Electric energy meter service life prediction method based on multi-source data

Abstract

The invention provides an electric energy meter service life prediction method based on multi-source data, and belongs to the technical field of metering detection equipment. Multi-source time sequence data of an electric energy meter is collected, a multi-source time sequence data set is formed, and a training data set and an online monitoring data set are constructed; constructing an improved double-stage attention mechanism convolutional long-short term memory network as a data driving model; constructing a physical information neural network based on a failure physical equation as a physical model, and performing joint training on the data driving model and the physical information neural network; and performing health index calculation on the online electric energy meter data by adopting the bidirectional driving fusion model subjected to joint training, and outputting a residual service life prediction result based on a residual consensus prediction mechanism. The method has the advantages that the data-driven model and the physical model are subjected to interactive training on the parameter level, so that the final model has the flexible learning ability for data features and the internal compliance for physical rules, and the rationality of electric energy meter health state evaluation is improved.

Description

Technical Field

[0001] This invention belongs to the field of metering and testing equipment technology, and specifically relates to a method for predicting the lifespan of electricity meters based on multi-source data. Background Technology

[0002] With the deepening of smart grid construction, the reliability of electricity meters, as key terminals for electricity metering and data acquisition, is crucial to the stability of the entire power system. Because electricity meters are typically installed outdoors or in complex industrial environments, they are subjected to various stresses, including electrical, thermal, and mechanical stresses, which gradually degrade the performance of their internal components, eventually leading to inaccurate metering or functional failure. Therefore, accurately predicting the remaining service life of electricity meters helps power grid companies develop scientific replacement and maintenance strategies, avoid economic losses from sudden failures, and facilitate the transition from periodic maintenance to condition-based maintenance, thus possessing significant engineering application value.

[0003] Currently, industry and academia have conducted extensive research in the field of equipment life prediction, resulting in several typical technical approaches: One type of method focuses on extracting statistical patterns or time-series features from the historical operating data of equipment to build predictive models. The effectiveness of this type of method largely depends on the size and completeness of the dataset.

[0004] Another approach attempts to establish mathematical models based on known failure physical mechanisms, with the advantage that the models have clear physical meanings.

[0005] However, in practical engineering applications, both relying solely on data and relying solely on physical models face inherent limitations. Especially in the complex operating environment of the power grid, the equipment degradation process is affected by the coupling of multiple stresses and factors, and the accuracy of assessing the health status of electricity meters needs to be improved. Summary of the Invention

[0006] In view of the above, this invention addresses the shortcomings of existing technologies by providing a method for predicting the lifespan of electricity meters based on multi-source data. To solve the aforementioned technical problems, the technical solution adopted by this invention includes: collecting multi-source time-series data from electricity meters to form a multi-source time-series dataset; constructing a training dataset and an online monitoring dataset; the multi-source data includes operational data, environmental data, and event data; constructing an improved two-stage attention mechanism convolutional long short-term memory network as a data-driven model to extract deep feature vectors reflecting the health status of the electricity meters from the multi-source time-series data; constructing a physical information neural network based on failure physics equations as a physical model, wherein physical laws are embedded in the network in the form of differential equation residual loss; establishing a bidirectional driving fusion training framework to jointly train the data-driven model and the physical information neural network; the joint training includes a forward physical correction path and a reverse physical constraint path to achieve deep interaction between the data-driven model and the physical model at the parameter level; using the jointly trained bidirectional driving fusion model to calculate health indicators from online electricity meter data and output the remaining lifespan prediction result based on a residual consensus prediction mechanism.

[0007] Furthermore, the operating data includes voltage, current, active power, reactive power, power factor, and metering error value; the environmental data includes the chip junction temperature collected by the built-in temperature sensor of the energy meter; and the event data includes logs of power outage events, zeroing events, time synchronization events, and fault alarm events.

[0008] Furthermore, the chip junction temperature is expressed as: in, For chip junction temperature, This refers to the case temperature or ambient temperature. This refers to the thermal resistance between the chip and the environment. This represents the typical power consumption of the chip.

