Transformer multi-parameter time series data prediction method and device, electronic equipment and storage medium

Abstract

The invention provides a transformer multi-parameter time series data prediction method and device, electronic equipment and a storage medium. The method comprises the steps that multi-parameter operation parameters in the working process of a transformer are collected; preprocessing the multi-parameter operation parameters to form a multi-parameter time sequence data set; performing feature extraction on data in the time series data set through a local convolution module and a global attention module to obtain stable local features and global features; based on the stable local features and global features, collaborative extraction of time domain local dynamic features and frequency domain frequency distribution features is carried out through a frequency domain enhanced expansion feedforward network, and time-frequency fusion features are generated; projecting the time-frequency fusion feature to a multi-parameter prediction space to obtain a multi-parameter prediction result; the method has the beneficial effect of high-precision prediction of non-stationary and multi-parameter time series data; the method is suitable for the technical field of power equipment intelligent monitoring and data analysis.

Description

Technical Field

[0001] This application relates to the technical field of intelligent monitoring and data analysis of power equipment, specifically to methods, devices, electronic equipment and storage media for predicting multi-parameter time-series data of transformers. Background Technology

[0002] Transformers are key equipment for power transmission and conversion, and their safe operation directly determines the continuity, reliability and stability of the power system. During the operation of transformers, multi-parameter sensors are usually used to collect data such as transformer temperature, partial discharge and oil gas concentration in real time, so as to predict the trend and assess the condition of transformer operation.

[0003] The transformer operation data collected by existing multi-parameter sensors are mostly time-series data. These time-series data have high noise, multivariate coupling and non-stationary characteristics. Traditional multi-parameter time-series prediction methods are difficult to capture complex time-series dependencies, resulting in insufficient prediction accuracy.

[0004] Specifically, existing multi-parameter time series prediction methods typically have the following technical limitations: (1) The temporal convolution model based on dilated convolution is an important model in the field of time series prediction. Its core advantage comes from the characteristic of dilated convolution to "expand the receptive field without adding parameters", that is, by controlling the sampling interval through the dilation rate, the receptive field of each convolution layer is exponentially superimposed without changing the configuration of the convolution kernel parameters. However, this model is limited by the inherent contradiction between the convolution structure and the characteristics of time series data, which makes it possible to capture a certain range of time dependence, but lacks the ability to model the dynamic interaction between parameters and cannot reflect the coupling relationship between temperature, pressure and discharge in the transformer. (2) Transformer has become one of the mainstream models in the field of time series prediction because of its global dependency capture capability through the self-attention mechanism. It performs particularly well in long sequences and complex dependencies (such as multi-factor interaction and nonlinear trends). However, although the traditional Transformer structure has the advantage of global modeling, its self-attention mechanism is not sensitive enough to local sudden signals and has high computational complexity, making it unsuitable for high-frequency continuous monitoring of transformers. In summary, existing methods do not address optical measurement data distributed across multiple sensors and locations, lack a collaborative feature fusion mechanism, leading to a separation between local patterns and global trends, and resulting in a lack of physical consistency in the predicted output.

[0005] Therefore, a method, device, electronic equipment, and storage medium for predicting transformer multi-parameter time-series data that can improve the availability and prediction accuracy of online transformer monitoring data is particularly important. Summary of the Invention

[0006] To address one of the aforementioned technical deficiencies, this application provides a method, apparatus, electronic device, and storage medium for predicting multi-parameter time-series data of transformers.

[0007] According to the first aspect of this application, a method for predicting multi-parameter time-series data of a transformer is provided, comprising: Collect multiple operating parameters during the transformer's operation; Preprocess the multi-parameter operating parameters to form a multi-parameter time series dataset; Stable local and global features are obtained by extracting features from the time-series dataset through local convolutional modules and global attention modules. Based on stable local and global features, a frequency-domain enhanced extended feedforward network is used to collaboratively extract time-domain local dynamic features and frequency-domain frequency distribution features to generate time-frequency fusion features. The time-frequency fusion features are projected onto the multi-parameter prediction space to obtain the multi-parameter prediction results.

