A mechanism-data collaborative driving-based direct-current charging pile error online estimation method

Abstract

The application discloses a mechanism-data collaborative driving-based direct-current charging pile error online estimation method, aiming to solve the problems of lack of error labels in existing transformer area data, incomplete transformer area working condition coverage, difficulty in remote calibration of charging piles and insufficient precision of traditional error estimation models. The method comprises the following steps: constructing a direct-current charging pile twin transformer area simulation data generation model for offline training and hyperparameter calibration of the error estimation model; proposing a transformer area line loss estimation model based on transient voltage compensation; proposing a charging pile time-varying efficiency dynamic tracking model based on linear Kalman filtering; and constructing an LSTM charging pile error estimation model based on embedded energy conservation physical constraints. The application realizes remote high-precision estimation of direct-current charging pile errors, and provides reliable data support for timely replacement and operation and maintenance of direct-current charging piles.

Description

Technical Field

[0001] This invention relates to the field of power metering and simulation technology, specifically to an online error estimation method for DC charging piles based on mechanism-data collaborative driving. Background Technology

[0002] With the increasing popularity of electric vehicles, DC charging piles, as core charging infrastructure, directly impact the fairness of trade settlements due to the accuracy of their electricity metering. During use, charging piles may experience increased errors due to aging, malfunction, or misalignment. According to national metrological regulations, charging piles must undergo regular verification and calibration. However, traditional manual on-site verification methods are inefficient, costly, and lack real-time monitoring capabilities, making them unsuitable for managing large-scale charging pile networks.

[0003] In recent years, remote error estimation methods have become a research hotspot. In existing technologies, some schemes employ data-driven models for error estimation. However, the training data for transformer substations lacks true error values ​​as labels, making it difficult to cover the complex operating conditions of the substations. Purely data-driven methods lack physical constraints, leading to unreliable estimation results. Furthermore, most existing line loss estimation models are based on the assumption of constant voltage, failing to fully consider voltage fluctuations (such as ±7% voltage deviation) that occur in actual operation of the substations, resulting in distorted line loss calculations and severely impacting the accuracy of error estimation.

[0004] Therefore, in response to the problems of difficulty in obtaining true error from existing transformer area data, incomplete coverage of transformer area operating conditions, difficulty in remote calibration of charging piles, and insufficient accuracy of traditional error estimation models, an online error estimation method for DC charging piles based on mechanism-data collaborative driving is proposed, which provides reliable data support for the timely replacement and operation and maintenance of DC charging piles. Summary of the Invention

[0005] The purpose of this invention is to propose an online estimation method for DC charging pile area error based on mechanism-data collaborative driving, in order to solve the problems mentioned in the background technology, such as difficulty in obtaining true error data for the area, incomplete coverage of the area's operating conditions, difficulty in remote calibration of charging piles, and insufficient accuracy of traditional error estimation models, and to provide reliable data support for the timely replacement and operation and maintenance of DC charging piles.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an online error estimation method for DC charging piles based on mechanism-data collaborative driving, comprising: establishing a DC charging pile area twin simulation data generation model for offline training and hyperparameter calibration of the error estimation model; proposing an area line loss estimation model based on transient voltage compensation; proposing a charging pile time-varying efficiency dynamic tracking model based on linear Kalman filtering; and constructing an LSTM charging pile error estimation model based on embedded energy conservation physical constraints.

[0007] Step S1: Establish a DC charging pile twin substation simulation data generation model for offline training and hyperparameter calibration of the error estimation model. Given that the actual measurement error (i.e., error truth label) of each charging pile cannot be directly obtained through physical measurement during actual substation operation, the supervised deep learning model cannot be directly trained on real data. This step constructs a twin substation simulation data generation model that is completely consistent with the topology of the real substation. This model is only used in the offline stage to generate a multi-sample dataset with accurate error labels to complete the pre-training of the error estimation model and the calibration of physical constraint coefficients. In the online application stage, this method only relies on real substation data for forward inference and does not need to run the simulation model again.

[0008] The twin substation simulation data generation model is built using Simulink. The actual DC charging pile substation includes one main power meter and 16 DC charging piles, each equipped with a power meter. The simulation data generation model is built based on the actual DC charging pile substation topology and is consistent with the actual substation topology. The substation simulation data generation model is as follows: Figure 1 As stated above.

