A power grid equipment residual life prediction method, system, device and storage medium

Abstract

The application discloses a power grid equipment residual life prediction method, system, device and storage medium, which is applied to the field of power grid equipment residual life prediction, acquires multiple types of data of power grid equipment within a preset time, and constructs a data set containing basic attribute data, operation data, maintenance record data and environment data; based on the data, equivalent thermal stress current, model-working condition matching degree coefficient, maintenance timing characteristics and high-dimensional time sequence dynamic feature set are obtained; the equivalent thermal stress current, the model-working condition matching degree coefficient, the maintenance timing characteristics and the high-dimensional time sequence dynamic feature set at each moment are spliced to generate the final high-dimensional feature vector at the corresponding moment; and finally, the final high-dimensional feature vectors at continuous multiple moments are input into an adaptive hybrid deep learning prediction model for prediction to obtain a residual life prediction value. The method combines multi-source feature fusion with deep learning, and improves the accuracy and reliability of power grid equipment residual life prediction.

Description

Technical Field

[0001] This invention relates to the field of power grid equipment remaining life prediction technology, and in particular to a method, system, device and storage medium for predicting the remaining life of power grid equipment. Background Technology

[0002] With the expansion of power grids and the acceleration of equipment aging, traditional equipment replacement models based on fixed lifecycles are no longer adequate for the needs of refined and intelligent management. Existing technologies mainly rely on failure physics models, such as the Paris crack propagation model and thermal aging model, as well as traditional statistical models, such as the Weibull distribution and the Cox proportional hazards model. However, existing methods are insufficient in capturing the local characteristics and long-term temporal dependencies in the equipment degradation process, resulting in large errors in remaining life prediction. Furthermore, most models are static and cannot be dynamically updated according to changes in equipment operating status, making it difficult to support real-time early warning and decision-making.

[0003] Therefore, designing a hybrid deep learning architecture that can capture short-term fluctuations, model long-term dependencies, and adapt to changes in the aging stage of equipment to achieve accurate prediction of remaining lifetime has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, embodiments of the present invention provide a method, system, device, and storage medium for predicting the remaining lifetime of power grid equipment. This addresses the issue that existing methods for predicting the remaining lifetime of power grid equipment lack the ability to capture local features and long-term time-series dependencies during equipment degradation, leading to significant errors in the prediction of remaining lifetime.

[0005] A first aspect of this invention provides a method for predicting the remaining life of power grid equipment, the method comprising: Acquire various types of data from power grid equipment within a preset time period to obtain a data set, wherein the data set includes at least basic attribute data, operational data, maintenance record data, and environmental data; Based on the basic attribute data, operational data, and environmental data at each moment within a preset time period, the equivalent thermal stress current and model-operating condition matching coefficient at each moment are obtained. The maintenance record data at each moment within a preset time period is transformed into time-series features to obtain maintenance time-series features. Statistical features, trend features, and frequency domain features are extracted from the time-series data in the dataset to obtain a high-dimensional time-series dynamic feature set. The equivalent thermal stress current, model-operating condition matching coefficient, maintenance timing characteristics, and high-dimensional time-series dynamic features at each moment are concatenated to obtain the final high-dimensional feature vector at each moment. The final high-dimensional feature vectors from multiple consecutive time points are input into an adaptive hybrid deep learning prediction model for processing to obtain the predicted remaining lifespan of the power grid equipment.

[0006] Furthermore, based on the basic attribute data, operational data, and environmental data at each moment within a preset time period, the equivalent thermal stress current and model-operating condition matching coefficient at each moment are obtained, including: Based on the basic attribute data and environmental data at each moment within a preset time period, the temperature rise coefficient and temperature data of the power grid equipment at each moment are filtered out. Based on the current data and temperature rise coefficient of the operating data at each moment, the temperature rise value is obtained. Based on the temperature rise value and the corresponding temperature data at each moment, the equivalent thermal stress current at each moment is obtained. Based on the operational data, the average load rate, total number of start-stop cycles, and maximum load rate within the preset time period are obtained; The model-operating condition matching coefficient is obtained based on the average load rate, total number of start-stop cycles, and maximum load rate.

[0007] Furthermore, based on the average load rate, total number of start-stop cycles, and maximum load rate, the model-operating condition matching coefficient is obtained, including: Based on the basic attribute data within a preset time period, the rated load rate and rated start-stop frequency of the power grid equipment are obtained. The rated load rate, rated start-stop count, average load rate, total start-stop count, and maximum load rate are input into the matching degree function for calculation to obtain the model-operating condition matching degree coefficient.

