A transformer drying time sequence prediction method and system

Abstract

The present application belongs to the technical field of transformer manufacturing, and provides a transformer drying time sequence prediction method and system, which comprises the following steps: obtaining transformer drying full-process information from multiple sources to form an original data set, obtaining and classifying the original data set according to corresponding classification information and performing good product screening to obtain a good product data group and perform data processing and integration to obtain a learning data set containing an input feature group, a drying time sequence vector group and an energy consumption data group, constructing a drying prediction model based on a hybrid neural network, training the drying prediction model using the learning data set, and judging whether the prediction accuracy of the drying prediction model meets the standard, if the standard is met, performing reasoning prediction to obtain a predicted drying time sequence vector and predicted energy consumption data, judging whether the drying process parameters need to be optimized, if the drying process parameters need to be optimized, performing collaborative optimization and adjustment of the drying process parameters based on the prediction results by using a self-adaptive control algorithm in combination with the drying prediction model.

Description

Technical Field

[0001] This invention belongs to the field of transformer manufacturing technology, specifically a method and system for predicting the drying sequence of transformers. Background Technology

[0002] In the field of power equipment manufacturing, the drying process of transformers is a crucial step in ensuring their stable performance and long-term reliable operation. During the manufacturing process, the insulating materials of transformers contain a certain amount of moisture. If this moisture is not thoroughly removed, it will seriously affect the insulation performance, service life, and even the safe and stable operation of the power system. Therefore, accurately predicting the transformer drying sequence and rationally controlling the drying process are essential for improving drying quality, reducing energy consumption, and increasing production efficiency.

[0003] In actual production, technicians have found that existing methods for predicting transformer drying sequence have many shortcomings. On the one hand, traditional methods often rely on a single data source or limited drying process information, making it difficult to comprehensively cover various influencing factors in the entire transformer drying process, such as structural differences between different transformer models, characteristics of insulation materials, operating status of drying equipment, and environmental conditions. This leads to inaccurate prediction of drying sequence, often resulting in drying times that are too long or too short. Too long a drying time increases energy consumption and production cycle, while too short a drying time may lead to incomplete drying and affect product quality.

[0004] On the other hand, existing technologies lack effective classification and screening mechanisms when processing data during the drying process. A large amount of low-quality or irrelevant data enters the model, which not only increases the difficulty and cost of data processing, but also reduces the reliability of the prediction model. At the same time, traditional prediction models often use a single neural network or other simple algorithms, which are insufficient for fitting the nonlinear characteristics of complex drying processes and cannot adapt to the dynamic changes of different drying stages, making it difficult to meet the prediction accuracy requirements of actual production. In addition, in terms of adjusting drying process parameters, existing methods are usually based on experience or fixed rules and cannot be adaptively optimized according to real-time prediction results, making it difficult to achieve fine control of the drying process.

[0005] To address the above problems, this invention proposes a method and system for predicting the drying sequence of transformers. Summary of the Invention

[0006] To address the shortcomings of existing technologies and solve at least one of the technical problems mentioned in the background art, the present invention provides a method and system for predicting the drying sequence of transformers.

[0007] The technical solution adopted by this invention to solve the technical problem is: a method for predicting the drying sequence of a transformer, comprising:

[0008] The original dataset is composed of information on the entire transformer drying process obtained from multiple sources. The original dataset is then classified and selected for high-quality products based on the corresponding classification information to obtain a high-quality data group.

[0009] The high-quality data set is processed and integrated to obtain a learning dataset that includes the input feature set, the drying time series vector set, and the energy consumption data set.

[0010] A drying prediction model is constructed based on a hybrid neural network. The model is trained using a learning dataset, and the accuracy of the model is judged. If the accuracy is met, the model is used to make inference predictions before the current drying process starts, and the predicted drying time series vector and predicted energy consumption data are used as the prediction results.

[0011] Based on the prediction results, determine whether the drying process parameters need to be optimized. If so, use an adaptive control algorithm combined with the drying prediction model to perform collaborative optimization and adjustment of the drying process parameters based on the prediction results.

[0012] The method for obtaining the premium data group is as follows:

[0013] With the help of SCADA and IoT technologies, during the data acquisition phase, the data acquisition module obtains the entire process information and corresponding classification information of transformer drying from multiple sources with different production architectures. For different sources, corresponding data acquisition interface programs that can parse data of different formats are set up, and the data of different formats are converted into a unified data format to obtain the raw dataset within the data acquisition period. Based on the corresponding classification information, the raw dataset is divided into drying data groups, and then filtered based on the high-quality judgment criteria to obtain the high-quality data group.

