Multi-step prediction method for superheated steam temperature of thermal power unit based on future load instruction

Abstract

The present application proposes a multi-step prediction method for the superheated steam temperature of a thermal power unit based on future load instructions. First, the parameters affecting the temperature of the superheater are screened for features, and the multiple parameters with the largest correlation with the superheated steam temperature are selected as the final input features. The future power instruction is combined and input into the CNN-BiLSTM-Attention network constructed. Then, the time sliding window is used for sliding, realizing the prediction of the future multi-step superheated steam temperature by using the historical data combined with the future power instruction. The prediction method of the present application combines the historical data and the future power instruction, effectively improving the problems of prediction jitter and lag.

Description

Technical Field

[0001] This invention belongs to the field of energy information technology, specifically relating to a multi-step prediction method for superheated steam temperature of thermal power units based on future load commands. Background Technology

[0002] Industrial boilers are crucial energy conversion devices in industries such as power generation, chemical engineering, and metallurgy. Their safe and stable operation is vital to the entire production process. However, issues such as furnace slagging and pipe ash accumulation can cause rapid changes in superheated steam parameters, potentially leading to serious accidents. Superheated steam temperature is one of the key parameters reflecting the boiler's operating status. Accurately predicting superheated steam temperature changes is of great significance for improving boiler combustion optimization control and preventing tube rupture accidents. However, steam temperature is influenced by multiple coupled factors, exhibiting significant nonlinear and dynamic characteristics, posing a challenge to accurate prediction.

[0003] In the safe operation of industrial boilers, superheater temperature control is particularly critical. Excessively high superheater tube wall temperatures can lead to tube cracking, while excessively low temperatures result in low thermal efficiency. Therefore, accurately predicting changes in the superheater outlet steam temperature is essential to ensuring that the wall temperature is controlled within a reasonable range. However, the steam superheating system has a complex structure, and its temperature is affected by many factors. For example, steam will generate varying degrees of temperature increase as it passes through different tube sections, and changes in steam flow rate will cause dynamic temperature fluctuations. This makes accurately predicting the superheated steam temperature extremely difficult. Current methods mainly employ traditional mathematical modeling. However, these methods rely on precise physical models and fail to fully consider the coupled dynamic changes of various influencing factors.

[0004] Predicting superheated steam temperature is crucial for the safe and stable operation of industrial boilers. Currently, the main prediction methods include:

[0005] Physical model-based methods: These methods establish physical relationship models between superheated steam parameters to predict steam temperature. However, these methods rely on accurate physical models, but the parameters in real-world scenarios have significant uncertainties, limiting prediction accuracy. Statistical model prediction methods: These methods use statistical models such as linear regression and ARIMA for temperature prediction. These methods do not consider the physical relationships between variables, and prediction accuracy decreases when the system dynamically changes. Machine learning methods: These methods use RNNs and LSTMs for time-series temperature prediction. These methods have the ability to learn and model historical data, but they are not accurate enough for predicting sudden and abnormal situations. Multi-sensor data fusion methods: These methods comprehensively utilize heterogeneous sensors from multiple sources, such as temperature, pressure, and flow, for integrated modeling and prediction. However, they face difficulties in data synchronization and complex model design. Multi-step prediction methods: These methods consider modeling and predicting temperature changes over a certain future time period. Existing methods have low accuracy in long-term trend prediction.

[0006] Therefore, developing advanced deep learning algorithms to model the complex spatiotemporal variations of superheated steam temperature and achieve accurate prediction of superheated steam is crucial. This not only ensures proper temperature control of the superheater but also provides a foundation for the prediction and optimized control of other parameters. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to provide a multi-step prediction method for superheated steam temperature of thermal power units based on future load commands, which integrates an attention mechanism, a convolutional neural network, and a bidirectional LSTM. This method can achieve better multi-step temperature prediction results, improve prediction accuracy, and enhance adaptability to dynamic anomalies.

