Data-driven transformer life evaluation method and device and storage medium
By constructing a transformer life prediction model based on the quantum Hilbert space coding layer, using multi-source timing data to capture the complex correlation of transformer life, the accuracy problem of transformer life prediction is solved, and more efficient transformer management is achieved.
Patent Information
- Application Number
- CN202510672238.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-23
- Publication Date
- 2025-08-19
AI Technical Summary
The prior art is difficult to effectively predict the life of the transformer, resulting in possible failures and excessive repair or replacement cycles, increasing economic losses and risks.
Using a data-driven method, a transformer life prediction model for the improved quantum Hilbert space coding layer is constructed by acquiring multi-source timing data such as insulating oil mass, furan concentration, local discharge and temperature, to capture the complex relationship between multi-source timing data and life, and to enhance feature extraction capabilities using quantum computing.
Improves the accuracy and efficiency of transformer life prediction, reduces the possibility of power outages, optimizes maintenance plans, and reduces economic losses.
Smart Images
Figure CN120509314A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of transformer life prediction, and in particular to a data-driven transformer life assessment method, device and storage medium. Background Art
[0002] Transformers play a vital role in power systems. Over time, they are subject to aging, degradation, and potential failures. These unexpected failures can lead to significant losses, while lengthy repair or replacement cycles exacerbate financial losses and risks. Assessing transformer health and predicting remaining life are crucial to ensuring efficient operation and developing targeted maintenance plans.
[0003] To mitigate these risks, effective lifespan prediction is essential to ensure power transformers remain in optimal operating condition, minimizing the likelihood of power outages. With the development of deep learning, convolutional neural networks and recurrent neural networks (RNNs) for lifespan prediction based on transformer data have emerged as a new direction in transformer prediction technology. For RNNs used for time series modeling, control gates are primarily implemented through linear layers and activation functions. Nonlinear relationships are primarily introduced through activation functions, and parameter associations are primarily established in linear space, approximating the relationships between complex real-world transformer data and between transformer data and transformer lifespan. A design that breaks through the limitations of RNN control gates is needed to improve prediction results. Summary of the Invention
[0004] In order to solve the above technical problems or at least partially solve the above technical problems, the present invention provides a data-driven transformer life assessment method, device and storage medium.
[0005] In a first aspect, the present invention provides a data-driven transformer life assessment method, comprising: During the service life of the transformer, multi-source time series data for transformer life assessment and the corresponding remaining life of the transformer are obtained to form a data set for training the transformer life prediction model. The multi-source time series data includes: dissolved gas time series, furan time series, partial discharge time series, temperature time series, and power time series combined with insulating oil quality coding. In the multi-source time series data, different types of time series data are aligned in time and standardized. A transformer life prediction model is constructed, comprising a time series modeling network formed by an improved gated recurrent unit and a fully connected network for transformer life prediction based on the output of the time series modeling network. In the improved gated recurrent unit, its linear layer is replaced with a quantum Hilbert space coding layer that supports nonlinear modeling. The quantum Hilbert space coding layer uses quantum computing to encode multi-source time series data and training parameters in the quantum Hilbert space to capture the temporal connections between the multi-source time series data and between the multi-source time series data and the transformer life. The transformer life prediction model is trained using the acquired data set, the mean square error loss between the predicted life and the actual life is used as a loss function, and the parameters of the transformer life prediction model are dynamically adjusted with the goal of minimizing the loss function; During the application phase, transformer parameters are collected in real time and preprocessed to obtain multi-source time series data. The trained transformer life prediction model uses multi-source time series data to predict life.
[0006] Furthermore, during the life cycle of a transformer, the insulating oil is replaced multiple times. For the regularly replaced insulating oil, the quality is coded according to the usage time to model the impact of insulating oil quality on different types of time series data reflecting the degree of transformer aging. The insulating oil quality is encoded using sine and cosine coding to obtain the insulating oil quality code: ; ; Where: d is the dimension after quality coding, K is the frequency regulation technology, t is the service life of the insulating oil, and j is the dimension index of the quality coding.
