A method for predicting internal pressure data of mobile steam energy storage tank

Abstract

The present invention proposes a method for predicting the internal pressure data of a mobile steam energy storage tank, and relates to the technical field of machine learning. The present invention proposes a TSPNet prediction model, which includes a time-frequency feature capture module, a spatiotemporal feature learning module, and a prediction layer. Specifically, the time-frequency feature capture module is used to capture the time and frequency features in the input internal pressure data of the mobile steam energy storage tank, the spatiotemporal feature learning module is used to process the spatial dependency of the internal pressure data of the mobile steam energy storage tank, and the prediction layer is used to map the refined features extracted previously to the target output space, thereby generating the final prediction result. Each module has a good effect in capturing the internal pressure data of the mobile steam energy storage tank with complex time and space relationships.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a method for predicting internal pressure data of a mobile steam energy storage tank. Background Art

[0002] In modern industry, mobile steam energy storage tanks, as an important energy storage and management equipment, are widely used to cope with peak load demands and improve energy utilization efficiency. The pressure inside the steam energy storage tank is one of the important parameters affecting its operating efficiency and safety. Due to the dynamic characteristics of steam and the complex changes in the external environment, the pressure inside the energy storage tank often exhibits high nonlinearity and non-stationarity. Traditional methods based on empirical formulas or simple statistical models are unable to cope with these complex time series data.

[0003] The prediction of the internal pressure data of mobile steam energy storage tanks requires the combination of multi-dimensional data analysis, including real-time equipment status data, historical equipment data, and external environmental factors. The prediction method of the internal pressure data of mobile steam energy storage tanks based on these data can be analyzed through machine learning algorithms, and can also be combined with real-time data obtained by sensor technology for model training. This type of method can help decision makers optimize production scheduling and operation modes based on the prediction results, and improve the utilization efficiency of equipment.

[0004] Considering the complexity of the internal pressure data of mobile steam energy storage tanks, the current traditional machine learning methods find it difficult to capture long-term time series dependencies in complex pressure data, resulting in inaccurate long-term pressure predictions, especially when predicting pressure trends over a long period of time in the future. Traditional methods may not be able to fully identify the changes in the internal pressure of mobile steam energy storage tanks over time. In addition, traditional models usually focus on processing time series data, but often ignore spatial dependencies, resulting in their inability to provide comprehensive and accurate predictions when faced with complex energy consumption data. This neglect of spatiotemporal dependencies further limits the application of traditional methods, making them perform poorly when processing complex multi-dimensional data such as the internal pressure of mobile steam energy storage tanks. Summary of the invention

[0005] The present invention provides a method for predicting the internal pressure data of a mobile steam energy storage tank. Aiming at the internal pressure data of a mobile steam energy storage tank with multi-periodicity and long-term dependence, a TSPNet model is proposed, which consists of a time-frequency feature capture module, a spatiotemporal feature learning module and a prediction layer.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose specifically includes the following steps:

[0007] S1. Collecting data related to the internal pressure of the mobile steam energy storage tank, including features and target variables, and preprocessing the collected data related to the internal pressure of the mobile steam energy storage tank;

[0008] S2. Use the pre-processed data on the internal pressure of the mobile steam energy storage tank The method performs normalization operations and divides the data set;

[0009] S3, constructing a time-frequency feature capture module for capturing the time and frequency features of the input internal pressure data of the mobile steam energy storage tank, specifically comprising the following steps:

[0010] S31. Input the internal pressure data of the mobile steam energy storage tank ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, The time step is the frequency analysis of the internal pressure data of the mobile steam energy storage tank by stationary wavelet transform to obtain the high-frequency coefficient components. and low frequency coefficient components ;

[0011] S32, calculate the query, key and value of high-frequency coefficient components and low-frequency coefficient components, propose dynamic autoregressive attention, and introduce external context dynamic weights , perform dynamic autoregressive attention calculations to generate attention score matrices for high-frequency coefficient components and low-frequency coefficient components , , and calculate the dynamic autoregressive attention output of the high-frequency coefficient component and the low-frequency coefficient component , ;

[0012] S33, perform inverse stationary wavelet transform on the dynamic autoregressive attention output to obtain the output of the time-frequency feature capture module ;

[0013] S4. Constructing a spatiotemporal feature learning module for processing the spatial dependency of the internal pressure data of the mobile steam energy storage tank, specifically comprising the following steps:

[0014] S41, input the output of the time-frequency feature capture module ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, and through embedding and convolution operations, the feature matrix is ​​obtained ;

[0015] S42, Introducing the adjacency matrix based on dynamic time warping The multi-head self-attention mechanism is used to calculate the weights between the internal pressure data collected by different sensors of the mobile steam energy storage tank to obtain the attention output , aggregate the attention outputs of multiple heads to get the attention aggregation output ;

[0016] S43, using adjacency matrix based on dynamic time warping Compute the Laplacian matrix , input attention aggregation output , and perform graph convolution operations through Chebyshev polynomial approximation to calculate the feature representation after graph convolution ;

[0017] S5. Build the prediction layer and represent the features The input time series convolutional network is further processed to obtain refined features , and finally output the prediction results of the internal pressure data of the mobile steam energy storage tank through the fully connected layer .