[0008] Technical solution: The multi-step prediction method for superheated steam temperature of thermal power units based on future load commands of the present invention includes the following steps:

[0009] S1. Collect parameters of the superheater changes caused by the unit's variable operating conditions, including header outlet temperature, A and B side outlet temperatures, steam flow rate, primary and secondary air volume, and high pressure heater inlet pressure, all of which will cause changes in superheated steam temperature.

[0010] S2. Perform feature filtering on the collected parameters, including signal denoising. Filter the parameters according to the degree of relationship between them, retain highly correlated features as one input variable, and use the future power command as another input variable to form the input sample data, and divide the data into training set and test set.

[0011] S3. The training set is used to train the CNN-BiLSTM-Attention model for predicting superheated steam temperature. This includes: first, normalizing the data in the training set; then, extracting the features between the normalized data using the CNN module; importing the features into the BiLSTM network and fitting and predicting them through the fully connected layer; and finally, multiplying the output with the weights obtained by the Attention mechanism to obtain the final output prediction result; and then using a time sliding window to slide the normalized training set data to iterate and train the CNN-BiLSTM-Attention model to obtain the trained CNN-BiLSTM-Attention model.

[0012] S4. Using the trained CNN-BiLSTM-Attention model, a time sliding window is used to perform sliding prediction on the test set to achieve continuous multi-step prediction of superheated steam temperature.

[0013] Furthermore, step S2 includes the following steps:

[0014] S21. Perform wavelet transform on the original collected parameter data. After wavelet transform, determine the wavelet coefficient threshold. Wavelet coefficients greater than the threshold are generated by signals and should be retained. Wavelet coefficients less than the threshold are generated by noise and should be set to zero.

[0015] S22. Use the correlation coefficient method to filter the denoised data. The correlation coefficient ranges from -1 to 1, where 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no linear relationship. Filter out independent variables as the final input feature variables.

[0016] Furthermore, in step S3, the normalization method adopts Min-max normalization:

[0017]

[0018] Where min(x) and max(x) are the minimum and maximum values ​​of the input data in the training set, respectively, x is the value of the feature parameter in the training set, and x′ is the normalized result of the feature parameter in the training set.

[0019] Furthermore, the specific training process of the CNN-BiLSTM-Attention model used to predict superheated steam temperature includes the following steps:

[0020] S31. Use a CNN network to extract features from the normalized training set: aggregate each feature output from the CNN network's convolutional layers into a feature vector, ensuring that the feature output dimension is the same as the number of input feature maps; specifically, the feature vector aggregation is as follows:

[0021]

[0022] Among them, c j This represents the feature vector aggregated by feature j. N represents the i-th feature vector of feature j. j Indicates the length of the feature map;

[0023] The specific convolutional layer operations are as follows:

[0024]

[0025] in, It is the j′th weight of the i-th convolutional kernel in the l-th layer. It is the j-th convolutional local region in the l-th layer, where W is the width of the convolution kernel, and y l(i.j) It is the convolution result of the l-th layer based on i convolution kernels and j′ weights. x is the weight of the i-th convolutional kernel in the l-th layer. l(j+j′)It is the (j+j′)th convolutional local region in the l-th layer;

[0026] S32. Prediction using a BiLSTM network: BiLSTM memorizes and filters the features extracted by the CNN, fully considering the feature relationships between the forward and backward data, and performs fitting prediction. Finally, the output is fed to a fully connected layer for weighted output to calculate the predicted data. The BiLSTM's memory state update is as follows:

[0027]

[0028] The output of the memory unit is as follows:

[0029] O t =σ(W o x t +U o h t-1 +V o c t +b o )

[0030] Among them, c t c represents the long-term memory state of a memory unit at time t. t-1 f represents the long-term memory state of a memory unit at time t-1. t i represents the forget gate output at time t of the memory unit. t The input gate output represents the memory cell at time t. W represents the product of corresponding elements. o U o V o W represents the weight matrix of the input gate, forget gate, and output gate of the current memory unit, respectively. c U c x represents the weight matrices of the input gate and forget gate stored in the memory unit, respectively. t Represents input, h t-1 b represents the hidden layer state at time t-1. c b represents the bias vector during memory unit updates. o The bias vector representing the output of the memory unit, O t σ represents the output of the memory unit, and σ represents the activation function.