[0007] Furthermore, an oil-gas separator installed in the transformer is used to carry out on-site quantitative insulating oil collection and oil-gas separation under constant temperature and pressure to perform dissolved gas detection and obtain a dissolved gas time series, wherein the oil-gas separator includes a separation chamber, an oil interface connected to the transformer is provided at the bottom of the separation chamber, and the oil interface is connected to the transformer through a pipeline with a valve; a stirrer is provided at the bottom of the separation chamber; a thermal resistor rod and a temperature sensor are provided in the separation chamber; a liquid level sensor is provided on the side wall of the separation chamber; an exhaust interface and an air inlet pipe are provided at the top of the separation chamber, the exhaust interface is connected to the gas detection chamber, the gas detection chamber is connected to a negative pressure machine for providing negative pressure, a tunable diode laser and a photodetector are provided in the gas detection chamber, the tunable diode laser and the photodetector perform dissolved gas detection on the transformer insulating oil, the dissolved gas detection results are constructed into a dissolved gas time series, and after standardization, combined with the insulating oil quality coding, one of the multi-source time series data is obtained.
[0008] Furthermore, based on the characteristic absorption peak of furan molecules in the near-infrared region, near-infrared lasers and fiber optic sensors were used to monitor the furan concentration in transformer oil online. The furan detection results were aligned with the dissolved gas time series at the same time interval to construct a furan time series. The furan time series was standardized and combined with the insulating oil quality code as one of the multi-source time series data. Based on the acoustic effect of partial discharge, a piezoelectric sensor installed in the transformer is used to detect partial discharges occurring in the transformer. At the same time interval, the dissolved gas time series are aligned to construct the partial discharge time series. The partial discharge time series is standardized and combined with the insulating oil quality code as one of the multi-source time series data. Multiple fiber optic temperature sensors are distributed throughout the transformer to detect the temperature of various parts of the transformer. At the same time interval, the dissolved gas time series are aligned to construct a temperature time series measured by the fiber optic temperature sensor group. The temperature time series is standardized and combined with the insulating oil quality code as one of the multi-source time series data. The power of the transformer is detected by a transmitter. The dissolved gas time series is aligned at the same time interval to construct the power time series. The power time series is combined with the insulating oil quality code as one of the multi-source time series data.
[0009] Furthermore, the multi-source time series data is segmented by a sliding window of a set size and then input into the transformer life prediction model.
[0010] Furthermore, the improved gated recurrent unit calculates the update gate at each time step T by the following formula , Reset Gate , candidate hidden states , and the hidden state of the output : ;in, represents the Sigmoid activation function, is the quantum Hilbert space encoding layer of the update gate, Represents a splicing operation, is the previous hidden state , is the current input; ;in, The quantum Hilbert space encoding layer for the reset gate; ;in, A quantum Hilbert space encoding layer for candidate hidden states; .
[0011] Furthermore, the quantum Hilbert space coding layer includes: Initialize the quantum bits corresponding to each type of time series data; configure several sets of trainable parameters corresponding to each quantum bit; Use each trainable parameter in the first set of trainable parameters as a rotation angle to construct the first rotation matrix around the X axis for each qubit: ; Each trainable parameter in the second set of trainable parameters is used as a rotation angle to construct the second rotation matrix corresponding to each qubit around the X axis: ; in, is any one of the first set of trainable parameters, is any one of the second set of trainable parameters, , is the index of the four trainable parameters in each set of trainable parameters, i is the imaginary unit; Each initial qubit is rotated about the X-axis using the corresponding first rotation matrix, so that all the resulting qubits are arranged in order. Recursive processing is performed using a controlled NOT gate to recursively establish associations between the qubits until a qubit cycle is completed. The controlled NOT gate is a two-qubit gate that selectively flips the state of the target qubit based on the state of the control qubit, thereby establishing associations between different quantum states. Then, the input of the quantum Hilbert space coding layer is used as the rotation angle to construct the rotation matrix corresponding to each quantum bit around the Z axis as the rotation coding matrix: ; in, is the input of the quantum Hilbert space coding layer corresponding to the quantum bit; Each new qubit is rotated by the corresponding rotation encoding matrix, thereby encoding the input of the quantum Hilbert space encoding layer into the qubit; After the input is encoded into the qubit, each newly obtained qubit is rotated by the corresponding second rotation matrix around the X axis, and all the qubits are arranged in order. Recursive processing is performed using controlled NOT gates to recursively establish associations between qubits until a qubit cycle is completed; Finally, the quantum bit is read out through the Pauli Y gate.
[0012] Furthermore, the timing modeling network includes multiple stacked improved gated recurrent units, the hidden state dimensions of the stacked improved gated recurrent units are gradually reduced, and a mapping layer for dimension adjustment is set between adjacent improved gated recurrent units; the final output of the timing modeling network is flattened and input into the fully connected network, and the fully connected network uses the fully connected layer to compress and expand the compressed feature dimension, and finally maps the feature into a one-dimensional transformer life.