[0014] The prediction results are obtained in the following way:

[0015] Before the current drying process starts, the basic data of transformer accessories, drying process parameters and environmental data of the transformer to be dried are collected to obtain the collection feature group. The original dimensions and normalized parameters recorded in the data normalization module are obtained and normalized to obtain the collection feature group to obtain the input feature vector. The input feature vector is input into the drying prediction model with the accuracy standard to obtain the predicted drying time series vector and predicted energy consumption data as the prediction result.

[0016] The method for determining whether the accuracy meets the standard is as follows:

[0017] The learning dataset is divided into a training set and a test set. The dryness prediction model is trained using the training set according to the set parameters. After training, the dryness prediction model is validated using the test set. The coefficient of determination and mean square error of the dryness prediction model in the test set are used as indicators to evaluate the prediction accuracy of the dryness prediction model. If the calculated coefficient of determination is greater than 0.9 and the mean square error is less than the preset error threshold, the prediction accuracy of the dryness prediction model is judged to meet the standard.

[0018] The drying prediction model is constructed as follows:

[0019] A drying prediction model is constructed using a hybrid neural network architecture, which includes an input layer, a feature fusion layer, a time series processing layer, and a fully connected output layer. The input layer receives input feature groups, the feature fusion layer uses a fully connected neural network, the time series processing layer introduces long short-term memory network units to take into account the time series characteristics of drying time series, and the fully connected output layer contains two branches, which output the drying time series vector and energy consumption data respectively.

[0020] The method for determining whether the drying process parameters need to be optimized is as follows:

[0021] Set the expected conditions, obtain the predicted energy consumption data, calculate the deviation between the predicted energy consumption data and the preset expected energy consumption data, and if the deviation is greater than the allowable deviation threshold, it is determined that the predicted energy consumption data does not meet the expected conditions.

[0022] If the predicted drying time vector does not meet the expected conditions, or the predicted energy consumption data does not meet the expected conditions, it is determined that the prediction result does not meet the expected conditions, and the drying process parameters need to be optimized.

[0023] The method for determining whether the predicted drying time series vector does not meet the expected conditions is as follows:

[0024] Obtain the predicted drying timeline vector and the preset expected drying timeline vector, calculate the predicted total drying time and the expected total drying time respectively. If the predicted total drying time is greater than the expected total drying time, calculate the Euclidean distance between the predicted drying timeline vector and the expected drying timeline vector, and compare it with the preset allowable deviation threshold. If it is greater than the allowable deviation threshold, it is determined that the predicted drying timeline vector does not meet the expected conditions.

[0025] The method of collaborative optimization and adjustment is as follows:

[0026] Combining the drying prediction model, the particle swarm optimization algorithm is used as the adaptive control algorithm. By continuously updating the particle position and particle velocity until the maximum number of iterations or the fitness function converges, the drying process parameters of the global optimum are obtained. The original dimensions and normalized parameters recorded in the data normalization module are obtained and the drying process parameters and prediction results are denormalized to obtain the actual drying process parameters, which are then sent to the controller to complete the optimization and adjustment.

[0027] The fitness function is obtained as follows:

[0028] The particle position is used as the input feature vector to modify the drying process parameters. Each particle position represents a set of candidate solutions for the drying process parameters. The modified input feature vector is input into the drying prediction model to obtain the predicted drying time series vector and the predicted energy consumption data. Based on the Euclidean distance between the predicted drying time series vector and the expected drying time series vector, as well as the deviation between the predicted energy consumption data and the expected energy consumption data, the fitness function is constructed by fusion calculation.

[0029] A transformer drying timing prediction system, comprising:

[0030] The drying dataset subsystem includes a data acquisition module, a data cleaning module, a data normalization module, and a dataset construction module. It can acquire information on the entire transformer drying process from multiple sources to form an original dataset, acquire and classify the original dataset according to the corresponding classification information, and select high-quality data groups. It can also process and integrate the high-quality data groups to obtain a learning dataset containing input feature groups, drying time series vector groups, and energy consumption data groups.