[0009] Furthermore, an improved two-stage attention mechanism convolutional long short-term memory network is constructed as a data-driven model. The method for extracting deep feature vectors reflecting the health status of electricity meters from the multi-source time-series data includes: The improved two-stage attention mechanism convolutional long short-term memory network sequentially includes a variable attention layer, a convolutional long short-term memory network layer, and a temporal attention layer. The variable attention layer is used to assign dynamic weights to variables from different data sources, and its calculation process is as follows: Calculate the attention score for each variable. : ;in, It is the value of the d-th variable at time t. It is the hidden state of the previous time step. When t=1, Initialize to a zero vector; These are trainable parameters; The scores are normalized to obtain the variable attention weight vector; The original input is multiplied element-wise by the weight vector to obtain the weighted input vector. ; The convolutional long short-term memory network layer receives a weighted input sequence. and output the hidden state sequence. ; The temporal attention layer is used to focus on key time segments in historical data, including: For the hidden state at time step t Their attention score for: ,in These are trainable parameters; The attention scores are normalized to obtain the attention weights at each time step; The entire hidden state sequence is weighted and summed according to the attention weights at each time step to obtain a deep feature vector reflecting the health status of the electricity meter.

[0010] Furthermore, the convolutional long short-term memory network layer receives a weighted input sequence. and output the hidden state sequence. The methods include: Input at each time step It is a vector of dimension D; each Reconstructed into a size of Two-dimensional feature map ,in H and W are preset positive integer hyperparameters; The calculation process of the convolutional long short-term memory network layer at each time step t is as follows: First, compute the input gate. Forgotten Gate Output gate and candidate cell status The calculation formula is as follows: Where * represents the convolution operation. It is the hidden state of the previous time step in the form of a feature map. For the corresponding convolution kernel weights, The bias term is σ, which is the sigmoid activation function; when t=1, the hidden state is... and cell state Initialize to zero; Then, update the cell state at the current time step. and hidden state : Where ⊙ represents element-wise multiplication; Finally, the hidden state obtained by updating all time steps. Flattening it into a vector form yields the hidden state sequence. ,in yes The flattened vector.

[0011] Furthermore, methods for constructing a physical information neural network based on failure physics equations as a physical model, where physical laws are embedded in the network in the form of differential equation residual loss, include: The physical information neural network takes time t and environmental stress S as inputs, and health status values ​​as inputs. This is the output; its loss function is... ; in, For data fitting loss, The observed health status; For physical constraint loss, in; The health status is predicted by the physical information neural network, with a value range of [0,1], where 1 represents complete health and 0 represents failure; T(t) is the temperature stress. represents the parameters of the physical model embedded in the network and can be learned, where k is the Boltzmann constant.

[0012] Furthermore, a bidirectional driven fusion training framework is established to jointly train the data-driven model and the physical information neural network; the joint training includes a forward physics correction path and a backward physics constraint path, and the method for achieving deep interaction between the data-driven model and the physical model at the parameter level includes: On the forward path, the extracted depth feature vectors are mapped to physical model parameters using a fully connected layer. The correction amount Δθ dimension space, using the corrected parameters Update the physics model in the physics neural network, where θ is the set of physics model parameters. ; On the reverse path, the health status prediction sequence calculated by the physical neural network based on the current parameters is used. Calculate the residuals with the true observation sequence, and then use the norm of these residuals. As a regularization loss Add to the total loss of the data-driven model; The total loss function for joint training is: in, It is the prediction loss of the data-driven model. , These are hyperparameters used to balance the contributions of the three losses; by minimizing... To synchronize and update the weights of the data-driven model and the physical information neural network.