[0013] In a second aspect, the present invention provides a data-driven transformer life assessment device, comprising: at least one processing unit, the processing unit being connected to a storage unit via a bus unit, the storage unit storing a computer program, and when the computer program is executed by the processing unit, the data-driven transformer life assessment method is implemented.
[0014] In a third aspect, the present invention provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the data-driven transformer life assessment method is implemented.
[0015] The above technical solution provided by the embodiment of the present invention has the following advantages compared with the prior art: The present invention obtains multi-source time series data for evaluating transformer life and the corresponding remaining life of the transformer as a data set during the use of the transformer. The multi-source time series data contains both the representation of transformer aging and the factors affecting transformer aging, and comprehensively integrates multiple aspects of data to evaluate the transformer life.
[0016] In the improved gated cyclic unit of the transformer life prediction model of the present invention, its linear layer is replaced by a quantum Hilbert space coding layer. The quantum Hilbert space coding layer encodes multi-source time series data in the quantum Hilbert space through quantum computing to capture the time series relationship between the multi-source time series data and the multi-source time series data and the transformer life. Through the above process, the quantum Hilbert space coding layer uses any trainable parameter as the rotation angle to obtain the rotation matrix of the quantum bit around the X-axis, and then recursively establishes connections between the quantum bits rotated around the X-axis by controlling the NOT gate until a quantum bit cycle is completed; the quantum Hilbert space coding layer input is used as the rotation angle to obtain the rotation matrix of the quantum bit around the Z-axis, and the quantum Hilbert space coding layer input is encoded into the quantum bit. The input is directly encoded into the Hilbert space through quantum rotation, and the relationship between different quantum states is established through multi-quantum state cyclic dependency, thereby enhancing the time series correlation; compared with the traditional linear layer, quantum parallelism and entanglement enhance feature extraction capabilities, especially when processing high-dimensional time series data. This application introduces nonlinear relationships through quantum rotation and selective entanglement. The nonlinear relationships required for modeling do not only rely on subsequent activation functions, thereby improving the effect of transformer life prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0019] Figure 1 A flowchart of a data-driven transformer life assessment method provided by an embodiment of the present invention; Figure 2 A schematic diagram of an oil-gas separator provided in an embodiment of the present invention; Figure 3 A schematic diagram of a transformer life prediction model provided by an embodiment of the present invention; Figure 4 A schematic diagram of an improved cyclic gating unit provided in an embodiment of the present invention; Figure 5 A schematic diagram of a data-driven transformer life assessment device provided by an embodiment of the present invention.
[0020] The numbers and meanings in the figure are as follows: 1. Oil-gas separator, 11. Separation chamber, 12. Exhaust interface, 13. Inlet pipe, 14. Liquid level sensor, 15. Thermal resistor rod, 16. Mixer, 17. Oil interface, 18. Temperature sensor. DETAILED DESCRIPTION
[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.
[0022] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.
[0023] Example 1 See Figure 1 As shown, this application provides a data-driven transformer life assessment method. To achieve accurate transformer life assessment, this application constructs a transformer life prediction model based on an improved gated cyclic unit. The overall solution includes model training and application of the trained model.
[0024] During the model training phase, the process includes: During the use of the transformer, multi-source time series data for evaluating the transformer life and the corresponding transformer remaining life are obtained.
[0025] From the time a transformer is put into operation to the end of its life, its multi-source time series data is collected, and the collected multi-source time series data is annotated with the remaining useful life of the transformer to form a data set for training a transformer life prediction model; the multi-source time series data are transformer-related parameters for evaluating transformer life. Specifically, the multi-source time series data include: dissolved gas time series, furan time series, partial discharge time series, temperature time series and power time series combined with insulating oil quality coding. In the multi-source time series data, different types of time series data are aligned in time and standardized.
[0026] From the time the transformer is put into operation to the end of its life, its multi-source time series data is collected, including: Dissolved gas detection is performed on the transformer insulating oil, and the dissolved gas detection results are constructed into a dissolved gas time series. After the dissolved gas time series is standardized, it is combined with the insulating oil quality coding to obtain one of the multi-source time series data.
