step 2.1: Deploy the Gated Loop Unit (GRU) layer, and generate the final hidden state of the current time step by resetting and updating the gates;

[0014] Step 2.2: Introduce a self-attention mechanism layer on the gated recurrent unit (GRU) layer to output the context vector;

[0015] Step 2.3: Map the context vector to the desired final output dimension through a fully connected layer, which is the specific physical prediction value of the unit power and main steam pressure at the next moment.

[0016] Furthermore, step 2.1 specifically includes:

[0017] For each time step t in the input sequence, the GRU unit updates its hidden state ht by the following calculation: First, the reset gate calculates the hidden state ht from the previous time step. t-1 The system takes the current input U(t) as input and outputs a vector between 0 and 1, which will determine the historical information h. t-1 How much is used to calculate the current candidate state;

[0018] Reset gate: r(t)=σ(W) r u(t)+U r h(t-1)+b r (2)

[0019] In the formula: r(t) is the reset gate; σ is the Sigmoid function; W r It resets the gate learning weights; U r It resets the gate input weights; b r It resets the gate bias; u is the input variable; h is the candidate state;

[0020] The update gate determines the extent to which "new" information is used to update "old" information, and this is determined by calculating h. t-1 And U(t), and output a vector between 0 and 1:

[0021] Update gate: z(t)=σ(W) z u(t)+U zh(t - 1) + b z )(3)

[0022] where z(t) is the update gate; σ is the Sigmoid function; W z is the update gate learning weight; U z is the update gate input weight; b z is the update gate bias; u is the input variable; h is the candidate state;

[0023] The model calculates a "candidate" hidden state based on the current input U(t) and the historical information filtered by the reset gate:

[0024] Candidate hidden state:

[0025] where h(t) is the current time hidden state; and represents element-wise multiplication. is the candidate hidden state; tanh is the hyperbolic tangent function; and represents element-wise multiplication; W h is the candidate hidden state learning weight; U h is the candidate hidden state input weight; b h is the candidate hidden state bias; u is the input variable; h is the candidate state;

[0026] Finally, the model uses the update gate z(t) to generate the final hidden state h t of the current time step:

[0027] Current time hidden state:

[0028] where h(t) is the current time hidden state; and represents element-wise multiplication.

[0029] After the GRU layer processes the entire input sequence of length T, a hidden state sequence H = [h1, h2,..., h T ] is generated, which encodes the dynamic evolution information of the coal-fired unit at the past T time steps.

[0030] Further, step 2.2 is specifically: mapping the hidden state sequence H to the query space, key space, and value space, respectively;

[0031] Query matrix: Q = HW Q (6)

[0032] Key matrix: K = HW K (7)

[0033] Value matrix: V = HW V (8)

[0034] where W Q represents the query weight matrix, W Kdenotes the key weight matrix, W V denotes the value weight matrix.

[0035] By calculating the dot product of the Q matrix and the transpose of the K matrix, a score matrix is obtained, where each element represents the degree of attention of one position in the sequence to another position, which is scaled;

[0036]

[0037] where d k is the dimension of the key vector;

[0038] By applying the Softmax function to each row of the score matrix, it is converted into a set of non-negative weight values that sum to 1; the non-negative weight values are the "attention weights", which represent the degree of attention given to all other positions in the input sequence when generating the output of the current position;

[0039] W A = softmax(Scores) (10)

[0040] where W A is the attention weight;

[0041] Finally, the calculated attention weight matrix is multiplied by the value matrix V; the result is a context vector C;

[0042] C = W A · V (11).

[0043] Further, step 2.3 is specifically: the context vector C output by the self-attention layer is a complete and highlighted high-dimensional representation of the entire input time sequence, which is mapped to the desired final output dimension through a standard feedforward neural network layer, i.e. a fully connected layer, and the linear transformation converts the high-dimensional context features into specific physical prediction values of the unit power and main steam pressure at the next time;

[0044]

[0045] where W out and b out are the learnable weight matrix and bias vector of the output layer, which is the final prediction result of the model for the two output variables at the next time.

[0046] Further, step 3 is specifically: an explicit optimization objective is defined by a loss function, and the mean square error is selected as the loss function for continuous value prediction problems, so that the model can quickly correct the deviation:

[0047]

[0048] Wherein: MSE is mean square error; N is sample quantity; Y i is actual value; is model prediction output value;

[0049] The model is optimized by using an Adam optimizer.