[0009] Each DC charging pile simulation module is equipped with two energy meter simulation modules: one energy meter simulation module without error sources, used to measure the actual output energy of the charging pile; and the other energy meter simulation module with error sources, used to simulate actual metering conditions. The two energy meter simulation modules operate synchronously under the same simulation conditions, outputting time-domain aligned actual energy data and metered energy data respectively. The difference between the two constitutes the error truth label. This design ensures that subsequent error estimation algorithms can obtain accurate training labels, and the two sets of data are completely aligned in the time domain, avoiding label deviation problems caused by time asynchrony or load changes in on-site data acquisition. The complete structure of a single DC charging pile is as follows: Figure 2 As stated above.

[0010] The electricity meter simulation module collects voltage and current signals from the circuit, calculates the voltage, current, power factor, active power, and active energy in the circuit, and exports the data to the workbench through the To Workspace module. The metering error comes from voltage metering error, current metering error, and phase metering error. The above three types of error sources are added to the electricity meter simulation module to simulate the metering situation of a real electricity meter.

[0011] The DC charging pile simulation module consists of a charging output submodule and a charging control submodule. The charging output submodule includes an AC-DC converter circuit and a BOOST boost circuit. The AC-DC converter circuit uses a bridge rectifier circuit to convert AC power into DC power, and then passes through a filter circuit to make the output DC voltage stable. The BOOST boost circuit changes the closing of the switching transistor and uses capacitors and inductors to power the circuit, so that the input low voltage rises to a high voltage.

[0012] The charging control submodule feeds back the current battery voltage to the charging circuit controller, enabling the charging control circuit to automatically adjust the charging strategy based on the current battery voltage. The charging control circuit controller consists of a PID controller and a PWM controller. The PID controller generates a control signal by comparing the current battery voltage with the battery voltage threshold, and then inputs the control signal to the PWM controller. The PWM controller generates a PWM wave to control the opening and closing of the circuit, thereby controlling the circuit voltage and current. After passing through a filtering circuit, a stable charging current is output.

[0013] Connect the output of the AC-DC circuit to the input of the BOOST boost circuit, connect the output of the BOOST boost circuit to the input of the charging control circuit, and connect the output of the charging control circuit to the input of the automotive lithium battery to complete the DC charging pile simulation model.

[0014] Based on the circuit characteristics of real DC charging pile areas, line losses are added to the twin area simulation data generation model. According to the voltage fluctuations existing in real areas, the voltage source of the area simulation data generation model is set to a 380V controlled AC voltage source based on normal distribution to simulate the voltage fluctuations of real areas.

[0015] The twin transformer area simulation data generation model covers voltage fluctuations (±7%) allowed by national standards, different battery states of charge, and three-stage charging switching conditions, ensuring that the simulation dataset covers the feasible domain boundary of the real transformer area. During training, a "pre-training-fine-tuning" strategy is adopted. First, the main parameters of the error estimation model are initialized using the simulated labeled dataset. After connecting to the real transformer area, the unlabeled running data collected from the real transformer area is used to fine-tune only the fully connected output layer of the model, thereby generalizing the physical time series features learned from the simulation to the real scene.

[0016] Step S2: Establish a transformer substation line loss estimation model based on transient voltage compensation. According to national standards, the voltage will fluctuate within ±7% of the standard voltage. The assumption of a constant voltage will lead to distortion in the line loss calculation. Based on Joule's law, a line loss estimation model is derived to improve the model's adaptability to voltage fluctuations and enhance its accuracy. The line loss estimation model is as follows:

[0017]

[0018] in Let the line loss be at time k. The line loss at time k-1, Let k be the output power of the DC charging pile at time k. The output power of the DC charging pile at time k-1 is... Let be the power factor of the DC charging pile at time k. Let be the power factor of the DC charging pile at time k-1. Let k be the output voltage of the DC charging pile at time k. The output voltage of the DC charging pile at time k-1 is given.

[0019] The model explicitly introduces the squared voltage ratio term in the line loss recursion, enabling the line loss calculation to respond in real time to grid voltage fluctuations. This breaks the error chain that propagates voltage disturbances to error estimation. Based on the known line loss value at the previous moment, the line loss value at each subsequent moment can be calculated using this recursive model.

[0020] Step 3: Establish a dynamic tracking model for the time-varying efficiency of charging piles based on linear Kalman filtering. According to the law of conservation of energy, the relationship between the total energy meter of the DC charging pile area and the energy of the DC charging pile is as follows:

[0021]

[0022] in Let K be the active power collected by the total power meter at time k. Let be the relative error of the i-th charging pile at time k. Let be the active power consumed by the i-th charging pile at time k. Let be the conversion efficiency of the i-th charging pile at time k.