[0013] Furthermore, the method of using a jointly trained bidirectional driven fusion model to calculate health indicators from online energy meter data and outputting remaining useful life prediction results based on a residual consensus prediction mechanism includes: The deep feature vector and health status value at time t are obtained from the trained data-driven model and physical information neural network, respectively. The deep feature vectors are mapped through a regression layer to the remaining useful life predicted by the data-driven model. The time required for the curve of the health status value changing over time to decay to a preset failure threshold is defined as the remaining service life predicted by the physical information neural network. ; Calculate the predicted residual ; like If τ is the preset consensus threshold, then the final RUL prediction value is Weight Based on the model's historical performance settings; like If the case retrieval arbitration is initiated, the deep feature vector at time t is concatenated with the health status value to form a fused health feature vector. The cosine similarity between the fused health feature vector and the case features in the historical fault case database is calculated. The K most similar cases are selected, and the weighted average of their RUL evolution trajectories is used as the arbitration benchmark. The final RUL prediction value is , where η is the arbitration weight, which increases as δ(t) increases; This is a weighted average of the predictions from the two models.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a bidirectional deep fusion framework of data-driven model and physical model, the limitations of single prediction method are overcome; the data-driven model can automatically mine complex nonlinear degradation laws from massive historical data, while the physical model introduces clear failure mechanisms as constraints; the interactive training of the two at the parameter level enables the final model to have both flexible learning ability of data features and inherent compliance with physical laws, thereby improving the rationality of the assessment of the health status of electricity meters.

[0015] 2. The residual consensus prediction mechanism adopted provides a dynamic credibility assessment for the lifetime prediction results. When the prediction results of the two models are consistent, a weighted average output is used to ensure the stability of the results. When there is a significant discrepancy in the prediction, the system can automatically identify and initiate an arbitration mechanism based on historical cases, and introduce prior knowledge to assist in decision-making. This design reduces the risk of prediction failure due to model misjudgment or encountering rare operating conditions, and enhances the robustness of the method in practical applications.

[0016] 3. The fusion processing capability of this application for multi-source time-series data enables it to comprehensively utilize multi-dimensional information such as operating parameters, environmental stress, and event records to more fully reflect the actual service status of the electricity meter. Compared with methods that rely solely on a single type of data, the health indicators constructed in this application can more sensitively capture early signs of performance degradation, providing a longer warning time for preventive maintenance and helping to optimize asset management. Attached Figure Description

[0017] The present invention will now be described in further detail with reference to the accompanying drawings.

[0018] Figure 1 : A schematic diagram of the overall process in Embodiment 1 of the present invention; Figure 2 : A schematic diagram of the online prediction process of this invention. Detailed Implementation

[0019] To better understand the present invention, the content of the invention is further clearly illustrated below with reference to embodiments and accompanying drawings. However, the scope of protection of the present invention is not limited to the embodiments described below. Numerous specific details are set forth in the following description to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the present invention can be practiced without one or more of these details.

[0020] Example 1: See Figure 1-2 This embodiment of a method for predicting the lifespan of an energy meter based on multi-source data includes: S1. Collect multi-source time-series data from electricity meters to form a multi-source time-series dataset, and construct a training dataset and an online monitoring dataset; the multi-source data includes at least operational data, environmental data, and event data; The multi-source data collected from the electricity meter includes: operational data such as voltage, current, active power, reactive power, power factor, and metering error; these operational parameters directly reflect the electrical operating status of the electricity meter in the actual power grid environment. Voltage and current monitoring is used to assess electrical stress load; the time-series changes in active and reactive power characterize load fluctuations; the power factor is an important indicator of power quality; and the metering error directly reflects the accuracy changes of the electricity meter's core metering function. Environmental data includes the chip junction temperature collected by the electricity meter's built-in temperature sensor; by monitoring junction temperature changes, the intensity of thermal stress experienced by the electricity meter during operation can be quantified, providing accurate temperature load input for a life prediction model based on failure physics, thereby establishing a quantitative relationship between temperature stress and component aging rate. Event data includes logs of power outage events, reset events, time synchronization events, and fault alarm events; power outage events may cause impact stress on the power module, reset and time synchronization events may affect the continuity of metering data, and fault alarms directly indicate potential functional abnormalities. By analyzing the correlation between event sequences and performance parameters, we can identify the impact of specific events on the health status of electricity meters and enhance the model's adaptability to sudden operating conditions.

