Charging station network loss evaluation method, system and device based on MFB-TEB model and medium

Abstract

The invention relates to the technical field of charging station network loss assessment, and discloses a charging station network loss assessment method, system and device based on an MFB-TEB model, and a medium. A Transform encoder module and a multi-receptive field feature extraction module are used for respectively extracting global time sequence dependence features and multi-scale nonlinear features, and the global time sequence dependence features and the multi-scale nonlinear features are fused. The method can comprehensively and accurately capture the complex characteristics of the network operation of the charging station, and effectively solves the problems in the prior art that the network loss evaluation of the charging station is inaccurate, the complex and changeable charging load and related influence factors are difficult to process, and the like.

Description

Technical Field

[0001] This invention relates to the field of charging station network loss assessment technology, and in particular to a charging station network loss assessment method, system, equipment and medium based on the MFB-TEB model. Background Technology

[0002] At present, as the penetration rate of electric vehicles becomes higher and higher, the electric vehicle charging stations required for electric vehicles have become a core node of a new type of power load. Therefore, the network loss problem of electric vehicle charging stations has become a serious problem in the power system.

[0003] Among these serious problems, one is that the spatiotemporal distribution characteristics of the charging load corresponding to electric vehicle charging stations are irregular, which can lead to the transformer operating under light or overload conditions, thus exacerbating no-load losses and copper losses. On the other hand, the nonlinear load of electric vehicle charging piles can generate harmonic pollution, which can further reduce the power factor. If the reduced power factor is combined with factors such as equipment aging and line impedance, it can cause even greater additional energy loss.

[0004] According to existing calculations, the losses at a typical electric vehicle fast charging station may account for 8%-15% of the total charging volume, and this figure increases non-linearly with scale expansion. Furthermore, mismatches between the layout of electric vehicle charging stations and their embedded power grid structure, or the lack of intelligent scheduling mechanisms, can further amplify the loss effect. This not only increases the operating costs of the power grid but also affects the stability of the power distribution system.

[0005] Therefore, a method for assessing network losses at charging stations is needed. Summary of the Invention

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] Therefore, this invention provides a charging station network loss assessment method, system, device, and medium based on the MFB-TEB model, which can solve the problems of inaccurate charging station network loss assessment and inability to effectively cope with complex and variable charging loads and related influencing factors in the prior art.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for evaluating the network loss of charging stations based on the MFB-TEB model, comprising: Acquire multi-source operational data of the charging station, including at least equipment parameters, real-time electrical quantity data, environmental monitoring data, and user charging behavior data; The multi-source runtime data is preprocessed to generate a standardized input dataset, which is a dataset that is time-consistent, logically complete, and free of anomalies. The standardized input dataset is input into the Transformer encoder module, and global temporal dependency features in the data are extracted through a self-attention mechanism. The standardized input dataset is simultaneously input into the multi-receptive field feature extraction module, which uses receptive fields of different scales to extract local and global nonlinear features in parallel. By fusing the temporal dependency features output by the Transformer encoder module with the multi-scale features output by the multi-receptive field feature extraction module, a joint feature representation is constructed. Based on the joint feature representation, the loss assessment result of the charging station network in the current period is output through a lightweight regression structure.

[0009] As a preferred embodiment of the charging station network loss assessment method based on the MFB-TEB model described in this invention, the preprocessing includes: Time alignment and format standardization are performed on raw data from vehicle networking platforms and electricity consumption information collection systems; Detect and remove abnormal data points that do not conform to physical laws or equipment operating characteristics; Missing data segments are filled using interpolation or the average of nearby time periods; Verify the logical consistency of voltage, current, and power data across timestamps.

[0010] As a preferred embodiment of the charging station network loss assessment method based on the MFB-TEB model described in this invention, the Transformer encoder module contains only an encoder structure and does not contain a decoder. Internally, it includes a position encoding layer, a multi-head self-attention sub-layer, and a feedforward neural network sub-layer. Each sub-layer is followed by a residual connection and a layer normalization operation to reduce computational complexity and retain the position information of the input sequence.

[0011] As a preferred embodiment of the charging station network loss assessment method based on the MFB-TEB model described in this invention, the equipment parameters in the multi-source operating data include transformer model, line impedance, CT ratio and equipment service life. The real-time electrical quantity data includes voltage, current, active power, and cumulative electricity consumption; The environmental monitoring data includes ambient temperature and relative humidity; The user charging behavior data includes charging start and end times, charging power curves, and charging pile usage frequency.

[0012] As a preferred embodiment of the charging station network loss assessment method based on the MFB-TEB model described in this invention, the multi-receptive field feature extraction module adopts a parallel multi-branch convolutional structure, with each branch configured with convolutional kernels of different sizes, to synchronously capture the local change patterns and long-term trend features of the input data at different time scales.

[0013] As a preferred embodiment of the charging station network loss assessment method based on the MFB-TEB model described in this invention, the convolutional kernels of different sizes include 3×3, 5×5 and 7×7 convolutional kernels. The feature maps output by each branch are concatenated through channels and then compressed through a dimensionality reduction convolutional layer. The dimensionality reduction convolutional layer is used to control the model complexity and improve the generalization ability.

[0014] This preferred approach can effectively extract multi-scale features without significantly increasing the computational load of the model.

