Condenser vacuum prediction method and device based on frequency domain information guidance
By combining variational modal decomposition and Transformer encoder, using attention mechanism to perform feature fusion, the shortcomings of the condenser vacuum degree prediction model in the prior art in the processing of complex relationships and timing relationships are achieved, and higher precision Vacuum degree prediction.
Patent Information
- Application Number
- CN202210538578.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-18
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2042-05-18
AI Technical Summary
The existing condenser vacuum degree prediction model has shortcomings in considering the complex relationship and time sequence relationship between data and vacuum degree, resulting in inaccurate prediction.
Using a feature fusion method based on frequency domain information guidance combining variational modal decomposition, Transformer encoder and attention mechanism, frequency domain features in historical vacuum sequences are extracted through variational modal decomposition, and the Transformer encoder extracts historical working conditions. Time domain features in the data, and fusion feature encoding is used to use attention mechanisms, and vacuum degree prediction is finally performed using a multi-layer perceptron.
The accuracy of condenser vacuum degree prediction is improved, and it is better than the Xgboost method in ensemble learning and the regression algorithm in machine learning, with higher prediction accuracy.
Smart Images

Figure CN114819388B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of thermal power plant equipment operation optimization, and in particular to a condenser vacuum prediction method and device. Background Art
[0002] As the main auxiliary equipment of the steam turbine in a thermal power plant, the condenser plays a very important role in the safe operation of the unit. The condenser vacuum is an indicator that comprehensively reflects the operating status of the condenser and is an important parameter for the operation of a thermal power generating unit. Whether in the thermal design of the condenser or in the operation of the cold end equipment of the steam turbine, the condenser vacuum is required to have an optimal value. Strict monitoring of it is the focus of every power plant.
[0003] As the condenser vacuum increases, the unit output increases, but the power consumption of the circulating water pump and vacuum pump also increases. When the difference between the turbine power and the circulating water pump and vacuum pump power reaches a maximum value, the condenser vacuum at this time is called the optimal vacuum. The condenser vacuum is mainly affected by many factors such as the circulating water flow rate, the temperature and pressure of the circulating water at the inlet and outlet of the condenser, and the current of the circulating water pump, and the relationship is relatively complex. Therefore, an effective method is needed to predict the condenser vacuum to ensure the safe, stable, economical and effective operation of the unit.
[0004] At present, there are many vacuum prediction models based on machine learning and deep learning algorithms, and some achievements have been made. They include: condenser vacuum prediction model based on ensemble learning, condenser vacuum prediction method for thermal power plant based on multi-layer LSTM, vacuum prediction model based on regression algorithm in machine learning, condenser vacuum prediction based on improved fruit fly algorithm to optimize SVM, condenser vacuum prediction based on fruit fly algorithm to optimize generalized regression neural network, multivariate regression prediction model based on K-means clustering, condenser vacuum target value prediction based on PSO-SVR model, condenser vacuum prediction model combining particle swarm algorithm with radial neural network, and double-pressure condenser vacuum prediction using PSO algorithm to improve Elman neural network, etc. Some of these models use machine learning algorithms, but do not consider the complex relationship between data and vacuum, some do not consider the temporal relationship of data, and some do not consider the role of historical vacuum sequence information, resulting in inaccurate prediction. Summary of the invention
[0005] In order to overcome the deficiencies in the prior art and improve the accuracy of prediction, the present invention provides a condenser vacuum prediction method that combines variational mode decomposition, Transformer encoder, and feature fusion based on frequency domain information guidance. The method first mines the implicit periodic features in the historical vacuum sequence data through variational mode decomposition, extracts the time domain features in the historical operating condition data through the Transformer encoder, then uses the attention mechanism to obtain the fusion features based on frequency domain information guidance, and finally uses a multi-layer perceptron to predict the vacuum.