[0018] Preferably, in step S1, the collected data related to the internal pressure of the mobile steam energy storage tank include top layer sensor data inside the energy storage tank, middle layer sensor data inside the energy storage tank, bottom layer sensor data inside the energy storage tank, internal temperature data of the energy storage tank, external temperature data of the energy storage tank, total volume data of the energy storage tank, flow data of the energy storage tank, internal pressure data of the energy storage tank, and pressure change rate data of the energy storage tank. The mean method is used to fill the missing values ​​of the internal pressure data of the mobile steam energy storage tank. The specific formula is as follows:

[0019] ;

[0020] In the formula, The mean internal pressure data of the mobile steam energy storage tank to be filled, is the total number of internal pressure data of the mobile steam energy storage tank, The internal pressure data of the mobile steam energy storage tank is data value.

[0021] Preferably, in step S2, using The method normalizes the internal pressure data of the mobile steam energy storage tank and divides the training set and the test set in proportion. The specific formula is as follows:

[0022] ;

[0023] In the formula, To obtain the internal pressure data of the mobile steam energy storage tank, is the normalized internal pressure data of the mobile steam energy storage tank, is the maximum value of the internal pressure data of the mobile steam energy storage tank, It is the minimum value in the internal pressure data of the mobile steam energy storage tank.

[0024] Preferably, in the steps S3 and S31, the internal pressure data of the mobile steam energy storage tank is input. ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, , is the time step, and the stationary wavelet transform is used to perform hierarchical frequency analysis on the internal pressure data of the mobile steam energy storage tank to obtain the high-frequency coefficient and the low-frequency coefficient. The specific formula is as follows:

[0025] ;

[0026] ;

[0027] In the formula, For scale and time point The high frequency coefficients of For scale and time point The low frequency coefficients of For scale The filter length is is the expanded high frequency filter coefficient, is the extended low-frequency filter coefficient, Internal pressure data of mobile steam storage tank In time The value at which the high frequency coefficients and low frequency coefficients Converted into high-frequency coefficient components and low-frequency coefficient components expressed in matrix form, the specific formula is as follows:

[0028] ;

[0029] ;

[0030] In the formula, For all time points In scale The high frequency coefficient components of For all time points In scale The low-frequency coefficient components of For scale The high frequency filter matrix, For scale The low frequency filter matrix, is the internal pressure data of the original mobile steam energy storage tank, which can be expressed as ,in is the internal pressure data of the original mobile steam energy storage tank, For all time points In scale The high frequency coefficient components of For all time points In scale The low frequency coefficient components.

[0031] Preferably, a stationary wavelet transform is used to perform a hierarchical frequency analysis on the pressure data inside the mobile steam energy storage tank. The data can be decomposed into high-frequency and low-frequency coefficient components, and the multi-scale features in the data are captured, thereby extracting the changing trends at different frequencies. At the same time, the stationary wavelet transform can effectively capture the periodicity and trend components in the data, and provide a multi-scale feature representation for subsequent attention mechanisms and predictions, which is particularly critical for processing complex data and can interpret the internal pressure changes of the steam energy storage tank from different time scales.

[0032] Preferably, the query, key and value of the high-frequency coefficient component and the low-frequency coefficient component are calculated in steps S3 and S32, and the specific formula is as follows:

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] In the formula, , , , , They are the query, key and value of high-frequency coefficient components and low-frequency coefficient components respectively. for The deformation of The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, for The deformation of The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, , , The weight matrices of query, key and value of high-frequency coefficient components and low-frequency coefficient components are respectively proposed, and then dynamic autoregressive attention is proposed to introduce external context dynamic weight , the specific formula is as follows:

[0040] ;

[0041] ;

[0042] In the formula, is an adjustable coefficient used to control the influence of context information on the weight. is the weighting function between context information, , is the time step and External context information, is the time step The transposition of the external context information, is a trainable weight matrix, and then the dynamic autoregressive attention is calculated to generate the attention score matrix and obtain the dynamic autoregressive attention output. The specific formula is as follows:

[0043] ;

[0044] ;

[0045] ;

[0046] In the formula, and are the attention score matrices of high-frequency coefficient components and low-frequency coefficient components, respectively. , They are respectively the queries of high-frequency coefficient components and low-frequency coefficient components. , are the key transpositions of the high-frequency coefficient components and the low-frequency coefficient components, respectively. is the external context dynamic weight, The query and key dimensions for high-frequency coefficient components and low-frequency coefficient components, To mask the matrix, ensure that each time step Only previous time steps can be seen The attention weights are calculated without letting future time steps affect the calculation of the current time step, and the dynamic autoregressive attention outputs of the high-frequency coefficient components and the low-frequency coefficient components are calculated. The specific formula is as follows:

[0047] ;

[0048] ;

[0049] In the formula, and are the attention score matrices of high-frequency coefficient components and low-frequency coefficient components, respectively. and They are the values ​​of the high-frequency coefficient component and the low-frequency coefficient component respectively.

[0050] Preferably, a dynamic autoregressive attention mechanism is proposed for querying, key and value calculation of high-frequency and low-frequency coefficient components. The mechanism introduces external context dynamic weights, so that the attention calculation can flexibly combine historical information and external information, thereby enhancing the adaptability of the model. The proposed dynamic autoregressive attention mechanism can better capture the dependencies in the time series, and adjust the attention distribution in combination with external dynamic weights, which can more effectively integrate historical data and external context information, improve the ability to capture the relationship between different frequency components, and enhance the robustness and prediction accuracy of the model in complex time series data.