[0008] Furthermore, the maintenance record data for each moment within a preset time period is transformed into time-series features to obtain maintenance time-series features, including: Construct a dictionary of defect types and a dictionary of handling measures; Based on the defect type dictionary and the treatment measure dictionary, the defect class description and treatment measure description in the maintenance record are converted into numerical codes respectively; Each numerical code is input into the scoring model to obtain a score for the severity of the defect; The maintenance sequence characteristics are obtained based on the timestamp, numerical code, and defect severity score corresponding to the maintenance record.

[0009] Furthermore, the final high-dimensional feature vectors from multiple consecutive time points are input into an adaptive hybrid deep learning prediction model for processing to obtain the predicted remaining lifetime of the power grid equipment, including: The adaptive multi-scale feature extraction module in the adaptive hybrid deep learning prediction model is used to extract features from each final high-dimensional feature vector to obtain local feature representations. The stage-aware attention mechanism in the adaptive hybrid deep learning prediction model is used to process the stage encoding vector and local feature representation to obtain the calibrated local feature representation; The calibrated local feature representation and stage encoding vector are input into the stage-aware temporal modeling module in the adaptive hybrid deep learning prediction model for processing to obtain global context features. The fully connected layer in the adaptive hybrid deep learning prediction model is used to process the global context features to obtain the remaining lifetime prediction value of the power grid equipment.

[0010] Furthermore, the adaptive multi-scale feature extraction module in the adaptive hybrid deep learning prediction model is used to extract features from each final high-dimensional feature vector to obtain local feature representations, including: The multi-scale deformable convolutional layer in the adaptive multi-scale feature extraction module is used to extract features from each final high-dimensional feature vector to obtain feature representations at different scales. The feature pyramid fusion layer in the adaptive multi-scale feature extraction module is used to fuse the feature representations at various scales to obtain local feature representations.

[0011] Furthermore, the stage encoding vector is obtained by processing the dataset using a lightweight fully connected network. The specific generation process is as follows: Based on the dataset, the cumulative operating years of the equipment, average load rate, and cumulative environmental stress index were obtained. The cumulative operating years of the equipment, average load rate, and cumulative environmental stress index are input into a lightweight fully connected network for processing to obtain the stage coding vector.

[0012] Another embodiment of the present invention provides a power grid equipment remaining life prediction system, comprising: The acquisition module is used to acquire various types of data from power grid equipment within a preset time period to obtain a data set, wherein the data set includes at least basic attribute data, operation data, maintenance record data and environmental data; The cross-source interactive feature construction module is used to obtain the equivalent thermal stress current and model-operating condition matching coefficient at each moment based on the basic attribute data, operation data and environmental data at each moment within a preset time. The high-dimensional time-series dynamic feature acquisition module is used to convert maintenance record data at each moment within a preset time into time-series features to obtain maintenance time-series features, and to extract statistical features, trend features and frequency domain features from the time-series data in the dataset to obtain a high-dimensional time-series dynamic feature set. The feature splicing module is used to splice the equivalent thermal stress current, model-operating condition matching coefficient, maintenance timing features and high-dimensional time-series dynamic feature set at each moment to obtain the final high-dimensional feature vector at each moment. The remaining lifetime prediction module is used to input the final high-dimensional feature vectors from multiple consecutive time points into an adaptive hybrid deep learning prediction model for processing, in order to obtain the remaining lifetime prediction value of the power grid equipment.

[0013] A third aspect of the present invention provides a computer device, comprising: Memory, used to store computer programs; A processor is used to execute computer programs to implement steps of the power grid equipment remaining life prediction method as described in the first aspect.

[0014] A fourth aspect of the present invention provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the power grid equipment remaining life prediction method of the first aspect.