[0027] During transformer operation, some abnormalities can cause the insulating oil to decompose and produce gases such as hydrogen, methane, ethane, ethylene, acetylene, carbon dioxide, and carbon monoxide. Acetylene production, for example, is often caused by arc discharge; methane and ethylene production is often caused by local overheating; and carbon monoxide and carbon dioxide are often generated by the thermal decomposition of the solid insulation material within the transformer. As the transformer ages, the probability of abnormal transformer behavior inevitably increases, and accordingly, the rate of gas generation in the insulating oil also accelerates. This gas generation rate can be measured through the time-series gas fraction, which can be used to assess the severity of transformer aging.
[0028] In the specific implementation process, the oil-gas separator 1 installed in the transformer is used to collect quantitative insulating oil in situ and separate oil and gas under constant temperature and pressure to perform dissolved gas detection and analysis to obtain the presence and percentage of hydrogen, methane, ethane, ethylene, acetylene, carbon dioxide and carbon monoxide gases in the oil sample. Figure 2 As shown, the oil-gas separator 1 includes a separation chamber 11. An oil port 17 for connecting to a transformer is provided at the bottom of the separation chamber 11. The oil port 17 is connected to the transformer via a valved pipeline. A stirrer 16 is provided at the bottom of the separation chamber 11 to promote oil-gas separation through stirring. A thermal resistor 15 and a temperature sensor 18 are provided within the separation chamber 11, which work together to maintain a constant temperature. A liquid level sensor 14 is provided on the sidewall of the separation chamber 11 to control the amount of oil introduced into the separation chamber 11, thereby achieving quantitative collection of insulating oil. An exhaust port 12 and an air inlet pipe 13 are provided at the top of the separation chamber 11. The exhaust port 12 is connected to a gas detection chamber, which is connected to a negative pressure generator for providing negative pressure. The gas detection chamber is equipped with a tunable diode laser and a photodetector. The tunable diode laser and photodetector perform absorption spectroscopy detection to determine the gas content in the quantitative insulating oil.
[0029] In addition to transformer aging, the factors that lead to the accelerated gas generation rate also include the deterioration of transformer insulating oil quality. In order to introduce a variable that characterizes the quality of transformer insulating oil, this application uses sine and cosine coding to encode the insulating oil quality to obtain the insulating oil quality code: ; ; Where d is the quality-encoded dimension, K is the frequency regulation technology, t is the age of the insulating oil, and j is the quality-encoded dimension index. Odd-numbered indices are coded using sin, and even-numbered indices are coded using cos. Insulating oil quality degrades with age, so age can be used to encode the quality of the insulating oil.
[0030] Furan detection was performed on transformer insulating oil. The furan detection results were aligned with the dissolved gas time series at the same time interval to construct a furan time series. In specific implementation, a fiber optic sensor was used to monitor the furan concentration in transformer oil online based on the characteristic absorption peak of furan molecules in the near-infrared region. Furans primarily result from the aging and decomposition of insulating materials. The aging rate of solid insulating materials is also affected by the quality of the insulating oil. The furan time series was standardized and combined with the insulating oil quality code to form part of the multi-source time series data.
[0031] Partial discharge is an important indicator of transformer insulation weakening. Based on the acoustic effect of partial discharge, piezoelectric sensors are used to detect partial discharge occurring in the transformer. At the same time interval, the dissolved gas time series are aligned to construct a partial discharge time series. Similarly, the increase in the rate of partial discharge occurrence is also affected by the quality of the insulating oil. Since the partial discharge time series and the dissolved gas time series are aligned, the corresponding insulating oil quality codes are consistent. After standardization, the partial discharge time series is combined with the insulating oil quality code as one of the multi-source time series data.
[0032] The thermal conditions within a transformer and its heat dissipation capacity are key factors affecting insulation degradation and ultimately the transformer's operating life. Multiple fiber optic temperature sensors are distributed throughout the transformer to monitor the temperature of various parts. At consistent time intervals, the dissolved gas time series are aligned to construct a temperature time series measured by the fiber optic temperature sensor group. This temperature time series, combined with the insulation oil quality code, is used as part of the multi-source time series data.
[0033] Transformer load is a key factor affecting its operating life. Transformer power is detected by a transmitter. At consistent time intervals, the dissolved gas time series are aligned to construct a power time series. This power time series, combined with the insulating oil quality code, is used as part of the multi-source time series data.