[0051] Preferably, the dynamic autoregressive attention output of the high-frequency coefficient component and the low-frequency coefficient component in steps S3 and S33 is , , perform inverse stationary wavelet transform to merge these components. The specific formula is as follows:

[0052] ;

[0053] ;

[0054] ;

[0055] In the formula, For the Layer all time points In scale The high frequency coefficient components of For the Layer all time points In scale The low-frequency coefficient components of is the dynamic autoregressive attention output of the high-frequency coefficient component, is the dynamic autoregressive attention output of the low-frequency coefficient component, is the inverse stationary wavelet transform operation, is a fully connected layer, which outputs dynamic autoregressive attention to high-frequency coefficient components and low-frequency coefficient components respectively. and Add the high-frequency coefficient components and low-frequency coefficient components of the current layer input. is the layer normalization operation.

[0056] Preferably, the high-frequency and low-frequency coefficient components calculated by the dynamic autoregressive attention mechanism are subjected to an inverse stationary wavelet transform to merge and restore the time-frequency features, ensuring the effective fusion of high-frequency and low-frequency information, and reconstructing the original data features through inverse transform. The inverse wavelet transform can reconstruct the time-frequency features in the original data, thereby providing a more refined and complete input feature representation for the subsequent spatiotemporal feature learning module, so that the model can fully retain the key features of high-frequency and low-frequency information, thereby improving the richness of data representation and prediction effect.

[0057] Preferably, the output of the time-frequency feature capture module is input in steps S4 and S41. ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, and embedding and convolution operations are performed through the embedding layer and the dilated convolution to obtain the feature matrix. The specific formula is as follows:

[0058] ;

[0059] In the formula, is the embedding operation of the embedding layer, is the convolution operation of the dilated convolution, The output of the time-frequency feature capture module The transpose of The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, is the weight matrix of the convolution kernel, is the hole rate, which is used to control the spacing between convolution kernel elements.

[0060] Preferably, the introduction of embedding operations and dilated convolutions can effectively process pressure data from different sensors. Dilated convolutions improve the model's ability to capture long-range dependencies by expanding the receptive field, mapping the pressure data of different sensors to a unified feature space, and further capturing the spatial dependencies between sensors, thereby enhancing the model's comprehensive processing capabilities for multi-sensor data and enabling the model to better learn the spatiotemporal dependencies between different sensors, thereby improving the comprehensive prediction capability of the internal pressure of the energy storage tank.

[0061] Preferably, the step S4, S42 introduces an adjacency matrix based on dynamic time warping, for the mobile steam energy storage tank sensor and sensors Collected internal pressure data of mobile steam energy storage tanks and , and the dynamic time warping distance calculation formula is as follows:

[0062] ;

[0063] In the formula, and Sensors for mobile steam storage tanks and sensors exist The internal pressure data of the mobile steam energy storage tank collected in time steps, is the number of alignment points on the path, It is a dynamic time warping path, which represents the nonlinear mapping between time steps. Then the dynamic time warping distance is used to calculate the adjacency matrix based on dynamic time warping. The specific formula is as follows:

[0064] ;

[0065] In the formula, is the dynamic time warping distance, and Sensors for mobile steam storage tanks and sensors The collected internal pressure data of the mobile steam energy storage tank is used to calculate the weights between the internal pressure data collected by different sensors of the mobile steam energy storage tank through the multi-head self-attention mechanism to obtain the attention output. The specific formula is as follows:

[0066] ;

[0067] In the formula, is the feature matrix, , For the The key and query weight matrix of the attention heads, is the dimension of the attention head, For the The learnable weight matrix in the attention head, is an element-by-element multiplication operation, The attention output of multiple heads is aggregated to obtain the attention aggregate output based on the adjacency matrix of dynamic time warping. .

[0068] An adjacency matrix based on dynamic time warping is introduced to describe the dynamic time dependency between different sensors. Dynamic time warping can handle the complexity of nonlinear mapping in time series. The nonlinear time dependency between different sensors is captured by calculating the dynamic time warping distance. It can effectively handle the time step differences in multi-sensor data, thereby improving the model's ability to model the spatiotemporal dependency of sensor data and significantly enhancing the model's ability to handle nonlinear mapping between data, enabling the model to better capture the complex spatiotemporal relationship between sensors and improving the extraction effect of spatiotemporal features.

[0069] Preferably, in steps S4 and S43, an adjacency matrix based on dynamic time warping is used. Calculate the Laplace matrix. The specific formula is as follows:

[0070] ;

[0071] In the formula, is the identity matrix with dimensions ,in The number of sensors arranged inside the mobile steam energy storage tank, is the degree matrix calculated based on dynamic time warping, where is the square root inverse matrix of the degree matrix, which is used to calculate the adjacency matrix based on dynamic time warping. Perform symmetric normalization, is the adjacency matrix based on dynamic time warping, and then inputs the attention aggregation output , and perform graph convolution operations through Chebyshev polynomial approximation to calculate the feature representation after graph convolution. The specific formula is as follows:

[0072] ;

[0073] In the formula, is the Laplace matrix, are learnable parameters, For the Chebyshev polynomials of order, is an element-by-element multiplication operation, To obtain the internal pressure data of the mobile steam energy storage tank, is the attention aggregation output, is the index of the attention head.