[0031] S33. Add an Attention mechanism to the output of the BiLSTM to weight the prediction results: based on the hidden layer states h of all encoders in the BiLSTM network. t and the current decoder hidden layer state h s The scoring function is obtained as follows:

[0032]

[0033] Among them, W a Let h be the weight matrix, and score(h) t h s ) represents the result of the scoring function. This is the transpose of the hidden state of all encoders. Given the current forward and backward states of the decoder, the corresponding attention weights are then calculated based on the score function results. Finally, the directly usable weights are calculated based on the context feature vectors, as shown in the following formula:

[0034]

[0035] Where, α ij For the weighted result, h i h n Let be the encoder hidden layer states at time i and time n, respectively. Let be the hidden layer states of the decoder in the forward and backward directions at time j. Let m be the forward and backward hidden layer states of the decoder at time m. Finally, the weighted result is used as the final prediction output, which is then inversely normalized and compared with the true value.

[0036] S34, determine the evaluation indicators for the final prediction results.

[0037] Furthermore, the evaluation index determined in step S34 is the root mean square error (RMSE), and the formula is as follows:

[0038]

[0039] Where n is the number of predicted data points, y i For the i-th predicted value, Let be the true value of the i-th data point, where i is the predicted data label.

[0040] Furthermore, the CNN-BiLSTM-Attention network model structure includes a convolutional layer, followed by a dropout layer, a BiLSTM layer, an attention layer, a flattening layer, and two fully connected layers.

[0041] Furthermore, the sliding window has a sliding step of one time step. When the time step of the input future power command is less than the time step of the feature parameters after feature filtering, the missing part of the input future power command is filled with the value of the power command input in the last step.

[0042] Based on the same inventive concept, the multi-step prediction system for superheated steam temperature of thermal power units based on future load commands of the present invention includes:

[0043] The data acquisition unit is used to collect parameters that change due to the unit's variable operating conditions, including header outlet temperature, A and B side outlet temperatures, steam flow rate, primary and secondary air volume, and high pressure heater inlet pressure, all of which will cause changes in superheated steam temperature.

[0044] The data processing unit is used to perform feature filtering on the collected parameters, including signal denoising, filtering based on the degree of correlation between features, retaining highly correlated features as an input variable, and using future power commands as another input variable to form the input sample data, and dividing the data into training and test sets.

[0045] The model training unit is used to train the CNN-BiLSTM-Attention model for predicting superheated steam temperature using the training set. The process includes: first, normalizing the data in the training set; then, extracting the features between the normalized data using the CNN module; importing the features into the BiLSTM network and fitting and predicting them through the fully connected layer; and finally, multiplying the output with the weights obtained by the Attention mechanism to obtain the final output prediction result. The normalized training set data is then slid using a time sliding window to iterate and train the CNN-BiLSTM-Attention model cyclically to obtain the trained CNN-BiLSTM-Attention model.

[0046] The multi-step prediction unit is used to perform sliding predictions on the test set using a time sliding window by employing a trained CNN-BiLSTM-Attention model, thereby achieving continuous multi-step predictions of superheated steam temperature.

[0047] Based on the same inventive concept, the present invention provides an electronic device, the device comprising:

[0048] Memory containing executable program code;

[0049] A processor coupled to the memory;

[0050] The processor calls the executable program code stored in the memory to execute the steps of the multi-step prediction method for superheated steam temperature of thermal power units based on future load instructions, as described above.

[0051] Based on the same inventive concept, the present invention provides a computer-readable storage medium storing computer instructions, which, when invoked, are used to execute the steps of the multi-step prediction method for superheated steam temperature of thermal power units based on future load instructions, as described above.

[0052] Beneficial effects: Compared with the prior art, the advantages of the present invention are:

[0053] (1) The prediction accuracy is significantly improved. With sufficient data, the loss can reach 0.047, indicating that the prediction results are significantly more accurate and reliable.

[0054] (2) It can predict rapid temperature changes. With the addition of future power commands, compared with the case without future commands, the prediction lag is significantly improved.