[0008] In an optional embodiment of this application, the step of extracting features from the time-series dataset using a local convolutional module and a global attention module to obtain stable local and global features includes: Stable local features of the time-series dataset are extracted using a multi-scale dilated causal convolution module and a residual gating module. By using adaptive positional encoding and an improved temporal cross-attention mechanism, long-range dependencies between different time slices and coupling relationships between different physical parameters are captured in stable local features, forming global features.

[0009] In an optional embodiment of this application, the extraction of stable local features from the time-series dataset using a multi-scale dilated causal convolution module and a residual gating module includes: A multi-scale dilated causal convolution module is used to extract scale features of time-series datasets in parallel at different dilation rates; and the scale features corresponding to each dilation rate are fused by channel to form local features; A residual gating module is used to suppress interference information in local features and dynamically select effective features from the local features to form stable local features.

[0010] In an optional embodiment of this application, the transformer multi-parameter time-series data prediction method further includes: The multi-parameter prediction space is trained by using a multi-scale weighted loss function to obtain a trained multi-parameter prediction space.

[0011] In one optional embodiment of this application, the multi-scale weighted loss function includes: giving higher weights to short-term prediction steps; and using smoothing weights for medium- and long-term prediction steps.

[0012] In one optional embodiment of this application, the preprocessing of the multi-parameter operating parameters includes: noise reduction, normalization, and sliding window construction.

[0013] In an optional embodiment of this application, the improved temporal cross-attention mechanism includes: In the time channel, a time-dimensional cross-attention mechanism is used to aggregate attention on stable local features in the time dimension, resulting in a global time-dependent representation: On the variable channel, a variable-dimensional cross-attention mechanism is used to perform cross-attention aggregation of different parameters on the variable dimension at the same time step to obtain a global variable dependency representation; By fusing the time channel output and the variable channel output using a weighted algorithm, the global features of the dual-channel fusion are obtained.

[0014] According to a second aspect of this application, a transformer multi-parameter time-series data prediction apparatus is provided, including a module for implementing the transformer multi-parameter time-series data prediction method as described in any of the preceding claims.

[0015] According to a third aspect of this application, an electronic device is provided, comprising: Memory; Processor; and Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the transformer multi-parameter timing data prediction method as described in any of the preceding claims.

[0016] According to a fourth aspect of this application, a computer-readable storage medium is provided, characterized in that: a computer program is stored thereon; the computer program is executed by a processor to implement the transformer multi-parameter timing data prediction method as described in any of the preceding claims.

[0017] Using the transformer multi-parameter time series data prediction method, apparatus, electronic equipment, and storage medium provided in this application, a complete modeling chain of "local → global → interaction → collaboration" is constructed: a local convolution module captures short-term abrupt signals to generate local features; a global attention module captures long-term dependencies across time scales and realizes dynamic coupling between multiple parameters to generate global features; a frequency-domain enhanced extended feedforward network performs collaborative extraction of time-domain local dynamic features and frequency-domain frequency distribution features based on local and global features to generate time-frequency fusion features; finally, a joint prediction result that satisfies physical consistency is output through multi-parameter prediction space, thereby achieving high-precision prediction of non-stationary, multi-parameter time series data.

[0018] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of what is pointed out in the written description and the accompanying drawings. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart of a transformer multi-parameter time-series data prediction method provided in one embodiment of this application; Figure 2 This is a flowchart of step S30 in a transformer multi-parameter time-series data prediction method provided in one embodiment of this application; Figure 3 This is a flowchart of step S301 in a transformer multi-parameter time-series data prediction method provided in one embodiment of this application; Figure 4 This is a flowchart of step S302 in a transformer multi-parameter time-series data prediction method provided in one embodiment of this application; Figure 5 This is a schematic diagram of the structure of a transformer multi-parameter time-series data prediction device provided in one embodiment of this application; Figure 6 This is a comparison chart of the predicted curve and the actual measured value of the temperature parameter in one embodiment of this application; In the picture: 10 is the acquisition module, 20 is the preprocessing module, 30 is the feature extraction module, 40 is the feature fusion module, and 50 is the collaborative prediction module. Detailed Implementation

[0020] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0021] In existing technologies, transformer multi-parameter time-series prediction is based on multi-dimensional time-series data during its operation. The goal is to predict the future trends of key state parameters (such as top oil temperature, winding temperature, and dissolved gas concentration in the oil) to support condition monitoring and fault early warning. Compared with multi-parameter time-series prediction in general scenarios, the strong operating condition coupling of power equipment, the complex interference in industrial sites, and the scarcity of fault data make it difficult for traditional prediction methods to capture complex time-series dependencies, resulting in insufficient prediction accuracy.