[0006] To achieve the above object, the technical solution adopted by the present invention is:
[0007] On the one hand, the present invention provides a condenser vacuum prediction method based on frequency domain information guidance, comprising the following steps:
[0008] Obtain the historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding time of each historical vacuum degree;
[0009] The vacuum degree prediction model is used to obtain the condenser vacuum degree at the time to be predicted, including: using the variational mode decomposition method and the one-dimensional convolutional network to extract the frequency domain features in the historical vacuum degree series, and using the Transformer encoder to extract the time domain features in the historical operating condition data;
[0010] The attention mechanism is used to fuse frequency domain features and time domain features. The attention scores of frequency domain features and time domain features are calculated to obtain the fusion feature encoding guided by frequency domain information.
[0011] The extracted fusion feature encoding is used through a multi-layer perceptron to obtain the condenser vacuum degree at the time to be predicted.
[0012] Further, obtaining the historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding time of each historical vacuum degree includes preprocessing the original historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding time of each historical vacuum degree, wherein the preprocessing includes:
[0013] The data of each sample is judged. If the data of the sample contains null values or abnormal values, it is deleted directly; and the retained data is normalized.
[0014] Further, the vacuum degree prediction model is trained, and the training process includes:
[0015] Obtain the historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding time of each historical vacuum degree;
[0016] Preprocess the acquired historical vacuum degree series and historical operating condition data;
[0017] For each time t to be predicted, construct a corresponding training sample (V t-b:t-1 ,X t-a:t , t ), where V t is the true vacuum degree at time t, V t-b:t-1 is the historical vacuum degree sequence, X t-a:t is the historical operating data, a is the length of the historical operating data, and b is the length of the historical sequence; V t is the actual condenser vacuum at time t;
[0018] The frequency domain features in the historical vacuum degree series are extracted using the variational mode decomposition method and one-dimensional convolutional network, and the time domain features in the historical operating condition data are extracted using the Transformer encoder.
[0019] The attention mechanism is used to fuse frequency domain features and time domain features. The attention scores of frequency domain features and time domain features are calculated to obtain the fusion feature encoding guided by frequency domain information.
[0020] For each moment t, the corresponding fusion feature code is encoded through a multi-layer perceptron to obtain the predicted vacuum output at the corresponding moment; the parameters are iteratively optimized through gradient descent and back propagation to minimize the loss until convergence or the maximum number of iterations is reached.
[0021] Furthermore, the loss is calculated as follows:
[0022]
[0023] Where l is the loss, is the predicted actual condenser vacuum at time t.
[0024] Furthermore, the steps of using the variational mode decomposition method and the one-dimensional convolutional network to extract the frequency domain features in the historical vacuum degree series are as follows:
[0025] (1) For time t, randomly initialize the Fourier transform parameters corresponding to the kth submode in the first iteration kth mode center frequency Langrange multipliers Assume the number of iterations n is 1; by introducing the quadratic penalty factor α and the Lagrange multiplier λ(t), the constrained variational problem is transformed into an unconstrained variational problem;
[0026] (2) Assign n+1 to n and use the following update formula:
[0027]
[0028]
[0029]
[0030] Get the Fourier transform parameters corresponding to the kth submode in the nth iteration kth mode center frequency and the Lagrange multiplier in, is the submodal set {u k}The result after Fourier transform; {ω k} is the set of center frequencies; ω is the center frequency; K is the number of intrinsic mode functions; δ(t) is the Dirac function; n is the number of iterations
[0031] (3)Judgment If it holds, the loop stops; otherwise, return to (2) and continue the iterative loop, where ∈ represents the convergence accuracy;
[0032] (4) For the historical space-time degree sequence V t-b:t-1 The K intrinsic mode functions obtained by variational mode decomposition are used to form the intrinsic mode function matrix U = [u 1 ,u 2 ,…,u N ] T , Among them, u 1 ,u 2 ,…,u N The Fourier transform parameters corresponding to the K modes obtained in the nth iteration The intrinsic mode function is obtained by inverse Fourier transform, where N represents the dimension of the intrinsic mode function; the frequency domain feature o of the historical vacuum degree sequence is obtained by passing the matrix U through a one-dimensional convolutional neural network. t .