[0031] The drying prediction model subsystem includes a model building module, a model training module, a prediction evaluation module, and a parameter optimization module. It realizes the construction of a drying prediction model based on a hybrid neural network, trains the drying prediction model using a learning dataset, and judges whether the prediction accuracy of the drying prediction model meets the standard. If it meets the standard, it performs inference prediction through the drying prediction model before the current drying process starts, and obtains the predicted drying time series vector and predicted energy consumption data as the prediction results. Based on the prediction results, it judges whether the drying process parameters need to be optimized. If so, it uses an adaptive control algorithm, combined with the drying prediction model, to perform collaborative optimization and adjustment of the drying process parameters based on the prediction results.

[0032] Drying model interface subsystem: includes interfaces between the drying prediction model and equipment, SCADA interaction interface and MES interaction interface, to achieve accurate interaction.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. This invention effectively improves data quality by acquiring full-process information from multiple sources and screening high-quality data sets, providing a reliable foundation for subsequent model training. The drying prediction model built on a hybrid neural network can fully utilize the advantages of different neural networks, improve prediction accuracy, and perform inference prediction before the drying process starts, obtaining drying time sequence vectors and energy consumption data in advance. This helps to plan production arrangements in advance, rationally allocate resources, and avoid production delays or resource waste caused by insufficient estimation of drying time or energy consumption.

[0035] 2. This invention determines whether to optimize drying process parameters based on prediction results, and uses an adaptive control algorithm combined with a prediction model for collaborative optimization and adjustment, thereby realizing intelligent control of the drying process. This dynamic optimization method can accurately adjust process parameters according to the drying needs and actual conditions of different transformers, improve drying efficiency, reduce energy consumption, and at the same time help improve the drying quality of transformers and reduce problems such as transformer performance degradation caused by improper drying processes, bringing significant economic benefits and quality improvement to enterprises. Attached Figure Description

[0036] Figure 1 This is a flowchart of the steps of the transformer drying timing prediction method according to an embodiment of the present invention;

[0037] Figure 2 This is a system module architecture diagram of the transformer drying timing prediction system described in an embodiment of the present invention. Detailed Implementation

[0038] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0039] Example 1

[0040] like Figure 1 As shown in the figure, a transformer drying timing prediction method according to an embodiment of the present invention includes the following steps:

[0041] S1: Obtain information on the entire process of transformer drying from multiple sources to form the original dataset. Obtain and classify the original dataset according to the corresponding classification information and select the best products to obtain the best product data group.

[0042] With the help of SCADA and IoT technologies, during the data acquisition phase, the data acquisition module obtains the entire process information and corresponding classification information of transformer drying from multiple sources with different production architectures. The data indicators of the entire process information of transformer drying include basic data of transformer accessories, drying process parameters, environmental data, drying sequence and energy consumption data. The corresponding classification information includes the corresponding drying tank model and the corresponding finished product experimental data.

[0043] The basic data for transformer components includes coil quality data and core quality data. Coil quality data includes coil insulation thickness and coil resistance. Coil insulation thickness is measured by a high-precision coil insulation thickness detector on the coil production line using non-contact measurement technology. Coil resistance is measured by applying a specific voltage and measuring the resistance of the coil using a resistance meter. Core quality data includes core permeability and core lamination gap. Core permeability is calculated by testing the permeability of the core under different magnetic field strengths using a core permeability tester. Core lamination gap is measured optically using a lamination gap measuring device.

[0044] The drying process parameters include drying temperature and drying vacuum, which are obtained by recording the drying process parameters set in the drying tank during the transformer drying process. The environmental data includes ambient temperature and ambient humidity, which are obtained by setting temperature and humidity sensors in the environment where the drying tank is located.

[0045] The drying sequence includes the duration of each drying stage during the drying process of the transformer in the drying tank. The drying stages include the preheating stage, the heating drying stage, the constant temperature drying stage, and the cooling drying stage. The drying sequence is obtained by the timing module automatically recording the timestamps at the beginning and end of each drying stage. The energy consumption data is collected by installing a high-precision power meter on the power supply line of the drying equipment.

[0046] Among them, the finished product test data includes the breakdown voltage and dielectric loss of transformer accessories. These are obtained by randomly selecting several transformer accessory samples from the same production batch, performing experiments and measurements after the drying process, and then combining the results with calculations. The calculated finished product test data is the finished product test data of all corresponding transformer accessories in the same production batch.

[0047] Since the transformer drying process information production system architecture collects data from different sources in the production architecture, corresponding data acquisition interface programs that can parse data of different formats are set up for different sources, and the different formats of data are converted into a unified data format. Through the data transmission network, the converted transformer drying process information is integrated to obtain the original dataset within the data acquisition period. Based on the drying tank model corresponding to each transformer drying process information in the original dataset, the transformer drying process information of the same drying tank model is divided into the same drying data group, resulting in several drying data groups.