[0022] To address some problems in existing technologies, this application proposes a prediction method that combines local and global approaches. This method is applicable to trend prediction and state assessment of high-frequency monitoring data with multiple parameters such as temperature, pressure, and partial discharge. Addressing the characteristics of transformer operating data, such as high noise, strong nonlinearity, non-stationarity, and multivariate coupling, this method integrates local time-series feature extraction and global dependency learning, along with cross-attention mechanisms and adaptive coding functions. It can accurately predict transformer multi-parameter optical sensing time-series data, offering advantages such as local-global feature fusion, multi-parameter collaborative interpretation, and long- and short-term error balancing.

[0023] In a first aspect, embodiments of this application provide a method for predicting transformer multi-parameter time-series data. This method can be executed by a transformer multi-parameter time-series data prediction device, or by components configured inside the transformer multi-parameter time-series data prediction device, such as chips or chip systems, or by logic modules or software having some or all of the functions of the transformer multi-parameter time-series data prediction device. This application does not limit this method.

[0024] For example, please refer to Figure 1 Transformer multi-parameter time-series data prediction methods include: S10 collects multiple operating parameters during the transformer's operation; S20, preprocess the multi-parameter operating parameters to form a multi-parameter time series dataset; S30 extracts features from the time-series dataset through local convolutional modules and global attention modules to obtain stable local and global features. S40, based on stable local and global features, uses a frequency-domain enhanced extended feedforward network to collaboratively extract time-domain local dynamic features and frequency-domain frequency distribution features, generating time-frequency fusion features; S50 projects the time-frequency fusion features onto the multi-parameter prediction space to obtain the multi-parameter prediction results.

[0025] In this embodiment of the application, the multi-parameter operating parameters include: temperature, partial discharge, and gas concentration in the oil.

[0026] Based on the above scheme, the transformer multi-parameter time series data prediction method provided in this application constructs a complete modeling link of "local → global → interaction → collaboration": a local convolution module captures short-term abrupt signals to generate local features, a global attention module captures long-term dependencies across time scales and realizes dynamic coupling between multiple parameters to generate global features; a frequency-domain enhanced extended feedforward network performs collaborative extraction of time-domain local dynamic features and frequency-domain frequency distribution features based on local and global features to generate time-frequency fusion features; finally, a joint prediction result that satisfies physical consistency is output through multi-parameter prediction space, thereby achieving high-precision prediction of non-stationary, multi-parameter time series data.

[0027] In some possible implementations of the first aspect, S20 preprocesses the multi-parameter operating parameters to form a multi-parameter time-series dataset, including: denoising, normalization, and sliding window construction.

[0028] Optionally, let the expression for a multi-parameter time series dataset be:

[0029] Where X represents the input matrix, T represents the sampling time, and N represents multiple parameter dimensions (such as temperature, pressure, partial discharge, etc.). Represents a number field; For example: T=5 represents sampling 5 points; N=3 represents three dimensions: temperature, pressure, and partial discharge, then X can be: 20 21 22 20 21 (℃) 10 11 12 10 11 (pa) 0.1 0.1 0 0.1 0 (Discharge quantity).

[0030] Optionally, in the embodiments of this application, wavelet transform, Savitzy-Golay filtering or adaptive filtering can be used for noise reduction to suppress high-frequency noise and instrument spikes and improve the signal-to-noise ratio.

[0031] Optionally, in this embodiment, each parameter is normalized, and its calculation expression is as follows:

[0032] in, , These are the mean and standard deviation of the nth parameter, respectively, used to eliminate dimensional differences and amplitude dominance issues.

