Construction Method of Natural Gas Pipeline Network Energy Transmission Difference Calculation System Integrated with Machine Learning

Abstract

The present invention proposes a construction method for a natural gas pipeline network energy transmission difference calculation system integrating machine learning. This energy transmission difference calculation system includes a source database unit, a prediction unit, and an optimization unit; the source database unit includes a sensor data source module and other data modules; the prediction unit includes a sufficient sample prediction module and a small sample prediction module; the optimization unit includes an actual measurement data module and a Bayesian parameter update module. The main steps are as follows: determining the influencing factors of the energy transmission difference to construct a basic database; performing clustering processing to establish a machine learning regression prediction model for the energy transmission difference of the natural gas pipeline network; inputting the real-time sensor data of the required pipeline section to be predicted in real time for energy transmission difference prediction; constructing an optimization database and updating the weights of the input parameters in the machine learning model. The present invention avoids the serious dependence on artificial experience in the traditional method, has a simple calculation process and short time consumption; solves the small sample problem in the field of pipeline transportation and improves the generalization performance of the model.

Description

Technical Field

[0001] The present invention relates to the field of natural gas transportation, and particularly to a construction method of a natural gas pipeline network energy transmission difference calculation system integrating machine learning. Background Art

[0002] How to ensure the full coverage of natural gas resources in the urban areas of our country is an important factor restricting the development of the natural gas industry. Among them, due to the significant advantages of pipeline transportation in terms of safety, transportation cost, construction period, and transportation volume, etc., it has become one of the main long-distance transportation methods for energy such as natural gas. However, affected by many uncertain factors such as metering devices, personnel measurement errors, inherent losses of aging pipelines, gas quality, gas flow state, temperature, etc., the long-distance pipeline transportation of natural gas will inevitably have the problem of transmission difference. This has a significant impact on the accurate calculation of the natural gas demand side, the formulation of the pipeline network gas supply plan, and the economic benefits of natural gas companies.

[0003] At present, the internationally common metering method for natural gas is the energy measurement unit. During the selected metering period, the energy transmission difference E of natural gas can be calculated by formula (1). The commonly used calculation methods for the energy transmission difference of natural gas pipeline networks mainly include: traditional SCADA systems, manual calculation, and dispatching experience models. However, due to the diversity and uncertainty of the above-mentioned influencing factors, the above energy transmission difference calculation methods often have problems such as low calculation efficiency and low cost performance.

[0004] E=(H C ) G (t 1 )Q (1)

[0005] Q=(V 1 +Q 1 )-(Q 2 +Q 3 +Q 4 +V 2 ) (2)

[0006] Where: E is the energy transmission difference, (H C ) G (t 1 ) is the actual high calorific value per mole of the gaseous fuel, kJ / kmol, Q is the difference in the balanced gas transmission volume in the gas transmission pipeline at a certain time, m 3 ; Q 1 is the input gas volume value at the same time, m 3 ; Q 2 is the output gas volume value at the same time, m 3 ; Q 3 is the production and domestic gas consumption of the gas transmission unit at the same time, m 3 ; Q 4The air release volume within the same time, m 3 ; V 1 is the gas storage volume within the calculated section of the pipeline at the start of the calculation time, m 3 ; V 2 is the gas storage volume within the calculated section of the pipeline at the end of the calculation time, m 3 .

[0007] For example, the invention patent with the application number CN202010958923.3 discloses a visual natural gas pipeline transmission difference optimization solution system that comprehensively considers the actual pipeline inventory and transmission difference problems, significantly reducing the calculation error of energy transmission difference. However, due to solving a system of multivariate differential equations, the solving process is cumbersome and time-consuming. At the same time, it has poor applicability for calculating energy transmission differences in different sections or different service environments.