[0015] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following: First, multiple types of data from power grid equipment within a preset time period are acquired to construct a dataset containing basic attribute data, operational data, maintenance record data, and environmental data. Based on the basic attribute, operational, and environmental data at each time point, the equivalent thermal stress current and the model-operating condition matching coefficient are calculated. The maintenance record data is converted into maintenance time-series features, and statistical, trend, and frequency domain features are extracted from the time-series data to form a high-dimensional time-series dynamic feature set. The equivalent thermal stress current, model-operating condition matching coefficient, maintenance time-series features, and high-dimensional time-series dynamic feature set at each time point are concatenated to generate the final high-dimensional feature vector for the corresponding time point. Then, the final high-dimensional feature vectors from multiple consecutive time points are input into an adaptive hybrid deep learning prediction model, and the model processes and outputs the predicted remaining life of the power grid equipment. This method improves the accuracy and reliability of the remaining life prediction of power grid equipment by combining multi-source feature fusion and deep learning. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps in the method for predicting the remaining life of power grid equipment in one embodiment of the present invention; Figure 2 This is a flowchart of the multi-source data fusion and model construction process for a power grid equipment remaining life prediction method in one embodiment of the present invention; Figure 3 This is a diagram of the adaptive hybrid deep learning model network structure of the power grid equipment remaining lifetime prediction method in one embodiment of the present invention; Figure 4 This is a dynamic change diagram of the training process of the adaptive hybrid deep learning model of the power grid equipment remaining life prediction method in one embodiment of the present invention; Figure 5 This is a structural block diagram of a power grid equipment remaining life prediction system according to one embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0019] One embodiment of the present invention provides a method for predicting the remaining life of power grid equipment. For details, please refer to [link / reference]. Figure 1 , Figure 1 The flowchart shown is a step-by-step flowchart of a method for predicting the remaining life of power grid equipment according to one embodiment of the present invention, including: S11. Obtain various types of data from power grid equipment within a preset time period to obtain a data set, wherein the data set includes at least basic attribute data, operation data, maintenance record data, and environmental data.

[0020] In this embodiment, power grid equipment refers to key electrical equipment in a power system used for power transmission, transformation, control, and protection. This mainly includes, but is not limited to, transformers, circuit breakers, switchgear, reactors, instrument transformers, cable lines, disconnect switches, and other transmission and transformation equipment, as well as various distribution terminals and high-voltage electrical installations. All of the above equipment are power assets that operate continuously under load and environmental stress, and exhibit aging and degradation characteristics. The preset time in this embodiment can be determined according to actual needs, such as one year, six months, or one month. Various types of data include, but are not limited to, basic attribute data, operational data, maintenance record data, and environmental data. Figure 2The flowchart for multi-source data fusion and model building describes how, after acquiring various types of data from power grid equipment within a preset time period, the unique ID of the power grid equipment and a high-precision unified timestamp are used as dual primary keys. Through data warehouse technology or a stream processing framework, seven categories of heterogeneous data from different business systems, such as ERP (Enterprise Resource Planning), SCADA (Supervisory Control and Data Acquisition), online monitoring, and meteorological platforms, are physically aligned and associated to form a data cube with equipment-time as the primary dimension (see Table 1). All data are aligned using a unified time-series alignment algorithm, such as dynamic time warping or linear interpolation, forming a multi-dimensional time-series data matrix with one file per device, serving as the basis for subsequent model input.

[0021] Table 1 Data Set Table

[0022] S12. Based on the basic attribute data, operation data and environmental data of each moment within a preset time period, obtain the equivalent thermal stress current and model-operating condition matching coefficient for each moment.

[0023] In one embodiment, based on the basic attribute data, operational data, and environmental data at each moment within a preset time period, the equivalent thermal stress current and model-operating condition matching coefficient at each moment are obtained, including: Based on the basic attribute data and environmental data at each moment within a preset time period, the temperature rise coefficient and temperature data of the power grid equipment at each moment are filtered out. Based on the current data and temperature rise coefficient of the operating data at each moment, the temperature rise value is obtained. Based on the temperature rise value and the corresponding temperature data at each moment, the equivalent thermal stress current at each moment is obtained. Based on the operational data, the average load rate, total number of start-stop cycles, and maximum load rate within the preset time period are obtained; The model-operating condition matching coefficient is obtained based on the average load rate, total number of start-stop cycles, and maximum load rate.

[0024] This embodiment constructs cross-source interactive features after obtaining the data set within a preset time period. Specifically, it filters the operating data within the preset time period to obtain the current data of the power grid equipment at any given moment and the corresponding temperature data at that moment. Then, based on the equipment model and material properties, it obtains the temperature rise coefficient of the power grid equipment from the basic attribute data. This coefficient represents the temperature rise caused by a unit current square. Using the temperature rise coefficient and the current data, it calculates the temperature rise caused by the current. The expression for the temperature rise value is: In the formula, The coefficient of temperature rise. for Current data at any given time, for Temperature rise at any given moment.

[0025] Then, the equivalent temperature of the equipment under the combined effects of ambient temperature and current is calculated. The formula for calculating the equivalent temperature is: In the formula, for Temperature data at any given time for Temperature rise at any moment This is the equivalent temperature.

[0026] According to the Arrhenius thermal aging model, the aging rate of equipment insulation materials has an exponential relationship with temperature. The relative aging rate is calculated using the following formula: in, For reference temperature, 90℃ is usually taken. Equivalent temperature This represents the relative aging rate.