[0015] As a preferred embodiment of the charging station network loss assessment method based on the MFB-TEB model described in this invention, the step of fusing joint feature representation includes: The feature vector output by the Transformer encoder is concatenated with the feature map output by the multi-receptive field feature extraction module by channel dimension, and the feature space is aligned by upsampling or linear mapping operation to form a unified high-dimensional feature representation. The lightweight regression structure includes a fully connected layer and an activation function connected in sequence, which is used to map the joint feature representation into a scalar form of network loss rate, wherein the loss rate characterizes the proportion of power loss of the charging station per unit time. When the charging station has complete topology information, the line length, power supply radius, and conductor type are added as optional input features to the standardized input dataset.

[0016] Secondly, the present invention provides a charging station network loss assessment system based on the MFB-TEB model, comprising: The data acquisition module is used to acquire multi-source operating data of the charging station. The multi-source operating data includes at least equipment parameters, real-time electrical quantity data, environmental monitoring data, and user charging behavior data. The preprocessing module is used to preprocess the multi-source running data to generate a standardized input dataset, wherein the standardized input dataset is a dataset that is time-consistent, logically complete, and free of anomalies. The first feature extraction module is used to input the standardized input dataset into the Transformer encoder module and extract global temporal dependency features from the data through a self-attention mechanism. The second feature extraction module is used to simultaneously input the standardized input dataset into the multi-receptive field feature extraction module, and extract local and global nonlinear features in parallel using receptive fields of different scales. The fusion module is used to fuse the temporal dependency features output by the Transformer encoder module with the multi-scale features output by the multi-receptive field feature extraction module to construct a joint feature representation; The evaluation module is used to output the loss evaluation result of the charging station network in the current period based on the joint feature representation and through a lightweight regression structure.

[0017] Thirdly, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0018] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0019] Compared with existing technologies, the beneficial effect of this invention is that it proposes a charging station network loss assessment method based on the MFB-TEB model. This method can fully utilize multi-source operational data, extracting global temporal dependency features and multi-scale nonlinear features through a Transformer encoder module and a multi-receptive field feature extraction module, respectively, and then fusing the two. This approach can comprehensively and accurately capture the complex characteristics of charging station network operation, effectively addressing the problems of inaccurate charging station network loss assessment and difficulty in handling complex and variable charging loads and related influencing factors in existing technologies. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 The present invention provides a flowchart of a method for evaluating the network loss of a charging station based on the MFB-TEB model, which is an embodiment of the present invention.

[0022] Figure 2 A Transformer architecture diagram of a charging station network loss assessment method based on the MFB-TEB model provided in one embodiment of the present invention. Figure 3This is a Transformer encoder architecture diagram of a charging station network loss assessment method based on the MFB-TEB model, provided as an embodiment of the present invention.

[0023] Figure 4 This is a diagram of a multi-receptive field feature extraction module for a charging station network loss assessment method based on the MFB-TEB model, provided as an embodiment of the present invention.

[0024] Figure 5 The diagram shows the TEB-MFB model architecture of a charging station network loss assessment method based on the MFB-TEB model, as provided in one embodiment of the present invention.

[0025] Figure 6 This is an internal structure diagram of an electronic device for a charging station network loss assessment method based on the MFB-TEB model, provided as an embodiment of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for evaluating the network loss of charging stations based on the MFB-TEB model, including: This invention provides a method that can effectively solve the problems mentioned above. The following will describe in detail how to implement the charging station network loss assessment method based on the MFB-TEB model with a number of embodiments. Figure 1 A flowchart of a method for evaluating charging station network losses based on the MFB-TEB model is shown, including: S101, acquire multi-source operation data of the charging station, which includes at least equipment parameters, real-time electrical quantity data, environmental monitoring data and user charging behavior data; In this embodiment of the invention, the equipment parameters in the multi-source operating data include transformer model, line impedance, CT ratio and equipment service life. Real-time electrical quantity data includes voltage, current, active power, and cumulative energy consumption; Environmental monitoring data includes ambient temperature and relative humidity; User charging behavior data includes charging start and end times, charging power curves, and frequency of charging station usage.

[0028] In some embodiments, the equipment parameters in the multi-source operating data include transformer model, line impedance, CT ratio, and equipment service life. These parameters directly determine the inherent loss characteristics of the charging station's electrical system. For example, if a charging station uses an SCB11 dry-type transformer, its no-load loss and load loss indicators are known. Combined with the information that the transformer has been in service for 8 years, the impact of its core aging on hysteresis and eddy current losses can be assessed. Line impedance reflects the resistance and reactance of the cable between the transformer and each charging pile. For example, a power supply line with a length of 120 meters using YJV22-3×95mm² copper core cable has a fixed impedance value per unit length, which can be used to calculate the Joule heat loss generated when current passes through it. The CT ratio is used to convert the large primary current into a small secondary current for metering and monitoring. For example, a current transformer with a CT ratio of 100 / 5 outputs a 4A signal when the current is 80A. If the ratio is incorrectly configured or the current transformer is aged, it will lead to distortion of the power acquisition, thereby affecting the accuracy of loss calculation.

[0029] In some embodiments, real-time electrical quantity data includes voltage, current, active power, and cumulative electricity consumption. This data is synchronously acquired from the electricity consumption information collection system at 15-minute granularities. For example, at 10:15 AM on June 3, 2025, the line voltage measured on the low-voltage side of the transformer substation was 379.5V, the three-phase current was 0.012kA, the corresponding active power was 9.1kW, and the cumulative electricity consumption was 15620.19kWh. These values ​​must satisfy P≈ The physical relationship is considered. If there is a situation where the voltage is normal but the current drops suddenly and the power does not change accordingly, it is considered a logical anomaly and is eliminated. The accumulated electricity is used to back-calculate the actual electricity consumption during the period. Combined with the charging electricity reported by the charging pile, the network loss boundary can be preliminarily estimated.