[0055] (3) It extends the long-term prediction of steam temperature. Using this invention, it is possible to predict the data for the next 10 steps using historical data from 30 steps, far exceeding the 5-minute prediction limit of traditional methods.

[0056] (4) Improved model adaptability to complex working conditions. This invention relies solely on sensor data and is not constrained by an accurate physical model, thus accurately adapting to complex working conditions such as different loads and different media.

[0057] (5) Reduce the risk of superheater accidents. The improved temperature prediction effect of this invention can effectively prevent steam pipe rupture accidents caused by large temperature changes.

[0058] (6) Optimize boiler operation, reduce energy consumption and extend service life. Based on accurate temperature forecasts, optimized scheduling schemes can be implemented for the safe and economical operation of the unit, reducing boiler operating energy consumption and extending boiler service life. Attached Figure Description

[0059] Figure 1 This is a flowchart of the multi-step prediction method for superheated steam temperature according to an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of the LSTM principle according to an embodiment of the present invention;

[0061] Figure 3 This is a diagram of the CNN-BiLSTM-Attention network structure according to an embodiment of the present invention;

[0062] Figure 4 This is a schematic diagram of the time sliding window principle in an embodiment of the present invention;

[0063] Figure 5 This is a multi-step prediction diagram of superheated steam temperature based on historical data and future power commands, according to an embodiment of the present invention. Detailed Implementation

[0064] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0065] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0066] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0067] This embodiment provides the main process of a multi-step prediction method for superheated steam temperature of thermal power units based on future load commands, as follows: Figure 1 As shown.

[0068] Step 1: Collect parameters related to changes in the superheater caused by the unit's variable operating conditions. These parameters include header outlet temperature, superheater A and B side outlet temperatures, steam flow rate, primary and secondary air volumes, and high-pressure heater inlet pressure, all of which contribute to changes in superheated steam temperature. Collecting each parameter affecting superheated steam temperature is beneficial for identifying factors with significant impact on temperature changes during subsequent feature screening.

[0069] Step 2: Feature filtering is performed on the collected parameters. This includes signal denoising and correlation coefficient filtering methods. Features are filtered based on their strong relationships, retaining highly correlated features as input variables. The specific process for feature filtering of the parameter data includes the following steps:

[0070] S21. Perform wavelet transform on the original collected parameter data. After wavelet transform, select an appropriate threshold. Wavelet coefficients greater than the threshold are considered to have generated a signal and should be retained. Wavelet coefficients less than the threshold are considered to have generated noise and should be set to zero to achieve the purpose of noise reduction.

[0071] S22. Use the correlation coefficient method to screen the parameter features after noise reduction. The correlation coefficient ranges from -1 to 1, where 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no linear relationship. Select independent parameter features as input variables.

[0072] To improve the response speed to changes in power commands and accurately predict changes in superheated steam temperature, this invention incorporates future power commands as another input variable of the model, combines the selected parameter features and future power commands to form input sample feature data, and divides the model into training and testing sets.

[0073] By taking future power commands into account, generator sets can operate more flexibly, adjusting superheated steam parameters in a timely manner according to changes in power commands, thereby improving their adaptability to load changes.

[0074] Step 3: Train the CNN-BiLSTM-Attention model for predicting superheated steam temperature using the training set. This includes: first, normalizing the feature data in the training set; then, extracting the features between the normalized data using the CNN module; importing the features into the BiLSTM network for fitting and prediction via the fully connected layer; and finally, multiplying the output with the weights obtained from the Attention mechanism to obtain the final output prediction result.

[0075] Normalization is performed using the Min-max normalization method:

[0076]

[0077] Where min(x) and max(x) are the minimum and maximum values ​​of the feature parameters in the training set of the input model, respectively, and x is the value of the feature parameter in the training set. ′ This represents the normalized result of the feature parameters in the training set.

[0078] The specific training process of the CNN-BiLSTM-Attention network used to predict superheated steam temperature includes the following steps:

[0079] S31. Extract feature data from the normalized training set using a CNN network: Aggregate each feature output from the CNN network's convolutional layers into a feature vector, ensuring that the feature output dimension is the same as the number of input feature maps. Specifically, the feature vector aggregation is as follows:

[0080]

[0081] Among them, c j This represents the feature vector aggregated by feature j. N represents the i-th feature vector of feature j. j Indicates the length of the feature map.