[0027] Using relative aging rate and current data, the equivalent thermal stress current is obtained. The formula for calculating the equivalent thermal stress current is as follows: In the formula, This refers to the relative aging rate. for Current data at any given time.

[0028] The equivalent thermal stress current characteristic integrates the effects of current magnitude and ambient temperature on the thermal aging of equipment, accurately reflecting the actual thermal stress level of equipment under specific environmental conditions.

[0029] Then, by quantifying the difference between the actual operating conditions of the equipment and its design expectations, a "model-operating condition matching coefficient" feature is constructed. Specifically, based on the operating data, the average load rate, total number of start-stop cycles, and maximum load rate within a preset time period are obtained. Then, the model-operating condition matching coefficient is obtained based on the average load rate, total number of start-stop cycles, and maximum load rate.

[0030] In one embodiment, the model-operating condition matching coefficient is obtained based on the average load rate, total number of start-stop cycles, and maximum load rate, including: Based on the basic attribute data within a preset time period, the rated load rate and rated start-stop frequency of the power grid equipment are obtained. The rated load rate, rated start-stop count, average load rate, total start-stop count, and maximum load rate are input into the matching degree function for calculation to obtain the model-operating condition matching degree coefficient.

[0031] In this embodiment, the device model, manufacturer, and rated parameters are extracted from the basic attribute data. The rated parameters include at least the rated load rate and the rated number of start-stop cycles. Long-term operational statistical characteristics of the power grid equipment are calculated from the operational data, such as the average load rate, number of start-stop cycles, and maximum load rate over a preset time period.

[0032] Define a matching degree function. Input the rated load rate, rated start / stop count, average load rate, total start / stop count, and maximum load rate into the matching degree function for calculation, obtaining the model-operating condition matching degree coefficient. The expression for the matching degree function is: In the formula, , , These are the weighting coefficients. Measure the degree of match between average load rate and rated load rate. To measure the degree of match between the actual number of start-stop cycles and the designed number of start-stop cycles. A penalty measure that determines whether the maximum load rate exceeds the rated load rate. For maximum load rate, Average load factor For the number of start-stop cycles, Rated load rate, The number of start-stop cycles is the rated number. For matching degree.

[0033] The matching degree function is normalized to obtain the model-operating condition matching degree coefficient: in, and These are the minimum and maximum values ​​of the matching degree function, respectively.

[0034] Model-Operating Condition Matching Coefficient: This coefficient ranges from 0 to 1. The closer it is to 1, the closer the operating condition of the equipment is to the design expectation.

[0035] S13. Transform the maintenance record data of each moment within the preset time into time-series features to obtain maintenance time-series features, and extract statistical features, trend features and frequency domain features from the time-series data in the dataset to obtain a high-dimensional time-series dynamic feature set.

[0036] In one embodiment, maintenance record data at each moment within a preset time period is converted into time-series features to obtain maintenance time-series features, including: Construct a dictionary of defect types and a dictionary of handling measures; Based on the defect type dictionary and the treatment measure dictionary, the defect class description and treatment measure description in the maintenance record are converted into numerical codes respectively; Each numerical code is input into the scoring model to obtain a score for the severity of the defect; The maintenance sequence characteristics are obtained based on the timestamp, numerical code, and defect severity score corresponding to the maintenance record.

[0037] In this embodiment, unstructured defect inspection records are transformed into structured time-series features. The specific steps are as follows: A defect type dictionary is established, encoding defect types. For example, partial discharge is mapped to 1, insulation degradation to 2, mechanical failure to 3, overheating to 4, and other defects to 5. This dictionary uniformly converts text-described defect types into numerical codes. A treatment measure dictionary is also established, encoding treatment measures. For example, monitoring is mapped to 1, maintenance to 2, component replacement to 3, and shutdown for maintenance to 4. This dictionary uniformly converts text-described treatment measures into numerical codes.

[0038] Then, a defect severity score is constructed. Based on the defect description and handling measures, the severity score is defined. ,in, For the scoring function, such as: in, and These are weighting coefficients, for example, α=2.0, β=2.5. Numerical encoding for defect types, The corresponding numerical code is used to describe the handling measures. For example, a serious defect that requires replacement of the part is scored as 10 points, while a minor defect that only needs to be monitored is scored as 1 point.

[0039] Transform defect events into time-series feature vectors. , , , ],in, These are defect timestamps. For time points without defects, these characteristics are set to 0 or default values. Specifically, for major repair or component replacement events, a status reset flag characteristic is introduced. Specifically, it takes the form of an exponential function that decays over time: in, For repair time, This is the decay time constant, such as 30 days. This feature is 1 immediately after maintenance and then gradually decays to 0, representing the recovery process of the equipment after maintenance.

