Coordination control method based on multi-state sensing of thermal power generating unit
Through the deep learning model, the feedforward signal data of thermal power units is predicted, and the main control feedforward volume of boiler is optimized, which solves the problem of energy imbalance in the process of rapid load change in traditional control systems, and improves the control performance and safety and speed of unit operations.
Patent Information
- Application Number
- CN202510345614.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-24
- Publication Date
- 2025-06-27
AI Technical Summary
The traditional thermal power unit coordination control system has energy imbalance problem during the rapid load change process, resulting in pressure deviation being transmitted to the PID controller, resulting in a degradation of control performance.
Using a multi-state perception coordination control method based on deep learning, the feedforward signal data of the thermal power unit is collected and normalized, and the trained deep learning model is input for prediction, and the main feedforward volume of the boiler is optimized.
It reduces the feedforward quantity deviation, enhances the robustness of the control system, and improves the safety and speed of unit operation.
Smart Images
Figure CN120215246A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of thermal power unit control. Specifically, it relates to a coordinated control method based on multi-state perception of thermal power units. Background Art
[0002] Under the background of the global energy structure transformation and increasingly strict environmental protection policies, the thermal power generation industry is facing severe challenges. On the one hand, the large-scale grid connection of renewable energy sources makes thermal power units need to undertake more frequent peak shaving tasks, and the operating conditions show significant complexity and variability; on the other hand, increasingly strict environmental protection regulations put forward higher requirements for the operating efficiency and emission control of thermal power units. The traditional coordinated control system of thermal power units mainly adopts the method of "PID + feedforward". During the rapid load change process of the unit, due to the characteristics of large inertia and large delay of the boiler, there is a large energy imbalance problem between the boiler and the steam turbine at the initial stage of load change. The energy imbalance is fed back to the boiler main control through the deviation of the pressure parameter. In order to solve the flexibility problem of thermal power units, many rapid load change feedforward signals are added to the coordinated control system. Although these feedforward signals effectively improve the rapidity of the unit, they bring new problems. Due to the inertia of the unit, the feedforward signal amount cannot pull back the main steam pressure in time. During the inertia period, the pressure deviation is transmitted to the PID controller, causing the PID controller to have incorrect actions and incorrect deviation accumulation, resulting in severe parameter fluctuations during the load change process of the unit.
[0003] The current optimization of the boiler main control feedforward is mainly achieved through mechanism, and its main implementation method is to construct through signals such as load command differentiation and pressure deviation differentiation.
[0004] The defect of the existing technology is that the operating state of thermal power units is restricted by many factors, and the traditional feedforward signals cannot fully represent the changes in the operating state of the units. The deviation is mainly compensated by the PID controller. During the rapid load change process, due to the large energy balance gap and the increase in the deviation of the feedforward signal, and the limited compensation ability of the PID controller, the control performance decreases. Summary of the Invention
[0005] Aiming at the problem of the accuracy deviation of the feedforward quantity in the PID + feedforward control method of the existing coordinated control method for thermal power units, the present invention provides a coordinated control method based on multi-state perception of thermal power units.
[0006] To achieve the above technical objectives, the technical solution adopted by the present invention is as follows:
[0007] A coordinated control method based on multi-state perception of thermal power units, comprising the steps of:
[0008] Collect the operating feedforward signal data of the thermal power unit;
[0009] Perform data normalization processing on the feedforward signal data of the thermal power unit before operation;
[0010] Input the data after normalization processing into the trained deep learning model for prediction to obtain the boiler master control instruction, dynamic feedforward, and static feedforward.
[0011] Furthermore, the feedforward signal data of the thermal power unit before operation includes the unit load instruction, actual unit load, main steam pressure set value, main steam pressure deviation, intermediate point temperature, integrated valve position opening instruction, and primary frequency modulation signal.
[0012] Furthermore, the calculation formula for data normalization processing:
[0013]
[0014] where is the data after normalization processing, data i is the data before normalization processing, data min and data max are respectively the minimum and maximum values in the data.
[0015] Furthermore, data min and data max should be set the same as the upper and lower limit thresholds of the parameters during the operation of the thermal power unit.
[0016] Furthermore, the deep learning model includes a data convolution processing module, a data cropping module, a fully connected layer mapping module, a GRU module, and a model prediction output module;
[0017] The data convolution processing module is three parallel convolutional neural network models, which are respectively used for extracting proportional, differential, and integral information;
[0018] The data cropping module crops the extracted data to form features with the same size and corresponding time information;
[0019] The fully connected layer mapping module is a 3-layer fully connected layer, which is used for non-linear mapping of the features corresponding to time information;
[0020] The GRU module is used for integrating the information on the time series of non-linear mapping of features;
[0021] The model prediction output module obtains the model prediction value output result through multiple iterative updates.