[0033] Furthermore, using the Transformer encoder to extract time domain features from historical operating condition data includes the following steps:
[0034] For a historical operating data X at time t t-a:t , divide the Transformer encoder based on the multi-head self-attention mechanism into m heads;
[0035] (1) Use the self-attention calculation formula in each head to get X t-a:t The precoding vector Will Concatenate all the items in to get X t-a:t Precoding E t , E t With X t-a:t Sum to get E t ′, E t ′ is obtained through a feedforward neural networkt ″, for E t ′ and E t ″Sum to get X t-a:t The single-layer encoding result h t ;
[0036] (2) For the L-layer encoder, the single-layer encoding result h in (2) is t As the input of the next encoder layer, execute (2) to obtain the time domain features H of the historical operating data t .
[0037] Furthermore, the attention mechanism is used to use frequency domain information to guide the fusion of features. The specific steps are as follows:
[0038] (1) For the frequency domain feature o at time t t , use linear transformation to get the 1×|x|-dimensional frequency domain feature vector o′ t , where x∈X t-a:t represents any historical data node, and a is the number of historical operating condition data;
[0039] (2) For the frequency domain eigenvector o′ t and the time domain feature H t , randomly initialize three matrices where x∈X t-a:t Represents any historical operating condition data node, represents the subspace encoding length;
[0040] (3) The frequency domain feature vector o′ t With W Q Multiply to obtain the frequency domain steering vector Q o , the time domain feature H t Respectively with W K , W V Multiply to get the time domain key vector K H , time domain value vector V H :
[0041] Q o =o′ t ·W Q
[0042] K H =H t ·W K
[0043] V H =H t ·W V
[0044] (4) Using the frequency domain guidance vector Q oThe time domain key vector Each element in is dot-product, where Indicates the time domain key corresponding to the i-th historical operating condition data;
[0045] Calculate the attention score S of each historical working condition data 1 ,s 2 ,…,s a};
[0046] (5) For each score s i ∈S is normalized to obtain
[0047] (6) All Respectively with the corresponding time domain value vector Multiply and sum to get the fusion feature code F guided by frequency domain information t :
[0048]
[0049] is the normalized i-th score, is the time domain value vector V H The i-th element in .
[0050] Furthermore, the normalization operation adopts the following formula:
[0051]
[0052] where d k is the scaling scale, scaling scale d k Set to the time domain key vector K H Dimension.
[0053] In a second aspect, the present invention further provides a condenser vacuum prediction device based on frequency domain information guidance, comprising a data acquisition module and a vacuum prediction model, wherein the vacuum prediction model comprises a frequency domain feature acquisition module, a time domain feature acquisition module, a feature fusion module and a multi-layer perceptron;
[0054] The data acquisition module is used to acquire the historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding moment of each historical vacuum degree;
[0055] The frequency domain feature acquisition module is used to extract frequency domain features in the historical vacuum degree sequence using a variational mode decomposition method and a one-dimensional convolutional network;
[0056] The time domain feature acquisition module is used to extract time domain features from historical operating condition data using a Transformer encoder;
[0057] The feature fusion module is used to fuse frequency domain features and time domain features using an attention mechanism, and obtain a fusion feature encoding guided by frequency domain information by calculating the attention scores of frequency domain features and time domain features;
[0058] The multi-layer perceptron is used to obtain the condenser vacuum degree at the time to be predicted according to the extracted fusion feature code.
[0059] Furthermore, a preprocessing module is included, which is used to preprocess the original condenser historical vacuum degree sequence and the historical operating condition data at the corresponding time of each historical vacuum degree, and the preprocessing includes:
[0060] The data of each sample is judged. If the data of the sample contains null values or abnormal values, it is deleted directly; and the retained data is normalized.
[0061] Beneficial technical effects achieved by the present invention:
[0062] The condenser vacuum prediction method based on frequency domain information guidance provided by the present invention captures the frequency domain features implicit in the historical vacuum sequence data by combining variational mode decomposition and convolutional neural network, obtains the time domain features in the historical operating condition data by using the Transformer encoder, and finally uses the attention mechanism to obtain the fusion features based on frequency domain information guidance. Compared with the Xgboost method in ensemble learning and the regression algorithm in machine learning, it has higher prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] Figure 1 It is a flow chart of a method according to an embodiment of the present invention. DETAILED DESCRIPTION
[0064] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.