[0023] Directly estimate conversion efficiency This leads to nonlinearity in the observation equations. To simplify the linearization process, the reciprocal of the transformation efficiency is defined as the state variable:

[0024]

[0025] Conversion efficiency It usually changes slowly, its reciprocal It also exhibits slow time-varying characteristics, and its dynamics are described using a random walk model:

[0026]

[0027] in For process noise, it represents the uncertainty of the state between two adjacent time periods.

[0028] The states of all charging stations can be represented in vector form:

[0029]

[0030] State equation of the time-varying efficiency dynamic tracking model for charging piles:

[0031]

[0032] in Let k be the state vector at time k. Here is the state transition matrix. For the state estimation at time k-1, This is the process noise matrix.

[0033] The observation equation for the dynamic tracking model of time-varying efficiency of charging piles is:

[0034]

[0035] in Let k be the observation vector at time k. For the observation matrix, This is the observation noise matrix.

[0036] Based on the state equation, a priori state equation is established, and the calculation formula is as follows:

[0037]

[0038] in This is the prior state estimate at time k.

[0039] The prior estimate covariance matrix is ​​calculated based on the state equation. The formula for calculating covariance is:

[0040]

[0041] in Let be the covariance matrix at time k-1. Process noise matrix The covariance matrix.

[0042] Calculate gain The calculation formula is:

[0043] in Noise matrix The covariance matrix.

[0044] Calculate the state estimate at time k The calculation formula is:

[0045]

[0046] in This represents the estimated conversion efficiency of each charging station at time k.

[0047] The updated covariance is calculated using the following formula:

[0048]

[0049] Where I is the identity matrix.

[0050] Step S4: Construct an LSTM charging pile error estimation model based on embedded energy conservation physical constraints. The error estimation model uses a two-layer LSTM neural network for error estimation. The LSTM neural network extracts the temporal features of the station area data and solves the data saturation problem through a forget gate to avoid insufficient recognition of new data.

[0051] An LSTM neural network consists of four parts: a forget gate, an input gate, a cell state update gate, and an output gate. The forget gate determines how much information is forgotten, and its calculation formula is as follows:

[0052]

[0053] in For the output of the forget gate, For activation function, For the weight of the Forgotten Gate, For historical information, The input data is at time k. This is an offset for the forget gate.

[0054] The output gate determines how much information to retain. It calculates which information needs to be retained based on the input gate control signal. The calculation formula is as follows:

[0055]

[0056] in The output is the input gate control signal. The weights of the input gate control signals. This is the bias of the input gate control signal.

[0057] Candidate cell states are generated to create new information to be stored. The calculation formula is as follows:

[0058]

[0059] in The output is the candidate cell state. For activation function, The weights for the candidate cell states. This is a bias for the candidate cell state.

[0060] Cell state updates are used to integrate historical and current information, and the calculation formula is as follows:

[0061]

[0062] in This represents the current state of the cell. This represents the cell state at the previous moment.

[0063] The output gate is used to generate the final output at the current moment. The first step is to calculate the output gate signal. Output gate signal The formula used to determine which information about a cell's state can be output is:

[0064]

[0065] in The weights of the output gate signal, This is the bias for the output gating signal.

[0066] Generate the current hidden state The calculation formula is:

[0067]

[0068] in , It is the model's prediction at the current moment.

[0069] Existing LSTM loss functions are based solely on the mean squared error between the predicted and labeled values:

[0070]

[0071] The current loss function is purely data-driven and lacks physical constraints. Therefore, we propose using the PINN method to embed the energy conservation equation as a physical constraint into the LSTM loss function, resulting in a new loss function. This avoids the unreliability of error estimation under purely data-driven conditions and improves the interpretability of the LSTM neural network. The loss function incorporating the energy conservation equation is as follows:

[0072]

[0073] LOSS is the loss function of the LSTM neural network. , These are parameters to be determined; set them according to your experimental conditions. Let k be the error estimate of the DC charging pile obtained by LSTM at time k.

[0074] The first term of the loss function is the mean square error between the error estimate and the true value, ensuring the model's ability to fit the training data; the second term is the residual of the energy conservation equation, reflecting the squared difference between the total electricity of the transformer area and the sum of the electricity of each charging pile (after error and efficiency correction). When the error estimate of the network output causes the energy conservation relationship to not hold, the loss function will increase, thereby guiding the network output to simultaneously satisfy both statistical fitting and physical conservation conditions during the training process.