[0021] The chip junction temperature data is preprocessed. Chip junction temperature is a direct driver of component performance degradation, while case temperature or ambient temperature is an indirect indicator of the external thermal environment. Due to heat generated by chip power consumption within the energy meter, its junction temperature is typically significantly higher than the ambient temperature. Directly using ambient or case temperature for lifespan assessment introduces significant errors and fails to accurately reflect the chip's true aging state. Therefore, establishing a thermal resistance model from the external temperature measurement point to the chip junction is a necessary prerequisite for accurate lifespan prediction. Specifically, based on the thermal resistance model of the energy meter's installation structure, the measured case temperature or ambient temperature is converted into the chip junction temperature, which is more directly relevant to lifespan prediction. The conversion formula is as follows: in, For chip junction temperature, For ambient temperature, This refers to the thermal resistance between the chip and the environment. This represents the typical power consumption of the chip.

[0022] S2. Construct an improved two-stage attention mechanism convolutional long short-term memory network as a data-driven model to extract deep feature vectors reflecting the health status of the electricity meter from the multi-source time-series data; including: The network consists of a variable attention layer, a convolutional long short-term memory network layer, and a temporal attention layer. This three-layer structure design achieves progressive refinement from multivariate inputs to deep features of the health state. The variable attention layer is responsible for filtering important signals in the feature dimension, the convolutional long short-term memory network layer is responsible for capturing local spatial patterns and temporal dependencies, and the temporal attention layer focuses on key degradation stages in the time dimension.

[0023] S21. The variable attention layer is used to assign dynamic weights to variables from different data sources. The calculation process is as follows: Let the input vector of the multivariate time series data at each time step t be denoted as... , where D is the variable dimension; by dynamically assigning weights to the multivariate inputs, this layer can adaptively enhance variable signals that are more relevant to the degradation process (such as temperature or error values ​​under specific operating conditions) and suppress the influence of irrelevant or noisy variables.

[0024] Calculate the attention score for each variable. : The attention scoring mechanism integrates information from the current input features and the network state at the previous time step, so that the allocation of variable weights not only depends on the current observations, but also takes into account historical context information, thereby achieving dynamic feature selection based on temporal context.

[0025] in, It is the value of the d-th variable at time t. It is the hidden state of the previous time step. When t=1, Initialize to a zero vector; These are trainable parameters; Normalize the scores to obtain the variable attention weight vector. : Original input With weight vector Element-wise multiplication yields a weighted input vector. : S22, The convolutional long short-term memory network layer receives the weighted input sequence. and output the hidden state sequence. Convolutional Long Short-Term Memory (LSTM) networks replace the fully connected operations in traditional LSTMs with convolutional operations, effectively capturing local correlations and spatial patterns in multivariate time series, including: The convolutional long short-term memory network layer receives a weighted input sequence. The input at each time step It is a vector of dimension D; the convolutional long short-term memory network layer receives the output of the variable attention layer, and uses multivariate time-series data weighted by variable importance as input to this layer to ensure that subsequent processing focuses on key signals related to degradation; each Reconstructed into a size of Two-dimensional feature map ,in H and W are preset positive integer hyperparameters. Reconstructing a one-dimensional vector into a two-dimensional feature map is to adapt to the convolution operation's requirement for capturing spatial local correlations. This reconstruction transforms the original variable dimension D into a spatially meaningful two-dimensional structure, enabling the convolution kernel to extract local feature patterns in the reconstructed spatial dimension. The settings of hyperparameters H and W determine the spatial layout of the feature map and affect the size of the local receptive field.

[0026] The calculation process of the convolutional long short-term memory network layer at each time step t is as follows: First, compute the input gate. Forgotten Gate Output gate and candidate cell status The calculation formula is as follows: Gating mechanisms are the core of Long Short-Term Memory (LSTM) networks. The input gate controls the degree to which current input information is adopted, the forget gate controls the degree to which historical information is retained, and the output gate controls the degree to which internal states influence external outputs. Candidate cell states are latent representations of the current input information. By replacing traditional fully connected layers with convolution operations, gating computation can effectively utilize the spatial local correlations of input data, enhancing the model's sensitivity to local feature patterns; where * represents the convolution operation. yes The reconstructed feature map, It is the hidden state of the previous time step in the form of a feature map. The corresponding convolutional kernel weights are used to process the current input. Hidden state from the previous moment Spatial features are extracted through convolution operations; The bias term provides the necessary offset capability for the model; σ is the sigmoid activation function, which compresses the gate signal to the [0,1] interval; when t=1, the hidden state... and cell state Initialization to zero; initialization ensures the determinism of the state at the start of sequence processing, providing a starting point for the zero-state assumption.