[0033] Optionally, in this embodiment of the application, supervised samples can be constructed using a sliding window to divide the time series into input-output pairs; The input is represented as: The output is represented as: ; Let τ represent the prediction window length.

[0034] Based on the above scheme, the embodiments of this application, through preprocessing, transform the original multi-parameter operating parameters that are "noisy, have disordered dimensions, and lack structure" into high-quality model input data, providing a solid foundation of data for improving prediction accuracy.

[0035] Please see Figure 2 In some possible implementations of the first aspect, S30 involves extracting features from the data in the time-series dataset using a local convolutional module and a global attention module to obtain stable local and global features; including: S301 extracts stable local features from time-series datasets through a multi-scale dilated causal convolution module and a residual gating module. S302 captures long-range dependencies between different time slices and coupling relationships between different physical parameters in stable local features through adaptive positional encoding and an improved temporal cross-attention mechanism, forming global features.

[0036] To fully capture short-term mutations, partial discharge impacts, and high-frequency perturbations, this application employs a multi-scale dilated causal convolution module to extract local features in parallel at different dilation rates.

[0037] Optionally, please refer to Figure 3 In step S301, stable local features of the time-series dataset are extracted through a multi-scale dilated causal convolution module and a residual gating module, including: S301-1 employs a multi-scale dilated causal convolution module to extract scale features of time-series datasets in parallel at different dilation rates; and then fuses the scale features corresponding to each dilation rate by channel to form local features. Specifically, the expression for the output of the s-th dilation branch in the l-th layer of the multi-scale dilated causal convolution module is: ； in: ; This represents the convolution kernel of the s-th dilatational branch in the l-th layer, where k represents the kernel length. This indicates the number of input channels (corresponding to transformer parameters such as temperature, partial discharge, and gas concentration in the oil). This represents the number of output channels after feature extraction; The difference between convolution kernels in different dilation branches is the dilation rate, which does not change the physical length k of the convolution kernel, but changes the sampling stride of the convolution kernel on the temporal sequence; the dilation factor of the s-th dilation branch is... , .

[0038] When the dilation rate d=1, it is a standard convolution that captures local features at the smallest scale (e.g., a single pulse of partial discharge, with a time span of 1-2 sampling points). When the expansion rate d=2, sampling at every 1 point expands the receptive field, enabling the capture of mesoscale fluctuations (such as the temperature change trend over 5-10 minutes). When the expansion rate d=4, sampling every 3 points results in a larger receptive field, which can capture large-scale trends (such as hourly temperature changes, periodicity of discharge activity, etc.).

[0039] Multiple expansion branches form a multi-scale local receptive field.

[0040] Furthermore, features at each scale can be spliced ​​or weighted and fused by channel, and the calculation expression is as follows: .

[0041] S301-2 uses a residual gating module to suppress interference information in local features and dynamically selects effective features from the local features to form stable local features.

[0042] Specifically, in the residual gating module, the expression for the gated linear unit (GLU) is:

[0043] Where σ represents the Sigmoid function, and ⊙ represents element-wise multiplication, used to dynamically select useful features.

[0044] Indicates input features, , Both represent linear transformation matrices; This means that the input features are transformed through two independent linear transformations, generating candidate features in one path. Another path generates gating weights. Ultimately, by multiplying elements, the gating weights adaptively retain useful features and suppress useless features.

[0045] Furthermore, by adjusting the channel dimension through 1×1 convolution, the input X and the local gate control output are made additive, forming a residual connection:

[0046] in, Indicates to 1×1 convolution adjustment.

[0047] In this embodiment, the residual gating module can ensure that local features are not "washed out" in deep stacking, providing a stable local basis for subsequent global and cross-parameter modeling, and enhancing the coherence of the overall algorithm.

[0048] Based on the above scheme, the multi-scale dilated causal convolution module can ensure that the convolution at any time depends only on the current and past samples, strictly avoiding the leakage of future information. Compared with the traditional single dilation rate, the multi-scale parallel structure can more finely characterize local patterns at different time scales such as "transient spikes - short-term fluctuations - medium-term fluctuations", which is beneficial for the subsequent global attention layer to model based on richer local semantics. The local features extracted by the multi-scale dilated causal convolution module can capture transient changes, short-term disturbances and high-frequency fluctuations in the same receptive field, making the model more sensitive to the rapidly changing parts of temperature, pressure and partial discharge signals, thereby significantly improving the accuracy and stability of short-term prediction.