[0082] The specific convolutional layer operations are as follows:

[0083]

[0084] in, It is the j′th weight of the i-th convolutional kernel in the l-th layer. It is the j-th convolutional local region in the l-th layer, where W is the width of the convolution kernel, and y l(i.j) It is the convolution result of the l-th layer based on i convolution kernels and j′ weights. x is the weight of the i-th convolutional kernel in the l-th layer. l(j+j′) It is the (j+j′)th convolutional local region in the l-th layer.

[0085] S32. Prediction using a BiLSTM network: BiLSTM is a bidirectional LSTM. The LSTM principle diagram provided in this embodiment is as follows: Figure 2 As shown, LSTM uses forget gates, input gates, and output gates to remember and weigh features in short- and long-term time series data. BiLSTM remembers and filters the features extracted by CNN, fully considers the feature relationships between forward and backward data, performs fitting and prediction, and finally outputs to a fully connected layer for weighted output to calculate the predicted data. The memory state update of BiLSTM is as follows:

[0086]

[0087] The output of the memory unit is as follows:

[0088] O t =σ(W o x t +U o h t-1 +V o c t +b o (5)

[0089] Among them, c t c represents the long-term memory state of a memory unit at time t. t-1 f represents the long-term memory state of a memory unit at time t-1. t i represents the output of the forget gate at time t. t The input gate output represents the input gate output at time t. W represents the product of corresponding elements. o U o V o W represents the weight matrix of the input gate, forget gate, and output gate of the current memory unit, respectively. o U o x represents the weight matrices of the input gate and forget gate stored in the memory unit, respectively. t Represents input, h t-1 b represents the hidden layer state at time t-1. c b represents the bias vector during memory unit updates. o The bias vector representing the output of the memory unit, O t σ represents the output of the memory unit, and σ represents the activation function.

[0090] S33. Add an Attention mechanism to the output of the BiLSTM to weight the prediction results: based on the hidden layer states h of all encoders in the BiLSTM network. t and the current decoder hidden layer state h s The scoring function can be obtained as follows:

[0091]

[0092] Among them, W a Given a weight matrix, score(h) t h s ) represents the result of the scoring function. This is the transpose of the hidden state of all encoders. Given the current forward and backward states of the decoder, the corresponding attention weights can be calculated based on the score function results. Finally, the directly usable weights are calculated based on the context feature vectors, as shown in the following formula:

[0093]

[0094] Where, α ij For the weighted result, h i h n Let i and n be the encoder hidden layer states. Let be the hidden layer states of the decoder in the forward and backward directions at time j. Let m be the forward and backward hidden layer states of the decoder at time m. Finally, the weighted result is used as the final prediction output, which is then inversely normalized and compared with the true value.

[0095] S34. Set evaluation indicators for the final prediction results: The evaluation indicator used is the root mean square error (RMSE), and the formula is as follows:

[0096]

[0097] Where n is the number of predicted data points, y i For the i-th predicted value, Let be the true value of the i-th data point, where i is the predicted data label.

[0098] The CNN-BiLSTM-Attention network structure diagram provided in this embodiment is as follows: Figure 3 As shown:

[0099] The final CNN-BiLSTM-Attention network model structure includes one convolutional layer, followed by a dropout layer with a ratio of 0.2, one BiLSTM layer with 128 hidden neurons, one attention layer, one flattening layer, and two fully connected layers.

[0100] The normalized feature data in the training set is slid using a time sliding window, and the CNN-BiLSTM-Attention model is trained iteratively to obtain the trained CNN-BiLSTM-Attention model.

[0101] The time-sliding window setting includes: setting the time steps of the feature parameters after feature filtering in step 2, the time steps of the input future power command, and the number of steps for the output future superheated steam temperature prediction. Specifically, the prediction accuracy is higher when the time steps of the feature parameters after feature filtering in step 2 are greater than or equal to the time steps of the input future power command, the number of steps for the output future superheated steam temperature prediction is the same as the time steps of the input future power command, and the time steps of the input future power command are less than the time steps of the feature parameters after feature filtering.