[0027] Then, update the cell state at the current time step. and hidden state : Cell state The update incorporates the Forgotten Gate's influence on historical states. Selective forgetting and input gate pair candidate states Selective memory enables the transfer and updating of information across time steps. Hidden state. It is generated based on the updated cell state and the regulation of the output gate, and serves as the output at the current moment; where ⊙ represents element-wise multiplication; Finally, the hidden state obtained by updating all time steps. Flattening it into a vector form yields the hidden state sequence. ,in yes The flattened vector has a length of The hidden state is represented as a two-dimensional feature map at each time step. Flattened into a one-dimensional vector This is to adapt to the processing requirements of the subsequent temporal attention layer for vector sequences and to complete the format conversion from spatial features to temporal features.

[0028] Convolutional Long Short-Term Memory (LSTM) network layers, through their unique gated convolution mechanism, preserve temporal dependencies during sequence processing, enhance the ability to extract features from the spatial structure of input data, and provide hidden state sequences rich in spatiotemporal information for subsequent processing.

[0029] S23. The time-series attention layer is used to focus on key time segments in historical data, including: For the hidden state sequence output by the convolutional long short-term memory network layer Each of them It is a vector of length L.

[0030] The temporal attention mechanism analyzes the hidden states of all time steps and can automatically identify historical moments related to key events such as performance degradation inflection points and abnormal state changes, thereby focusing on the key time periods that contribute most to the prediction of the current health status in long-term time series data.

[0031] For the hidden state at time step t Their attention score Calculated using the following formula: The scores are normalized to obtain the attention weights at each time step. : in, These are trainable parameters; The entire hidden state sequence is weighted and summed according to the temporal attention weights to obtain a fixed depth feature vector. : The depth feature vector This refers to the deep feature vector reflecting the health status of the electricity meter extracted by the data-driven model; By combining a two-stage attention mechanism with a convolutional long short-term memory network, this data-driven model adaptively extracts deep features related to performance degradation from multivariate, long-term data. It can take into account the relative importance differences of different monitoring variables, highlight the historical impact of key time segments, and provide information-rich and discriminative feature representations for subsequent lifetime prediction.

[0032] S3. Construct a physical information neural network based on failure physics equations as a physical model, where physical laws are embedded in the network in the form of differential equation residual loss; including: The physical information neural network takes time t and environmental stress S as inputs, and health status as input. For output, the network structure takes time variables and environmental stresses as basic inputs, and learns the evolution of health status over time through neural network parameterization; its loss function... It consists of two parts: , Ensure that the network output matches the actual observation data. This forces the network to obey physical laws, hyperparameters Used to adjust the relative importance between the two; in, For data fitting loss, The observed health status; the data fitting loss is achieved by minimizing the network predictions. Compared with actual observed values The mean square error between the two is used to ensure that the network output is consistent with the actual measurement data.

[0033] The physical constraint loss is constructed based on physical equations describing component degradation. It embeds physical laws into the network training process as soft constraints, enabling the network to not only learn data features but also follow known physical laws. The physical constraint loss is defined as: This physical constraint, based on the Arrhenius model and power-law relations, describes the component performance parameters. The dynamic process of degradation over time. The network's predicted health status. The evolution of the network's learning process is compared with that predicted by the physical equations, and the difference between the two is minimized to make the network learning process conform to the physical laws.

[0034] in; Let C(t) be the health status predicted by the physical information neural network, with a value range of [0,1], where 1 represents complete health and 0 represents failure; C(t) is the normalized physical quantity related to the health status predicted by the network, defined as follows: T(t) is the temperature stress. θ represents the embedded and learnable physical model parameters of the network. By setting the physical model parameters as embedded learnable parameters of the network, the network can automatically learn the most suitable physical model parameters from the data while maintaining the structural constraints of the physical equations; θ is the set of physical model parameters. , where k is the Boltzmann constant.