[0075] The operating data of the twin simulation data generation model of the transformer substation under various operating conditions were made into an offline training benchmark dataset, which was divided into training set and validation set in a 7:3 ratio. In this offline stage, the improved dynamic model was first used to solve the line loss, and then the time-varying efficiency dynamic tracking model was used to solve the conversion efficiency. Finally, the line loss and conversion efficiency were used as known quantities to input the LSTM+PINN model for supervised training. During online deployment, the power data collected from the real transformer substation was input into the trained model in real time, and the model could directly output the error estimate of each charging pile without the need for error ground truth labels.

[0076] This invention provides an online error estimation method for DC charging pile distribution areas based on mechanism-data collaborative driving, which has the following beneficial effects: First, the twin simulation data generation model breaks through the supervised training bottleneck of "data available but no labels" in real distribution areas. Combined with the online fine-tuning mechanism, it realizes the effective transfer of simulation knowledge to the real physical world, solving the engineering problem that pure field data-driven models cannot be deployed due to the lack of labels; Second, by adjusting the parameters of the twin simulation data generation model, the operating conditions of distribution areas under different working conditions can be simulated, generating various types of power data; Third, the distribution area line loss estimation model explicitly introduces the voltage ratio square term, enabling line loss calculation to respond in real time to grid voltage fluctuations (±7%), cutting off the error chain of voltage disturbances propagating to error estimation; Fourth, time-varying efficiency dynamic tracking. The model sets the state variables as the reciprocal of the conversion efficiency, linearizing the observation equation and enabling precise online tracking of time-varying efficiency. Fifth, the energy conservation equation is embedded as a physical constraint term in the loss function, ensuring that the error estimation model output simultaneously satisfies both statistical fitting and physical conservation conditions. This fundamentally differs from purely data-driven solutions and addresses the industry-wide problems of poor generalization and weak interpretability of traditional models. Sixth, the above four modules constitute a complete technical chain of "simulation data generation → analytical model guidance → data model correction." Data transmission and coupling exist between each step, working synergistically to solve the problems of incomplete data coverage and insufficient error estimation accuracy in the charging station area. This achieves high-precision estimation of DC charging pile errors and provides reliable data support for the timely replacement and operation and maintenance of DC charging piles. Attached Figure Description

[0078] Figure 1This is an overall structural diagram of the DC charging pile substation simulation model of the present invention;

[0079] Figure 2 This is a schematic diagram of the complete structure of a single DC charging pile according to the present invention;

[0080] Figure 3 This is a complete system block diagram of the present invention. Detailed Implementation

[0082] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0083] This invention provides a technical solution: an online estimation method for DC charging pile substation errors based on mechanism-data collaborative driving, comprising: establishing a DC charging pile substation twin simulation data generation model for offline training and hyperparameter calibration of error estimation models; proposing a substation line loss estimation model based on transient voltage compensation; proposing a charging pile time-varying efficiency dynamic tracking model based on linear Kalman filtering; and constructing an LSTM charging pile error estimation model based on embedded energy conservation physical constraints.

[0084] The specific implementation of the electricity meter simulation module consists of three parts: a data acquisition unit, a data processing unit, and a data transmission unit. It is used to measure electrical energy data in the metering circuit. The data acquisition unit collects voltage and current signals from the circuit. The data processing unit extracts the magnitude and phase angle of the voltage and current signals to obtain the voltage value, current value, and voltage-current phase difference. Then, it calculates the active power and active energy. The active power calculation formula is as follows:

[0085]

[0086] Where P is the active power, U is the voltage value, and I is the current value. This represents the phase difference between voltage and current.

[0087] The formula for calculating active electrical energy is:

[0088]

[0089] in It is active electrical energy.

[0090] The data transmission unit is used to upload electrical energy data such as voltage, current, power factor, active energy, and active power to the workbench.

[0091] The electricity meter simulation module contains error sources. The error sources are added to the electricity meter simulation module to introduce measurement errors. In reality, the measurement errors come from voltage measurement errors, current measurement errors, and phase measurement errors, simulating the measurement errors that exist in DC charging piles under real conditions.

[0092] The DC charging pile simulation module is implemented by consisting of a charging output circuit and a charging control circuit. The charging output circuit includes an AC-DC converter and a BOOST boost circuit. The AC-DC circuit converts the AC input of the DC charging pile into DC power for charging the car's lithium battery. A bridge rectifier circuit is used to convert the AC to DC, and then a filter circuit is used to ensure a stable DC output. The BOOST boost circuit raises the 380V low-voltage DC output from the AC-DC circuit to 800V high-voltage DC. This is achieved by controlling the switching transistor to close. When the transistor is on, the input voltage powers the inductor, which stores energy, and the capacitor powers the lithium battery. When the transistor is off, the inductor generates an induced electromotive force to maintain a constant current. This induced electromotive force, along with the input voltage, powers the capacitor and the lithium battery, increasing the average charging voltage of the lithium battery and raising the low-voltage DC to high-voltage DC. The voltage increase is achieved by controlling the duty cycle of the switching transistor. The duty cycle is calculated using the following formula:

[0093]

[0094] in The duty cycle for the switching transistor to open and close. Input voltage, For output voltage, For conversion efficiency.