[0050] The application also discloses a computer device, including a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to realize the steps of the method.

[0051] The application also discloses a computer readable storage medium, which stores a computer program / instruction, and the computer program / instruction is executed by a processor to realize the steps of the method.

[0052] The application also discloses a computer program product, which includes a computer program / instruction, and the computer program / instruction is executed by a processor to realize the steps of the method.

[0053] Beneficial effects: compared with the prior art, the application has the following remarkable advantages:

[0054] 1. The self-attention mechanism and the gated recurrent unit neural network are deeply fused and applied to the coordinated modeling of the coal-fired unit, the dynamic focusing of the key historical information is realized, and the capture ability of the model to the long-range dependence is significantly improved;

[0055] 2. The time sequence feature extraction and the importance recognition are deeply fused, and the prediction precision and the generalization performance of the model are greatly improved;

[0056] 3. An end-to-end nonlinear dynamic model is constructed, and the complexity and the limitation of the traditional mechanism modeling are effectively avoided;

[0057] 4. The model structure considers the performance and the efficiency, is more suitable for the actual application requirements of the industrial process, can be directly embedded into the existing control or optimization system, provides a unified and reliable state estimation interface for the advanced process control, the economic scheduling and the digital twin.

[0058] 5、Self-attention mechanism can effectively highlight the most relevant historical information for the current output through dynamic weight distribution, and its advantages have been verified in natural language processing, load prediction and other fields. Therefore, constructing a gated recurrent unit neural network fused with a self-attention mechanism can realize high-precision and fast-response modeling of the dynamic multi-variable strong coupling of the unit, which has important influence on improving the flexibility of the unit and supporting deep peak regulation and intelligent control. The coordinated modeling method of the gated recurrent unit neural network fused with the self-attention mechanism provided by the present application can significantly improve the modeling accuracy and real-time performance of the coal-fired unit under complex working conditions such as deep peak regulation and rapid load change, and has important significance for ensuring the safe, efficient and flexible operation of the unit. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A typical coal-fired unit and key control variable schematic diagram are provided for the present application.

[0060] Figure 2 A GRU model unit load identification result diagram fused with a self-attention mechanism is provided for the present application.

[0061] Figure 3 A GRU model unit main steam pressure identification result diagram fused with a self-attention mechanism is provided for the present application. DETAILED DESCRIPTION

[0062] The technical solutions of the present application will be further described below in combination with the drawings.

[0063] A gated recurrent unit coal-fired unit coordinated modeling method fused with a self-attention mechanism, comprising the following steps:

[0064] S1, coordinated system key variable data acquisition and preprocessing

[0065] A typical coal-fired unit coordinated control system Figure 1 The system is the core control loop of the generator unit, and its dynamic behavior can be approximately described by a two-input two-output structure. In order to accurately capture the main dynamic characteristics of the coordinated control system, the present application selects two core manipulated variables of the system as the inputs of the model. Let the input vector be the fuel quantity given signal on the boiler side and the regulating door opening degree command signal on the turbine side. At the same time, select two key controlled state variables of the system: output power and injection pressure as the outputs of the model.

[0066] Data is the basis of the model. From the distributed control system or monitoring information system of the power plant, the historical data of the above four variables in the long-term operation process of the unit are collected in a fixed sampling period. The collected data should cover as many working conditions as possible, such as steady-state operation, load increase, load decrease, primary frequency disturbance, etc., to ensure the generalization ability of the model.

[0067] The original operation data has a huge difference in dimension and numerical range. If it is directly used for neural network training, it will cause unstable model training process, slow convergence, and variables with large numerical range will have an disproportionate impact on model parameters. Therefore, the data must be normalized. In this embodiment, the maximum-minimum normalization method is adopted, which can linearly map all data to the interval [0, 1], thereby eliminating the dimension effect.

[0068] The specific mathematical transformation formula is as follows:

[0069]

[0070] In the formula, x represents the original data point; x min and x max respectively represent the minimum and maximum values recorded by the variable in the entire data set. x norm is the normalized data obtained after calculation.