[0065] Variational mode decomposition is an adaptive, completely non-recursive method of modal variation and signal processing. This technology has the advantage of being able to determine the number of modal decompositions. Its adaptability is reflected in determining the number of modal decompositions of a given sequence according to actual conditions. In the subsequent search and solution process, it can adaptively match the optimal center frequency and finite bandwidth of each mode, and can achieve effective separation of intrinsic mode functions, frequency domain division of signals, and then obtain effective decomposition components of a given signal, and finally obtain the optimal solution to the variational problem. The vacuum degree of the condenser will show a certain periodicity after being converted to the frequency domain. The variational mode decomposition method can extract the stationary characteristics of the condenser vacuum degree time series data and effectively improve the prediction accuracy.
[0066] The attention mechanism is widely used in natural language processing and has shown good results. The self-attention mechanism can effectively solve the problem that the neural network cannot fully utilize the relationship between inputs during actual training, resulting in extremely poor model training results. The Transformer encoder stacks many self-attention modules, which can fully explore the correlation between data at different times and better explore the time domain features in the data. Moreover, in the prediction models of other fields, the frequency domain features and time domain features are simply connected or added for fusion. The feature fusion method based on the attention mechanism allows the time domain features and frequency domain features to interact, so as to obtain fusion features guided by frequency domain information and improve the final prediction accuracy.
[0067] The condenser vacuum prediction method based on frequency domain information guidance provided by the present invention captures the frequency domain features implicit in the historical vacuum sequence data by combining variational mode decomposition and convolutional neural network, obtains the time domain features in the historical operating condition data by using the Transformer encoder, and finally uses the attention mechanism to obtain the fusion features based on frequency domain information guidance. Compared with the Xgboost method in ensemble learning and the regression algorithm in machine learning, it has higher prediction accuracy.
[0068] Embodiment: A condenser vacuum prediction method based on frequency domain information guidance, such as Figure 1 As shown, the method includes three steps: data preprocessing and training set generation, feature extraction and fusion, and prediction.
[0069] The problem can be described as follows: Use V = {V 1 ,V 2 …V N} represents the historical condenser vacuum degree series data, and X={x 1 ,x 2 …x N} represents the operating data. The task of condenser vacuum prediction is to predict the vacuum at the current moment based on the previous b historical vacuum sequence data V and the operating data X.
[0070] 1. Data preprocessing and training set generation
[0071] Step 1: In the data preprocessing and training set generation stage, the collected data at each moment is judged. If the data at that moment contains null values or abnormal values, it is directly deleted; the data is normalized; for each time t to be predicted, the corresponding training sample (V t-b:t-1 ,X t-a:t , t ), where V t is the true vacuum degree at time t, V t-b:t-1is the historical vacuum degree sequence, is the length of the historical sequence, X t-a:t is the historical working condition data, and a is the number of historical working condition data. Finally, the time before November is used as the training set, and the time after November is used as the test set.
[0072] 2. Feature Extraction and Fusion
[0073] Step 2, in the feature extraction stage of the model, the steps of using variational mode decomposition and Transformer encoder to extract frequency domain features and time domain features are as follows:
[0074] (1) For time t, randomly initialize the Fourier transform parameters corresponding to the kth submode in the first iteration kth mode center frequency Langrange multipliers The number of iterations n=1; by introducing the quadratic penalty factor α and the Lagrange multiplier λ(t), the constrained variational problem can be transformed into an unconstrained variational problem.
[0075] (2) Assign n+1 to n and use the following update formula:
[0076]
[0077]
[0078]
[0079] Get the Fourier transform parameters corresponding to the kth submode in the nth iteration kth mode center frequency and the Lagrange multiplier in, is the submodal set {u k}The result after Fourier transform; {ω k} is the set of center frequencies; ω is the center frequency; K is the number of intrinsic mode functions; δ(t) is the Dirac function; n is the number of iterations;
[0080] (3)Judgment If true, the loop stops; otherwise, return to (2) and continue the iterative loop.