[0074] Preferably, Chebyshev polynomial approximation is combined with graph convolution, and the feature representation of data collected by different sensors is calculated by processing the adjacency matrix. At the same time, a multi-head self-attention mechanism is used to calculate the weight distribution between sensors, so as to more accurately capture the spatiotemporal dependencies between sensors, ensuring that the model can dynamically focus on important sensor information, thereby further improving the expressive power of feature extraction. The combination of graph convolution and self-attention mechanism significantly enhances the model's ability to model complex spatial dependencies, thereby improving the overall prediction performance of the model.

[0075] Preferably, in step S5, the feature representation Input the temporal convolutional network, use dilated convolution and residual connection for further processing, and obtain the refined features. The specific formula is as follows:

[0076] ;

[0077] In the formula, is the activation function, is the time step, is the current time step, is the convolution kernel size, is the initial index for summation, is the weight of the convolution kernel, is the bias term after convolution, For the time step The characteristic representation of is the expansion factor, which controls the expansion of the receptive field. For the time step The feature representation at the time of outputting the prediction result of the internal pressure data of the mobile steam energy storage tank through the fully connected layer is as follows:

[0078] ;

[0079] In the formula, is the fully connected layer, After refining.

[0080] Preferably, a temporal convolutional network is introduced in combination with dilated convolution to process the time series features of sensor data through residual connection. Dilated convolution can capture long-term dependencies by expanding the receptive field while avoiding information loss. It can effectively extract temporal features in sensor data and capture long-term dependencies, ensuring that the model effectively models time series data.

[0081] In summary, due to the adoption of the technical solution, the beneficial effects of the present invention are as follows: the present invention proposes a TSPNet prediction model, which is applied to the prediction scenario of the internal pressure data of a mobile steam energy storage tank, including a time-frequency feature capture module, a spatiotemporal feature learning module and a prediction layer. The time-frequency feature capture module is used to capture the time and frequency characteristics of the input internal pressure data of the mobile steam energy storage tank, the spatiotemporal feature learning module is used to process the spatial dependency of the internal pressure data of the mobile steam energy storage tank, and the prediction layer is used to map the refined features extracted previously to the target output space, thereby generating the final prediction result. Each module has a good effect in capturing the internal pressure data of the mobile steam energy storage tank with complex time and space relationships. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 A step diagram of a method for predicting internal pressure data of a mobile steam energy storage tank.

[0083] Figure 2 This is the structure diagram of the TSPNet prediction model.

[0084] Figure 3 This is the structure diagram of the time-frequency feature capture module.

[0085] Figure 4 This is the structure diagram of the spatiotemporal feature learning module.

[0086] Figure 5 The TSPNet prediction model realizes the prediction fitting effect diagram of the internal pressure data of the mobile steam energy storage tank. DETAILED DESCRIPTION

[0087] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0088] See also Figure 1-Figure 5 The present invention provides a technical solution: a method for predicting the internal pressure data of a mobile steam energy storage tank, which can capture the time and frequency characteristics of the input internal pressure data of the mobile steam energy storage tank by constructing a time-frequency feature capture module, and the spatiotemporal feature learning module can process the spatial dependency of the internal pressure data of the mobile steam energy storage tank. The prediction layer is used to map the extracted refined features to the target output space to complete the accurate prediction of the internal pressure data of the mobile steam energy storage tank in the future. The specific steps are as follows: Figure 1 shown.

[0089] Construct a TSPNet prediction model, whose structure is as follows Figure 2As shown, the specific steps are as follows:

[0090] S1. Collect data related to the internal pressure of the mobile steam energy storage tank, including features and target variables, and preprocess the collected data related to the internal pressure of the mobile steam energy storage tank.

[0091] Furthermore, the collected data related to the internal pressure of the mobile steam energy storage tank include the top layer sensor data inside the energy storage tank, the middle layer sensor data inside the energy storage tank, the bottom layer sensor data inside the energy storage tank, the internal temperature data of the energy storage tank, the external temperature data of the energy storage tank, the total volume data of the energy storage tank, the flow data of the energy storage tank, the internal pressure data of the energy storage tank, and the pressure change rate data of the energy storage tank. The mean method is used to fill the missing values ​​of the internal pressure data of the mobile steam energy storage tank. The specific formula is as follows:

[0092] ;

[0093] In the formula, The mean internal pressure data of the mobile steam energy storage tank to be filled, is the total number of internal pressure data of the mobile steam energy storage tank, The internal pressure data of the mobile steam energy storage tank is data value.

[0094] S2. Use the pre-processed data on the internal pressure of the mobile steam energy storage tank The method performs normalization and divides the data set.

[0095] Further, using The method normalizes the internal pressure data of the mobile steam energy storage tank and divides the training set and the test set in proportion. The specific formula is as follows:

[0096] ;

[0097] In the formula, To obtain the internal pressure data of the mobile steam energy storage tank, is the normalized internal pressure data of the mobile steam energy storage tank, is the maximum value of the internal pressure data of the mobile steam energy storage tank, It is the minimum value in the internal pressure data of the mobile steam energy storage tank.