[0040] For continuously monitored time-series data, such as vibration signals, oil chromatography gas concentrations, and temperatures, not only are instantaneous values ​​extracted, but the statistical characteristics, trend characteristics, and frequency domain characteristics over multiple preset time windows, such as 1 hour, 24 hours, and 7 days, are systematically calculated. This transforms the one-dimensional time-series signal into a high-dimensional feature vector reflecting its short-term fluctuations and long-term evolution patterns. The specific calculations are as follows, using vibration signals... For example, other signals are similar, such as statistical characteristics (taking the time window length T as an example), which include mean, standard deviation, skewness, and kurtosis. Specifically, the formula for calculating the mean is: In the formula, Let be the vibration signal at time t, and T be the preset time window. The mean value of the vibration signal within the preset time window T.

[0041] The formula for calculating standard deviation is: In the formula, The vibration signal at time t, The mean value of the vibration signal within a preset time window T is given, where T is the preset time window. The standard deviation of the vibration signal within the preset time window T is given.

[0042] The formula for calculating skewness is: In the formula, The vibration signal at time t, The mean value of the vibration signal within a preset time window T is given, where T is the preset time window. The standard deviation of the vibration signal within the preset time window T. The skewness of the vibration signal within a preset time window T.

[0043] The formula for calculating kurtosis is: In the formula, The vibration signal at time t, The mean value of the vibration signal within a preset time window T is given, where T is the preset time window. The standard deviation of the vibration signal within the preset time window T. The kurtosis of the vibration signal within a preset time window T.

[0044] The trend feature is calculated by performing a linear fit on the data points within the window; the slope of the resulting straight line represents the trend feature, reflecting the upward or downward trend of the signal within that time period. The frequency domain feature is calculated by performing a Fourier transform on the signal within the window to extract the amplitude of the main frequency components (such as the 50Hz power frequency and its harmonics) to discover hidden periodic or resonant modes in the signal.

[0045] S14. The equivalent thermal stress current, model-operating condition matching coefficient, maintenance sequence characteristics and high-dimensional time-series dynamic feature set at each time moment are concatenated to obtain the final high-dimensional feature vector at each time moment.

[0046] In this embodiment, raw data from different sources and of different natures are fused and reconstructed into a unified, high-dimensional feature vector to comprehensively characterize the device's state at a given moment. At any time point t, the device's complete state vector F(t) is composed of three types of features: for example, for time point t, the feature vector includes cross-source interaction features, event features, and temporal dynamic features. These features are fully aligned in the time dimension, collectively describing the device's overall state at time t. Through the above steps, multi-source heterogeneous data is actively transformed into a unified, high-dimensional time-series dataset with a strongly representative feature vector for each time point. Finally, the feature vector for each time point t is represented as: Then, a sliding window technique is used to convert the unified time series after fusion and reconstruction into supervised learning samples that can be processed by a deep learning model. Let the window length be L (e.g., L = 30 time steps, corresponding to approximately 7.5 years), and the feature dimension be D, then the... a sample pair Defined as: in, This includes the device's status information at L consecutive time points. This represents the true value of the remaining device lifetime at the end of the window.

[0047] Through the above processing method, each high-dimensional time-series dynamic feature set not only contains the sequence of device state changes (time dependence) over a historical period, but also incorporates the correlation information of multi-source data through rich feature dimensions, providing high-quality structured input for subsequent deep learning models to make accurate predictions.

[0048] S14. Input the final high-dimensional feature vectors from multiple consecutive time points into the adaptive hybrid deep learning prediction model for processing to obtain the remaining life prediction value of the power grid equipment.

[0049] In this embodiment, the adaptive hybrid deep learning prediction model consists of an adaptive multi-scale feature extraction module, a stage-aware attention mechanism, and a stage-aware temporal modeling module, such as... Figure 3 As shown, Figure 3 This is a diagram of the network structure of the adaptive hybrid deep learning model. The final high-dimensional feature vectors from multiple consecutive time points are input into the adaptive hybrid deep learning prediction model for processing, yielding the predicted remaining lifetime of the power grid equipment.