[0048] Within each drying data group, the finished product experimental data corresponding to the full-process information of transformer drying are used to screen for superior products. Specifically, according to the enterprise's production standards and product quality requirements, the criteria for judging superior products are set, including the breakdown voltage and dielectric loss of superior products. For each transformer drying process information, the corresponding finished product experimental data are compared with the criteria for judging superior products.

[0049] If the breakdown voltage corresponding to the full-process information of transformer drying is greater than the breakdown voltage of superior product, and the corresponding dielectric loss is less than the dielectric loss of superior product, then the superior product judgment criteria are met, and the full-process information of transformer drying is marked as superior product data.

[0050] Within each drying data group, only all premium data are retained to obtain the premium data group corresponding to each drying data group;

[0051] It should be noted that this step involves collecting information from multiple sources throughout the entire transformer drying process, encompassing basic component data, process parameters, environmental data, and other multi-dimensional indicators. This ensures the comprehensiveness and completeness of the data. Combined with SCADA and IoT technologies, real-time and accurate data collection is achieved. Data is categorized and high-quality data is selected based on classification information, effectively eliminating substandard data, improving data quality, providing a reliable data foundation for subsequent analysis, reducing prediction errors caused by data issues, and ensuring that all data used for analysis is of high quality. This makes the models and predictions built upon this data more aligned with actual production needs, helping to improve production efficiency, reduce scrap rates, and minimize resource waste.

[0052] S2: Process and integrate the high-quality data set to obtain a learning dataset containing the input feature set, the drying time series vector set, and the energy consumption data set.

[0053] The obtained high-quality data set is then cleaned using the data cleaning module.

[0054] Specifically, for noise, outliers, and missing values ​​in all transformer drying process information in the premium data group, the data cleaning module uses a median filtering algorithm to remove noise, a density-based local anomaly factor (LOF) algorithm to identify outliers and detect abnormal data, and a nearest neighbor imputation method to supplement the detected abnormal data and missing values.

[0055] For the cleaned and processed premium data set, the data normalization module is used to normalize the premium data set.

[0056] Specifically, the data normalization module targets each data indicator in the entire transformer drying process information in the premium data group. Within each data indicator, the maximum-minimum normalization method is used to map the data indicator to the [0,1] interval. After normalization, the original dimensional information and normalization parameters of each data indicator are recorded to ensure the accuracy of the normalization process.

[0057] The cleaned and normalized high-quality data set is labeled as the normalized high-quality data set. The dataset construction module integrates the normalized high-quality data set into a learning dataset containing the input feature set, the drying time series vector set, and the energy consumption data set.

[0058] Specifically, the drying process information of each transformer in the Guiyi Youpin data group is integrated. Based on the drying process information of any transformer, the basic data of transformer accessories, drying process parameters and environmental data in the drying process information are integrated into an input feature vector. The drying time sequence in the drying process information is integrated into a drying time sequence vector. The duration of each drying stage in the drying time sequence vector is sorted according to the time sequence. The drying time sequence vector and energy consumption data are used as output features. All input feature vectors, drying time sequence vectors and energy consumption data are integrated in the same order according to the one-to-one correspondence to obtain the input feature group, the drying time sequence vector group and the energy consumption data group. That is, the learning dataset containing the input feature group, the drying time sequence vector group and the energy consumption data group is obtained.

[0059] It should be noted that the purpose of this step is to comprehensively utilize multiple data processing algorithms to form a complete data processing workflow, adopt corresponding solutions for different types of data problems, record the original dimensional information and normalization parameters during the normalization process, provide a basis for the denormalization of data, ensure the traceability and accuracy of data before and after processing, and enable the processed data to better adapt to subsequent model training, so that the model can learn the patterns in the data more accurately and improve the prediction accuracy of the model.

[0060] S3: Construct a drying prediction model based on a hybrid neural network, train the drying prediction model using a learning dataset, and determine whether the prediction accuracy of the drying prediction model meets the standard. If it does not meet the standard, supplement the learning dataset. If it meets the standard, perform inference prediction through the drying prediction model before the current drying process starts to obtain the predicted drying time series vector and predicted energy consumption data.

[0061] A drying prediction model is constructed, which adopts a hybrid neural network architecture, including an input layer, a feature fusion layer, a temporal processing layer, and a fully connected output layer.