[0071] S2, construct a GRU network model architecture fused with self-attention mechanism

[0072] At the bottom of the model architecture, a gated recurrent unit (GRU) layer is first deployed. GRU, as an advanced variant of recurrent neural network (RNN), can effectively learn the dependencies in time series data, has high computational efficiency, and is very suitable for processing real-time data streams in industrial processes. GRU determines how information is passed and updated between time steps through its internal sophisticated "gate" structure.

[0073] For each time step t in the input sequence, the GRU unit updates its hidden state ht through the following calculation. First, the model needs to decide how much to "forget" the past information. The reset gate examines the previous moment's hidden state h t-1 and the current moment's input U(t), and outputs a vector between 0 and 1. This vector will determine how much historical information h t-1 can be used to calculate the current candidate state.

[0074]

[0075] In the formula, r(t) is the reset gate; σ is the Sigmoid function; W r is the reset gate learning weight; U r is the reset gate input weight; b r is the reset gate bias; u is the input variable; h is the candidate state.

[0076] Secondly, the model needs to decide how much "new" information to update the "old" information. The update gate is responsible for this decision, which also examines h t-1And U(t), and output a vector between 0 and 1:

[0077]

[0078] In the formula: z(t) is the update gate; σ is the Sigmoid function; W z It updates the gate learning weights; U z It updates the gate input weights; b z It updates the gate bias; u is the input variable; h is the candidate state.

[0079] Next, the model calculates a "candidate" hidden state based on the current input U(t) and some historical information filtered by the reset gate:

[0080] Candidate hidden state:

[0081] In the formula: It represents the candidate hidden state; tanh is the hyperbolic tangent function; ⊙ represents element-wise multiplication; W h It is the candidate hidden state learning weight; U h b is the input weight of the candidate hidden state; h is the candidate hidden state bias; u is the input variable; h is the candidate state.

[0082] Finally, the model uses an update gate to generate the final hidden state h at the current time step. t :

[0083] Current hidden state:

[0084] In the formula: h(t) is the hidden state at the current time; ⊙ represents element-wise multiplication.

[0085] After the GRU layer processes the entire input sequence of length T, it generates a hidden state sequence H = [h1, h2, ..., h...]. T This sequence initially encodes the dynamic evolution information of the coal-fired unit over the past T time points.

[0086] The dynamic response of coal-fired units has the characteristics of large time delay and long-range dependence, that is, the change of current output may be closely related to some input operation a long time ago. Although the traditional GRU can capture the time sequence relationship, it has limited ability to allocate weights for such long-range dependence. Therefore, a self-attention mechanism layer is introduced on top of the GRU layer, which aims to enable the model to learn the importance of information at different times in the input sequence and pay more attention to the features at key times. In order to perform attention calculation, three different roles need to be derived from the same source (i.e. the hidden state sequence H) first. This is achieved by three independent linear transformations (i.e. fully connected layers without activation functions) that map H to the query space, key space and value space, respectively.

[0087] Query matrix: Q = HW Q (6)

[0088] Key matrix: K = HW K (7)

[0089] Value matrix: V = HW V (8)

[0090] The core of attention is to calculate the similarity between queries and keys. By calculating the dot product of the transpose of the Q matrix and the K matrix, a score matrix is obtained, where each element represents the attention degree of one position to another position in the sequence. In order to prevent the dot product result from being too large and causing gradient vanishing, which affects model training, scaling processing is needed.

[0091]

[0092] In the formula: d k is the dimension of the key vector;

[0093] The original score needs to be converted into a valid weight distribution. By applying the Softmax function to each row of the score matrix, it can be converted into a set of non-negative weight values that sum to 1. These values are the "attention weights", which intuitively represent the degree of attention that should be given to all other positions in the input sequence when generating the output at the current position.

[0094] W A = softmax(Scores) (10)

[0095] In the formula: W A is the attention weight.

[0096] Finally, the computed attention weight matrix is multiplied with the value matrix V. The result is a context vector C, which is a highly condensed summary of information that integrates the features of the entire input sequence and has highlighted the information at key moments according to their intrinsic relevance.