[0081] Among them, ∈ represents the convergence accuracy;
[0082] (4) For the historical vacuum degree series data W t-b:t-1 The K intrinsic mode functions obtained by variational mode decomposition can form the eigenmode function matrix U = [u 1 ,u 2 ,…,u N ] T, Among them, u 1 ,u 2 ,…,u N The Fourier transform parameters corresponding to the K modes obtained in the nth iteration The intrinsic mode function obtained by inverse Fourier transform represents the dimension of the intrinsic mode function. The matrix U is passed through a one-dimensional convolutional neural network to obtain the frequency domain characteristics of the historical vacuum degree sequence o t ;
[0083] (5) For a historical operating condition data X at time t t-a:t , first divide the Transformer encoder based on the multi-head self-attention mechanism into m heads;
[0084] (6) Use the self-attention calculation formula in each head to get X t-a:t The precoding vector Will Concatenate all the items in to get X t-a:t Precoding E t , E t With X t-a:t Sum to get E t ′, then E t ′ is obtained through a feedforward neural network t ″, for E t ′ and E t ″Sum to get X t-a:t The single-layer encoding result h t ;
[0085] (7) For the L-layer encoder, the single-layer encoding result in (6) is used as the input of the next-layer encoder, and (6) is executed to obtain the time domain feature encoding H of the historical working condition data. t .
[0086] Step 3: In the feature fusion stage of the model, the attention mechanism is used to use frequency domain information to guide the fusion of features. The specific steps are as follows:
[0087] (1) For the frequency domain feature o at time t t , first use linear transformation to get the 1×|x|-dimensional vector o′ t , where x∈X t-a:t Represents any historical data node;
[0088] (2) For the frequency domain feature o′ t and the time domain feature H t , randomly initialize three matrices where x∈X t-a:t Represents any historical operating condition data node, represents the subspace encoding length;
[0089] (3) Change o′ t With W Q Multiply to obtain the frequency domain steering vector Q o , H t Respectively with W K , W V Multiply to get the time domain key vector K H , time domain value vector V H :
[0090] Q o =o′ t ·W Q
[0091] K H =H t ·W K
[0092] V H =H t ·W V
[0093] (4) Using the frequency domain guidance vector Q o Respectively with all time domain key vectors Point multiplication can calculate the attention score S = {s 1 ,s 2 ,…,s a};
[0094] (5) For each score s i ∈S is normalized to obtain Optionally, the following normalization operation is used in this embodiment:
[0095]
[0096] where d k is the scaling scale, usually set to K H Dimension.
[0097] (6) Then all Respectively with the corresponding time domain value vector Multiply and sum to get the fusion feature code F guided by frequency domain information t :
[0098]
[0099] 3. Prediction
[0100] Step 4: For each time t, the corresponding fusion feature encoding F tThe predicted vacuum output at the corresponding moment is obtained through a multi-layer perceptron Multi-Layer perceptron (MLP) is the basic algorithm of Deep Neural Networks (DNN). It is a prior art and will not be described in detail in this application.
[0101] The process of the condenser vacuum prediction method based on frequency domain information is as follows:
[0102]
[0103]
[0104] Corresponding to the above embodiment, the embodiment of the present invention further provides a condenser vacuum prediction device based on frequency domain information guidance, including a data acquisition module and a vacuum prediction model, wherein the vacuum prediction model includes a frequency domain feature acquisition module, a time domain feature acquisition module, a feature fusion module and a multi-layer perceptron;
[0105] The data acquisition module is used to acquire the historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding moment of each historical vacuum degree;
[0106] The frequency domain feature acquisition module is used to extract frequency domain features in the historical vacuum degree sequence using a variational mode decomposition method and a one-dimensional convolutional network;
[0107] The time domain feature acquisition module is used to extract time domain features from historical operating condition data using a Transformer encoder;
[0108] The feature fusion module is used to fuse frequency domain features and time domain features using an attention mechanism, and obtain a fusion feature encoding guided by frequency domain information by calculating the attention scores of frequency domain features and time domain features;
[0109] The multi-layer perceptron is used to obtain the condenser vacuum degree at the time to be predicted according to the extracted fusion feature code.