[0062] Specifically, the input layer is responsible for receiving the input feature group. The number of neurons in the input layer is consistent with the dimension of the input feature group. Since the input feature vector in the input feature group contains basic data of transformer accessories, drying process parameters, and environmental data, the basic data of transformer accessories includes coil insulation layer thickness, coil resistance value, core permeability, and core lamination gap. The drying process parameters include drying temperature and drying vacuum degree, and the environmental data includes ambient temperature and ambient humidity. Therefore, the dimension of the input feature group is set to 8, and the number of neurons in the input layer is 8.

[0063] The feature fusion layer uses a fully connected neural network (FCN) to perform deep feature extraction on the input feature set. The FCN learns the non-linear relationship between different data indicators through the full connection between multiple layers of neurons, and maps the 8-dimensional input feature vector to a 16-dimensional feature vector.

[0064] The temporal processing layer introduces Long Short-Term Memory (LSTM) network units to capture the temporal dependencies between the durations of each drying stage, taking into account the temporal characteristics of the drying time sequence.

[0065] The fully connected output layer contains two branches, which output the drying time vector respectively. Energy consumption data ;

[0066] ;

[0067] in, , , , Let i represent the duration of the preheating stage, the heating and drying stage, the isothermal drying stage, and the cooling and drying stage, respectively, and let i represent the drying time sequence vector. Energy consumption data For the numbers in the training set, the mean squared error (MSE) is used as the loss function, and the formula is:

[0068] ;

[0069] Where N is the number of training samples in the training set. and These are the predicted drying time series vector and predicted energy consumption data obtained from the pre-inference prediction of the drying prediction model, respectively.

[0070] Based on any given learning dataset, the dataset is divided into training and testing sets in an 8:2 ratio. The Adam optimizer is used, with parameters such as learning rate, batch size, and training epoch set. The network weights are updated in each training round. ;

[0071] ;

[0072] Where j represents the current training round, Indicates the learning rate. Represents the gradient of the loss function;

[0073] The drying prediction model was trained using the training set according to the set parameters, and then validated using the test set after training. The coefficient of determination R² and mean square error RMSE of the drying prediction model in the test set were used as indicators to evaluate the prediction accuracy of the drying prediction model.

[0074] It should be noted that the coefficient of determination R² reflects the ability of the drying prediction model to explain data variation. The closer the coefficient of determination R² is to 1, the better the drying prediction model fits the data. RMSE measures the average error between the predicted result and the actual value. The smaller the RMSE value, the more accurate the prediction result.

[0075] Set the prediction accuracy criteria. If the calculated coefficient of determination is greater than 0.9 and the mean square error is less than the preset error threshold, the drying prediction model meets the prediction accuracy criteria and is judged to have met the criteria. Otherwise, the drying prediction model is judged to have failed to meet the criteria.

[0076] If the accuracy of the drying prediction model is not up to standard, the data supplementation mechanism is triggered. By expanding the time range of the data collection phase, the high-quality data set is supplemented, and the supplemented high-quality data set is used to construct a learning dataset to retrain the drying prediction model.

[0077] If the accuracy of the drying prediction model meets the standard, then proceed to the prediction analysis stage;

[0078] Before the current drying process starts, basic data of transformer accessories, drying process parameters and environmental data of the transformer to be dried are collected to obtain a collection feature group. The original dimensions and normalized parameters recorded in the data normalization module are used to normalize the collection feature group to obtain the input feature vector. The input feature vector is input into the drying prediction model to obtain the predicted drying time series vector and predicted energy consumption data as the prediction results.

[0079] It should be noted that this step involves building a drying prediction model based on a hybrid neural network. Combining the advantages of fully connected neural networks (FCN) and long short-term memory networks (LSTM), it effectively extracts data features, captures the temporal dependencies of the drying time series, and achieves accurate predictions of drying time series vectors and energy consumption data. Scientific training and evaluation methods are employed, using training and testing sets, and the Adam optimizer for training. The coefficient of determination (R²) and mean squared error (RMSE) are used as evaluation metrics to ensure the model's prediction accuracy and generalization ability. If the model does not meet the standards, it is retrained with supplementary data to continuously optimize its performance. An accurate drying prediction model can provide enterprises with predicted drying time series vectors and predicted energy consumption data before the drying process begins, helping them plan production processes in advance, allocate resources rationally, and reduce production costs. Through continuous optimization, the model's applicability in actual production environments is improved, providing a scientific basis for enterprise production decisions.