[0050] In one embodiment, the final high-dimensional feature vectors from multiple consecutive time points are input into an adaptive hybrid deep learning prediction model for processing to obtain the predicted remaining lifetime of the power grid equipment, including: The adaptive multi-scale feature extraction module in the adaptive hybrid deep learning prediction model is used to extract features from each final high-dimensional feature vector to obtain local feature representations. The stage-aware attention mechanism in the adaptive hybrid deep learning prediction model is used to process the stage encoding vector and local feature representation to obtain the calibrated local feature representation; The calibrated local feature representation and stage encoding vector are input into the stage-aware temporal modeling module in the adaptive hybrid deep learning prediction model for processing to obtain global context features. The fully connected layer in the adaptive hybrid deep learning prediction model is used to process the global context features to obtain the remaining lifetime prediction value of the power grid equipment.

[0051] In this embodiment, the adaptive multi-scale feature extraction module includes a multi-scale deformable convolutional layer and a feature pyramid fusion layer. The multi-scale design aims to simultaneously capture degradation patterns at different time granularities, from daily operational fluctuations and monthly load cycles to annual aging trends. The outputs of each scale are fused from top to bottom through a feature pyramid network to generate a unified and rich local feature representation. Then, an independent stage encoding vector is introduced and input into an attention sub-network to dynamically recalibrate each channel of the local feature representation, resulting in a calibrated local feature representation. For example, in the early stages of aging, the model may automatically increase the weight of feature channels reflecting long-term slow trends; in the later stages of aging, it may increase the attention to feature channels associated with short-term mutations and defects. This achieves dynamic focusing on the features most relevant to the current aging stage based on domain knowledge priors.

[0052] The calibrated local feature representations are input into the stage-aware temporal modeling module. Specifically, the stage-aware temporal modeling module designs a stage information deep injection gating mechanism, which deeply integrates the aforementioned stage encoding vectors into the core gating unit computation of the bidirectional long short-term memory network. Specifically, the stage encoding vectors are concatenated with the current input and the previous hidden state, jointly participating in the sigmoid activation calculations of the forget gate and the input gate. This allows the model to dynamically adjust its historical memory retention strategy and the intensity of new information absorption based on the device's macro-aging stage; for example, it pays more attention to recently occurring anomalous patterns at the end of the device's lifespan.

[0053] To achieve deep collaboration between local features and global temporal context, a closed-loop optimization path was designed. The forward path feeds the local feature representation into a Bi-LSTM for temporal modeling. The feedback path takes the global context feature vector, which integrates forward and backward information at the final time step of the Bi-LSTM, and passes it forward through a feedback connection to an intermediate layer of the adaptive multi-scale feature extraction module. For example, it performs channel addition with the output feature map of a convolutional layer.

[0054] It should be noted that during training, this adaptive hybrid deep learning prediction model uses an RMSProp (Root Mean Square Propagation) optimizer with a learning rate of 0.001 and a batch size of 200. An early stopping strategy (patience=10) is also employed to prevent overfitting. Evaluation metrics include: Where RMSE is the root mean square error and MAPE is the mean absolute percentage error. For the goodness of fit, For the true value, For predicted values, This is the mean.

[0055] Dynamic changes during the training process, such as Figure 4 As shown in the figure, three key curves are plotted: training loss, validation loss, and mean absolute error on the validation set. It is clearly observed from the figure that all metrics decrease rapidly during the first 30 training epochs, indicating that the model is learning effectively. After 30 epochs, the metrics gradually stabilize, and the curves for the training and validation sets do not show significant separation, verifying the model's good generalization ability and stable convergence characteristics.

[0056] In one embodiment, the adaptive multi-scale feature extraction module in the adaptive hybrid deep learning prediction model is used to extract features from each final high-dimensional feature vector to obtain local feature representations, including: The multi-scale deformable convolutional layer in the adaptive multi-scale feature extraction module is used to extract features from each final high-dimensional feature vector to obtain feature representations at different scales. The feature pyramid fusion layer in the adaptive multi-scale feature extraction module is used to fuse the feature representations at various scales to obtain local feature representations.

[0057] In this embodiment, the adaptive multi-scale feature extraction module employs a set of parallel one-dimensional deformable convolutional layers with kernel sizes of {3, 5, 7}. Deformable convolution can adaptively learn the offset of non-uniform sampling points on the input sequence, making it particularly suitable for capturing local anomalous signals that are unevenly distributed along the time axis during device degradation, such as sudden discharges. After extracting features from each final high-dimensional feature vector through the multi-scale deformable convolutional layers to obtain feature representations at different scales, these representations are fused from top to bottom using a feature pyramid network to generate local feature representations.

[0058] In one embodiment, the stage coding vector is obtained by processing the dataset using a lightweight fully connected network. The specific generation process is as follows: Based on the dataset, the cumulative operating years of the equipment, average load rate, and cumulative environmental stress index were obtained. The cumulative operating years of the equipment, average load rate, and cumulative environmental stress index are input into a lightweight fully connected network for processing to obtain the stage coding vector.