[0110] Optionally, a preprocessing module and a training module are also included;
[0111] The preprocessing module is used to preprocess the original condenser historical vacuum degree sequence and the historical operating condition data at the corresponding time of each historical vacuum degree, and the preprocessing includes:
[0112] The data of each sample is judged. If the data of the sample contains null values or abnormal values, it is deleted directly; and the retained data is normalized.
[0113] The training module is used to train the entire vacuum prediction model. The training process includes:
[0114] Obtain the historical vacuum degree sequence of the condenser and the historical operating condition data of each historical vacuum degree at the corresponding time;
[0115] Preprocess the acquired historical vacuum degree series and historical operating condition data;
[0116] For each time t to be predicted, construct a corresponding training sample (V t-b:t-1 ,X t-a:t ,V t ), where V t is the true vacuum degree at time t, V t-b:t-1 is the historical vacuum degree sequence, X t-a:t is the historical operating data, a is the length of the historical operating data, and b is the length of the historical sequence; V t is the actual condenser vacuum at time t;
[0117] The frequency domain features in the historical vacuum degree series are extracted using the variational mode decomposition method and one-dimensional convolutional network, and the time domain features in the historical operating condition data are extracted using the Transformer encoder.
[0118] The attention mechanism is used to fuse frequency domain features and time domain features. The attention scores of frequency domain features and time domain features are calculated to obtain the fusion feature encoding guided by frequency domain information.
[0119] For each moment t, the corresponding fusion feature code is encoded through a multi-layer perceptron to obtain the predicted vacuum output at the corresponding moment; the parameters are iteratively optimized through gradient descent and back propagation to minimize the loss until convergence or the maximum number of iterations is reached.
[0120] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0121] The condenser vacuum prediction method based on frequency domain information guidance used in the present invention is trained and tested on the measured data of a power plant in Jiangsu. The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators. The lower the index, the better the effect. In the experiment, the RMSE of the prediction method in this paper is 0.1174 and the MAE is 0.0917, which are better than the 0.7786 and 0.6081 of the optimal Decision Tree Regression method in the baseline model.
[0122] The present invention combines variational mode decomposition and convolutional neural network to capture the frequency domain features implicit in the historical vacuum degree sequence data, obtains the time domain features in the historical working condition data through the Transformer encoder, and finally uses the attention mechanism to obtain the fusion features guided by the frequency domain information. Compared with the Xgboost method in ensemble learning and the regression algorithm in machine learning, it has higher prediction accuracy.
[0123] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0124] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0125] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0126] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0127] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.
Claims
1. A condenser vacuum prediction method based on frequency domain information guidance, characterized in that: The following steps are involved: Obtain the historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding time of each historical vacuum degree; The vacuum degree prediction model is used to obtain the condenser vacuum degree at the time to be predicted, including: using the variational mode decomposition method and the one-dimensional convolutional network to extract the frequency domain features in the historical vacuum degree sequence, and using the Transformer encoder to extract the time domain features in the historical operating condition data; wherein, using the Transformer encoder to extract the time domain features in the historical operating condition data includes the following steps: For a historical operating data X at time t t-a:t , divide the Transformer encoder based on the multi-head self-attention mechanism into m heads; (1) Use the self-attention calculation formula in each head to get X t-a:t The precoding vector Will Concatenate all the items in to get X t-a:t Precoding E t , E t With X t-a:t Sum to get E t ′, E t ′ is obtained through a feedforward neural network t ″, for E t ′ and E t ″Sum to get X t-a:t The single-layer encoding result h t ; (2) For the L-layer encoder, the single-layer encoding result h in (2) is t As the input of the next encoder layer, execute (2) to obtain the time domain features H of the historical operating data t ; The attention mechanism is used to fuse frequency domain features and time domain features. The attention scores of frequency domain features and time domain features are calculated to obtain the fusion feature encoding guided by frequency domain information. Among them, the attention mechanism is used to use frequency domain information to guide the fusion of features. The specific steps are as follows: (1) For the frequency domain feature o at time t t , use linear transformation to get the 1×|x|-dimensional frequency domain feature vector o′ t , where x∈X t-a:t represents any historical data node, and a is the number of historical operating condition data; (2) For the frequency domain eigenvector o′ t and the time domain feature H t , randomly initialize three matrices where x∈X t-a:t Represents any historical operating condition data node, represents the subspace encoding length; (3) The frequency domain feature vector o′ t With W Q Multiply to obtain the frequency domain steering vector Q o , the time domain feature H t Respectively with W K , W V Multiply to get the time domain key vector K H , time domain value vector V H : Q o =o′ t ·W Q K H =H t ·W K V H =H t ·W V (4) Using the frequency domain guidance vector Q o The time domain key vector Each element in is dot-product, where Indicates the time domain key corresponding to the i-th historical operating condition data; Calculate the attention score S = {s1, s2, ..., s a }; (5) For each score s i ∈S is normalized to obtain (6) All Respectively with the corresponding time domain value vector Multiply and sum to get the fusion feature code F guided by frequency domain information t : is the normalized i-th score, is the time domain value vector V H The i-th element in ; The extracted fusion feature encoding is used through a multi-layer perceptron to obtain the condenser vacuum degree at the time to be predicted.