[0034] Insulating oil is a replaceable insulating material in transformers. Its lifespan is inconsistent with that of the transformer, and is much shorter than that of the transformer. During the transformer's life cycle, the insulating oil will undergo multiple replacements. Therefore, when evaluating the life of the transformer, this application encodes the quality of the regularly replaced insulating oil according to its usage time. Insulating oil quality coding is introduced to better model the impact of insulating oil quality on multi-source time series data reflecting the degree of transformer aging.
[0035] Multi-source time series data contains both the characteristics of transformer aging and the factors that affect transformer aging. Transformer life is evaluated by integrating multiple aspects of data.
[0036] In order to estimate the life of transformer, this application builds a transformer life prediction model, such as Figure 3As shown, the transformer life prediction model includes: a timing modeling network formed by an improved gated cyclic unit and a fully connected network for transformer life prediction based on the output of the timing modeling network; the timing modeling network contains multiple stacked improved gated cyclic units, the hidden state dimensions of the stacked improved gated cyclic units are gradually reduced, and a mapping layer for dimensional adjustment is set between adjacent improved gated cyclic units; the final output of the timing modeling network is flattened and input into the fully connected network, and the fully connected network uses the fully connected layer to compress and expand the compressed feature dimension, and finally maps the feature into a one-dimensional transformer life.
[0037] In the improved gated recurrent unit (IGRU), its linear layer is replaced with a quantum Hilbert space encoding layer that supports nonlinear modeling. This layer uses quantum computing to encode multi-source time series data and training parameters within the quantum Hilbert space to capture the temporal connections between multi-source time series data and between multi-source time series data and transformer life. Specifically, the transformer life prediction model includes multiple stacked IGRUs. The hidden state dimensions of these stacked IGRUs gradually decrease. To accommodate this dimensionality change, a dimensionality-adjusting mapping layer is provided between adjacent IGRUs.
[0038] In the specific implementation process, Figure 4 As shown, the improved gated recurrent unit calculates the update gate at each time step T by the following formula , Reset Gate , candidate hidden states , and the hidden state of the output : ;in, represents the Sigmoid activation function, is the quantum Hilbert space encoding layer of the update gate, Represents a splicing operation, is the previous hidden state , is the current input; ;in, The quantum Hilbert space encoding layer for the reset gate; ;in, A quantum Hilbert space encoding layer for candidate hidden states; .
[0039] The quantum Hilbert space coding layer includes: Initialize the quantum bits corresponding to each type of time series data; configure several sets of trainable parameters corresponding to each quantum bit; Use each trainable parameter in the first set of trainable parameters as a rotation angle to construct the first rotation matrix around the X axis for each qubit: ; Each trainable parameter in the second set of trainable parameters is used as a rotation angle to construct the second rotation matrix corresponding to each qubit around the X axis: ; in, is any one of the first set of trainable parameters, is any one of the second set of trainable parameters, , is the index of the four trainable parameters in each set of trainable parameters, i is the imaginary unit; Each initial quantum bit is rotated around the X-axis by the corresponding first rotation matrix, and all the quantum bits are arranged in order. Recursive processing is performed using a controlled NOT gate to recursively establish associations between the quantum bits until a quantum bit cycle is completed. Given five qubits, QAQBQCQDQE, first use QA as the control qubit and QB as the target qubit. A controlled NOT gate, acting as a two-qubit gate, selectively flips the target qubit's state based on the control qubit's state, updating QB. Then, using the updated QB as the control qubit and QC as the target qubit, a controlled NOT gate selectively flips the target qubit's state based on the control qubit's state, updating QC. Then, using the updated QC as the control qubit and QD as the target qubit, a controlled NOT gate selectively flips the target qubit's state based on the control qubit's state, updating QD. Then, using the updated QD as the control qubit and QF as the target qubit, a controlled NOT gate selectively flips the target qubit's state based on the control qubit's state, updating QF. Then, using the updated QF as the control qubit and QA as the target qubit, a controlled NOT gate selectively flips the target qubit's state based on the control qubit's state, updating QA. This forms a multi-quantum state cyclic dependency to establish connections between different quantum states and enhance temporal correlation.
[0040] Then, the input of the quantum Hilbert space coding layer is used as the rotation angle to construct the rotation matrix corresponding to each quantum bit around the Z axis as the rotation coding matrix: ; in, is the input of the quantum Hilbert space coding layer corresponding to the quantum bit; Each new qubit is rotated by the corresponding rotation encoding matrix, thereby encoding the input of the quantum Hilbert space coding layer into the qubit.