[0080] S4: Determine whether the drying process parameters need to be optimized based on the prediction results. If so, use an adaptive control algorithm combined with the drying prediction model to perform collaborative optimization and adjustment of the drying process parameters based on the prediction results.

[0081] Set the expected conditions, which include the expected drying time vector, expected energy consumption data and allowable deviation threshold. The expected conditions are combined with the original dimensions, normalization parameters and production expectation settings recorded in the data normalization module to determine whether the prediction results meet the expected conditions.

[0082] Specifically, based on the obtained predicted drying time series vector and expected drying time series vector, the predicted total drying time and expected total drying time are calculated respectively. If the predicted total drying time is greater than the expected total drying time, the Euclidean distance Dt between the predicted drying time series vector and the expected drying time series vector is calculated and compared with the preset allowable deviation threshold. If the calculated Euclidean distance is greater than the allowable deviation threshold, it is determined that the predicted drying time series vector does not meet the expected conditions.

[0083] Conversely, if the Euclidean distance is less than or equal to the allowable deviation threshold, or the predicted total drying time is less than or equal to the expected total drying time, the predicted drying time series vector is judged to meet the expected condition.

[0084] Based on the obtained predicted energy consumption data, the difference between the predicted energy consumption data and the expected energy consumption data is calculated as the deviation De. If the deviation is greater than the allowable deviation threshold, it is determined that the predicted energy consumption data does not meet the expected conditions. Conversely, if the deviation is less than or equal to the allowable deviation threshold, it is determined that the predicted energy consumption data meets the expected conditions.

[0085] If both the predicted drying time series vector and the predicted energy consumption data meet the expected conditions, the prediction result is judged to meet the expected conditions. Based on the original dimensions and normalization parameters recorded in the data normalization module, the prediction result is inversely normalized to obtain the actual predicted drying time series and the actual predicted energy consumption data.

[0086] If the predicted drying time series vector does not meet the expected conditions, or the predicted energy consumption data does not meet the expected conditions, it is determined that the prediction result does not meet the expected conditions. In this case, the drying process parameters need to be optimized. An adaptive control algorithm is used in conjunction with the drying prediction model to perform coordinated optimization and adjustment of the drying process parameters.

[0087] Specifically, Particle Swarm Optimization (PSO) algorithm is used as the adaptive control algorithm. In PSO, each particle position represents a set of candidate solutions for drying process parameters. ;

[0088] ;

[0089] in The normalized drying temperature. The normalized vacuum degree of the dried material;

[0090] Particle position The input feature vector is changed as a drying process parameter. The changed input feature vector is input into the drying prediction model to obtain the predicted drying time series vector and the predicted energy consumption data. The fitness function is constructed based on the Euclidean distance Dt between the predicted drying time series vector and the expected drying time series vector, and the deviation De between the predicted energy consumption data and the expected energy consumption data.

[0091] The particle velocity update formula is:

[0092] ;

[0093] in This represents the particle velocity at the current step k. This indicates the particle position at the current step k. Inertial weights are used to balance global and local search capabilities. and As an acceleration factor, it controls the particle's trajectory towards the individual optimal solution. and the global optimal solution The step size of the movement, and Use random numbers in the interval [0,1] to increase the randomness of the search;

[0094] The particle swarm optimization algorithm continuously updates particle positions and velocities until the maximum number of iterations or the fitness function converges, obtaining the globally optimal drying process parameters. The drying process parameters of the input feature vector are then changed to the globally optimal drying process parameters. The modified input feature vector is input into the drying prediction model to obtain the predicted drying time sequence vector and predicted energy consumption data as the prediction results. Based on the original dimensions and normalized parameters recorded in the data normalization module, both the drying process parameters and the prediction results are denormalized to obtain the actual drying process parameters, the actual predicted drying time sequence, and the actual predicted energy consumption data. The actual drying process parameters are then sent to the controller to complete the optimization and adjustment.

[0095] It should be noted that the purpose of this step is to set expected conditions and rigorously judge the prediction results to accurately identify whether the prediction results meet production expectations. When the prediction results do not meet the expectations, the particle swarm optimization (PSO) algorithm is used in conjunction with the drying prediction model to collaboratively optimize and adjust the drying process parameters. This allows for dynamic adjustment of process parameters according to actual production needs, making the drying process more efficient and energy-saving, improving product quality, and reducing energy costs. In actual production, this step can promptly identify potential problems in the production process and solve them by optimizing and adjusting the process parameters. The optimized process parameters make the drying process more in line with production requirements, improve production efficiency, reduce energy consumption, reduce production costs, and enhance the company's economic benefits and market competitiveness.