2. The condenser vacuum prediction method based on frequency domain information guidance according to claim 1 is characterized in that: Acquiring the historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding time of each historical vacuum degree, including preprocessing the original historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding time of each historical vacuum degree, wherein the preprocessing includes: The data of each sample is judged. If the data of the sample contains null values or abnormal values, it is deleted directly; and the retained data is normalized.
3. The condenser vacuum prediction method based on frequency domain information guidance according to claim 1 is characterized in that: Including training the vacuum degree prediction model, the training process includes: Obtain the historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding time of each historical vacuum degree; Preprocess the acquired historical vacuum degree series and historical operating condition data; For each time t to be predicted, construct a corresponding training sample (V t-b:T-1 , X t-a:T , V t ), where V t is the true vacuum degree at time t, V t-b:t-1 is the historical vacuum degree sequence, X t-a:t is the historical operating data, a is the length of the historical operating data, and b is the length of the historical sequence; V t is the actual condenser vacuum at time t; The frequency domain features in the historical vacuum degree series are extracted using the variational mode decomposition method and one-dimensional convolutional network, and the time domain features in the historical operating condition data are extracted using the Transformer encoder. The attention mechanism is used to fuse frequency domain features and time domain features. The attention scores of frequency domain features and time domain features are calculated to obtain the fusion feature encoding guided by frequency domain information. For each moment t, the corresponding fusion feature code is encoded through a multi-layer perceptron to obtain the predicted vacuum output at the corresponding moment; the parameters are iteratively optimized through gradient descent and back propagation to minimize the loss until convergence or the maximum number of iterations is reached.
4. The condenser vacuum prediction method based on frequency domain information guidance according to claim 3 is characterized in that: The loss is calculated as follows: Where l is the loss, is the predicted actual condenser vacuum at time t.
5. The condenser vacuum prediction method based on frequency domain information guidance according to any one of claims 1 to 4, characterized in that: The steps of using variational mode decomposition method and one-dimensional convolutional network to extract frequency domain features in historical vacuum degree series are as follows: (1) For time t, randomly initialize the Fourier transform parameters corresponding to the kth submode in the first iteration kth mode center frequency Langrange multipliers Assume the number of iterations n is 1; By introducing the quadratic penalty factor α and the Lagrange multiplier λ(t), the constrained variational problem is transformed into an unconstrained variational problem. (2) Assign n+1 to n and use the following update formula: Get the Fourier transform parameters corresponding to the kth submode in the nth iteration kth mode center frequency and the Lagrange multiplier in, is the submodal set {u k }The result after Fourier transform; {ω k } is the set of center frequencies; ω is the center frequency; K is the number of intrinsic mode functions; δ(t) is the Dirac function; n is the number of iterations; (3)Judgment If it holds, the loop stops; otherwise, return to (2) and continue the iterative loop, where ∈ represents the convergence accuracy; (4) For the historical space-time degree sequence V t-b:t-1 The K intrinsic mode functions obtained by variational mode decomposition are used to form the intrinsic mode function matrix U = [u1, u2, ..., u N ] T , Among them, u1, u2, ..., u N The Fourier transform parameters corresponding to the K modes obtained in the nth iteration The intrinsic mode function is obtained by inverse Fourier transform, where N represents the dimension of the intrinsic mode function; the frequency domain feature o of the historical vacuum degree sequence is obtained by passing the matrix U through a one-dimensional convolutional neural network. t .