[0041] After the input is encoded into the quantum bit, each newly obtained quantum bit is rotated by the corresponding second rotation matrix around the X-axis, and all the quantum bits are arranged in sequence. Recursive processing is performed using controlled NOT gates to recursively establish associations between quantum bits until a quantum bit cycle is completed.
[0042] Finally, the quantum bit is read out through the Pauli Y gate.
[0043] Through the above process, the quantum Hilbert space encoding layer uses any trainable parameter as the rotation angle to obtain the rotation matrix of the qubit about the X-axis. Then, by controlling the NOT gate, recursively establishes connections between the qubits rotated about the X-axis until a qubit cycle is completed. The quantum Hilbert space encoding layer input is used as the rotation angle to obtain the rotation matrix of the qubit about the Z-axis. The quantum Hilbert space encoding layer input is encoded into the qubit. The input is directly encoded into the Hilbert space through quantum rotation, supporting the capture of more complex patterns. The qubit is read out using the Pauli Y gate. In Hilbert space, quantum computing, combined with trainable parameters and input, is used to transform quantum states, effectively modeling the complex connections between individual quantum states, thereby capturing factors that characterize transformer life or the complex correlations between variables affecting transformer life and transformer life. Compared to traditional linear layers, quantum parallelism and entanglement enhance feature extraction capabilities, especially when processing high-dimensional time series data. This application introduces nonlinear relationships through quantum rotation and selective entanglement. The required nonlinear relationships are not solely dependent on subsequent activation functions, resulting in better results.
[0044] The transformer life prediction model is trained using the acquired data set, and the mean square error loss between the predicted life and the actual life is used as the loss function. The parameters of the transformer life prediction model are dynamically adjusted with the goal of minimizing the loss function.
[0045] During the application phase, transformer parameters are collected in real time and preprocessed to obtain multi-source time series data. The trained transformer life prediction model uses multi-source time series data to predict life.
[0046] This application captures the complex correlation of multi-source time series data in quantum Hilbert space, requires fewer training parameters than conventional recursive neural networks, and optimizes transformer prediction results.
[0047] Example 2 See Figure 5As shown, an embodiment of the present invention provides a data-driven transformer life assessment device, comprising: at least one processing unit, the processing unit connected to a storage unit via a bus unit, the storage unit being a computer-readable storage medium that can be used to store software programs, computer executable programs, and modules, such as the software programs, computer executable programs, and modules corresponding to a data-driven transformer life assessment method in an embodiment of the present invention. The processing unit implements the aforementioned data-driven transformer life assessment method by running the software programs, computer executable programs, and modules stored in the storage unit, comprising: During the service life of the transformer, multi-source time series data for transformer life assessment and the corresponding remaining life of the transformer are obtained to form a data set for training the transformer life prediction model. The multi-source time series data includes: dissolved gas time series, furan time series, partial discharge time series, temperature time series, and power time series combined with insulating oil quality coding. In the multi-source time series data, different types of time series data are aligned in time and standardized. A transformer life prediction model is constructed, comprising a time series modeling network formed by an improved gated recurrent unit and a fully connected network for transformer life prediction based on the output of the time series modeling network. In the improved gated recurrent unit, its linear layer is replaced with a quantum Hilbert space coding layer that supports nonlinear modeling. The quantum Hilbert space coding layer uses quantum computing to encode multi-source time series data and training parameters in the quantum Hilbert space to capture the temporal connections between the multi-source time series data and between the multi-source time series data and the transformer life. The transformer life prediction model is trained using the acquired data set, the mean square error loss between the predicted life and the actual life is used as a loss function, and the parameters of the transformer life prediction model are dynamically adjusted with the goal of minimizing the loss function; During the application phase, transformer parameters are collected in real time and preprocessed to obtain multi-source time series data. The trained transformer life prediction model uses multi-source time series data to predict life.
[0048] Of course, the computer program stored in the storage unit of the data-driven transformer life assessment device provided in an embodiment of the present invention is not limited to the method operations described above, and can also execute related operations in the data-driven transformer life assessment method provided in any embodiment of the present invention.