[0098] S31. Input the internal pressure data of the mobile steam energy storage tank ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, The time step is the frequency analysis of the internal pressure data of the mobile steam energy storage tank by stationary wavelet transform to obtain the high-frequency coefficient components. and low frequency coefficient components .

[0099] Furthermore, a time-frequency feature capture module is constructed, and its structure is as follows: Figure 3 As shown, input the internal pressure data of the mobile steam energy storage tank ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, , is the time step, and the stationary wavelet transform is used to perform hierarchical frequency analysis on the internal pressure data of the mobile steam energy storage tank to obtain the high-frequency coefficient and the low-frequency coefficient. The specific formula is as follows:

[0100] ;

[0101] ;

[0102] In the formula, For scale and time point The high frequency coefficients of For scale and time point The low frequency coefficients of For scale The filter length is is the expanded high frequency filter coefficient, is the extended low-frequency filter coefficient, Internal pressure data of mobile steam storage tank In time The value at which the high frequency coefficients and low frequency coefficients Converted into high-frequency coefficient components and low-frequency coefficient components expressed in matrix form, the specific formula is as follows:

[0103] ;

[0104] ;

[0105] In the formula, For all time points In scale The high frequency coefficient components of For all time points In scale The low-frequency coefficient components of For scale The high frequency filter matrix, For scale The low frequency filter matrix, is the internal pressure data of the original mobile steam energy storage tank, which can be expressed as ,in is the internal pressure data of the original mobile steam energy storage tank, For all time points In scale The high frequency coefficient components of For all time points In scale The low frequency coefficient components.

[0106] S32, calculate the query, key and value of high-frequency coefficient components and low-frequency coefficient components, propose dynamic autoregressive attention, and introduce external context dynamic weights , perform dynamic autoregressive attention calculations to generate attention score matrices for high-frequency coefficient components and low-frequency coefficient components , , and calculate the dynamic autoregressive attention output of the high-frequency coefficient component and the low-frequency coefficient component , .

[0107] Furthermore, the query, key and value of the high-frequency coefficient component and the low-frequency coefficient component are calculated. The specific formula is as follows:

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] In the formula, , , , , They are the query, key and value of high-frequency coefficient components and low-frequency coefficient components respectively. for The deformation of The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, for The deformation of The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, , , The weight matrices of query, key and value of high-frequency coefficient components and low-frequency coefficient components are respectively proposed, and then dynamic autoregressive attention is proposed to introduce external context dynamic weight , the specific formula is as follows:

[0115] ;

[0116] ;

[0117] In the formula, is an adjustable coefficient used to control the influence of context information on the weight. is the weighting function between context information, , is the time step and External context information, is the time step The transposition of the external context information, is a trainable weight matrix, and then the dynamic autoregressive attention is calculated to generate the attention score matrix and obtain the dynamic autoregressive attention output. The specific formula is as follows:

[0118] ;

[0119] ;

[0120] ;

[0121] In the formula, and are the attention score matrices of high-frequency coefficient components and low-frequency coefficient components, respectively. , They are respectively the queries of high-frequency coefficient components and low-frequency coefficient components. , are the key transpositions of the high-frequency coefficient components and the low-frequency coefficient components, respectively. is the external context dynamic weight, The query and key dimensions for high-frequency coefficient components and low-frequency coefficient components, To mask the matrix, ensure that each time step Only previous time steps can be seen The attention weights are calculated without letting future time steps affect the calculation of the current time step, and the dynamic autoregressive attention outputs of the high-frequency coefficient components and the low-frequency coefficient components are calculated. The specific formula is as follows:

[0122] ;

[0123] ;

[0124] In the formula, and are the attention score matrices of high-frequency coefficient components and low-frequency coefficient components, respectively. and They are the values ​​of the high-frequency coefficient component and the low-frequency coefficient component respectively.

[0125] S33, perform inverse stationary wavelet transform on the dynamic autoregressive attention output to obtain the output of the time-frequency feature capture module .

[0126] Furthermore, the dynamic autoregressive attention output of the high-frequency coefficient components and the low-frequency coefficient components , , perform inverse stationary wavelet transform to merge these components. The specific formula is as follows:

[0127] ;

[0128] ;

[0129] ;

[0130] In the formula, For the Layer all time points In scale The high frequency coefficient components of For the Layer all time points In scale The low-frequency coefficient components of is the dynamic autoregressive attention output of the high-frequency coefficient component, is the dynamic autoregressive attention output of the low-frequency coefficient component, is the inverse stationary wavelet transform operation, is a fully connected layer, which outputs dynamic autoregressive attention to high-frequency coefficient components and low-frequency coefficient components respectively. and Add the high-frequency coefficient components and low-frequency coefficient components of the current layer input. is the layer normalization operation.

[0131] S41, input the output of the time-frequency feature capture module ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, and through embedding and convolution operations, the feature matrix is ​​obtained .

[0132] Furthermore, we construct a spatiotemporal feature learning module, whose structure is as follows: Figure 4 As shown, the output of the input time-frequency feature capture module ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, and embedding and convolution operations are performed through the embedding layer and the dilated convolution to obtain the feature matrix. The specific formula is as follows:

[0133] ;

[0134] In the formula, is the embedding operation of the embedding layer, is the convolution operation of the dilated convolution, The output of the time-frequency feature capture module The transpose of The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, is the weight matrix of the convolution kernel, is the hole rate, which is used to control the spacing between convolution kernel elements.