[0096] The technical solution of this invention is as follows: A raw dataset is formed by acquiring information on the entire transformer drying process from multiple sources. The raw dataset is then classified and selected for high-quality data based on corresponding classification information to obtain a high-quality data group. This high-quality data group is processed and integrated to obtain a learning dataset containing an input feature group, a drying time-series vector group, and an energy consumption data group. A drying prediction model is constructed based on a hybrid neural network. The drying prediction model is trained using the learning dataset, and the accuracy of the prediction is determined. If the accuracy is not met, the learning dataset is supplemented. If the accuracy is met, the drying prediction model is used to perform inference prediction before the current drying process starts, obtaining predicted drying time-series vectors and predicted energy consumption data. Based on the prediction results, it is determined whether the drying process parameters need to be optimized. If so, an adaptive control algorithm, combined with the drying prediction model, is used to collaboratively optimize and adjust the drying process parameters based on the prediction results.

[0097] Example 2

[0098] like Figure 2 As shown in the embodiment of the present invention, a transformer drying timing prediction system includes:

[0099] The drying dataset subsystem includes a data acquisition module, a data cleaning module, a data normalization module, and a dataset construction module. It can acquire information on the entire transformer drying process from multiple sources to form an original dataset, acquire and classify the original dataset according to the corresponding classification information, and select high-quality data groups. It can also process and integrate the high-quality data groups to obtain a learning dataset containing input feature groups, drying time series vector groups, and energy consumption data groups.

[0100] The drying prediction model subsystem includes a model building module, a model training module, a prediction evaluation module, and a parameter optimization module. It realizes the construction of a drying prediction model based on a hybrid neural network, trains the drying prediction model using a learning dataset, and judges whether the prediction accuracy of the drying prediction model meets the standard. If it meets the standard, it performs inference prediction through the drying prediction model before the current drying process starts, and obtains the predicted drying time series vector and predicted energy consumption data as the prediction results. Based on the prediction results, it judges whether the drying process parameters need to be optimized. If so, it uses an adaptive control algorithm, combined with the drying prediction model, to perform collaborative optimization and adjustment of the drying process parameters based on the prediction results.

[0101] Drying model interface subsystem: includes interfaces between the drying prediction model and equipment, SCADA interaction interface and MES interaction interface, to achieve accurate interaction.

[0102] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and various changes and modifications can be made without departing from the spirit and scope of the invention; all such changes and modifications will fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the drying time sequence of a transformer, characterized in that, include: The original dataset is composed of information on the entire transformer drying process obtained from multiple sources. The original dataset is then classified and selected for high-quality products based on the corresponding classification information to obtain a high-quality data group. The high-quality data set is processed and integrated to obtain a learning dataset containing the input feature set, the drying time series vector set, and the energy consumption data set. A drying prediction model is constructed based on a hybrid neural network. The model is trained using a learning dataset, and the accuracy of the model is judged. If the accuracy is met, the model is used to make inference predictions before the current drying process starts, and the predicted drying time series vector and predicted energy consumption data are used as the prediction results. Based on the prediction results, determine whether the drying process parameters need to be optimized. If so, use an adaptive control algorithm combined with the drying prediction model to perform collaborative optimization and adjustment of the drying process parameters based on the prediction results. The method for determining whether the drying process parameters need to be optimized is as follows: Set the expected conditions, obtain the predicted energy consumption data, calculate the deviation between the predicted energy consumption data and the preset expected energy consumption data, and if the deviation is greater than the allowable deviation threshold, it is determined that the predicted energy consumption data does not meet the expected conditions. If the predicted drying time vector does not meet the expected conditions, or the predicted energy consumption data does not meet the expected conditions, it is determined that the prediction result does not meet the expected conditions, and the drying process parameters need to be optimized. The method for determining whether the predicted drying time series vector does not meet the expected conditions is as follows: Obtain the predicted drying timeline vector and the preset expected drying timeline vector, calculate the predicted total drying time and the expected total drying time respectively. If the predicted total drying time is greater than the expected total drying time, calculate the Euclidean distance between the predicted drying timeline vector and the expected drying timeline vector, and compare it with the preset allowable deviation threshold. If it is greater than the allowable deviation threshold, it is determined that the predicted drying timeline vector does not meet the expected conditions. The method of collaborative optimization and adjustment is as follows: Combining the drying prediction model, the particle swarm optimization algorithm is used as the adaptive control algorithm. By continuously updating the particle position and particle velocity until the maximum number of iterations or the fitness function converges, the drying process parameters of the global optimum are obtained. The original dimensions and normalized parameters recorded in the data normalization module are obtained and the drying process parameters and prediction results are denormalized to obtain the actual drying process parameters, which are then sent to the controller to complete the optimization and adjustment. The fitness function is obtained as follows: The particle position is used as the input feature vector to modify the drying process parameters. Each particle position represents a set of candidate solutions for the drying process parameters. The modified input feature vector is input into the drying prediction model to obtain the predicted drying time series vector and the predicted energy consumption data. Based on the Euclidean distance between the predicted drying time series vector and the expected drying time series vector, as well as the deviation between the predicted energy consumption data and the expected energy consumption data, the fitness function is constructed by fusion calculation.