[0049] Example 3 An embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed, the data-driven transformer life assessment method is implemented, including: During the service life of the transformer, multi-source time series data for transformer life assessment and the corresponding remaining life of the transformer are obtained to form a data set for training the transformer life prediction model. The multi-source time series data includes: dissolved gas time series, furan time series, partial discharge time series, temperature time series, and power time series combined with insulating oil quality coding. In the multi-source time series data, different types of time series data are aligned in time and standardized. A transformer life prediction model is constructed, comprising a time series modeling network formed by an improved gated recurrent unit and a fully connected network for transformer life prediction based on the output of the time series modeling network. In the improved gated recurrent unit, its linear layer is replaced with a quantum Hilbert space coding layer that supports nonlinear modeling. The quantum Hilbert space coding layer uses quantum computing to encode multi-source time series data and training parameters in the quantum Hilbert space to capture the temporal connections between the multi-source time series data and between the multi-source time series data and the transformer life. The transformer life prediction model is trained using the acquired data set, the mean square error loss between the predicted life and the actual life is used as a loss function, and the parameters of the transformer life prediction model are dynamically adjusted with the goal of minimizing the loss function; During the application phase, transformer parameters are collected in real time and preprocessed to obtain multi-source time series data. The trained transformer life prediction model uses multi-source time series data to predict life.
[0050] An embodiment of the present invention provides a computer-readable storage medium, in which the computer program stored is not limited to the method operations described above, but can also execute related operations in a data-driven transformer life assessment method provided by any embodiment of the present invention.
[0051] In the embodiments provided by the present invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, structure or unit, which can be electrical, mechanical or other forms.
[0052] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.
[0053] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.
[0054] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is intended to be accorded the widest scope consistent with the principles and novel features claimed herein.
Claims
1. A data-driven transformer life assessment method, characterized in that: include: During the service life of the transformer, multi-source time series data for transformer life assessment and the corresponding remaining life of the transformer are obtained to form a data set for training the transformer life prediction model. The multi-source time series data includes: dissolved gas time series, furan time series, partial discharge time series, temperature time series, and power time series combined with insulating oil quality coding. In the multi-source time series data, different types of time series data are aligned in time and standardized. A transformer life prediction model is constructed, comprising a time series modeling network formed by an improved gated recurrent unit and a fully connected network for transformer life prediction based on the output of the time series modeling network. In the improved gated recurrent unit, its linear layer is replaced with a quantum Hilbert space coding layer that supports nonlinear modeling. The quantum Hilbert space coding layer uses quantum computing to encode multi-source time series data and training parameters in the quantum Hilbert space to capture the temporal connections between the multi-source time series data and between the multi-source time series data and the transformer life. The transformer life prediction model is trained using the acquired data set, the mean square error loss between the predicted life and the actual life is used as a loss function, and the parameters of the transformer life prediction model are dynamically adjusted with the goal of minimizing the loss function; During the application phase, transformer parameters are collected in real time and preprocessed to obtain multi-source time series data. The trained transformer life prediction model uses multi-source time series data to predict life.
2. The data-driven transformer life assessment method according to claim 1, characterized in that: For insulating oil, which will be replaced regularly during the life cycle of the transformer, the quality of the insulating oil is coded according to the usage time to model the impact of insulating oil quality on different types of time series data reflecting the degree of transformer aging. The insulating oil quality is coded using sine and cosine coding to obtain the insulating oil quality code: ； ； Where: d is the dimension after quality coding, K is the frequency regulation technology, t is the service life of the insulating oil, and j is the dimension index of the quality coding.
3. The data-driven transformer life assessment method according to claim 1, characterized in that: An oil-gas separator provided on a transformer is used to collect quantitative insulating oil in situ and separate oil and gas at constant temperature and pressure, so as to detect dissolved gas and obtain a dissolved gas time series. The oil-gas separator (1) comprises a separation chamber (11), an oil-passing interface (17) for connecting to the transformer is provided at the bottom of the separation chamber (11), and the oil-passing interface (17) is connected to the transformer via a pipeline provided with a valve; a stirrer (16) is provided at the bottom of the separation chamber (11); a thermal resistor rod (15) and a temperature sensor (18) are provided in the separation chamber (11); the separation chamber (11) is provided with a temperature sensor (18); and the separation chamber (11) is provided with a temperature sensor (18). A liquid level sensor (14) is provided on the side wall of the chamber (11); an exhaust interface (12) and an air inlet pipe (13) are provided on the top of the separation chamber (11); the exhaust interface (12) is connected to a gas detection chamber; the gas detection chamber is connected to a negative pressure machine for providing negative pressure; a tunable diode laser and a photodetector are provided in the gas detection chamber; the tunable diode laser and the photodetector perform dissolved gas detection on the transformer insulating oil; the dissolved gas detection result is constructed into a dissolved gas time series; after standardization, one of the multi-source time series data is obtained in combination with the insulating oil quality coding.