[0135] S42, Introducing the adjacency matrix based on dynamic time warping The multi-head self-attention mechanism is used to calculate the weights between the internal pressure data collected by different sensors of the mobile steam energy storage tank to obtain the attention output , aggregate the attention outputs of multiple heads to get the attention aggregation output .

[0136] Furthermore, the adjacency matrix based on dynamic time warping is introduced for the mobile steam energy storage tank sensor and sensors Collected internal pressure data of mobile steam energy storage tanks and , and the dynamic time warping distance calculation formula is as follows:

[0137] ;

[0138] In the formula, and Sensors for mobile steam storage tanks and sensors exist The internal pressure data of the mobile steam energy storage tank collected in time steps, is the number of alignment points on the path, It is a dynamic time warping path, which represents the nonlinear mapping between time steps. Then the dynamic time warping distance is used to calculate the adjacency matrix based on dynamic time warping. The specific formula is as follows:

[0139] ;

[0140] In the formula, is the dynamic time warping distance, and Sensors for mobile steam storage tanks and sensors The collected internal pressure data of the mobile steam energy storage tank is used to calculate the weights between the internal pressure data collected by different sensors of the mobile steam energy storage tank through the multi-head self-attention mechanism to obtain the attention output. The specific formula is as follows:

[0141] ;

[0142] In the formula, is the feature matrix, , For the The key and query weight matrix of the attention heads, is the dimension of the attention head, For the The learnable weight matrix in the attention head, is an element-by-element multiplication operation, The attention output of multiple heads is aggregated to obtain the attention aggregate output based on the adjacency matrix of dynamic time warping. .

[0143] S43, using adjacency matrix based on dynamic time warping Compute the Laplacian matrix , input attention aggregation output , and perform graph convolution operations through Chebyshev polynomial approximation to calculate the feature representation after graph convolution .

[0144] Furthermore, using the adjacency matrix based on dynamic time warping Calculate the Laplace matrix. The specific formula is as follows:

[0145] ;

[0146] In the formula, is the identity matrix with dimensions ,in The number of sensors arranged inside the mobile steam energy storage tank, is the degree matrix calculated based on dynamic time warping, where is the square root inverse matrix of the degree matrix, which is used to calculate the adjacency matrix based on dynamic time warping. Perform symmetric normalization, is the adjacency matrix based on dynamic time warping, and then inputs the attention aggregation output , and perform graph convolution operations through Chebyshev polynomial approximation to calculate the feature representation after graph convolution. The specific formula is as follows:

[0147] ;

[0148] In the formula, is the Laplace matrix, are learnable parameters, For the Chebyshev polynomials of order, is an element-by-element multiplication operation, To obtain the internal pressure data of the mobile steam energy storage tank, is the attention aggregation output, is the index of the attention head.

[0149] S5. Build the prediction layer and represent the features The input time series convolutional network is further processed to obtain refined features , and finally output the prediction results of the internal pressure data of the mobile steam energy storage tank through the fully connected layer .

[0150] Furthermore, the feature representation Input the temporal convolutional network, use dilated convolution and residual connection for further processing, and obtain the refined features. The specific formula is as follows:

[0151] ;

[0152] In the formula, is the activation function, is the time step, is the current time step, is the convolution kernel size, is the initial index for summation, is the weight of the convolution kernel, is the bias term after convolution, For the time step The characteristic representation of is the expansion factor, which controls the expansion of the receptive field. For the time step The feature representation at the time of outputting the prediction result of the internal pressure data of the mobile steam energy storage tank through the fully connected layer is as follows:

[0153] ;

[0154] In the formula, is the fully connected layer, After refining.

[0155] Furthermore, the TSPNet prediction model is written in Python. The experiment runs on the Windows operating system. Pytorch is selected as the framework in the CUDA11.27 environment. The training is performed on the Google Colab Pro+ GPU. The data set is the internal pressure data of the mobile steam energy storage tank collected by the sensor for one month. After preprocessing, it is input into the TSPNet prediction model. The training batch is set to 32 and the training is performed using As the optimizer, the learning rate is set to 0.001.

[0156] Furthermore, the TSPNet prediction model realizes the prediction and fitting effect of the internal pressure data of the mobile steam energy storage tank. Figure 5 As shown in the figure, the horizontal axis is time (days), the vertical axis is the pressure inside the energy storage tank (unit: MPa), the black solid line and dots represent the actual data, and the gray dotted line and cross represent the predicted data of the model. It can be seen from the figure that the fit between the predicted value and the actual value is high, and the overall trend of the predicted data is very consistent with the actual data, especially in the early stage with a time span of 1 to 5 days, the predicted value of the model almost completely coincides with the actual value. At this stage, the model can accurately capture the rapid upward trend of pressure, indicating that it is stable when the early data fluctuates greatly. In the subsequent stage from the 6th to the 10th day, although there are some fluctuations and slight deviations between the actual data and the predicted data, the overall trend is still consistent, indicating that the model has a strong adaptability in dealing with fluctuating data. Secondly, throughout the prediction process, the predicted value of the model basically keeps close to the actual data, indicating that the model has good stability and robustness, and can maintain accurate prediction of long-term trends in the face of short-term fluctuations in the data. From the experimental results, it can be seen that the model performs well in predicting the internal pressure of mobile steam energy storage tanks.