2. The method for predicting the drying sequence of a transformer according to claim 1, characterized in that, The method for obtaining the premium product data set is as follows: With the help of SCADA and IoT technologies, during the data acquisition phase, the data acquisition module obtains the entire process information and corresponding classification information of transformer drying from multiple sources with different production architectures. For different sources, corresponding data acquisition interface programs are set up to parse data of different formats and convert the different data formats into a unified data format to obtain the raw dataset within the data acquisition period. Based on the corresponding classification information, the raw dataset is divided into drying data groups and filtered based on the superior product judgment criteria to obtain the superior product data group.

3. The method for predicting the drying sequence of a transformer according to claim 1, characterized in that, The prediction results are obtained in the following way: Before the current drying process starts, basic data of transformer accessories, drying process parameters and environmental data of the transformer to be dried are collected to obtain a collection feature group. The original dimensions and normalized parameters recorded in the data normalization module are obtained and normalized to obtain the collection feature group to obtain the input feature vector. The input feature vector is input into the drying prediction model with the required accuracy to obtain the predicted drying time series vector and predicted energy consumption data as the prediction results.

4. The method for predicting the drying sequence of a transformer according to claim 3, characterized in that, The method for determining whether the accuracy meets the standard is as follows: The learning dataset is divided into a training set and a test set. The dryness prediction model is trained using the training set according to the set parameters. After training, the dryness prediction model is validated using the test set. The coefficient of determination and mean square error of the dryness prediction model in the test set are used as indicators to evaluate the prediction accuracy of the dryness prediction model. If the calculated coefficient of determination is greater than 0.9 and the mean square error is less than the preset error threshold, the prediction accuracy of the dryness prediction model is judged to meet the standard.

5. The method for predicting the drying sequence of a transformer according to claim 1, characterized in that, The drying prediction model is constructed as follows: A drying prediction model is constructed using a hybrid neural network architecture, which includes an input layer, a feature fusion layer, a time series processing layer, and a fully connected output layer. The input layer receives input feature groups, the feature fusion layer uses a fully connected neural network, the time series processing layer introduces long short-term memory network units to take into account the time series characteristics of drying time series, and the fully connected output layer contains two branches, which output the drying time series vector and energy consumption data respectively.

6. A transformer drying timing prediction system, the system being used to implement the prediction method as described in any one of claims 1-5, characterized in that, include: The drying dataset subsystem includes a data acquisition module, a data cleaning module, a data normalization module, and a dataset construction module. It can acquire information on the entire transformer drying process from multiple sources to form an original dataset, acquire and classify the original dataset according to the corresponding classification information, and select high-quality data groups. It can also process and integrate the high-quality data groups to obtain a learning dataset containing input feature groups, drying time series vector groups, and energy consumption data groups. The drying prediction model subsystem includes a model building module, a model training module, a prediction evaluation module, and a parameter optimization module. It realizes the construction of a drying prediction model based on a hybrid neural network, trains the drying prediction model using a learning dataset, and judges whether the prediction accuracy of the drying prediction model meets the standard. If it meets the standard, it performs inference prediction through the drying prediction model before the current drying process starts, and obtains the predicted drying time series vector and predicted energy consumption data as the prediction results. Based on the prediction results, it judges whether the drying process parameters need to be optimized. If so, it uses an adaptive control algorithm, combined with the drying prediction model, to perform collaborative optimization and adjustment of the drying process parameters based on the prediction results. Drying model interface subsystem: includes interfaces between the drying prediction model and equipment, SCADA interaction interface and MES interaction interface, to achieve accurate interaction.

Images