[0022] Further, the input data of the data convolution processing module performs data processing for each feature in the time dimension. The parallel convolutional neural networks respectively use 1, 2, and 8 convolutional kernels for convolution operations, corresponding to the ratio, differential, and integral information in the data extraction process; their calculation formulas are respectively:
[0023] X1 = cnn1(x)
[0024] X2 = cnn2(x)
[0025] X8 = cnn8(x)
[0026] Among them, b represents the batch size, T represents the time step of the data, m is the number of features of the input data of the model, m = 7; cnn1 represents a convolutional neural network with a convolutional kernel length of 1; cnn2 represents a convolutional neural network with a convolutional kernel length of 2; cnn8 represents a convolutional neural network with a convolutional kernel length of 8.
[0027] Further, the data cropping module forms features with consistent sizes and corresponding time information by performing data cropping operations on X1, X2, and X8. Among them, the data cropping part only retains the data of the T - 7 moments closest to the current moment, and only performs cropping processing on X1 and X2 in the data cropping part; the calculation formula:
[0028] X1' = Crop(X1)
[0029] X2' = Crop(X2)
[0030] Among them, Crop(x) represents the data cropping operation, and the cropped data is
[0031] Further, the neuron parameters corresponding to the 3 - layer fully - connected layer are {32, 64, 32}, and the calculation formula for each layer:
[0032] y = sigmoid(wx + b)
[0033]
[0034] Among them, w and b are respectively the weight and bias of the fully - connected layer, and the data after passing through the 3 - layer fully - connected layer is
[0035] Further, the calculation formula of the GRU module:
[0036] Z t = sigmoid(W Z h t-1 + U Z xt )
[0037] r t = sigmoid(W r h t-1 + U r x t )
[0038]
[0039] where W Z 、W r 、W, U Z 、U r 、U are weights, which are training parameters of the model. ⊙ is the Hadamard product, indicating element-wise multiplication of matrices. The neuron parameter of the GRU model is 4, and the output of the GRU model
[0040] Furthermore, in the model training stage of the model prediction output module, the overall prediction output of the model is the vector x T-7 at the last moment of the time dimension in X. The calculation formula is:
[0041] y = sigmoid(x 1+p W + B)
[0042] where the weight W has a size of 4×1 and the bias B has a size of 1×1;
[0043] In the model prediction stage, the target value at time t is additionally input into the model Therefore, the final output is determined by x T-8 and x T-7 The calculation formula is:
[0044] y 1 = sigmoid(x T-8 W + B)
[0045] The weights W and B are updated through the following formula, and the number of updates is 4 times;
[0046]
[0047] y 1 = sigmoid(x T-8 W + B)
[0048] After the update, calculate the final output of the model;
[0049] y = sigmoid(x T-7 W + B)
[0050] The predicted value Pre is obtained through the following formula, where Pre is the predicted value, data max , data min is the same as that in the previous text:
[0051] Pre = y(data max - data min ) + data min .
[0052] Compared with the prior art, the present invention has the following beneficial effects:
[0053] By considering various variable factors, the optimization of the feedforward quantity of the boiler main control in the coordinated control system is realized, the feedforward quantity deviation is reduced, the robustness of the control system is enhanced, and the safety and rapidity of the unit operation are improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 is the overall flowchart of a coordinated control method based on multi-state perception of a thermal power unit in an embodiment of the present invention;
[0055] Figure 2 is the prediction structure block diagram of the deep learning model in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0056] For the convenience of those skilled in the art to understand, the present invention will be further described below in conjunction with the embodiments and the drawings. The content mentioned in the embodiments does not limit the present invention.
[0057] As Figure 1 shown, this embodiment provides a coordinated control method based on multi-state perception of a thermal power unit, including the steps of:
[0058] S1. Collect the feedforward signal data of the thermal power unit during operation;
[0059] S2. Perform data normalization processing on the feedforward signal data of the thermal power unit during operation;
[0060] S3. Input the data after normalization processing into the trained deep learning model for prediction to obtain the boiler main control command, dynamic feedforward, and static feedforward.
[0061] The feedforward signal data of the thermal power unit during operation includes the unit load command, the actual unit load, the main steam pressure set value, the main steam pressure deviation, the intermediate point temperature, the integrated valve position opening command, and the primary frequency modulation signal.