Claims

1. A method for predicting internal pressure data of a mobile steam energy storage tank, characterized in that: The following steps are involved: S1. Collecting data related to the internal pressure of the mobile steam energy storage tank, including features and target variables, and preprocessing the collected data related to the internal pressure of the mobile steam energy storage tank; S2. Use the pre-processed data on the internal pressure of the mobile steam energy storage tank The method performs normalization operations and divides the data set; S3, constructing a time-frequency feature capture module for capturing the time and frequency features of the input internal pressure data of the mobile steam energy storage tank, specifically comprising the following steps: S31. Input the internal pressure data of the mobile steam energy storage tank ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, The time step is the frequency analysis of the internal pressure data of the mobile steam energy storage tank by stationary wavelet transform to obtain the high-frequency coefficient components. and low frequency coefficient components ; S32, calculate the query, key and value of high-frequency coefficient components and low-frequency coefficient components, propose dynamic autoregressive attention, and introduce external context dynamic weights , perform dynamic autoregressive attention calculations to generate attention score matrices for high-frequency coefficient components and low-frequency coefficient components , , and calculate the dynamic autoregressive attention output of the high-frequency coefficient component and the low-frequency coefficient component , ; S33, perform inverse stationary wavelet transform on the dynamic autoregressive attention output to obtain the output of the time-frequency feature capture module ; S4. Constructing a spatiotemporal feature learning module for processing the spatial dependency of the internal pressure data of the mobile steam energy storage tank, specifically comprising the following steps: S41, input the output of the time-frequency feature capture module ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, and through embedding and convolution operations, the feature matrix is ​​obtained ; S42, Introducing the adjacency matrix based on dynamic time warping The multi-head self-attention mechanism is used to calculate the weights between the internal pressure data collected by different sensors of the mobile steam energy storage tank to obtain the attention output , aggregate the attention outputs of multiple heads to get the attention aggregation output ; S43, using adjacency matrix based on dynamic time warping Compute the Laplacian matrix , input attention aggregation output , and perform graph convolution operations through Chebyshev polynomial approximation to calculate the feature representation after graph convolution ; S5. Build the prediction layer and represent the features The input time series convolutional network is further processed to obtain refined features , and finally output the prediction results of the internal pressure data of the mobile steam energy storage tank through the fully connected layer .

2. A method for predicting internal pressure data of a mobile steam energy storage tank according to claim 1, characterized in that: In step S31, the internal pressure data of the mobile steam energy storage tank is input. ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, , is the time step, and the stationary wavelet transform is used to perform hierarchical frequency analysis on the internal pressure data of the mobile steam energy storage tank to obtain the high-frequency coefficient and the low-frequency coefficient. The specific formula is as follows: ； ； In the formula, For scale and time point The high frequency coefficients of For scale and time point The low frequency coefficients of For scale The filter length is is the expanded high frequency filter coefficient, is the extended low-frequency filter coefficient, Internal pressure data of mobile steam storage tank In time The value at which the high frequency coefficients and low frequency coefficients Converted into high-frequency coefficient components and low-frequency coefficient components expressed in matrix form, the specific formula is as follows: ； ； In the formula, For all time points In scale The high frequency coefficient components of For all time points In scale The low-frequency coefficient components of For scale The high frequency filter matrix, For scale The low frequency filter matrix, is the internal pressure data of the original mobile steam energy storage tank, which can be expressed as ,in is the internal pressure data of the original mobile steam energy storage tank, For all time points In scale The high frequency coefficient components of For all time points In scale The low frequency coefficient components.

3. A method for predicting internal pressure data of a mobile steam energy storage tank according to claim 2, characterized in that: The query, key and value of the high-frequency coefficient component and the low-frequency coefficient component are calculated in step S32, and the specific formula is as follows: ； ； ； ； ； ； In the formula, , , , , They are the query, key and value of high-frequency coefficient components and low-frequency coefficient components respectively. for The deformation of The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, for The deformation, , , The weight matrices of query, key and value of high-frequency coefficient components and low-frequency coefficient components are respectively proposed, and then dynamic autoregressive attention is proposed to introduce external context dynamic weight , the specific formula is as follows: ； ； In the formula, is an adjustable coefficient used to control the influence of context information on the weight. is the weighting function between context information, , is the time step and External context information, is the time step The transposition of the external context information, is a trainable weight matrix, and then the dynamic autoregressive attention is calculated to generate the attention score matrix and obtain the dynamic autoregressive attention output. The specific formula is as follows: ； ； ； In the formula, and are the attention score matrices of high-frequency coefficient components and low-frequency coefficient components, respectively. , They are respectively the queries of high-frequency coefficient components and low-frequency coefficient components. , are the key transpositions of the high-frequency coefficient components and the low-frequency coefficient components, respectively. is the external context dynamic weight, The query and key dimensions for high-frequency coefficient components and low-frequency coefficient components, To mask the matrix, ensure that each time step Only previous time steps can be seen The attention weights are calculated without letting future time steps affect the calculation of the current time step, and the dynamic autoregressive attention outputs of the high-frequency coefficient components and the low-frequency coefficient components are calculated. The specific formula is as follows: ； ； In the formula, and are the attention score matrices of high-frequency coefficient components and low-frequency coefficient components, respectively. and They are the values ​​of the high-frequency coefficient component and the low-frequency coefficient component respectively.