[0008] Data-driven models with high-dimensional non-linear modeling capabilities and flexible data processing capabilities have been widely used in various actual engineering regression prediction and classification problems in recent years. In the field of natural gas transportation, the rapid iteration of advanced sensing devices has provided a large amount of accurate data, and the emergence of supercomputers has solved the problem of insufficient computing power. Based on this, artificial intelligence calculation methods provide a feasible solution for the rapid and accurate prediction of energy transmission differences in natural gas pipeline transportation.

[0009] Therefore, based on the monitoring data of existing sensor devices, there is an urgent need to develop a calculation method for energy transmission differences in natural gas pipelines with a wide application range, short time consumption, and high accuracy. Summary of the Invention

[0010] The present invention proposes a construction method for a natural gas pipeline energy transmission difference calculation system integrating machine learning, which can realize the intelligent regression prediction of energy transmission differences in any section and any working condition during the natural gas pipeline transportation process.

[0011] The technical solution of the present invention is realized as follows: A construction method for a natural gas pipeline energy transmission difference calculation system integrating machine learning, the energy transmission difference calculation system includes three units: a source database unit, a prediction unit, and an optimization unit;

[0012] The source database unit includes a sensor data source module and other data modules;

[0013] The prediction unit includes a sufficient sample prediction module and a small sample prediction module;

[0014] The optimization unit includes an actual measurement data module and a Bayesian parameter update module;

[0015] The construction process of the natural gas pipeline energy transmission difference calculation system integrating machine learning mainly includes the following steps:

[0016] Step 1: Determine the influencing factors of energy transfer difference. The source database unit collects the historical data of energy transfer difference in natural gas pipelines in various engineering cases and constructs a basic database.

[0017] Step 2: The prediction unit first performs clustering on the basic data, dividing it into small-sample and sufficient-sample sub-datasets. Then, the training set and test set are respectively divided for the two sub-datasets, and a machine learning regression prediction model for the energy transfer difference in the natural gas pipeline is established using the training set.

[0018] Step 3: Real-time input the real-time sensor data and other working data of the pipe section to be predicted, and use the trained model in Step 2 to predict the energy transfer difference.

[0019] Step 4: The measured data module in the optimization unit collects the actual data during each energy transfer difference detection. Based on the error between the prediction result and the actual data, the real sample data with an error of less than 10% is retained and an optimization database is constructed. Define the weights of each particle in the particle filter based on the comparison between the samples in the optimization database and the prediction results, and update the weights of the input parameters in the machine learning model to further reduce the error rate.

[0020] Preferably, the influencing factors in Step 1 include the internal air pressure, pressure drop, gas flow rate, gas temperature, actual high calorific value, initial flow rate, terminal flow rate, pipe section length, and external air pressure and temperature in the natural gas pipeline section. The above variables in the historical cases are constructed into a feature set X = {x11, x12, x13,..., xij}, where xij represents the jth feature of the ith sample number.

[0021] The energy transfer difference in the historical data collected in Step 1 is used as the label set Y = {Y1, Y2, Y3,..., Yi} of the basic data set.

[0022] Preferably, in Step 2, the basic data is divided into small-sample and sufficient-sample data sub-datasets using the K-means clustering method.

[0023] The machine learning models in Step 2 include a sufficient-sample machine learning prediction model and a small-sample machine learning prediction model. Among them, the sufficient-sample machine learning model uses the neural network method, and the small-sample machine learning model uses the transfer learning method.

[0024] Preferably, the sufficient samples use the above-mentioned feature set as the input and are given an initial weight factor, and a corresponding neural network model is established using the training set.

[0025] The transfer learning method is used to train the small-sample training set. The above transfer learning selects the feature transfer method, which aligns the data distributions in different domains based on feature mapping.

[0026] Preferably, in the transfer learning, the physical field of the migrated-out data is used as the source domain M, and the physical field of the migrated-in data is used as the target domain N. The feature sets and label sets in the source domain and the target domain are respectively defined as Xm, Ym, Xn, and Yn. On the basis of classical feature transfer, the transfer learning model optimizes the combined clustering hyperparameters derived from data fusion.