[0062] The calculation formula for data normalization processing:
[0063]
[0064] Where is the data after normalization, data i is the data before normalization, data min and data max are respectively the minimum and maximum values in the data.
[0065] data min and data max should be set the same as the upper and lower threshold values of the parameters during the operation of the thermal power unit.
[0066] The deep learning model includes a data convolution processing module, a data cropping module, a fully connected layer mapping module, a GRU module, and a model prediction output module;
[0067] As Figure 2 shown, the data convolution processing module is three parallel convolutional neural network models, which are respectively used for the extraction of proportional, differential, and integral information;
[0068] The data cropping module crops the extracted data to form features with the same size and corresponding time information;
[0069] The fully connected layer mapping module is a 3-layer fully connected layer, which is used for the non-linear mapping of the features corresponding to the time information;
[0070] The GRU module is used for the information integration on the time series of the non-linear mapping of the features;
[0071] The model prediction output module obtains the model prediction value output result through multiple iterative updates.
[0072] The input data of the data convolution processing module performs data processing for each feature in the time dimension. The parallel convolutional neural networks respectively use 1, 2, and 8 convolutional kernels for convolution operations, which respectively correspond to the proportional, differential, and integral information in the data extraction process; their calculation formulas are respectively:
[0073] X1 = cnn1(x)
[0074] X2 = cnn2(x)
[0075] X8 = cnn8(x)
[0076] Among them, b represents the batch size, which is mainly used for the model to perform batch processing to improve the operation efficiency; T represents the time step of the data, that is, the time sequence size of the data; m is the number of features of the input data of the model, m = 7; cnn1 represents the convolutional neural network with a convolutional kernel length of 1; cnn2 represents the convolutional neural network with a convolutional kernel length of 2; cnn8 represents the convolutional neural network with a convolutional kernel length of 8.
[0077] The data clipping module forms features with consistent sizes and corresponding time information from X1, X2, and X8 through data clipping operations. Among them, the data clipping part only retains the data of the nearest T - 7 moments to the current moment, and only X1 and X2 are clipped in the data clipping part; calculation formula:
[0078] X1′ = Crop(X1)
[0079] X2′ = Crop(X2)
[0080] Among them, Crop(x) represents the data clipping operation, and the clipped data is
[0081] The neuron parameters corresponding to the 3 - layer fully - connected layer are {32, 64, 32}, and the calculation formula for each layer is:
[0082] y = sigmoid(wx + b)
[0083]
[0084] Among them, w and b are the weights and biases of the fully - connected layer respectively. After passing through the 3 - layer fully - connected layer, the data is
[0085] The calculation formula of the GRU module is:
[0086] Z t = sigmoid(W Z h t-1 + U Z x t )
[0087] r t = sigmoid(W r h t-1 + U r x t )
[0088]
[0089] Among them, W Z 、W r 、W、U Z 、U r 、U are weights, which are the training parameters of the model. ⊙ is the Hadamard product, indicating the element - by - element multiplication of matrices. The neuron parameter of the GRU model is 4, and the output of the GRU model
[0090] During the model training phase, the overall predicted output of the model is the vector x at the last moment in the time dimension of X T-7 , and the calculation formula is:
[0091] y = sigmoid(x 1+p W + B)
[0092] where the weight W has a size of 4×1 and the bias B has a size of 1×1;
[0093] During the model prediction phase, the target value at time t is additionally input into the model Therefore, the final output is determined by x T-8 and x T-7 , and the calculation formula is:
[0094] y 1 = sigmoid(x T-8 W + B)
[0095] The weights W and B are updated through the following formula, and the number of updates is 4 times;
[0096]
[0097]
[0098] y 1 = sigmoid(x T-8 W + B)
[0099] After the update, calculate the final output of the model;
[0100] y = sigmoid(x T-7 W + B)
[0101] The predicted value Pre is obtained through the following formula. Here, Pre is the predicted value, and data max , data min are the same as those in the previous text:
[0102] Pre = y(data max - data min ) + data min .
[0103] Compared with the prior art, the present invention has the following beneficial effects:
[0104] By considering multiple variable factors, the optimization of the boiler master feedforward quantity in the coordinated control system is achieved, the feedforward quantity deviation is reduced, the robustness of the control system is enhanced, and the safety and rapidity of the unit operation are improved.
[0105] In the existing solution, the feedforward signal adopted is obtained by adding the processed results of multiple single variables. It cannot meet the operation state characterization of thermal power units under extreme operation conditions, thereby reducing the control performance.
[0106] The above has introduced in detail a coordinated control method based on multi-state perception of thermal power units provided by the present application. The description of specific embodiments is only used to help understand the method and its core idea of the present application. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.