[0102] The time sliding window principle diagram provided in this embodiment is as follows: Figure 4 As shown:

[0103] Since the combined parameter features selected through feature filtering and the future power command are used as the final feature input, the input future power command must have the same number of time steps as the parameter features selected through feature filtering. However, considering the prediction accuracy issue, the prediction result is more accurate when predicting less future data using more historical data. Therefore, the time sliding window is set as follows: the time steps for inputting the parameter features selected through feature filtering are 30, and the input future power command is 10. The last 20 steps of the input future power command are filled with the value of the power command input in the last step. An attention mechanism is set to ensure that the model pays more attention to the future power command in the first 10 steps. The output future superheated steam temperature prediction steps are 10, and the sliding step at each time step is one time step, thus realizing continuous multi-step prediction of future superheated steam temperature. The prediction results output after softmax of the model are inversely normalized and compared with the true values. The comparison results are used for loss calculation and index evaluation. Figure 5 It can be seen that this method has a good effect on temperature prediction, and the predicted value is close to the actual value.

[0104] Step 4: Using the trained CNN-BiLSTM-Attention model, perform sliding predictions on the test set using a time-sliding window to achieve continuous multi-step predictions of superheated steam temperature. The time-sliding window settings are the same as in step S3.

[0105] In summary, this invention provides a multi-step prediction method for superheated steam temperature of thermal power units based on future load commands. The specific implementation process of this method is as follows:

[0106] First, parameter data affecting superheated steam temperature changes are collected. Then, signal denoising and correlation coefficient methods are used for feature selection to obtain the most relevant input variables. Future power commands are then added as one of the model inputs. Next, to eliminate the influence of data dimensions, the input data is normalized, and a CNN-BiLSTM-Attention prediction model is constructed. The CNN extracts features from the input data, the BiLSTM memorizes historical information and performs fitting predictions, and the Attention mechanism assigns different weights to different features in the output. Finally, a time sliding window is set to achieve continuous prediction of future temperatures over multiple steps.

[0107] Experimental comparisons show that the method of the present invention can significantly improve the prediction accuracy of superheated steam temperature, respond rapidly to temperature changes, expand the prediction time range, adapt to complex operating conditions, effectively reduce the risk of tube rupture accidents, optimize boiler operation, and reduce energy consumption.

[0108] This invention provides a novel multi-step temperature prediction method that integrates an attention mechanism, CNN, and BiLSTM, achieving good prediction results and effectively improving the safe and stable operation capability of industrial boilers.

[0109] Based on the same inventive concept, the multi-step prediction system for superheated steam temperature of thermal power units based on future load commands of the present invention includes:

[0110] The data acquisition unit is used to collect parameters that change due to the unit's variable operating conditions, including header outlet temperature, A and B side outlet temperatures, steam flow rate, primary and secondary air volume, and high pressure heater inlet pressure, all of which will cause changes in superheated steam temperature.

[0111] The data processing unit is used to perform feature filtering on the collected parameters, including signal denoising, filtering based on the degree of correlation between features, retaining highly correlated features as an input variable, and using future power commands as another input variable to form the input sample data, and dividing the data into training and test sets.

[0112] The model training unit is used to train the CNN-BiLSTM-Attention model for predicting superheated steam temperature using the training set. The process includes: first, normalizing the data in the training set; then, extracting the features between the normalized data using the CNN module; importing the features into the BiLSTM network and fitting and predicting them through the fully connected layer; and finally, multiplying the output with the weights obtained by the Attention mechanism to obtain the final output prediction result. The normalized training set data is then slid using a time sliding window to iterate and train the CNN-BiLSTM-Attention model cyclically to obtain the trained CNN-BiLSTM-Attention model.