[0027] Preferably, the key point of the feature transfer is to construct a mapping φ, so that the source domain and the target domain are close to the same conditional distribution. In the present invention, the maximum mean discrepancy, abbreviated as MMD, is selected to evaluate the distribution distance between the source domain and the target domain. Its calculation expression is:

[0028]

[0029] where: n 1 、n 2 are the numbers of data in the source domain and the target domain respectively, x mi ∈X m , x nj ∈X n ;

[0030] After the feature mapping, a clustering method is introduced to construct an optimization index, and the regularization parameter is further optimized by combining with the random search method.

[0031] Preferably, when optimizing the hyperparameters of the feature mapping, the K-means clustering method is selected, which is specifically used for the feature matrix after the feature transfer mapping. At this time, a clustering center will be generated for the source domain and the target domain respectively, and the distance between these two clustering centers will represent the distribution difference between the source domain features and the target domain features;

[0032] The distance between the two clustering centers can be defined by the following formula:

[0033] Distance(K) = ||c m -c n || 2 (4)

[0034] where, c m and c n are the unique clustering centers of the source domain and target domain features after the feature mapping respectively.

[0035] Preferably, based on the introduction of the above K clustering, the optimization problem of the hyperparameters is transformed into the minimization problem of Distance(K), and the regularization parameter is optimized within a given range by using the random search method.

[0036] Preferably, when predicting the energy transfer difference of unknown pipeline segments in step three, it is necessary to first analyze the similarity between the input features to be predicted and the features in the basic dataset to determine whether the problem to be predicted belongs to a small-sample or sufficient-sample problem; after determining the classification, input the feature set of the sample to be predicted, and the current energy transfer difference prediction value can be obtained using the corresponding machine learning model.

[0037] Preferably, regularly sample and inspect some pipeline segments, and calculate the actual energy transfer difference based on the measured signals and other data; that is, new samples are obtained; at the same time, within the sampled pipeline segments, samples with a relative error less than 10% are retained, and samples with an error greater than 10% are excluded.

[0038] The sampled new samples can be used to define the particle filter parameters, thereby further optimizing the weight factors of the feature engineering in the machine learning model.

[0039] Compared with the prior art, the advantages of the present invention are as follows:

[0040] (1) The present invention avoids the serious dependence on manual experience in traditional methods; at the same time, compared with various analytical calculations, the calculation process of the present invention is simple and time-consuming.

[0041] (2) The introduction of transfer learning enables the present invention to well solve the small-sample problem in the field of pipeline transportation, improves the generalization performance of the model, and makes the method applicable to a wider range.

[0042] (3) The establishment of the optimization system can further reverse-optimize the machine learning regression prediction model established based on existing data, can update the model parameters in real time, and overcomes the problem of low prediction accuracy of the machine learning model under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the structure of the natural gas pipeline energy transfer difference calculation system integrating machine learning;

[0044] Figure 2 Flowchart of the application of the natural gas pipeline energy transfer difference calculation system integrating machine learning. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0046] The present invention provides a natural gas pipeline network energy transfer difference calculation system integrating machine learning.

[0047] The calculation of the energy transfer difference includes three units: a source database unit, a prediction unit, and an optimization unit. Among them, the database unit includes a sensor data source module and other data modules; the prediction unit includes a sufficient sample prediction module and a small sample prediction module; the optimization unit includes an actual measurement data module and a Bayesian parameter update module.

[0048] Furthermore, the present invention provides a construction process for a natural gas pipeline network energy transfer difference calculation system integrating machine learning, mainly including the following steps:

[0049] Step 1: Determine the influencing factors of the energy transfer difference. The sensor data source module and other data modules of the source database unit collect historical data on the energy transfer difference of natural gas pipeline networks in various engineering cases to construct a basic database.