Claims
1. A coordinated control method based on multi-state perception of thermal power units, characterized in that: Includes steps: Collect feedforward signal data of thermal power unit operation; Perform data normalization on the feedforward signal data of the thermal power unit operation; The normalized data is input into the trained deep learning model for prediction to obtain the boiler master control instructions, dynamic feedforward and static feedforward.
2. A coordinated control method based on multi-state perception of thermal power units according to claim 1, characterized in that: The feedforward signal data of the operation of the thermal power unit includes the unit load command, the actual load of the unit, the main steam pressure set value, the main steam pressure deviation, the intermediate point temperature, the comprehensive valve position opening command and the primary frequency modulation signal.
3. A coordinated control method based on multi-state perception of thermal power units according to claim 2, characterized in that: The calculation formula for data normalization is: in is the data after normalization. i is the data before normalization, data min With data max are the minimum and maximum values in the data respectively; data min With data max The upper and lower limit threshold settings of the parameters during the operation of the thermal power unit should be the same.
4. A coordinated control method based on multi-state perception of thermal power units according to claim 3, characterized in that: The deep learning model includes a data convolution processing module, a data clipping module, a fully connected layer mapping module, a GRU module, and a model prediction output module; The data convolution processing module consists of three parallel convolutional neural network models, which are used to extract proportional, differential and integral information respectively; The data clipping module clips the extracted data into features with consistent size and corresponding time information; The fully connected layer mapping module consists of three fully connected layers, which are used for nonlinear mapping of features corresponding to time information; GRU module, used to integrate information on the time series of feature nonlinear mapping; The model prediction output module obtains the model prediction value output result through multiple iterative updates.
5. A coordinated control method based on multi-state perception of thermal power units according to claim 4, characterized in that: The input data of the data convolution processing module is processed for each feature in the time dimension. The parallel convolutional neural network uses 1, 2, and 8 convolution kernels for convolution operations, which correspond to the proportion, differential, and integral information in the data extraction process respectively; The calculation formulas are: X1=cnn1(x) X2=cnn2(x) X8=cnn8(x) in, b represents the batch size, T represents the time step of the data, m represents the number of features of the input data of the model, m=7; cnn1 represents a convolutional neural network with a convolution kernel length of 1; cnn2 represents a convolutional neural network with a convolution kernel length of 2; cnn8 represents a convolutional neural network with a convolution kernel length of 8.
6. A coordinated control method based on multi-state perception of thermal power units according to claim 5, characterized in that: The data clipping module forms features of the same size and corresponding time information from X1, X2, and X8 through data clipping operations. The data clipping part only retains the data of the T-7 moments closest to the current moment, and only X1 and X2 are clipped in the data clipping part; calculation formula: X1'=Crop(X1) X2'=Crop(X2) in, Crop(x) represents the data cropping operation. The cropped data is 7. A coordinated control method based on multi-state perception of thermal power units according to claim 6, characterized in that: The neuron parameters corresponding to the 3-layer fully connected layer are {32, 64, 32}, and the calculation formula for each layer is: y=sigmoid(wx+b) Among them, w and b are the weight and bias of the fully connected layer respectively. The data after three layers of fully connected layers is 8. A coordinated control method based on multi-state perception of thermal power units according to claim 7, characterized in that: The calculation formula of the GRU module is: Z t =sigmoid(W Z h t-1 +U Z x t ) r t =sigmoid(W r h t-1 +U r x t ) Among them, W Z , W r , W, U Z , U r , U is the weight, is the training parameter of the model, ⊙ is the Hadamard product, which means the corresponding elements of the matrix are multiplied, the neuron parameter of the GRU model is 4, and the output of the GRU model 9. A coordinated control method based on multi-state perception of thermal power units according to claim 8, characterized in that: Model prediction output module During the model training phase, the overall prediction output of the model is composed of the vector x at the last moment of the time dimension in X T-7 , calculation formula: y=sigmoid(x 1+p W+B) Among them, the weight W size is 4×1, and the bias B size is 1×1; In the model prediction stage, the target value at time t is additionally input into the model. Therefore, the final output is x T-8 and x T-7 Determine, calculation formula: y 1 =sigmoid(x T-8 W+B) The weights W and B are updated 4 times using the following formula; y 1 =sigmoid(x T-8 W+B) After the update, calculate the final output of the model: y=sigmoid(x T-7 W+B) The predicted value Pre is converted by the following formula, where Pre is the predicted value, data max 、data min Consistent with the previous article: Pre=y(data max -data min )+data min 。