4. The data-driven transformer life assessment method according to claim 1, characterized in that: Based on the characteristic absorption peak of furan molecules in the near-infrared region, near-infrared lasers and fiber optic sensors were used to monitor the furan concentration in transformer oil online. The furan detection results were aligned with the dissolved gas time series at the same time interval to construct a furan time series. The furan time series was standardized and combined with the insulating oil quality code to form one of the multi-source time series data. Based on the acoustic effect of partial discharge, a piezoelectric sensor installed in the transformer is used to detect partial discharges occurring in the transformer. At the same time interval, the dissolved gas time series are aligned to construct the partial discharge time series. The partial discharge time series is standardized and combined with the insulating oil quality code as one of the multi-source time series data. Multiple fiber optic temperature sensors are distributed throughout the transformer to detect the temperature of various parts of the transformer. At the same time interval, the dissolved gas time series are aligned to construct a temperature time series measured by the fiber optic temperature sensor group. The temperature time series is standardized and combined with the insulating oil quality code as one of the multi-source time series data. The power of the transformer is detected by a transmitter. The dissolved gas time series is aligned at the same time interval to construct the power time series. The power time series is combined with the insulating oil quality code as one of the multi-source time series data.
5. The data-driven transformer life assessment method according to claim 1, characterized in that: Multi-source time series data are segmented by a sliding window of a set size and then input into the transformer life prediction model.
6. The data-driven transformer life assessment method according to claim 1, characterized in that: The improved gated recurrent unit calculates the update gate at each time step T by the following formula , Reset Gate , candidate hidden states , and the hidden state of the output : ;in, represents the Sigmoid activation function, is the quantum Hilbert space encoding layer of the update gate, Represents a splicing operation, is the previous hidden state , is the current input; ;in, The quantum Hilbert space encoding layer for the reset gate; ;in, A quantum Hilbert space encoding layer for candidate hidden states; 。 7. The data-driven transformer life assessment method according to claim 6, characterized in that: The quantum Hilbert space coding layer includes: Initialize the quantum bits corresponding to each type of time series data; configure several sets of trainable parameters corresponding to each quantum bit; Use each trainable parameter in the first set of trainable parameters as a rotation angle to construct the first rotation matrix around the X axis for each qubit: ； Each trainable parameter in the second set of trainable parameters is used as a rotation angle to construct the second rotation matrix corresponding to each qubit around the X axis: ； in, is any one of the first set of trainable parameters, is any one of the second set of trainable parameters, , is the index of the four trainable parameters in each set of trainable parameters, i is the imaginary unit; Each initial qubit is rotated about the X-axis using the corresponding first rotation matrix, so that all the resulting qubits are arranged in order. Recursive processing is performed using a controlled NOT gate to recursively establish associations between the qubits until a qubit cycle is completed. The controlled NOT gate is a two-qubit gate that selectively flips the state of the target qubit based on the state of the control qubit, thereby establishing associations between different quantum states. Then, the input of the quantum Hilbert space coding layer is used as the rotation angle to construct the rotation matrix corresponding to each quantum bit around the Z axis as the rotation coding matrix: ； in, is the input of the quantum Hilbert space coding layer corresponding to the quantum bit; Each new qubit is rotated by the corresponding rotation encoding matrix, thereby encoding the input of the quantum Hilbert space encoding layer into the qubit; After the input is encoded into the qubit, each newly obtained qubit is rotated by the corresponding second rotation matrix around the X axis, and all the qubits are arranged in order. Recursive processing is performed using controlled NOT gates to recursively establish associations between qubits until a qubit cycle is completed; Finally, the quantum bit is read out through the Pauli Y gate.
8. The data-driven transformer life assessment method according to claim 1, characterized in that: The timing modeling network includes multiple stacked improved gated recurrent units, the hidden state dimensions of the stacked improved gated recurrent units are gradually reduced, and a mapping layer for dimension adjustment is set between adjacent improved gated recurrent units; the final output of the timing modeling network is flattened and input into the fully connected network, and the fully connected network uses the fully connected layer to compress and expand the compressed feature dimension, and finally maps the feature into a one-dimensional transformer life.
9. A data-driven transformer life assessment device, characterized in that: include: At least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the data-driven transformer life assessment method according to any of claims 1-8 is implemented.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the data-driven transformer life assessment method according to any one of claims 1 to 8 is implemented.