[0035] S4. Establish a bidirectional driven fusion training framework to jointly train the data-driven model and the physical information neural network; the joint training includes a forward physics correction path and a backward physics constraint path, realizing deep interaction between the data-driven model and the physical model at the parameter level; including: On the forward path, the extracted depth features c are mapped to physical model parameters using a fully connected layer. The correction amount Δθ dimension space, using the corrected parameters The physical model in the Physics-Informed Neural Networks is updated to personalize it; the forward path establishes a directed information flow from data-driven features to physical model parameters. The deep feature c contains device-specific degradation features learned from real-world data, and generates physical parameter corrections Δθ through fully connected layers, enabling the general physical model to adapt to the personalized degradation characteristics of specific devices. On the reverse path, the health status prediction sequence calculated by the physical neural network based on the current parameters is used. Calculate the residuals with the true observation sequence, and then use the norm of these residuals. As a regularization loss This is added to the total loss of the data-driven model; this reverse path feeds the predictions of the physical model back to the data-driven model as constraints. By using the difference between the physical model's predictions and the actual observations as a regularization term, the training process of the data-driven model is guided, ensuring that its predictions not only conform to statistical laws but also align with physical mechanisms, thus enhancing the model's physical rationality and generalization ability.

[0036] The total loss function for joint training is: in, It is the prediction loss of the data-driven model. , These are hyperparameters used to balance the contributions of the three losses; by minimizing... To synchronize and update the weights of the data-driven model and the physical information neural network.

[0037] The joint training framework achieves co-optimization of the two models through a unified loss function. Hyperparameters β and γ respectively regulate the relative weights of the physical model loss and the regularization term. By balancing the contributions of the three loss terms, the two models can achieve deep interaction while maintaining their respective advantages, ultimately resulting in a fusion prediction model that conforms to both the data distribution and physical laws.

[0038] S5, see reference Figure 2 A jointly trained bidirectional driven fusion model is used to calculate health indicators from online energy meter data, and the remaining useful life prediction results are output based on a residual consensus prediction mechanism; including: Health status information at time t is obtained from both the trained data-driven model and the physical information neural network; the health status information of the data-driven model is its output deep feature vector c, and the health status information of the physical information neural network is its output health status value. The deep feature vector *c* contains the degradation pattern features learned by the data-driven model from historical data, while *d* represents the health status assessment derived by the physical model based on mechanisms. These two representations reflect the health status of the device from different perspectives, providing a multi-perspective information foundation for subsequent fusion decisions.

[0039] The deep feature vector c is mapped through a regression layer to the remaining useful life predicted by the data-driven model. The health status value The time required for the curve to decay to a preset failure threshold over time is defined as the remaining useful life predicted by the physical information neural network. This step transforms abstract health status information into specific remaining lifespan predictions. The regression layer maps from a high-dimensional feature space to a lifespan scalar, while the failure threshold-based calculation directly links the health status trajectory with lifespan. These two transformation methods preserve the characteristics of their respective models: the data-driven model focuses on statistical regularities, while the physical model focuses on mechanistic derivation.

[0040] Calculate the predicted residual Smaller residuals indicate that the predictions obtained by the two models based on different principles corroborate each other, enhancing the credibility of the predictions; larger residuals suggest that there may be differences in model cognition or special operating conditions, requiring further analysis and judgment.

[0041] like If τ is the preset consensus threshold, then the final RUL prediction value is Weight The consensus mechanism can be configured based on the historical performance of the models, and adopts a weighted fusion strategy when the models make consistent predictions. The weight configuration allows for dynamic adjustment based on the historical accuracy of each model, giving a higher weight to the more stable model in the fusion result, thereby optimizing prediction accuracy while maintaining a balance between models. like Then initiate case retrieval arbitration: combine the depth feature vector c at time t with the health status value The data are concatenated into a fused health feature vector. The cosine similarity between this vector and the features of cases in the historical failure case database is calculated. The K most similar cases are selected, and the weighted average of their RUL evolution trajectories is used as the arbitration benchmark. The arbitration mechanism resolves significant disagreements between models by incorporating prior historical knowledge. It integrates the representational advantages of health feature vectors (comprehensive data-driven approach) and physical models, and case retrieval ensures that arbitration outcomes are based on real historical experience. Similarity weighting allows historical cases more closely related to the current state to have a greater impact on the arbitration results.