[0059] In this embodiment, the independent stage encoding vector is not learned from the current input signal, but rather extracted from external multi-source data using statistical indicators with clear physical meaning, such as the equipment's cumulative operating years, recent average load rate, and environmental stress accumulation index. These are then generated through a lightweight fully connected network to represent the prior information of the equipment's current aging stage. This clearly distinguishes whether the equipment is in the early, middle, or late stages of aging, providing prior guidance consistent with the equipment's aging mechanism for subsequent deep learning models, thereby improving the model's adaptability and prediction accuracy for different aging stages. Specifically, the equipment's cumulative operating years, recent average load rate, and environmental stress accumulation index are input into a pre-constructed lightweight fully connected network for mapping and transformation. Through linear transformation and nonlinear activation processing of the fully connected network, a fixed-dimensional stage encoding vector is output.

[0060] This invention also provides a power grid equipment remaining lifetime prediction system for performing the above-described power grid equipment remaining lifetime prediction method. Figure 5This is a block diagram of a power grid equipment remaining life prediction system according to an embodiment of the present invention. The system includes: The acquisition module 501 is used to acquire various types of data from power grid equipment within a preset time period to obtain a data set, wherein the data set includes at least basic attribute data, operation data, maintenance record data and environmental data; The cross-source interaction feature construction module 502 is used to obtain the equivalent thermal stress current and model-operating condition matching coefficient at each moment based on the basic attribute data, operation data and environmental data at each moment within a preset time. The high-dimensional time-series dynamic feature acquisition module 503 is used to convert the maintenance record data at each moment within a preset time into time-series features to obtain maintenance time-series features, and to extract statistical features, trend features and frequency domain features from the time-series data in the dataset to obtain a high-dimensional time-series dynamic feature set. The feature splicing module 504 is used to splice the equivalent thermal stress current, model-operating condition matching coefficient, maintenance timing features and high-dimensional time-series dynamic feature set at each moment to obtain the final high-dimensional feature vector at each moment. The remaining lifetime prediction module 505 is used to input the final high-dimensional feature vectors from multiple consecutive time points into the adaptive hybrid deep learning prediction model for processing, so as to obtain the remaining lifetime prediction value of the power grid equipment.

[0061] In one embodiment of this application, a computer device is provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above steps. The implementation principle and technical effects of the computer device provided in this embodiment are similar to those of the above method embodiments, and will not be repeated here.

[0062] In one embodiment of this application, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, it performs the above steps; the implementation principle and technical effects of the computer-readable storage medium provided in this embodiment are similar to those of the above method embodiments, and will not be repeated here.

[0063] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0064] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting the remaining life of power grid equipment, characterized in that, include: Acquire various types of data from power grid equipment within a preset time period to obtain a data set, wherein the data set includes at least basic attribute data, operational data, maintenance record data, and environmental data; Based on the basic attribute data and environmental data at each moment within the preset time period, the temperature rise coefficient and temperature data of the power grid equipment at each moment are filtered out. Based on the current data and temperature rise coefficient of the operating data at each moment, the temperature rise value is obtained. Based on the temperature rise value and the temperature data corresponding to each moment, the equivalent thermal stress current at each moment is obtained. Based on the operating data, the average load rate, total number of start-stop cycles and maximum load rate within the preset time period are obtained. Based on the average load rate, the total number of start-stop cycles and the maximum load rate, the model-operating condition matching coefficient is obtained. The maintenance record data at each moment within the preset time period is converted into time-series features to obtain maintenance time-series features. Statistical features, trend features, and frequency domain features are extracted from the time-series data in the dataset to obtain a high-dimensional time-series dynamic feature set. The equivalent thermal stress current, the model-operating condition matching coefficient, the maintenance timing characteristics, and the high-dimensional time-series dynamic feature set at each time point are concatenated to obtain the final high-dimensional feature vector at each time point. The adaptive multi-scale feature extraction module in the adaptive hybrid deep learning prediction model is used to extract features from each of the final high-dimensional feature vectors to obtain local feature representations. The stage encoding vector and the local feature representation are processed using the stage-aware attention mechanism in the adaptive hybrid deep learning prediction model to obtain the calibrated local feature representation. The stage encoding vector is obtained by processing the dataset based on a lightweight fully connected network. Specifically, the generation process is as follows: based on the dataset, the cumulative operating years of the equipment, the average load rate, and the cumulative environmental stress index are obtained. The cumulative operating years of the equipment, the average load rate, and the cumulative environmental stress index are input into the lightweight fully connected network for processing to obtain the stage encoding vector. The calibrated local feature representation and the stage encoding vector are input into the stage-aware temporal modeling module of the adaptive hybrid deep learning prediction model for processing to obtain global context features; The global context features are processed using the fully connected layer in the adaptive hybrid deep learning prediction model to obtain the predicted remaining lifetime of the power grid equipment.