[0030] In some embodiments, environmental monitoring data includes ambient temperature and relative humidity. These two data points are collected in real time by sensors deployed near the charging station's power distribution room and outdoor charging piles. For example, at 11:00 on June 3, 2025, the ambient temperature rose to 23.6°C and the relative humidity dropped to 38%. High temperatures exacerbate the increase in conductor resistance, leading to increased copper losses in the lines. While low humidity reduces the risk of condensation, it may increase the impact of electrostatic interference on the acquisition of weak electrical signals. After 19:00 in the evening, when the temperature drops to 17.6°C and the humidity rises to 51%, the heat dissipation conditions of the equipment improve, the transformer efficiency increases, and the loss rate decreases accordingly. This coupling relationship between environmental factors and electrical quantities is explicitly modeled in the model.

[0031] In some embodiments, user charging behavior data includes charging start and end times, charging power curves, and charging pile usage frequency. This data originates from the vehicle-to-everything (V2X) platform and the charging pile controller. For example, a DC fast charging pile starts charging an electric taxi at 18:30, with an initial power of 60kW maintained for 20 minutes. Subsequently, due to the increase in battery SOC, the power gradually decreases to 30kW, forming a typical stepped power decay curve. During the same period, 11 out of 16 charging piles at the charging station are in operation, with a usage frequency as high as 68.75%, indicating concentrated load, which can easily lead to transformer overload and harmonic superposition. The dense distribution of charging start and end times also reveals load peak and valley characteristics. For example, 19:00 to 20:00 is the second peak charging period of the day, during which the loss rate is generally higher than during the morning off-peak period. The model achieves accurate dynamic loss assessment by capturing the correlation between such behavioral patterns and electrical responses.

[0032] The multi-source operational data here constitutes the input basis of the MFB-TEB model. Equipment parameters provide static topology constraints, real-time electrical quantities reflect instantaneous operating status, environmental monitoring data characterizes external disturbance conditions, and user charging behavior data reveals the load driving mechanism. After the four types of data are strictly aligned with timestamps, they are preprocessed to generate a standardized input sequence, ensuring that the subsequent feature extraction module can integrate physical laws and data-driven information to output a high-confidence network loss rate.

[0033] It should be noted that step S101 provides comprehensive, heterogeneous, but complementary raw inputs for the entire evaluation process. If any type of data is missing, the subsequent model will be unable to capture the coupling relationship between losses and external disturbances or load patterns, resulting in one-sided feature extraction and distorted evaluation results.

[0034] S102, preprocess the multi-source running data to generate a standardized input dataset. The standardized input dataset is a dataset that is consistent in time, logically complete, and free of anomalies. In this embodiment of the invention, the preprocessing includes: Time alignment and format standardization are performed on raw data from vehicle networking platforms and electricity consumption information collection systems; Detect and remove abnormal data points that do not conform to physical laws or equipment operating characteristics; Missing data segments are filled using interpolation or the average of nearby time periods; Verify the logical consistency of voltage, current, and power data across timestamps.

[0035] In some embodiments, preprocessing includes time alignment and format standardization of raw data from the vehicle networking platform and the electricity information collection system. These two systems typically use different sampling frequencies and time bases. For example, the vehicle networking platform reports charging pile power and charging status at 30-second intervals, while the electricity information collection system records the voltage, current, and cumulative electricity consumption of the transformer substation at a 15-minute granularity. During preprocessing, both types of data need to be resampled to a 15-minute granularity and aligned with the UTC+8 standard timestamp to ensure the comparability of multi-source data within the same time window. At the same time, the naming and units of data fields output by different systems also need to be standardized. For example, "kW", "KW", and "kilowatt" should be uniformly converted to "kW", and "V" and "volt" should be uniformly converted to "V" to avoid errors caused by format ambiguity in subsequent feature extraction.

[0036] In some embodiments, preprocessing includes detecting and removing abnormal data points that do not conform to physical laws or equipment operating characteristics. These abnormalities may originate from sensor malfunctions, communication packet loss, or human input errors. For example, during a 15-minute period, the system records a current of 0.012kA but the cumulative power suddenly increases by 47kWh. According to P = ΔE / Δt, the calculated power far exceeds the station's maximum power supply capacity of 120kW, which clearly violates the law of conservation of energy. Another example is that the ambient temperature is displayed as -10℃ but the relative humidity is as high as 98%, which is impossible in a southern city in June. Such data points that violate thermodynamic principles are automatically identified and removed. In addition, if a charging pile reports that it has maintained a full power output of 60kW for 8 consecutive hours without any interruption, it does not conform to the operating characteristics of the electric vehicle battery management system to actively reduce power during the high SOC stage, and is also regarded as an abnormality and cleared.