6. The condenser vacuum prediction method based on frequency domain information guidance according to claim 1 is characterized in that: The normalization operation adopts the following formula: where d k is the scaling scale, scaling scale d k Set to the time domain key vector K H Dimension.
7. A condenser vacuum prediction device based on frequency domain information guidance, characterized in that: It includes a data acquisition module and a vacuum degree prediction model, wherein the vacuum degree prediction model includes a frequency domain feature acquisition module, a time domain feature acquisition module, a feature fusion module and a multi-layer perceptron; The data acquisition module is used to acquire the historical vacuum degree sequence of the condenser and the historical operating condition data at the corresponding moment of each historical vacuum degree; The frequency domain feature acquisition module is used to extract frequency domain features in the historical vacuum degree sequence using a variational mode decomposition method and a one-dimensional convolutional network; The time domain feature acquisition module is used to extract the time domain features in the historical operating condition data using the Transformer encoder; wherein, extracting the time domain features in the historical operating condition data using the Transformer encoder comprises the following steps: For a historical operating data X at time t t-a:t , divide the Transformer encoder based on the multi-head self-attention mechanism into m heads; (1) Use the self-attention calculation formula in each head to get X t-a:t The precoding vector Will Concatenate all the items in to get X t-a:t Precoding E t , E t With X t-a:t Sum to get E t ′, E t ′ is obtained through a feedforward neural network t ″, for E t ′ and E t ″Sum to get X t-a:t The single-layer encoding result h t ; (2) For the L-layer encoder, the single-layer encoding result h in (2) is t As the input of the next encoder layer, execute (2) to obtain the time domain features H of the historical operating data t ; The feature fusion module is used to fuse frequency domain features and time domain features using the attention mechanism, and obtains fusion feature encoding guided by frequency domain information by calculating the attention scores of frequency domain features and time domain features; wherein, the attention mechanism is used to use frequency domain information to guide the fusion of features, and the specific steps are as follows: (1) For the frequency domain feature o at time t t , use linear transformation to get the 1×|x|-dimensional frequency domain feature vector o′ t , where x∈X t-a:t represents any historical data node, and a is the number of historical operating condition data; (2) For the frequency domain eigenvector o′ t and the time domain feature H t , randomly initialize three matrices where x∈X t-a:t Represents any historical operating condition data node, represents the subspace encoding length; (3) The frequency domain feature vector o′ t With W Q Multiply to obtain the frequency domain steering vector Q o , the time domain feature H t Respectively with W K , W V Multiply to get the time domain key vector K H , time domain value vector V H : Q o =o′ t ·W Q K H =H t ·W K V H =H t ·W V (4) Using the frequency domain guidance vector Q o The time domain key vector Each element in is dot-product, where Indicates the time domain key corresponding to the i-th historical operating condition data; Calculate the attention score S = {s1, s2, ..., s a }; (5) For each score s i ∈S is normalized to obtain 1≤i≤a: (6) All Respectively with the corresponding time domain value vector Multiply and sum to get the fusion feature code F guided by frequency domain information t : is the normalized i-th score, is the time domain value vector V H The i-th element in ; The multi-layer perceptron is used to obtain the condenser vacuum degree at the time to be predicted according to the extracted fusion feature code.
8. The condenser vacuum prediction device based on frequency domain information guidance according to claim 7, characterized in that: It also includes a preprocessing module for preprocessing the original condenser historical vacuum degree sequence and the historical operating condition data at the corresponding moment of each historical vacuum degree, and the preprocessing includes: The data of each sample is judged. If the data of the sample contains null values or abnormal values, it is deleted directly; and the retained data is normalized.