4. A method for predicting internal pressure data of a mobile steam energy storage tank according to claim 3, characterized in that: The dynamic autoregressive attention output of the high-frequency coefficient component and the low-frequency coefficient component in step S33 , , perform inverse stationary wavelet transform to merge these components. The specific formula is as follows: ； ； ； In the formula, For the Layer all time points In scale The high frequency coefficient components of For the Layer all time points In scale The low-frequency coefficient components of is the dynamic autoregressive attention output of the high-frequency coefficient component, is the dynamic autoregressive attention output of the low-frequency coefficient component, is the inverse stationary wavelet transform operation, is a fully connected layer, which outputs dynamic autoregressive attention to high-frequency coefficient components and low-frequency coefficient components respectively. and Add the high-frequency coefficient components and low-frequency coefficient components of the current layer input. is the layer normalization operation.

5. A method for predicting internal pressure data of a mobile steam energy storage tank according to claim 4, characterized in that: The output of the time-frequency feature capture module is input in step S41 ,in The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, and embedding and convolution operations are performed through the embedding layer and the dilated convolution to obtain the feature matrix. The specific formula is as follows: ； In the formula, is the embedding operation of the embedding layer, is the convolution operation of the dilated convolution, The output of the time-frequency feature capture module The transpose of The number of sensors arranged inside the mobile steam energy storage tank, For the The number of features in the layer, is the time step, is the weight matrix of the convolution kernel, is the hole rate, which is used to control the spacing between convolution kernel elements.

6. A method for predicting internal pressure data of a mobile steam energy storage tank according to claim 5, characterized in that: The step S42 introduces an adjacency matrix based on dynamic time warping, for the mobile steam energy storage tank sensor and sensors Collected internal pressure data of mobile steam energy storage tanks and , and the dynamic time warping distance calculation formula is as follows: ； In the formula, and Sensors for mobile steam storage tanks and sensors exist The internal pressure data of the mobile steam energy storage tank collected in time steps, is the number of alignment points on the path, It is a dynamic time warping path, which represents the nonlinear mapping between time steps. Then the dynamic time warping distance is used to calculate the adjacency matrix based on dynamic time warping. The specific formula is as follows: ； In the formula, is the dynamic time warping distance, and Sensors for mobile steam storage tanks and sensors The collected internal pressure data of the mobile steam energy storage tank is used to calculate the weights between the internal pressure data collected by different sensors of the mobile steam energy storage tank through the multi-head self-attention mechanism to obtain the attention output. The specific formula is as follows: ； In the formula, is the feature matrix, , For the The key and query weight matrix of the attention heads, is the dimension of the attention head, For the The learnable weight matrix in the attention head, is an element-by-element multiplication operation, The attention output of multiple heads is aggregated to obtain the attention aggregate output based on the adjacency matrix of dynamic time warping. .

7. A method for predicting internal pressure data of a mobile steam energy storage tank according to claim 6, characterized in that: In step S43, the adjacency matrix based on dynamic time warping is used. Calculate the Laplace matrix. The specific formula is as follows: ； In the formula, is the identity matrix with dimensions ,in The number of sensors arranged inside the mobile steam energy storage tank, is the degree matrix calculated based on dynamic time warping, where is the square root inverse matrix of the degree matrix, which is used to calculate the adjacency matrix based on dynamic time warping. Perform symmetric normalization, is the adjacency matrix based on dynamic time warping, and then inputs the attention aggregation output , and perform graph convolution operations through Chebyshev polynomial approximation to calculate the feature representation after graph convolution. The specific formula is as follows: ； In the formula, is the Laplace matrix, are learnable parameters, For the Chebyshev polynomials of order, is an element-by-element multiplication operation, To obtain the internal pressure data of the mobile steam energy storage tank, is the attention aggregation output, is the index of the attention head.

8. A method for predicting internal pressure data of a mobile steam energy storage tank according to claim 7, characterized in that: In step S5, the feature representation Input the temporal convolutional network, use dilated convolution and residual connection for further processing, and obtain the refined features. The specific formula is as follows: ； In the formula, is the activation function, is the time step, is the current time step, is the convolution kernel size, is the initial index for summation, is the weight of the convolution kernel, is the bias term after convolution, For the time step The characteristic representation of is the expansion factor, which controls the expansion of the receptive field. For the time step The feature representation at the time of outputting the prediction result of the internal pressure data of the mobile steam energy storage tank through the fully connected layer is as follows: ； In the formula, is the fully connected layer, After refining.

9. A method for predicting internal pressure data of a mobile steam energy storage tank according to claim 1, characterized in that: In order to predict the internal pressure data of mobile steam energy storage tanks, the collected data related to the internal pressure of mobile steam energy storage tanks include the top-layer sensor data inside the energy storage tank, the middle-layer sensor data inside the energy storage tank, the bottom-layer sensor data inside the energy storage tank, the internal temperature data of the energy storage tank, the external temperature data of the energy storage tank, the total volume data of the energy storage tank, the energy storage tank flow data, the internal pressure data of the energy storage tank, and the energy storage tank pressure change rate data. The collected relevant data are preprocessed to ensure that there are no missing values ​​in the data, and then the processed data are divided into training sets and test sets for training and evaluating the performance of the mobile steam energy storage tank internal pressure data prediction model.

Images