[0097] C = W A V (11)

[0098] The context vector C output from the self-attention layer is a high-dimensional representation of the entire input time sequence that is informationally complete and highlights key moments. The last step is to map this context vector to the desired final output dimension through a standard feedforward neural network layer, i.e., a fully connected layer. This linear transformation converts the high-dimensional context features into specific physical prediction values for the next time step of unit power and main steam pressure.

[0099]

[0100] where W out and b out are the output layer's learnable weight matrix and bias vector, which are the final prediction results of the model for the next time step of the two output variables.

[0101] S3, Model Training and Optimization

[0102] To drive the model learning, an explicit optimization objective is defined through a loss function. The mean squared error is chosen as the loss function for continuous value prediction problems, which is conducive to the model's rapid correction of deviations:

[0103]

[0104] where MSE is the mean squared error; N is the number of samples; Y i is the actual value; is the model's predicted output value.

[0105] After selecting the loss function, an efficient algorithm is needed to automatically and iteratively update all learnable parameters in the network, including all weights and biases of GRU layers, attention layers, and output layers, according to the gradient calculated by the loss function. The Adam optimizer widely used in the current deep learning field is adopted in this embodiment. The Adam algorithm combines first-order moment estimation and second-order moment estimation, which can calculate an adaptive learning rate for each parameter, thereby achieving fast and stable convergence. The preprocessed data set is divided into a training set, a validation set, and a test set according to a certain proportion. The model is trained on the training set for multiple rounds of iteration. After each round of training, the model performance is evaluated on the validation set. The error of the validation set can be used as a basis for judging whether the model has overfitting. Once the validation error stops declining or starts to rise, the training should be stopped in time, and the model parameters at this time or before the validation performance is optimal are saved.

[0106] S4, model verification and application

[0107] After training, the final selected model is comprehensively evaluated using test set data that has not participated in the training process at all. By calculating the mean square error performance index between the predicted value and the true value, the generalization ability and prediction accuracy of the model on new data are objectively evaluated. Once the performance of the model is verified and meets the engineering precision requirements, the trained model can be solidified and deployed to practical applications. Figure 2 and Figure 3 The GRU model with fusion self-attention mechanism provided by the present application provides a unit load and main steam pressure identification result graph. As shown in the figure, as a high-fidelity coal-fired unit coordination system dynamic model, the new model can receive real-time or simulated control input sequences and quickly predict the dynamic response of unit power and main steam pressure, which is significantly better than traditional identification methods. This provides an accurate object model for more advanced predictive control, reinforcement learning control, etc., and can be used for offline simulation optimization of control schemes, evolution deduction of complex working conditions, and high-simulation operation training for operating personnel.

Claims

1. A coordinated modeling method for gated cyclic unit coal-fired power plants that integrates a self-attention mechanism, characterized in that, Includes the following steps: Step 1: Collect operating data of the coal-fired power unit and preprocess the data; Step 2: Using the preprocessed data as input and the predicted values ​​of coal-fired unit power and main steam pressure as output, construct a GRU network model that incorporates a self-attention mechanism. Step 3: Train and optimize the GRU network model that incorporates the self-attention mechanism; Step 4: Based on the trained GRU network model with fusion self-attention mechanism, predict the power and main steam pressure of coal-fired units.

2. The method for coordinated modeling of gated cyclic unit coal-fired power plants incorporating a self-attention mechanism as described in claim 1, characterized in that, Step 1 specifically involves: collecting operating data of the coal-fired power unit, including steady-state operation, load increase, load decrease, and primary frequency regulation disturbances; preprocessing uses the maximum-minimum normalization method to linearly map all data to the [0,1] interval, eliminating the influence of dimensions; the specific mathematical transformation formula is as follows: In the formula: x represents the original data point; x min and x max These represent the minimum and maximum values ​​of the variable recorded in the entire dataset, respectively. norm This is the normalized data obtained after calculation.

3. The method for coordinated modeling of gated cyclic unit coal-fired power plants incorporating a self-attention mechanism as described in claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Deploy the Gated Loop Unit (GRU) layer, and generate the final hidden state of the current time step by resetting and updating the gates; Step 2.2: Introduce a self-attention mechanism layer on the gated recurrent unit (GRU) layer to output the context vector; Step 2.3: Map the context vector to the desired final output dimension through a fully connected layer, which is the specific physical prediction value of the unit power and main steam pressure at the next moment.