[0050] Step 2: First, perform clustering processing on the basic data, dividing it into a small sample and a sufficient sample sub-dataset. Then, divide the training set and the test set for the two subsets respectively, and use the training set to establish a machine learning regression prediction model for the energy transfer difference of the natural gas pipeline network.

[0051] Step 3: Real-time input the real-time sensor data and other working data of the pipeline section to be predicted, and use the model trained in Step 2 to predict the energy transfer difference.

[0052] Step 4: Collect the actual data during each energy transfer difference detection. Based on the error between the prediction result, retain the real sample data with an error of less than 10% and construct an optimization database. Define the weights of each particle in the particle filter based on the comparison between the samples in the optimization database and the prediction results, and update the weights of the input parameters in the machine learning model to further reduce the error rate.

[0053] Preferably, the influencing factors in Step 1 include the internal air pressure, pressure drop, gas flow rate, gas temperature, actual high - order molar calorific value, initial flow rate, terminal flow rate, pipe section length, and external air pressure, temperature, etc. in the natural gas pipeline section. Construct the above variables in the historical cases into a feature set X = {x11, x12, x13,..., xij}, where xij represents the jth feature of the ith sample number.

[0054] The energy transfer difference in the historical data collected in Step 1 is used as the label set Y = {Y1, Y2, Y3,..., Yi} of the basic data set.

[0055] Preferably, in Step 2, the basic data is divided into a small sample and a sufficient sample data subset using the K - means clustering method.

[0056] The machine learning model in Step 2 includes a sufficient-sample machine learning prediction model (corresponding to the sufficient-sample prediction module) and a small-sample machine learning prediction model (corresponding to the small-sample prediction module). Among them, the sufficient-sample machine learning model uses the neural network method, and the small-sample machine learning model uses the transfer learning method.

[0057] For the sufficient-sample problem, use the feature set described in Step 1 as the input, assign an initial weight factor, and use the training set to establish a corresponding neural network model.

[0058] For the small-sample data problem, the prediction accuracy of traditional machine learning methods is often low. However, transfer learning can achieve the mutual transfer and utilization of data in different domains, and can effectively solve the small-sample problem. In addition, appropriate transfer criteria enable transfer learning to reduce the difference in data distributions in different domains and can solve the multi-distribution problem of data. Therefore, the present invention selects the transfer learning method to train the small-sample training set.

[0059] The above transfer learning selects the feature transfer method, which is based on feature mapping, can align the data distributions in different domains, and has good applicability to the small-sample problem.

[0060] In the said transfer learning, the physical field of the migrated data is the source domain M, and the physical field of the migrated-in data is the target domain N. The feature sets and label sets in the source domain and the target domain are respectively defined as Xm, Ym, Xn, and Yn.

[0061] Preferably, the above transfer learning model further optimizes the combined clustering hyperparameters by fusing the source data on the basis of classical feature transfer.

[0062] The key point of feature transfer is to construct the mapping φ, so that the source domain and the target domain are close to the same conditional distribution. In the present invention, the Maximum Mean Discrepancy (MMD) is selected to evaluate the distribution distance between the source domain and the target domain, and its calculation expression is:

[0063]

[0064] Where: n 1 、n 2 are the numbers of data in the source domain and the target domain respectively,

[0065] Based on the above formula, the construction of the feature transfer mapping can be determined by minimizing the maximum mean discrepancy. Similar to other machine learning methods, by introducing a kernel function and a regularization term, the minimization problem of the maximum mean discrepancy can be transformed into an optimization problem.

[0066] Preferably, in order to achieve a better feature migration effect, it is necessary to optimize the parameters of the regularization term. Considering the differences in data distribution, the present invention also introduces a clustering method after feature mapping to construct an optimization index, and further combines the random search method to optimize the regularization parameters.

[0067] When optimizing the hyperparameters of feature mapping, the K-means clustering method is selected, which is specifically used for the feature matrix after feature migration mapping. At this time, a clustering center will be generated for the source domain and the target domain respectively, and the distance between these two clustering centers will represent the distribution difference between the source domain features and the target domain features.