[0042] The final RUL prediction value is , where η is the arbitration weight, which increases as δ(t) increases; The system uses a weighted average of the predictions from the two models; an adaptive weighting mechanism enables confidence-based decision fusion. The arbitration weights are positively correlated with the residual magnitude, ensuring greater reliance on historical experience when model discrepancies are large, while retaining the contribution of the original predictions when discrepancies are small. This design allows the system to dynamically adjust its decision-making strategy based on the level of uncertainty in the prediction results.

[0043] Technical effects of this embodiment: 1. By constructing a bidirectional deep fusion framework of data-driven model and physical model, the limitations of single prediction method are overcome; the data-driven model can automatically mine complex nonlinear degradation laws from massive historical data, while the physical model introduces clear failure mechanisms as constraints; the interactive training of the two at the parameter level enables the final model to have both flexible learning ability of data features and inherent compliance with physical laws, thereby improving the rationality of the assessment of the health status of electricity meters.

[0044] 2. The residual consensus prediction mechanism adopted provides a dynamic credibility assessment for the lifetime prediction results. When the prediction results of the two models are consistent, a weighted average output is used to ensure the stability of the results. When there is a significant discrepancy in the prediction, the system can automatically identify and initiate an arbitration mechanism based on historical cases, and introduce prior knowledge to assist in decision-making. This design reduces the risk of prediction failure due to model misjudgment or encountering rare operating conditions, and enhances the robustness of the method in practical applications.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention, as long as they do not depart from the spirit and scope of the technical solutions of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the lifespan of an energy meter based on multi-source data, characterized in that, include: Multi-source time-series data from electricity meters are collected to form a multi-source time-series dataset, and a training dataset and an online monitoring dataset are constructed. The multi-source data includes operational data, environmental data, and event data. An improved two-stage attention mechanism convolutional long short-term memory network is constructed as a data-driven model to extract deep feature vectors reflecting the health status of electricity meters from the multi-source time-series data. A physical information neural network based on failure physics equations is constructed as a physical model, in which physical laws are embedded in the network in the form of differential equation residual loss; A bidirectional driven fusion training framework is established to jointly train the data-driven model and the physical information neural network; the joint training includes a forward physical correction path and a reverse physical constraint path, realizing deep interaction between the data-driven model and the physical model at the parameter level. A jointly trained bidirectional driven fusion model is used to calculate health indicators from online energy meter data, and the remaining service life prediction results are output based on the residual consensus prediction mechanism.

2. The method for predicting the lifespan of an energy meter based on multi-source data as described in claim 1, characterized in that, The operational data includes voltage, current, active power, reactive power, power factor, and metering error values; the environmental data includes the chip junction temperature collected by the built-in temperature sensor of the energy meter; and the event data includes logs of power outage events, zeroing events, time synchronization events, and fault alarm events.

3. The method for predicting the lifespan of an energy meter based on multi-source data as described in claim 2, characterized in that, The chip junction temperature is expressed as follows: in, For chip junction temperature, This refers to the case temperature or ambient temperature. This refers to the thermal resistance between the chip and the environment. This represents the typical power consumption of the chip.

4. The method for predicting the lifespan of an energy meter based on multi-source data as described in claim 1, characterized in that, A method for constructing an improved two-stage attention mechanism convolutional long short-term memory network as a data-driven model to extract deep feature vectors reflecting the health status of electricity meters from the multi-source time-series data includes: The improved two-stage attention mechanism convolutional long short-term memory network sequentially includes a variable attention layer, a convolutional long short-term memory network layer, and a temporal attention layer. The variable attention layer is used to assign dynamic weights to variables from different data sources, and its calculation process is as follows: Calculate the attention score for each variable. : ;in, It is the value of the d-th variable at time t. It is the hidden state of the previous time step. When t=1, Initialize to a zero vector; These are trainable parameters; The scores are normalized to obtain the variable attention weight vector; The original input is multiplied element-wise by the weight vector to obtain the weighted input vector. ; The convolutional long short-term memory network layer receives a weighted input sequence. and output the hidden state sequence. ; The temporal attention layer is used to focus on key time segments in historical data, including: For the hidden state at time step t Their attention score for: ,in These are trainable parameters; The attention scores are normalized to obtain the attention weights at each time step; The entire hidden state sequence is weighted and summed according to the attention weights at each time step to obtain a deep feature vector reflecting the health status of the electricity meter.