[0049] Furthermore, after multi-scale dilated causal convolution, this application embodiment constructs a residual gating module to ensure the stable transmission of local information in the deep network and suppress irrelevant features, further guaranteeing the stable transmission of local information in the deep network, so that the extracted stable local features can enter the global learning stage without loss, forming a more coherent temporal representation.

[0050] Optionally, please refer to Figure 4 In S302, the improved temporal cross-attention mechanism includes: S302-1, in the time channel, employs a time-dimensional cross-attention mechanism to aggregate attention on stable local features in the time dimension, obtaining a global time-dependent representation: S302-2, on the variable channel, a variable dimension cross-attention mechanism is used to perform cross-attention aggregation of different parameters on the variable dimension at the same time step to obtain a global variable dependency representation; S302-3 uses a weighted algorithm to fuse the time channel output and the variable channel output to obtain the global features of dual-channel fusion.

[0051] Optionally, considering the characteristics of transformer operating data such as day-night cycles, gradual load changes, and seasonal overlap, in this embodiment of the application, a learnable positional encoding method, rather than a fixed trigonometric function encoding, is used to map the time positional information into a trainable vector; the expression for the adaptive positional encoding is:

[0052] in, represents adaptive location encoding; P represents adaptive update during training, which can automatically learn complex time patterns such as day-night cycles, weekend / weekday differences, and temperature-pressure hysteresis relationships. Based on the above scheme, compared with the fixed position encoding of the traditional Transformer, the embodiments of this application significantly enhance the ability to characterize long-term, slowly changing time series patterns through adaptive position encoding, which is more suitable for the operating data of slow time-varying equipment such as transformers.

[0053] In this embodiment, based on the traditional attention modeling only time dependence, a dual-channel structure of "time dimension cross attention + variable dimension cross attention" is introduced to form explicit modeling; including: long-range dependence between different time slices (such as multi-day load mode) and coupling relationship between different physical parameters (such as temperature rise → pressure change, partial discharge enhancement → local temperature fluctuation).

[0054] Optionally, in S302-1, the temporal cross-attention mechanism uses the output of the residual gating module as the Value and the features with positional encoding as the Query and Key, respectively, to perform attention aggregation in the temporal dimension to obtain a global temporal dependency representation, the calculation expression of which is:

[0055] in, These represent the query, key, and value that are aggregated for attention over the time dimension, respectively. The proposed time-dimensional cross-attention mechanism calculates and aggregates attention weights with the "time dimension" as its core, thereby breaking through the local receptive field limitation of convolutional / recurrent models and capturing the global temporal dependencies across all time points in a time sequence.

[0056] Optionally, in S302-2, the variable-dimensional cross-attention mechanism treats different parameters at the same time step as "variable nodes," captures the nonlinear coupling relationship between different monitored variables in the variable dimension, and aggregates the correlation information of the variable dimension through adaptive weight allocation to obtain a global variable dependency representation that integrates multi-parameter interaction features (explicitly learning the correlation and influence weights between temperature, pressure, and partial discharge); its calculation expression is:

[0057] in, These represent queries, keys, and values ​​that perform attention aggregation along the variable dimension, respectively.

[0058] Based on the above scheme, the time-dimensional cross-attention mechanism can focus on the global time dependence of the same variable at different times (such as oil temperature at...). ~ The correlation at different times); the variable-dimensional cross-attention mechanism can focus on the global variable dependencies of different variables at the same time (such as...). Constant load current, oil temperature, (Concentration correlation); the two mechanisms complement each other, perfectly addressing the core issue of multi-parameter time-series prediction for transformers, and solving the complex coupling problem between heterogeneous variables such as current, voltage, oil temperature, and dissolved gases in the oil.

[0059] Furthermore, in step S302-3, the time channel output and the variable channel output are fused using a weighted algorithm to obtain the global features of the dual-channel fusion, the calculation expression of which is:

[0060] in, α represents the global features of dual-channel fusion, and α represents the weights that can be learned or manually set.