[0068] The distance between the two clustering centers in the previous step can be defined by the following formula:

[0069] Distance(K)=||c m -c n || 2 (4)

[0070] Where c m and c n are the unique clustering centers of the source domain and target domain features after feature mapping respectively.

[0071] Based on the introduction of the above K-clustering, the optimization problem of hyperparameters is transformed into the minimization problem of Distance(K). The random search method can be used to optimize the regularization parameters within a given range.

[0072] Preferably, when predicting the energy transfer difference of the unknown pipe section in step three, it is necessary to first judge whether the problem to be predicted belongs to a small sample or a sufficient sample problem according to the similarity analysis between the input features to be predicted and the features in the basic dataset.

[0073] After determining the classification, input the feature set of the sample to be predicted, and the current energy transfer difference prediction value can be obtained by using the corresponding machine learning model.

[0074] Preferably, due to the importance of the energy transfer difference in natural gas pipelines, it is inevitable to regularly sample some pipe sections and calculate the actual energy transfer difference based on the measured signals and other data. That is, new samples are obtained. At the same time, for the sampled pipe sections, there will inevitably be a certain error between the predicted value and the true value. Samples with a relative error less than 10% are retained, and samples with an error greater than 10% are excluded.

[0075] Furthermore, the sampled new samples can be used to define the particle filter parameters, thereby further optimizing the weight factors of the feature engineering in the machine learning model.

[0076] Preferably, repeating the above steps one to four can obtain the energy transfer difference of any section of the natural gas pipeline.

[0077] Compared with the existing energy transfer difference calculation method, the beneficial effects of the present invention are:

[0078] (1) The present invention avoids the heavy reliance of traditional methods on manual experience. At the same time, compared with various analytical calculations, the present invention has a simple calculation process and takes less time.

[0079] (2) The introduction of transfer learning enables the present invention to effectively solve the small sample problem in the field of pipeline transportation and improve the generalization performance of the model, making the method more applicable.

[0080] (3) The establishment of an optimization system can further reversely optimize the machine learning regression prediction model established based on existing data, and can update the model parameters in real time, thus overcoming the problem of low prediction accuracy of the machine learning model under complex working conditions.

[0081] The concept and application process of the present invention are further described in detail below in conjunction with the embodiments of the drawings, so that those skilled in the art can implement them according to the description.

[0082] like Figure 1 As shown, the natural gas pipeline energy transmission difference calculation system integrating machine learning includes a source database unit, a prediction unit and an optimization unit.

[0083] like Figure 2 The following is an implementation process of a natural gas pipeline energy transmission difference calculation system integrating machine learning. Figure 1 , Figure 2 The specific implementation process of the system includes the following steps:

[0084] S1: Determine the factors that affect the energy transmission difference based on manual experience and existing research. Other data modules of the source database unit collect historical structured and unstructured data sets containing energy transmission difference from various public engineering cases, published literature, natural gas company databases, etc. Take each influencing factor as a feature set and the energy transmission difference result as a label value to build a natural gas pipeline network energy transmission difference source database.

[0085] Preprocess the source database data to remove missing and abnormal data. At the same time, due to the large number of influencing factors involved, the distribution ranges of different types of data vary greatly, and even have magnitude differences. Therefore, the source data set after screening is standardized and normalized.

[0086] The K-means clustering method is used to divide the source data set into small sample data sets and sufficient sample data sets.

[0087] For sufficient sample data sets and small sample data sets, the normalized features are constructed into feature matrices to facilitate the subsequent training of machine learning models.

[0088] S2: Based on the clustering division of the dataset in S1, machine learning model training is carried out for the two types of data respectively.

[0089] Among them, for the sufficient sample dataset, the training set and the test set are divided according to the ratio of 3:7. With each influencing factor as the input, the energy transfer difference as the output, and the root mean square error as the loss function. Initialize the neural network weights and bias terms. Use the random search method to optimize the hyperparameters within a certain range.