4. The method for coordinated modeling of gated cyclic unit coal-fired power plants incorporating a self-attention mechanism according to claim 3, characterized in that, Step 2.1 specifically involves: For each time step t in the input sequence, the GRU unit updates its hidden state ht by the following calculation: First, the reset gate calculates the hidden state ht from the previous time step. t-1 The system takes the current input U(t) as input and outputs a vector between 0 and 1, which will determine the historical information h. t-1 How much is used to calculate the current candidate state; Reset gate: r(t)=σ(W) r u(t)+U r h(t-1)+b r (2) In the formula: r(t) is the reset gate; σ is the Sigmoid function; W r It resets the gate learning weights; U r It resets the gate input weights; b r It resets the gate bias; u is the input variable; h is the candidate state; The update gate determines the extent to which "new" information is used to update "old" information, and this is determined by calculating h. t-1 And U(t), and output a vector between 0 and 1: Update gate: z(t)=σ(W) z u(t)+U z h(t-1)+b z (3) In the formula: z(t) is the update gate; σ is the Sigmoid function; W z It updates the gate learning weights; U z It updates the gate input weights; b z This updates the gate bias; u is the input variable; h is a candidate state; The model calculates a "candidate" hidden state based on the current input U(t) and historical information filtered by the reset gate: In the formula: It is a candidate hidden state; tanh is the hyperbolic tangent function; ⊙ represents element-wise multiplication; W h It is the candidate hidden state learning weight; U h b is the input weight of the candidate hidden state; h It represents the candidate hidden state bias; u is the input variable; h is the candidate state; Finally, the model uses the update gate z(t) to generate the final hidden state h at the current time step. t : In the formula: h(t) is the hidden state at the current time; ⊙ represents element-wise multiplication; After the GRU layer processes the entire input sequence of length T, it generates a hidden state sequence H = [h1, h2, ..., h...]. T The sequence encodes the dynamic evolution information of the coal-fired unit over the past T time points.

5. The method for coordinated modeling of gated cyclic unit coal-fired power plants incorporating a self-attention mechanism according to claim 3, characterized in that, Step 2.2 specifically involves mapping the hidden state sequence H to the query space, key space, and value space, respectively. Query matrix: Q = HW Q (6) Bond matrix: K = HW K (7) Value matrix: V = HW V (8) Among them, W Q W represents the query weight matrix. K W represents the key weight matrix. V Represents the value weight matrix; A fractional matrix is ​​obtained by calculating the dot product of the transposes of the Q matrix and the K matrix, where each element represents the degree of attention one position in the sequence gives to another position, and this matrix is ​​then scaled. In the formula: d k It is the dimension of the key vector; By applying the Softmax function to each row of the score matrix, it is transformed into a set of non-negative weight values ​​that sum to 1; the non-negative weight values ​​are called "attention weights", which represent the level of attention given to all other positions in the input sequence when generating the output at the current position; W A =softmax(Scores) (10) In the formula: W A It is attention weight; Finally, the calculated attention weight matrix is ​​multiplied by the value matrix V; the result is a context vector C. C=W A ·V (11)。 6. The method for coordinated modeling of gated cyclic unit coal-fired power plants incorporating a self-attention mechanism according to claim 5, characterized in that, Step 2.3 specifically involves the following: The context vector C output by the self-attention layer is a high-dimensional representation of the entire input time series that is complete in information and highlights the key points. Through a standard feedforward neural network layer, i.e., a fully connected layer, the context vector is mapped to the desired final output dimension. The linear transformation converts the high-dimensional context features into specific physical predictions of the unit power and main steam pressure at the next moment. In the formula: W out and b out These are the learnable weight matrix and bias vector of the output layer. This is the model's final prediction of the two output variables for the next time step.

7. The method for coordinated modeling of gated cyclic unit coal-fired power plants incorporating a self-attention mechanism according to claim 1, characterized in that, Step 3 specifically involves defining a clear optimization objective through a loss function, selecting mean squared error as the loss function for continuous value prediction problems, and enabling the model to quickly correct biases. Where: MSE is the mean squared error; N is the sample size; Y i This is the actual value; Predict the output value for the model; The Adam optimizer was used to optimize the model.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

Images