[0061] Compared to ordinary multi-head attention which only unfolds in the time dimension, the embodiments of this application solve the modeling problems of "local + global" and "time + quantity" simultaneously through dual-channel cross attention.

[0062] Based on the above scheme, in terms of global modeling, this application embodiment introduces adaptive position encoding and an improved time cross-attention mechanism, enabling the model to actively learn the day-night cycle, load trend and slow time-varying pattern in transformer operation, thereby effectively enhancing the modeling ability for long-term dependencies. Compared with the traditional Transformer, it is more adaptable to non-stationary and highly periodic operating data.

[0063] Meanwhile, the variable-dimensional cross-attention proposed in this application can explicitly model the dynamic coupling relationship between multiple parameters such as temperature, pressure, and partial discharge, enabling the model to learn real physical mechanisms rather than relying solely on statistical correlation. This avoids physical contradictions in the prediction results of different parameters and significantly improves the consistency and interpretability of the overall prediction.

[0064] In some possible implementations of the first aspect, based on the fused global features, embodiments of this application introduce a frequency-domain enhanced extended feedforward network, which can take into account information from different frequency bands. First, a one-dimensional FFT or short-time FFT is performed on the features in the time dimension to extract the main frequency band components, used to emphasize periodic / oscillatory information. Then, a nonlinear mapping is performed through a two-layer dilated convolution + GELU activated FFN structure. Its calculation expression is:

[0065] in, This represents the enhanced feature obtained by nonlinear transformation of the features from the previous layer. , Let represent the convolution weight matrices with expansion factors of 1 and 2, respectively.

[0066] Based on the above scheme, after global feature fusion, this application embodiment introduces a frequency domain enhanced feedforward network, which draws on the idea of ​​frequency domain enhanced TCN. It retains local information in the time domain and amplifies the dominant periodic components in the frequency domain. This enables the model to characterize data characteristics from both the time and frequency domains. In particular, it shows higher accuracy and trend stability in the prediction of temperature and pressure with obvious diurnal cycles, which is beneficial for the prediction of parameters such as transformer temperature and pressure with significant diurnal cycle characteristics.

[0067] In some possible implementations of the first aspect, S50, the computational expression for projecting the time-frequency fusion features onto the multi-parameter prediction space is:

[0068] in, Indicates the predicted output. This represents the predicted output mapping weight matrix.

[0069] In this embodiment of the application, by capturing the time-frequency characteristics of the signal and mapping them to the prediction space, more accurate multi-parameter joint prediction can be achieved.

[0070] In some possible implementations of the first aspect, the transformer multi-parameter time-series data prediction method further includes: The multi-parameter prediction space is trained by using a multi-scale weighted loss function to obtain a trained multi-parameter prediction space.

[0071] Based on the above scheme, in the embodiments of this application, a multi-scale weighted loss function is introduced to optimize the model performance during the training process of the multi-parameter prediction space, so that the model can differentiate the weighting of the prediction error at different scales, thereby improving the overall accuracy and robustness of multi-parameter prediction.

[0072] Optionally, the multi-scale weighted loss function includes: giving higher weights to short-term prediction steps; and using smoothing weights for medium- and long-term prediction steps.

[0073] Based on the above scheme, giving higher weight to the short-term forecast step can highlight the real-time early warning capability; using smoothing weights for the medium- and long-term forecast steps can ensure that the trend does not diverge.

[0074] Specifically, the training process includes: Calculate the prediction error for each prediction step; By introducing prediction step weight coefficients to weight the errors at different time steps, a multi-scale weighted loss function is formed.

[0075] Furthermore, to improve convergence stability, the weights of the prediction step weight coefficients vary with the number of training rounds, and the weight calculation expression for the weight coefficients is as follows:

[0076] in, Indicates the weight coefficient of the first The weights of each prediction step; This represents the weighting factor, which gradually decreases with the number of training rounds.