[0113] A multi-step prediction unit is used to perform sliding predictions on the test set using a trained CNN-BiLSTM-Attention model with a time sliding window, achieving continuous multi-step prediction of superheated steam temperature. Based on the same inventive concept, an electronic device of the present invention includes:

[0114] Memory containing executable program code;

[0115] A processor coupled to the memory;

[0116] The processor calls the executable program code stored in the memory to execute the steps of the multi-step prediction method for superheated steam temperature of thermal power units based on future load instructions, as described above.

[0117] Based on the same inventive concept, the present invention provides a computer-readable storage medium storing computer instructions, which, when invoked, are used for the steps of the above-described multi-step prediction method for superheated steam temperature of thermal power units based on future load instructions.

Claims

1. A multi-step prediction method for superheated steam temperature of thermal power units based on future load commands, characterized in that, Includes the following steps: S1. Collect parameters of the superheater changes caused by the unit's variable operating conditions, including header outlet temperature, A and B side outlet temperatures, steam flow rate, primary and secondary air volume, and high pressure heater inlet pressure, all of which will cause changes in superheated steam temperature. S2. Perform feature filtering on the collected parameters, including signal denoising. Filter the parameters according to the degree of relationship between them, retain highly correlated features as one input variable, and use the future power command as another input variable to form the input sample data, and divide the data into training set and test set. S3. The training set is used to train the CNN-BiLSTM-Attention model for predicting superheated steam temperature. This includes: first, normalizing the data in the training set; then, extracting the features between the normalized data using the CNN module; importing the features into the BiLSTM network and fitting and predicting them through the fully connected layer; and finally, multiplying the output with the weights obtained by the Attention mechanism to obtain the final output prediction result; and then using a time sliding window to slide the normalized training set data to iterate and train the CNN-BiLSTM-Attention model to obtain the trained CNN-BiLSTM-Attention model. S4. Using the trained CNN-BiLSTM-Attention model, a time sliding window is used to perform sliding prediction on the test set to achieve continuous multi-step prediction of superheated steam temperature.

2. The multi-step prediction method for superheated steam temperature of thermal power units based on future load commands as described in claim 1, characterized in that, Step S2 includes the following steps: S21. Perform wavelet transform on the original collected parameter data. After wavelet transform, determine the wavelet coefficient threshold. Wavelet coefficients greater than the threshold are generated by signals and should be retained. Wavelet coefficients less than the threshold are generated by noise and should be set to zero. S22. Use the correlation coefficient method to filter the denoised data. The correlation coefficient ranges from -1 to 1, where 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no linear relationship. Filter out independent parameter features as input variables.

3. The multi-step prediction method for superheated steam temperature of thermal power units based on future load commands according to claim 1, characterized in that, In step S3, the normalization method is Min-max normalization: Where min(x) and max(x) are the minimum and maximum values ​​of the feature parameters in the training set, respectively, x is the value of the feature parameters in the training set, and x′ is the normalized result of the feature parameters in the training set.