[0037] In some embodiments, preprocessing includes using interpolation or the average of nearby time periods to complete missing data segments. Missing data often occurs during communication interruptions or device restarts. For example, during the period from 12:00 to 13:00 on June 3, 2025, the vehicle network platform did not upload any charging behavior data due to a 4G signal interruption. If there are stable charging records before and after this time period, linear interpolation is used to estimate the charging power curve of the middle time period. For occasional single-point missing data in the electricity information collection system, such as the voltage value at 14:15 being missing, but the data at 14:00 and 14:30 being complete and with gentle fluctuations, the average of these two nearby times is directly used as the replacement value for 14:15. If the missing data lasts for more than two consecutive time periods without reliable context, the time period is marked as unusable and not included in the model input.

[0038] In some embodiments, preprocessing includes verifying the logical consistency of voltage, current, and power data at timestamps. This is a crucial step in ensuring the physical reliability of loss calculations. For example, at 10:30 AM on June 3, 2025, the system recorded an average three-phase voltage of 381.3V, a current of 0.01kA, and an active power of 6.6kW. The cumulative power increased by 30kWh compared to the previous period. If calculated over 15 minutes, the theoretical increase should be approximately 1.65kWh. However, the actual power jump of 30kWh is clearly inconsistent with the instantaneous power, indicating that there is a code skipping or reset error in the power data. Similarly, if the current is zero during a certain period but the cumulative power is still increasing, or if the voltage is below 320V (below the national standard lower limit) but the normal charging power is still maintained, these are considered logical conflicts and require backtracking to the original logs for verification or marking as invalid data.

[0039] The preprocessing process here constitutes the input quality assurance mechanism of the MFB-TEB model. It achieves multi-source collaboration through time alignment, maintains physical authenticity through anomaly removal, improves data continuity through missing data completion, and ensures electrical consistency through logical verification. Finally, it generates a standardized input dataset that is time-consistent, logically complete, and free of anomalies, providing a reliable foundation for the subsequent Transformer encoder and multi-receptive field feature extraction module.

[0040] In this embodiment of the invention, when the charging station has complete topology information, the line length, power supply radius, and conductor type are added as optional input features to the standardized input dataset.

[0041] It should be noted that step S102 ensures that the data input to the deep learning module is of high quality and computable. If timestamps are not aligned, the Transformer and convolution modules will process misaligned temporal signals; if outliers or logically contradictory data are retained, it will cause the feature extraction module to learn incorrect patterns, or even lead to training divergence.

[0042] S103, the standardized input dataset is input into the Transformer encoder module, and global temporal dependency features in the data are extracted through the self-attention mechanism; In this embodiment of the invention, the Transformer encoder module contains only an encoder structure and no decoder. Internally, it includes a position encoding layer, a multi-head self-attention sub-layer, and a feedforward neural network sub-layer. Each sub-layer is followed by a residual connection and a layer normalization operation to reduce computational complexity and preserve the position information of the input sequence.

[0043] In some embodiments, the Transformer encoder module contains only an encoder structure and no decoder. This design focuses on deep feature extraction from input multi-source time-series data rather than sequence generation tasks. Internally, it includes a position encoding layer, a multi-head self-attention sub-layer, and a feedforward neural network sub-layer. Each sub-layer is followed by a residual connection and a layer normalization operation to reduce computational complexity and preserve the positional information of the input sequence. For example, when processing a 96-dimensional input sequence with a 15-minute granularity (covering voltage, current, power, ambient temperature and humidity, device parameters, and charging behavior characteristics), the position encoding layer first embeds a unique position vector for each time step to ensure that the model can distinguish between "high load at 18:00" and "the same value at 02:00", avoiding temporal confusion caused by the Transformer's lack of sequence awareness. In some embodiments, the multi-head self-attention sublayer dynamically captures long-range dependencies between different time steps by computing multiple attention heads in parallel. For example, during the evening peak period from 19:00 to 20:00, the model can automatically identify the implicit association between the surge in the frequency of charging pile use and the increase in transformer loss rate, even if the two are separated by multiple feature dimensions in the original input. In some embodiments, the feedforward neural network sublayers perform nonlinear transformations on the attention output at each time step to enhance feature representation capabilities. For example, the original current value is cross-combined with the line impedance to map a potential Joule heat loss representation. Each sublayer output is residually connected to the input and then normalized by the layer. This not only alleviates the gradient vanishing problem in deep networks but also stabilizes the training process, enabling the model to efficiently model complex temporal patterns even when using only the encoder structure.

[0044] The location encoding here uses a learnable absolute location embedding method instead of a fixed sine function to adapt to the time characteristics of non-uniform sampling or occasional jumps in the charging station operation data.

[0045] It should be noted that step S103 provides a high-order temporal representation with long-range context awareness for the fusion stage.

[0046] S104 simultaneously inputs the standardized input dataset into the multi-receptive-field feature extraction module, and extracts local and global nonlinear features in parallel using receptive fields of different scales; In this embodiment of the invention, the multi-receptive field feature extraction module adopts a parallel multi-branch convolutional structure, with each branch configured with a convolutional kernel of a different size, to simultaneously capture the local change patterns and long-term trend features of the input data at different time scales.