[0077] Based on the above scheme, during the training process, the prediction error is first calculated for each prediction step. This prediction error reflects the degree to which the model fits the actual operating state at that time scale. Subsequently, prediction step weight coefficients are introduced to weight the errors of different time steps, forming a multi-scale prediction loss. In addition, during the training process, parameter constraint terms can be added to penalize combinations that obviously violate physical laws (e.g., temperature and pressure changes in opposite directions over a long period of time). This application maintains short-term prediction sensitivity while taking into account the stability of long-term trends through multi-parameter collaborative prediction and multi-scale weighted loss, and naturally ensures the logical consistency between different physical parameters.

[0078] It should be understood that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order constraint on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the diagram may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0079] Secondly, this application provides an apparatus for predicting transformer multi-parameter time-series data, which includes a module for implementing the aforementioned transformer multi-parameter time-series data prediction method.

[0080] For example, such as Figure 5 As shown, the transformer multi-parameter time-series data prediction device includes: Acquisition module 10 acquires multiple operating parameters during the transformer's operation. Preprocessing module 20 preprocesses the multi-parameter running parameters to form a multi-parameter time series dataset; The feature extraction module 30 extracts features from the data in the time series dataset through the local convolution module and the global attention module to obtain stable local features and global features. The feature fusion module 40, based on stable local and global features, performs collaborative extraction of time-domain local dynamic features and frequency-domain frequency distribution features through a frequency-domain enhanced extended feedforward network to generate time-frequency fusion features; The collaborative prediction module 50 projects the time-frequency fusion features onto the multi-parameter prediction space to obtain the multi-parameter prediction results.

[0081] Specific limitations regarding the aforementioned transformer multi-parameter time-series data prediction device can be found in the limitations of the transformer multi-parameter time-series data prediction method described above, and will not be repeated here. Each module in the aforementioned transformer multi-parameter time-series data prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0082] The following is an explanation based on actual 500kV transformer operation data through a specific embodiment, wherein: the multi-parameter operating parameters are acquired by a multi-parameter optical sensor; the multi-parameter operating parameters include temperature, pressure and partial discharge signal, the sampling frequency is 1Hz, and it has high noise, nonlinearity, time delay and multivariable coupling characteristics.

[0083] (1) First, the multi-parameter operating parameters are preprocessed, including: noise reduction, normalization, and construction of input-output pairs; specifically: Wavelet transform or adaptive filtering is used to remove sensor noise; The data from each sensor were normalized, with an average temperature of 50 degrees Celsius, an average pressure of 150 kPa, and an average discharge of 0.5 pC. Divide the time series into input-output pairs; The prediction is constructed using a sliding window, with 100 data points as input and the next 5 data points as predictions.

[0084] (2) After completing the data preprocessing, the multi-parameter time series data in the standardized time series dataset is input into the “local convolution module - global attention module - frequency domain enhanced extended feedforward network - parameter prediction space” structure in the embodiment of this application; the prediction process includes: multi-scale dilated causal convolution, residual gated fusion, adaptive position coding, improved temporal cross attention frequency domain enhanced extended feedforward network and multi-parameter prediction.

[0085] Specifically, local mutations and short-term dynamics are captured through multi-scale dilated convolution, long-term dependencies are extracted through adaptive positional encoding and temporal cross-attention mechanism, and the physical coupling relationship between temperature, pressure and partial discharge is learned through variable-dimensional cross-attention mechanism; periodic trends are modeled in a frequency-domain enhanced feedforward network; after about 100 rounds of training, the parameters converge, resulting in a stable prediction device.

[0086] Taking temperature as an example: Comparison between the predicted curve and the actual measured value in this application. Figure 6 As shown in the figure, this application can simultaneously capture short-term fluctuations and long-term trends. The prediction results are highly consistent with the actual operating conditions, verifying the effectiveness and engineering applicability of this algorithm in actual transformer operation scenarios.

[0087] Experimental results show that the prediction accuracy of the present invention on actual 500kV transformer multi-parameter monitoring data is significantly better than that of the traditional model, and it is suitable for scenarios such as online monitoring, operation status assessment and fault early warning.

[0088] Thirdly, this application provides a device that can be any device capable of predicting multi-parameter time-series data of a transformer. The device can be various terminal devices, such as desktop computers, laptops, tablets, handheld devices, etc., and can be implemented through software and / or hardware.