[0090] The above neural network hyperparameters include activation function, regularization coefficient, learning rate, maximum number of iterations, etc. According to the sample data volume, input volume, etc., the hyperparameter optimization interval is given, and the optimal hyperparameters are selected by comprehensively considering the accuracy and stability performance.

[0091] For the small sample problem, transfer learning is applied to construct a regression prediction model. The small sample source dataset is divided into the source domain and the target domain, and the source domain data and the target domain data are mapped into two closer feature matrices by using feature mapping.

[0092] In the above mapping process, K-means clustering is introduced at the same time to realize problem transformation, and with the goal of minimizing the distance between cluster centers, the regularization hyperparameters of feature mapping are automatically optimized in combination with the random search method.

[0093] For the new source domain and target domain datasets after mapping, the neural network is also used as a predictor, and the optimization of the weights of each particle in the particle filter is defined. And with the root mean square error and the coefficient of determination as the evaluation indexes, the optimal transfer learning model is obtained. Using the measured sensor data and other non-sensor data of the pipe segment as the input, the energy transfer difference of the current pipe segment is predicted.

[0094] It should be noted that using transfer learning can improve the prediction accuracy of small sample working conditions. At the same time, the use of transfer learning makes it possible to accurately predict the energy transfer difference of natural gas pipe segments under different pipe segments and different service environments.

[0095] S3: The sensor data source module collects the regular sampling inspection datasets of each pipeline, and uses the relative error with the prediction result as the evaluation index, and retains the sampling inspection samples with a relative error less than 10%. Combine with the Bayesian parameter update module for optimization, define the weights of each particle in the particle filter, with the goal of minimizing the relative error, reverse transmit to change the network node weights, update the machine learning model, and realize the automatic update of the regression model.

[0096] Repeat the above steps to achieve the regression prediction of the energy transfer difference of any pipe segment.

[0097] The above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for constructing a natural gas pipeline network energy transmission difference calculation system integrating machine learning, characterized in that: The energy transmission difference calculation system includes three units: source database unit, prediction unit and optimization unit; The source database unit includes a sensor data source module and other data modules; The prediction unit includes a sufficient sample prediction module and a small sample prediction module; The optimization unit includes a measured data module and a Bayesian parameter updating module; The construction process of the natural gas pipeline network energy transmission difference calculation system integrating machine learning mainly includes the following steps: Step 1: Determine the factors affecting energy transmission difference. The source database unit collects the historical data of energy transmission difference of natural gas pipeline networks in various engineering cases and builds a basic database. Step 2: The prediction unit first performs clustering processing on the basic data, dividing it into a small sample and a sufficient sample sub-data set, and then divides the two subsets into a training set and a test set respectively, and uses the training set to establish a machine learning regression prediction model for energy transmission difference of the natural gas pipeline network; the machine learning regression prediction model in step 2 includes a sufficient sample machine learning prediction model and a small sample machine learning prediction model; Step 3: Input the real-time sensor data and other working data of the pipeline segment to be predicted in real time, and use the model trained in step 2 to predict the energy transmission difference; when predicting the energy transmission difference of the unknown pipeline segment in step 3, it is necessary to first determine whether the problem to be predicted belongs to a small sample or a sufficient sample problem based on the similarity analysis between the input features to be predicted and the features in the basic data set; after determining the classification, input the feature set of the sample to be predicted, and use the corresponding machine learning regression prediction model to obtain the current energy transmission difference prediction value; Step 4: The measured data module in the optimization unit collects the actual data during each energy transmission difference detection, and uses the error with the predicted result as the judgment basis. The Bayesian parameter update module retains the real sample data with an error below 10% and builds an optimization database; based on the comparison between the samples in the optimization database and the predicted results, the weights of each particle in the particle filter are defined, and the weights of the input parameters in the machine learning regression prediction model are updated to further reduce the error rate.