[0047] In some embodiments, the multi-receptive-field feature extraction module employs a parallel multi-branch convolutional structure, with each branch configured with a convolutional kernel of different sizes. This allows for the simultaneous capture of local change patterns and long-term trend features of the input data at different time scales. For example, this module includes three parallel convolutional branches, employing one-dimensional convolutions with kernel sizes of 3, 7, and 15, respectively. The corresponding receptive fields cover time windows of 45 minutes, 105 minutes, and 225 minutes, enabling simultaneous analysis of short-term mutations, intraday fluctuations, and cross-time-period evolution patterns. Among these, the 3-point convolutional kernel branch focuses on capturing instantaneous events in charging behavior, such as a charging pile suddenly starting at 18:15, causing a step increase in current, or a rapid drop in ambient temperature within 10 minutes before a thunderstorm. The 7-point convolutional kernel branch effectively identifies typical load cycles, such as the stable high-power charging platform period formed from 19:00 to 21:00 on weekday evenings, whose corresponding loss rate exhibits a consistently high characteristic; while the 15-point convolutional kernel branch can model more macroscopic trends, such as the efficiency decay caused by the cumulative effect of transformer temperature rise under three consecutive days of high temperatures, or the systematic difference in user usage frequency between weekends and weekdays; each branch is followed by batch normalization and ReLU activation function after the convolution operation to ensure the nonlinear representation capability of the features, and the multi-scale outputs are fused into a unified feature map through channel splicing, which is used for subsequent fully connected regression head to predict the loss rate. The specific size and level can be set by relevant technicians according to actual needs.

[0048] The different convolution kernel sizes are not set arbitrarily, but are determined based on the actual sampling granularity of the charging station operation data and the duration of typical events.

[0049] In this embodiment of the invention, convolutional kernels of different sizes include 3×3, 5×5 and 7×7 convolutional kernels. The feature maps output by each branch are concatenated through channels and then compressed through a dimensionality reduction convolutional layer. The dimensionality reduction convolutional layer is used to control the model complexity and improve the generalization ability.

[0050] It should be noted that step S104 provides a fine-grained characterization of local mutations and medium- to long-term fluctuations. These multi-scale features complement the global view of the Transformer, ensuring that the fused joint representation is neither too smooth nor too fragmented.

[0051] S105, integrates the temporal dependency features output by the Transformer encoder module with the multi-scale features output by the multi-receptive field feature extraction module to construct a joint feature representation; In this embodiment of the invention, the step of fusing joint feature representations includes: The feature vector output by the Transformer encoder is concatenated with the feature map output by the multi-receptive field feature extraction module by channel dimension, and the feature space is aligned by upsampling or linear mapping operation to form a unified high-dimensional feature representation. In some embodiments, the feature vector output by the Transformer encoder and the feature map output by the multi-receptive-field feature extraction module are concatenated by channel dimension, and the feature space is aligned through upsampling or linear mapping operations to form a unified high-dimensional feature representation. For example, the Transformer encoder outputs a temporal feature matrix of shape [96, 128] for the input 96 time steps (corresponding to 24 hours and 15 minutes granularity), where each time step is encoded as a 128-dimensional vector, preserving global dependencies and positional information; while the feature map output by the multi-receptive-field feature extraction module after three parallel convolutions still maintains a temporal structure of [96, C] due to the use of different convolution kernel sizes but the same padding strategy, where C is the sum of the number of channels in each branch, for example, 32+32+32=96 dimensions; since the two have the same length in the time dimension but mismatched in the channel dimension, it is necessary to first concatenate the Transformer encoder with the feature map output by the multi-receptive-field feature extraction module. Linear mapping is performed on either the rmer output or the convolutional output. For example, a learnable fully connected layer can be used to project 128-dimensional Transformer features to 96-dimensional features, or vice versa, 96-dimensional convolutional features can be mapped to 128-dimensional features. After ensuring channel alignment, the features are concatenated along the channel axis to obtain a fused feature sequence of [96,224] or [96,256]. If downsampling is introduced due to convolution operations, a large unfilled convolutional kernel can be used, which reduces the time step. In this case, linear interpolation or transposed convolution upsampling is performed on the shorter sequence to restore it to 96 time steps, ensuring temporal alignment. The resulting unified high-dimensional feature representation contains both long-range semantic associations captured by the Transformer and local dynamic patterns extracted by multi-scale convolutions, providing discriminative input with both global context and multi-granular details for subsequent regression heads. The specific dimensions and layers can be set by technicians according to actual needs.

[0052] In this process, feature space alignment is not a simple copy or truncation, but rather a semantic-level adaptation achieved through parametric mapping. The linear mapping layer automatically learns how to compress or expand feature dimensions during training to minimize information loss. The channel splicing operation preserves the independence of the original features, avoids feature confusion caused by early fusion, and enables the model to autonomously weight the feature contributions from different sources in subsequent fully connected layers. This fusion strategy significantly improves the ability to model complex loss mechanisms. For example, it can identify both the long-term trend of "continuous high temperature causing a slow increase in transformer iron loss" and the local event of "a sudden full-power charging of a certain pile causing a surge in instantaneous copper loss".

[0053] It should be noted that step S105 forms a unified high-dimensional representation that combines global semantics and local details, greatly improving the information completeness of the regression task. If only a single feature source is used, such as only using Transformer or only using convolution, the lightweight regression structure will be difficult to accurately fit the complex nonlinear loss mechanism due to insufficient input information.

[0054] S106, based on joint feature representation, outputs the loss assessment results of the charging station network in the current period through a lightweight regression structure.