[0089] For example, the device includes: Memory; Processor; and Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the method as described in any of the preceding methods.

[0090] Fourthly, this application provides a computer-readable storage medium, which may be a ROM, RAM, disk, or optical disk, etc.

[0091] For example, the computer-readable storage medium stores a computer program; the computer program is executed by a processor to implement the method as described in any of the preceding methods.

[0092] In summary, this application employs a joint modeling strategy involving local, global, cross-parameter, and multi-scale approaches, which significantly improves the prediction accuracy, stability, and engineering applicability of transformer multi-parameter optical sensor data compared to traditional TCN, Transformer, or simple combinations thereof. This approach can more effectively support transformer condition assessment, trend analysis, and operational early warning, demonstrating outstanding practical value and technological advancement.

[0093] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as C, VHDL, Verilog, the object-oriented programming language Java, and the interpreted scripting language JavaScript.

[0094] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0097] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0098] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for predicting multi-parameter time-series data of transformers, characterized in that, include: Collect multiple operating parameters during the transformer's operation; Preprocess the multi-parameter operating parameters to form a multi-parameter time series dataset; Stable local and global features are obtained by extracting features from the time-series dataset through local convolutional modules and global attention modules. Based on stable local and global features, a frequency-domain enhanced extended feedforward network is used to collaboratively extract time-domain local dynamic features and frequency-domain frequency distribution features to generate time-frequency fusion features. The time-frequency fusion features are projected onto the multi-parameter prediction space to obtain the multi-parameter prediction results.

2. The method for predicting transformer multi-parameter time-series data according to claim 1, characterized in that, The process involves extracting features from the time-series dataset using local convolutional modules and global attention modules to obtain stable local and global features; including: Stable local features of the time-series dataset are extracted using a multi-scale dilated causal convolution module and a residual gating module. By using adaptive positional encoding and an improved temporal cross-attention mechanism, long-range dependencies between different time slices and coupling relationships between different physical parameters are captured in stable local features, forming global features.

3. The method for predicting transformer multi-parameter time-series data according to claim 2, characterized in that, The method extracts stable local features from the time-series dataset using a multi-scale dilated causal convolution module and a residual gating module, including: A multi-scale dilated causal convolution module is used to extract scale features of time-series datasets in parallel at different dilation rates; and the scale features corresponding to each dilation rate are fused by channel to form local features; A residual gating module is used to suppress interference information in local features and dynamically select effective features from the local features to form stable local features.

4. The method for predicting transformer multi-parameter time-series data according to claim 2, characterized in that, The improved temporal cross-attention mechanism includes: In the time channel, a time-dimensional cross-attention mechanism is used to aggregate attention on stable local features in the time dimension, resulting in a global time-dependent representation: On the variable channel, a variable-dimensional cross-attention mechanism is used to perform cross-attention aggregation of different parameters on the variable dimension at the same time step to obtain a global variable dependency representation; By fusing the time channel output and the variable channel output using a weighted algorithm, the global features of the dual-channel fusion are obtained.

5. The method for predicting transformer multi-parameter time-series data according to claim 1, characterized in that, Also includes: The multi-parameter prediction space is trained by using a multi-scale weighted loss function to obtain a trained multi-parameter prediction space.

6. The method for predicting transformer multi-parameter time-series data according to claim 5, characterized in that, The multi-scale weighted loss function includes: Give higher weight to short-term forecasting steps; Smoothing weights are used for medium- and long-term forecasting steps.

7. The method for predicting transformer multi-parameter time-series data according to claim 1, characterized in that, The preprocessing of multi-parameter operating parameters includes: noise reduction, normalization, and sliding window construction.

8. A transformer multi-parameter time-series data prediction device, characterized in that: It includes a module for implementing the transformer multi-parameter time-series data prediction method according to any one of claims 1 to 7.

9. An electronic device, characterized in that: include: Memory; processor; as well as Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the transformer multi-parameter time-series data prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: It stores a computer program; the computer program is executed by a processor to implement the transformer multi-parameter time-series data prediction method as described in any one of claims 1 to 7.

Images