2. The method for constructing a natural gas pipeline network energy transmission difference calculation system integrating machine learning according to claim 1 is characterized in that: The influencing factors in step 1 include internal gas pressure, pressure drop, gas flow, gas temperature, actual high-order molar calorific value, initial flow, terminal flow, pipe section length, and external gas pressure and temperature in the natural gas pipeline network section. The above influencing factors in the historical case are constructed as a feature set X = {x 11 ,x 12 ,x 13 ,…,x ij }, x ij Represents the jth feature of the i-th sample number; The energy input difference in the historical data collected in the step 1 is used as the label set Y={Y1, Y2, Y3, ..., Yi} of the basic data set.

3. The method for constructing a natural gas pipeline network energy transmission difference calculation system integrating machine learning according to claim 2 is characterized in that: In the step 2, the basic data is divided into small sample and sufficient sample data subsets using the K-means clustering method; wherein the sufficient sample machine learning prediction model adopts a neural network method, and the small sample machine learning prediction model adopts a transfer learning method.

4. The method for constructing a natural gas pipeline network energy transmission difference calculation system integrating machine learning according to claim 3 is characterized in that: The sufficient samples take the feature set X in step 1 as input, assign an initial weight factor, and use the training set to establish a corresponding neural network model; A transfer learning method is used to train the small sample training set. The transfer learning method uses a feature transfer method, which is based on feature mapping and aligns data distribution in different domains.

5. The method for constructing a natural gas pipeline network energy transmission difference calculation system integrating machine learning according to claim 4 is characterized in that: In the transfer learning method, the physical field of the outgoing data is taken as the source domain M, the physical field of the incoming data is taken as the target domain N, and the feature set and label set in the source domain and the target domain are defined as Xm, Ym, Xn, and Yn respectively; The transfer learning method, based on classic feature migration, fuses source domain data and further combines clustering hyperparameter optimization.

6. The method for constructing a natural gas pipeline network energy transmission difference calculation system integrating machine learning according to claim 5 is characterized in that: The key point of the feature migration is to construct a mapping φ so that the source domain and the target domain are close to the same conditional distribution. This scheme uses the maximum mean difference, that is, Maximum Mean Discrepancy, abbreviated as MMD, to evaluate the distribution distance between the source domain and the target domain. The calculation expression is: Where: n1 and n2 are the number of source domain and target domain data respectively. After mapping, the clustering method is introduced to construct the optimization index, and the random search method is further combined to optimize the regularization parameters.

7. The method for constructing a natural gas pipeline network energy transmission difference calculation system integrating machine learning according to claim 6 is characterized in that: When optimizing feature mapping hyperparameters, K-means clustering is used, which is specifically used for the feature matrix after feature migration mapping. At this time, a cluster center will be generated for the source domain and the target domain respectively, and the distance between the two cluster centers will represent the distribution difference between the source and target domain features. The distance between the two cluster centers is defined by the following formula: Distance(K′)=||c m -c n || 2 (4) Among them, c m and c n They are the unique clustering centers of the source domain and target domain features after feature mapping.

8. The method for constructing a natural gas pipeline network energy transmission difference calculation system integrating machine learning according to claim 7 is characterized in that: Based on the introduction of the above-mentioned K-means clustering method, the optimization problem of the hyperparameters is transformed into the minimization problem of Distance(K′), and the regularization parameter is optimized within a given range using the random search method.

9. The method for constructing a natural gas pipeline network energy transmission difference calculation system integrating machine learning according to claim 1, characterized in that: The measured data module periodically checks some pipe sections and calculates the actual energy transmission difference based on the measured signals and other data; That is, newly added samples are obtained; at the same time, samples with a relative error of less than 10% are retained within the sampled pipe section, and samples with an error of more than 10% are eliminated; The particle filter parameters are defined by sampling new samples, so that the weight factors of feature engineering in the machine learning model can be further optimized through the Bayesian parameter update module.

Images