[0055] In an embodiment of the present invention, the lightweight regression structure includes a fully connected layer and an activation function connected in sequence, which are used to map the joint feature representation into a scalar form of network loss rate, where the loss rate characterizes the proportion of power loss of the charging station per unit time. In some embodiments, the lightweight regression structure includes a fully connected layer and an activation function connected in sequence to map the joint feature representation to a scalar form of network loss rate, where the loss rate characterizes the proportion of energy loss at the charging station per unit time. For example, after data processing by the Transformer encoder and the multi-receptive field feature extraction module, a fused feature sequence of shape [96, 256] is obtained. This sequence is first passed through a global average pooling layer. Pooling is used to compress the sequence, averaging the 256-dimensional feature vectors at each time step to obtain a 256-dimensional overall feature representation. This step aims to summarize information from the entire time period and reduce subsequent computational complexity. Next, this overall feature is input into a first fully connected layer with 128 neurons, where the ReLU activation function is used to introduce a non-linear transformation, enhancing the model's ability to learn complex patterns. Subsequently, the output passes through a second fully connected layer with 64 neurons to further refine the features, also using the ReLU activation function. Finally, the output layer passes through a single-neuron output layer, using a linear activation function (i.e., no activation), directly outputting a scalar form of the network loss rate. This value accurately reflects the proportion of electrical energy lost by the charging station per unit time. The specific dimensions and layers can be set by technicians according to actual needs.

[0056] For example, in practical applications, the operating data of a charging station on June 3, 2025, after being processed by the above process, predicts that the average network loss rate for that day is 7.8%. This indicates that for every 100kWh of electricity consumed, approximately 7.8kWh cannot be effectively transmitted to the electric vehicle battery due to factors such as line resistance and transformer losses. During the morning peak hours (e.g., 07:00-09:00), the loss rate may drop to 7.3% due to concentrated load and lower temperature. However, during the evening peak hours (e.g., 19:00-21:00), the loss rate rises to 8.2% as the load increases and the ambient temperature rises. Furthermore, if a sudden high-power charging event occurs at a certain moment, the local loss rate may instantly exceed 10%. In this case, the lightweight regression structure can quickly respond and adjust the prediction results, providing an immediate and accurate loss assessment.

[0057] The lightweight design here is reflected in limiting the number and width of fully connected layers, avoiding overfitting while ensuring sufficient expressive power. This allows the model to not only perform well on the training set but also maintain high generalization performance on unseen real-world data, ensuring the real-time performance and accuracy of charging station network loss assessment.

[0058] In summary, the beneficial effects of this invention are that it proposes a charging station network loss assessment method based on the MFB-TEB model, which can fully utilize multi-source operational data. It extracts global temporal dependency features and multi-scale nonlinear features through a Transformer encoder module and a multi-receptive field feature extraction module, respectively, and then fuses the two. This approach can comprehensively and accurately capture the complex characteristics of charging station network operation, effectively addressing the problems of inaccurate charging station network loss assessment and difficulty in handling complex and variable charging loads and related influencing factors in existing technologies.

[0059] Example 2, refer to Figures 2-5 Based on the above embodiments, and in accordance with Tables 1 and 2, a specific implementation of a charging station network loss assessment method based on the MFB-TEB model can be designed as follows: Collect equipment parameters, operating data, environmental data, and user behavior data from the charging station's power distribution network; Furthermore, raw data is acquired and preprocessed from the vehicle-to-everything (V2X) platform and the electricity consumption information collection system. The preprocessing steps are as follows: ① Data needs to be collected from different sources. This data may include time-series data, geographic location information, etc. Integrating this data into a unified data framework ensures consistency in data format and time series.

[0060] ②Remove abnormal data from the original data that does not conform to the actual usage mode of charging piles to improve data quality.

[0061] ③ The main meter of the transformer substation may have missing data, such as meter data at the start, end and transition stages of charging.

[0062] ④ Ensure that the data is logically consistent. For example, are the voltage data after data processing from the same time?

[0063] Furthermore, a Transformer encoder module is designed. The Transformer is a deep learning model based on an attention mechanism that uses self-attention to capture long-range dependencies in sequence data. It can be used to handle various sequence tasks, and its architecture is shown in the attached figure. Figure 2 As shown.

[0064] It's important to note that the complete Transformer architecture includes both an encoder and a decoder, with the decoder significantly increasing computational complexity. When dealing with large-scale measurement data, using only the encoder can reduce the computational load during model training and feature extraction.

[0065] Furthermore, for the sake of model portability, this invention only uses a Transformer encoder that employs an attention mechanism for computation, the architecture of which is shown in the attached figure. Figure 3 As shown; Furthermore, the Transformer encoder consists of a multi-head self-attention module, a feedforward neural network module, residual connections, and normalization. The multi-head self-attention module calculates the correlation between the serialized local features using a self-attention mechanism and uses these correlations as attention weights to integrate all local features, thereby obtaining global contextual information. The feedforward neural network performs nonlinear transformations on the contextual information at each position.

[0066] Furthermore, to accurately capture positional features in sequence data, the Transformer model introduces positional encoding in the form of sine and cosine functions, as follows: (1.1) (1.2) In the formula, pos, c, and dmodel represent the position, dimension, and number of dimensions of the input data, respectively.

[0067] Furthermore, the Transformer possesses powerful features extraction capabilities for time-series data thanks to its self-attention mechanism, which modifies the input sequence by applying a weight matrix to each position. , , Multiplying these vectors and performing a projection transformation yields vectors Q, K, and V. The values ​​of the attention mechanism are obtained by calculating the inner product between Q and K at each point. (1.3) In the formula, Attention scores representing Q, K, and V; d represents the normalized exponential function; d is the dimension of the k vector.