5. The method for predicting the lifespan of an energy meter based on multi-source data as described in claim 4, characterized in that, The convolutional long short-term memory network layer receives a weighted input sequence. and output the hidden state sequence. The methods include: Input at each time step It is a vector of dimension D; each Reconstructed into a size of Two-dimensional feature map ,in H and W are preset positive integer hyperparameters; The calculation process of the convolutional long short-term memory network layer at each time step t is as follows: First, compute the input gate. Forgotten Gate Output gate and candidate cell status The calculation formula is as follows: Where * represents the convolution operation. It is the hidden state of the previous time step in the form of a feature map. For the corresponding convolution kernel weights, The bias term is σ, which is the sigmoid activation function; when t=1, the hidden state is... and cell state Initialize to zero; Then, update the cell state at the current time step. and hidden state : Where ⊙ represents element-wise multiplication; Finally, the hidden state obtained by updating all time steps. Flattening it into a vector form yields the hidden state sequence. ,in yes The flattened vector.

6. The method for predicting the lifespan of an energy meter based on multi-source data as described in claim 4, characterized in that, Methods for constructing a physical information neural network based on failure physics equations as a physical model, where physical laws are embedded in the network in the form of differential equation residual loss, include: The physical information neural network takes time t and environmental stress S as inputs, and health status values ​​as inputs. This is the output; its loss function is... ; in, For data fitting loss, The observed health status; For physical constraint loss, in; The health status is predicted by the physical information neural network, with a value range of [0,1], where 1 represents complete health and 0 represents failure; T(t) is the temperature stress. represents the parameters of the physical model embedded in the network and can be learned, where k is the Boltzmann constant.

7. The method for predicting the lifespan of an energy meter based on multi-source data as described in claim 6, characterized in that, A bidirectional driven fusion training framework is established to jointly train the data-driven model and the physical information neural network. The joint training includes a forward physics correction path and a backward physics constraint path. The method for achieving deep interaction between the data-driven model and the physical model at the parameter level includes: On the forward path, the extracted depth feature vectors are mapped to physical model parameters using a fully connected layer. The correction amount Δθ dimension space, using the corrected parameters Update the physics model in the physics neural network, where θ is the set of physics model parameters. ; On the reverse path, the health status prediction sequence calculated by the physical neural network based on the current parameters is used. Calculate the residuals with the true observation sequence, and then use the norm of these residuals. As a regularization loss Add to the total loss of the data-driven model; The total loss function for joint training is: in, It is the prediction loss of the data-driven model. , These are hyperparameters used to balance the contributions of the three losses; by minimizing... To synchronize and update the weights of the data-driven model and the physical information neural network.

8. The method for predicting the lifespan of an energy meter based on multi-source data as described in claim 7, characterized in that, The method of using a jointly trained bidirectional driven fusion model to calculate health indicators from online energy meter data and outputting remaining useful life prediction results based on a residual consensus prediction mechanism includes: The deep feature vector and health status value at time t are obtained from the trained data-driven model and physical information neural network, respectively. The deep feature vectors are mapped through a regression layer to the remaining useful life predicted by the data-driven model. The time required for the curve of the health status value changing over time to decay to a preset failure threshold is defined as the remaining service life predicted by the physical information neural network. ; Calculate the predicted residual ; like If τ is the preset consensus threshold, then the final RUL prediction value is Weight Based on the model's historical performance settings; like If the case retrieval arbitration is initiated, the deep feature vector at time t is concatenated with the health status value to form a fused health feature vector. The cosine similarity between the fused health feature vector and the case features in the historical fault case database is calculated. The K most similar cases are selected, and the weighted average of their RUL evolution trajectories is used as the arbitration benchmark. The final RUL prediction value is , where η is the arbitration weight, which increases as δ(t) increases; This is a weighted average of the predictions from the two models.

Images