4. The multi-step prediction method for superheated steam temperature of thermal power units based on future load commands according to claim 1, characterized in that, The specific training process of the CNN-BiLSTM-Attention model used to predict superheated steam temperature includes the following steps: S31. Extract feature data from the normalized training set using a CNN network: Aggregate each feature output from the CNN network's convolutional layers into a feature vector, ensuring that the feature output dimension is the same as the number of input feature maps; specifically, the feature vector aggregation is as follows: Among them, c j This represents the feature vector aggregated by feature j. N represents the i-th feature vector of feature j. j Indicates the length of the feature map; The specific convolutional layer operations are as follows: in, It is the j′th weight of the i-th convolutional kernel in the l-th layer. It is the j-th convolutional local region in the l-th layer, where W is the width of the convolution kernel, and y l(i.j) It is the convolution result of the l-th layer based on i convolution kernels and j′ weights. x is the weight of the i-th convolutional kernel in the l-th layer. l(j+j′) It is the (j+j′)th convolutional local region in the l-th layer; S32. Prediction using a BiLSTM network: BiLSTM memorizes and filters the features extracted by the CNN, fully considering the feature relationships between the forward and backward data, and performs fitting prediction. Finally, the output is fed to a fully connected layer for weighted output to calculate the predicted data. The BiLSTM's memory state update is as follows: The output of the memory unit is as follows: O t =σ(W o x t +U o h t-1 +V o c t +b o ) Among them, c t c represents the long-term memory state of a memory unit at time t. t-1 f represents the long-term memory state of a memory unit at time t-1. t i represents the forget gate output at time t of the memory unit. t The input gate output represents the memory cell at time t. W represents the product of corresponding elements. o U o V o W represents the weight matrix of the input gate, forget gate, and output gate of the current memory unit, respectively. c U c x represents the weight matrices of the input gate and forget gate stored in the memory unit, respectively. t Represents input, h t-1 b represents the hidden layer state at time t-1. c b represents the bias vector during memory unit updates. o The bias vector representing the output of the memory unit, O t σ represents the output of the memory unit, and σ represents the activation function. S33. Add an Attention mechanism to the output of the BiLSTM to weight the prediction results: based on the hidden layer states h of all encoders in the BiLSTM network. t The score function is obtained by combining the current decoder hidden state h1: Among them, W a Let h be the weight matrix, and score(h) t h1) represents the result of the scoring function. This is the transpose of the hidden state of all encoders. Given the current forward and backward states of the decoder, the corresponding attention weights are then calculated based on the score function results. Finally, the directly usable weights are calculated based on the context feature vectors, as shown in the following formula: Where, α ij For the weighted result, h i h n Let be the encoder hidden layer states at time i and time n, respectively. Let be the hidden layer states of the decoder in the forward and backward directions at time j. Let m be the forward and backward hidden layer states of the decoder at time m, and finally the weighted result is used as the final prediction output. S34, determine the evaluation indicators for the final prediction results.

5. The multi-step prediction method for superheated steam temperature of thermal power units based on future load commands according to claim 4, characterized in that, The evaluation index determined in step S34 is the root mean square error (RMSE), and the formula is as follows: Where n is the number of predicted data points, y i For the i-th predicted value, Let be the true value of the i-th data point, where i is the predicted data label.

6. The multi-step prediction method for superheated steam temperature of thermal power units based on future load commands according to claim 1, characterized in that, The CNN-BiLSTM-Attention network model structure includes one convolutional layer, followed by a dropout layer, one BiLSTM layer, one attention layer, one flattening layer, and two fully connected layers.

7. The multi-step prediction method for superheated steam temperature of thermal power units based on future load commands according to claim 1, characterized in that, The sliding window of time has one time step. When the time step of the input future power command is less than the time step of the feature parameters after feature filtering, the missing part of the input future power command is filled with the value of the power command input in the last step.

8. A multi-step prediction system for superheated steam temperature of thermal power units based on future load commands, characterized in that, include: The data acquisition unit is used to collect parameters that change due to the unit's variable operating conditions, including header outlet temperature, A and B side outlet temperatures, steam flow rate, primary and secondary air volume, and high pressure heater inlet pressure, all of which will cause changes in superheated steam temperature. The data processing unit is used to perform feature filtering on the collected parameters, including signal denoising, filtering based on the degree of correlation between features, retaining highly correlated features as an input variable, and using future power commands as another input variable to form the input sample data, and dividing the data into training and test sets. The model training unit is used to train the CNN-BiLSTM-Attention model for predicting superheated steam temperature using the training set. The process includes: first, normalizing the data in the training set; then, extracting the features between the normalized data using the CNN module; importing the features into the BiLSTM network and fitting and predicting them through the fully connected layer; and finally, multiplying the output with the weights obtained by the Attention mechanism to obtain the final output prediction result. The normalized training set data is then slid using a time sliding window to iterate and train the CNN-BiLSTM-Attention model cyclically to obtain the trained CNN-BiLSTM-Attention model. The multi-step prediction unit is used to perform sliding predictions on the test set using a time sliding window by employing a trained CNN-BiLSTM-Attention model, thereby achieving continuous multi-step predictions of superheated steam temperature.

9. An electronic device, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the steps of the multi-step prediction method for superheated steam temperature of thermal power units based on future load commands as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, which, when invoked, are used to perform the steps of the multi-step prediction method for superheated steam temperature of thermal power units based on future load instructions as described in any one of claims 1-7.

Images