2. The method for predicting the remaining life of power grid equipment as described in claim 1, characterized in that, The model-operating condition matching coefficient is obtained based on the average load rate, the total number of start-stop cycles, and the maximum load rate, including: Based on the basic attribute data within the preset time period, the rated load rate and rated start-stop frequency of the power grid equipment are obtained; The rated load rate, the rated number of start-stop cycles, the average load rate, the total number of start-stop cycles, and the maximum load rate are input into the matching degree function for calculation to obtain the model-operating condition matching degree coefficient.

3. The method for predicting the remaining life of power grid equipment as described in claim 1, characterized in that, The step of converting maintenance record data at each moment within the preset time period into time-series features to obtain maintenance time-series features includes: Construct a dictionary of defect types and a dictionary of handling measures; Based on the defect type dictionary and the handling measure dictionary, the defect class description and handling measure description in the maintenance record are respectively converted into numerical codes; Each of the numerical codes is input into the scoring model to obtain a score for the severity of the defect; Based on the timestamp corresponding to the maintenance record, the numerical code, and the defect severity score, the maintenance sequence characteristics are obtained.

4. The method for predicting the remaining life of power grid equipment as described in claim 1, characterized in that, The adaptive multi-scale feature extraction module in the adaptive hybrid deep learning prediction model is used to extract features from each of the final high-dimensional feature vectors to obtain local feature representations, including: The multi-scale deformable convolutional layer in the adaptive multi-scale feature extraction module is used to extract features from each of the final high-dimensional feature vectors to obtain feature representations at different scales. The feature pyramid fusion layer in the adaptive multi-scale feature extraction module is used to fuse the feature representations at various scales to obtain local feature representations.

5. A system for predicting the remaining life of power grid equipment, characterized in that, include: The acquisition module is used to acquire various types of data from power grid equipment within a preset time period to obtain a data set, wherein the data set includes at least basic attribute data, operation data, maintenance record data, and environmental data; The cross-source interaction feature construction module is used to filter the temperature rise coefficient and temperature data of the power grid equipment at each moment of the preset time based on the basic attribute data and the environmental data. Based on the current data and the temperature rise coefficient of the operating data at each moment, the temperature rise value is obtained. Based on the temperature rise value and the temperature data corresponding to each moment, the equivalent thermal stress current at each moment is obtained. Based on the operating data, the average load rate, the total number of start-stops and the maximum load rate within the preset time are obtained. Based on the average load rate, the total number of start-stops and the maximum load rate, the model-operating condition matching coefficient is obtained. The high-dimensional time-series dynamic feature acquisition module is used to convert the maintenance record data at each moment within the preset time into time-series features to obtain maintenance time-series features, and to extract statistical features, trend features and frequency domain features from the time-series data in the dataset to obtain a high-dimensional time-series dynamic feature set. The feature splicing module is used to splice the equivalent thermal stress current, the model-operating condition matching coefficient, the maintenance timing features, and the high-dimensional timing dynamic feature set at each time point to obtain the final high-dimensional feature vector at each time point. The remaining lifetime prediction module is used to extract features from each of the final high-dimensional feature vectors using the adaptive multi-scale feature extraction module in the adaptive hybrid deep learning prediction model to obtain local feature representations. The stage-aware attention mechanism in the adaptive hybrid deep learning prediction model is then used to process the stage encoding vector and the local feature representations to obtain calibrated local feature representations. The stage encoding vector is obtained by processing the dataset using a lightweight fully connected network. Specifically, the generation process is as follows: based on the dataset, the cumulative operating years of the equipment, the average load rate, and the cumulative environmental stress index are obtained. The cumulative operating years of the equipment, the average load rate, and the cumulative environmental stress index are then input into the lightweight fully connected network for processing to obtain the stage encoding vector. The calibrated local feature representation and the stage encoding vector are input into the stage-aware temporal modeling module of the adaptive hybrid deep learning prediction model for processing to obtain global context features. The fully connected layer in the adaptive hybrid deep learning prediction model is then used to process the global context features to obtain the remaining lifetime prediction value of the power grid equipment.

6. A computer device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the power grid equipment remaining life prediction method as described in any one of claims 1 to 4.

7. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the steps of the power grid equipment remaining life prediction method as described in any one of claims 1 to 4.

Images