[0068] Furthermore, to avoid performing single attention, equation (1.3) is repeated h times, and the results are then concatenated to form a multi-head self-attention mechanism, i.e. (1.4) In the formula, Wi represents the undetermined weight matrix (i=1, 2, 3...).

[0069] It should be noted that a feedforward neural network typically contains an input layer, one or more hidden layers, and an output layer, operating as a simple fully connected neural network. It transforms the input vector into an output vector through nonlinear transformations to extract more complex features.

[0070] (1.5) In the formula, Let W1 and W2 represent the input to the FFN, where W1 and W2 are the weight matrices to be determined, and b1 and b2 are the bias terms to be determined. =max(0, x).

[0071] Furthermore, each sub-layer of each encoder (Self-Attention layer and FFN layer) is followed by a residual connection and a layer normalization operation. The entire computation process can be represented as: (1.6) In the formula, express Layer input, This represents the input from the previous layer. μ and σ represent respectively The mean and standard deviation.

[0072] It should be noted that the Multi-Receptive Field Feature Extraction Block (MFB) is an important component in deep learning used to capture features at different scales. Its core idea is to enhance the model's adaptability to multi-scale targets by fusing features from different receptive fields through parallel or serial multi-branch structures.

[0073] Furthermore, to improve the feature extraction capability of the line loss estimation model from measurement data, this invention utilizes convolutional kernels of different sizes in different convolutional neural network layers to extract deep-level features of time-series and power data. Large receptive field convolutional kernels can capture a wider range of contextual information and global features; small receptive field convolutional kernels are better at extracting local information. Fusing the features extracted by different convolutional kernels allows the model to grasp both the global picture and the details. Simultaneously, to maintain the model's lightweight nature and reduce the depth and complexity of the convolutional neural network architecture, this invention uses convolutional kernels of 3x3, 5x5, and 7x7 sizes for data feature extraction and fusion to obtain better nonlinear fitting results, which is beneficial for the training and deployment of the line loss estimation model. The module structure is attached. Figure 4 As shown.

[0074] The expression describing the above process is: (1.7) In the formula: Output feature map; This indicates feature map channel splicing; This is a convolution operation; , , These correspond to 3x3, 5x5, and 7x7 convolution operations, respectively. Their specific expressions are as follows: (1.8) In the formula, Input feature map; and Let represent the weights and biases of the i×i convolution, respectively.

[0075] Furthermore, to reduce the computational cost and complexity of the model, the channel dimension needs to be reduced. After obtaining the feature map from Equation (1.7), the multi-receptive field features are combined through a concatenation operation, and then input into a 3x3 convolutional layer to reduce the dimensionality of the concatenated feature map.

[0076] Furthermore, low-voltage residential transformer substations have complex structures, and many factors influence line losses. These factors can be divided into two categories: inherent line parameters, such as line length, line type, and power supply radius; and electrical information, such as user load, voltage, and current values. Electrical information is relatively easy to obtain and is a mandatory input to the model. Inherent line parameters are optional; if the substation has complete inherent line parameters, they can be used as input to the model, as shown in the attached figure. Figure 5 As shown.

[0077] This invention leverages the advantages of both the Transformer model and convolutional computation, and applies residual connections to ensure model performance, thereby improving the estimation results of the line loss estimation model. It can be represented as: (1.9) (1.10) (1.11) (1.12) (1.13) (1.14) (1.15) (1.16) In the formula, , These represent the feature information extracted by the Transformer and MFB modules, respectively. Indicates feature splicing, Indicates upsampling, This represents a linear activation function.

[0078] In one specific embodiment, a pilot application was conducted at a charging station in a certain city. Relevant data from the charging station's transformer substation and charging piles were collected during a certain period on June 3rd. The collected data was processed to obtain the following results: Table 1. Results of data processing on the collected data.

[0079] Table 2. Loss assessment of the processed data using the MFB-TEB model.

[0080] As can be clearly seen from the two tables, the technical solution of the present invention can clearly assess the loss of the charging station network during actual operation.

[0081] Example 3, referring to Figure 6 This embodiment also provides a charging station network loss assessment system based on the MFB-TEB model, including: The data acquisition module is used to acquire multi-source operating data of the charging station. The multi-source operating data includes at least equipment parameters, real-time electrical quantity data, environmental monitoring data, and user charging behavior data. The preprocessing module is used to preprocess multi-source runtime data to generate a standardized input dataset. The standardized input dataset is a dataset that is time-consistent, logically complete, and free of anomalies. The first feature extraction module is used to input the standardized input dataset into the Transformer encoder module and extract global temporal dependency features from the data through a self-attention mechanism. The second feature extraction module is used to simultaneously input the standardized input dataset into the multi-receptive field feature extraction module, and extract local and global nonlinear features in parallel using receptive fields of different scales; The fusion module is used to fuse the temporal dependency features output by the Transformer encoder module with the multi-scale features output by the multi-receptive field feature extraction module to construct a joint feature representation. The evaluation module is used to output the loss evaluation results of the charging station network in the current period based on the joint feature representation and through a lightweight regression structure.

[0082] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0083] This embodiment also provides an electronic device, which can be a terminal, and its internal structure diagram can be as follows: Figure 6 As shown, the electronic device includes a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a charging station network loss assessment method based on the MFB-TEB model. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the device's casing, or an external keyboard, touchpad, or mouse.

[0084] This embodiment also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it performs the following steps: Acquire multi-source operational data of the charging station, which includes at least equipment parameters, real-time electrical quantity data, environmental monitoring data, and user charging behavior data; The multi-source runtime data is preprocessed to generate a standardized input dataset, which is a dataset that is time-consistent, logically complete, and free of anomalies. The standardized input dataset is fed into the Transformer encoder module, and global temporal dependency features in the data are extracted through the self-attention mechanism. The standardized input dataset is simultaneously input into the multi-receptive-field feature extraction module, which uses receptive fields of different scales to extract local and global nonlinear features in parallel. By fusing the temporal dependency features output by the Transformer encoder module with the multi-scale features output by the multi-receptive field feature extraction module, a joint feature representation is constructed. Based on joint feature representation, the loss assessment results of the charging station network in the current period are output through a lightweight regression structure.

[0085] 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. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0086] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0087] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for evaluating charging station network losses based on the MFB-TEB model, characterized in that, include: Acquire multi-source operational data of the charging station, including at least equipment parameters, real-time electrical quantity data, environmental monitoring data, and user charging behavior data; The multi-source runtime data is preprocessed to generate a standardized input dataset, which is a dataset that is time-consistent, logically complete, and free of anomalies. The standardized input dataset is input into the Transformer encoder module, and global temporal dependency features in the data are extracted through a self-attention mechanism. The standardized input dataset is simultaneously input into the multi-receptive field feature extraction module, which uses receptive fields of different scales to extract local and global nonlinear features in parallel. By fusing the temporal dependency features output by the Transformer encoder module with the multi-scale features output by the multi-receptive field feature extraction module, a joint feature representation is constructed. Based on the joint feature representation, the loss assessment result of the charging station network in the current period is output through a lightweight regression structure.

2. The charging station network loss assessment method based on the MFB-TEB model as described in claim 1, characterized in that, The preprocessing includes: Time alignment and format standardization are performed on raw data from vehicle networking platforms and electricity consumption information collection systems; Detect and remove abnormal data points that do not conform to physical laws or equipment operating characteristics; Missing data segments are filled using interpolation or the average of nearby time periods; Verify the logical consistency of voltage, current, and power data across timestamps.

3. The charging station network loss assessment method based on the MFB-TEB model as described in claim 2, characterized in that, The Transformer encoder module contains only the encoder structure and does not contain a decoder. Internally, it includes a position encoding layer, a multi-head self-attention sub-layer, and a feedforward neural network sub-layer. Each sub-layer is followed by a residual connection and a layer normalization operation to reduce computational complexity and preserve the positional information of the input sequence.

4. The charging station network loss assessment method based on the MFB-TEB model as described in claim 3, characterized in that, The equipment parameters in the multi-source operation data include transformer model, line impedance, CT ratio, and equipment service life. The real-time electrical quantity data includes voltage, current, active power, and cumulative electricity consumption; The environmental monitoring data includes ambient temperature and relative humidity; The user charging behavior data includes charging start and end times, charging power curves, and charging pile usage frequency.

5. The charging station network loss assessment method based on the MFB-TEB model as described in claim 4, characterized in that, The multi-receptive-field feature extraction module adopts a parallel multi-branch convolutional structure, with each branch configured with a convolutional kernel of different sizes, to simultaneously capture local change patterns and long-term trend features of the input data at different time scales.

6. The charging station network loss assessment method based on the MFB-TEB model as described in claim 5, characterized in that, The different sizes of convolutional kernels include 3×3, 5×5 and 7×7 convolutional kernels. The feature maps output by each branch are concatenated by channels and then compressed by a dimensionality reduction convolutional layer. The dimensionality reduction convolutional layer is used to control the model complexity and improve the generalization ability.

7. The charging station network loss assessment method based on the MFB-TEB model as described in claim 6, characterized in that, The steps of fusing joint feature representation include: The feature vector output by the Transformer encoder is concatenated with the feature map output by the multi-receptive field feature extraction module by channel dimension, and the feature space is aligned by upsampling or linear mapping operation to form a unified high-dimensional feature representation. The lightweight regression structure includes a fully connected layer and an activation function connected in sequence, which are used to map the joint feature representation into a scalar form of network loss rate, wherein the loss rate characterizes the proportion of power loss of the charging station per unit time. When the charging station has complete topology information, the line length, power supply radius, and conductor type are added as optional input features to the standardized input dataset.

8. A charging station network loss assessment system based on the MFB-TEB model, using the method described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire multi-source operating data of the charging station. The multi-source operating data includes at least equipment parameters, real-time electrical quantity data, environmental monitoring data, and user charging behavior data. The preprocessing module is used to preprocess the multi-source running data to generate a standardized input dataset, wherein the standardized input dataset is a dataset that is time-consistent, logically complete, and free of anomalies. The first feature extraction module is used to input the standardized input dataset into the Transformer encoder module and extract global temporal dependency features from the data through a self-attention mechanism. The second feature extraction module is used to simultaneously input the standardized input dataset into the multi-receptive field feature extraction module, and extract local and global nonlinear features in parallel using receptive fields of different scales. The fusion module is used to fuse the temporal dependency features output by the Transformer encoder module with the multi-scale features output by the multi-receptive field feature extraction module to construct a joint feature representation; The evaluation module is used to output the loss evaluation result of the charging station network in the current period based on the joint feature representation and through a lightweight regression structure.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the charging station network loss assessment method based on the MFB-TEB model according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the charging station network loss assessment method based on the MFB-TEB model according to any one of claims 1 to 7.

Images