[0095] The charging control circuit automatically adjusts the charging mode based on the current voltage of the car's lithium battery. The current battery voltage is input to the feedback circuit and compared with a set threshold. The comparison result is then input to the PID controller to quickly adjust the control signal and eliminate steady-state errors. The control signal is then input to the PWM controller in the charging circuit to adjust the PWM wave, achieving precise control of the circuit voltage and current. This meets the different needs of the three-stage charging of the car's lithium battery. When the current battery voltage is less than threshold 1, a pre-charging mode is used, employing a small current to prevent damage to the battery due to excessive current charging when the voltage is too low. When the current battery voltage is greater than threshold 1 but less than threshold 2, a constant current charging mode is used, employing a large current for rapid charging. When the current battery voltage is greater than threshold 2, a constant voltage charging mode is used to prevent voltage drop and incomplete charging.

[0096] Connect the output of the AC-DC circuit to the input of the BOOST boost circuit, connect the output of the BOOST boost circuit to the input of the charging control circuit, and connect the output of the charging control circuit to the input of the automotive lithium battery to complete the DC charging pile simulation module.

[0097] The specific construction of the simulation data generation model for DC charging pile area: The real DC charging pile area includes one main power meter and 16 DC charging piles. Each DC charging pile is equipped with one power meter to measure the power data of the DC charging pile. The simulation data generation model is built according to the structure of the real DC charging pile area to simulate the actual power consumption of the charging pile area.

[0098] In the twin-area simulation data generation model, each DC charging pile simulation module is equipped with two energy meter simulation modules. One energy meter simulation module does not contain an error source and is used to measure the actual output energy of the charging pile, while the other energy meter simulation module contains an error source and is used to simulate the actual metering situation.

[0099] In reality, the input voltage of the charging station area fluctuates. According to national standards, the voltage will fluctuate within ±7% of the standard voltage. The AC voltage source of the DC charging station area simulation is set as a controlled AC voltage source. The amplitude of the AC voltage source is adjusted according to the input signal. The signal source is set as a normally distributed source to simulate the voltage fluctuation in real-world conditions.

[0100] The specific derivation process of the transformer substation line loss estimation model based on transient voltage compensation involves sorting the power data according to the simulation time, preprocessing the power data, removing simulation data of transient power and incomplete power data, and calculating the active power per unit time using the following formula:

[0101]

[0102] in Let be the energy consumption per unit time from time k to k+1. Let be the active electrical energy at time k+1. Let be the active electrical energy at time k.

[0103] Line loss is the electrical energy lost due to the resistance of a circuit conductor. According to Joule's law, the formula for calculating line loss is:

[0104]

[0105] in For line loss, The current flowing through the wire, For the resistance of the wire, For time.

[0106] Current flowing through the wire The calculation formula is:

[0107]

[0108] in For charging pile power, For voltage, The power factor.

[0109] Current Substituting the original calculation formula into the line loss calculation formula yields a new line loss calculation formula, which is:

[0110]

[0111] The line loss ratio at different times is:

[0112]

[0113] in Let the line loss be at time k. The line loss at time k-1, Let k be the output power of the charging gun at time k. The output power of the charging gun at time k-1. Let be the power factor of the charging gun at time k. Let be the power factor of the charging gun at time k-1. Let k be the output voltage of the charging gun at time k. Let be the output voltage of the charging gun at time k-1.

[0114] The final formula for calculating line loss is:

[0115]

[0116] The specific iterative steps of the dynamic tracking model for time-varying efficiency of charging piles based on linear Kalman filtering are as follows: According to the law of conservation of energy, the relationship between the total energy meter of the DC charging pile area and the energy of the DC charging pile is as follows:

[0117]

[0118] in Let be the active energy at time k of the total energy meter. The line loss at time k on the main path.

[0119] Directly estimate conversion efficiency This leads to nonlinearity in the observation equations. To simplify the linearization process, the reciprocal of the transformation efficiency is defined as the state variable:

[0120]

[0121] Conversion efficiency It usually changes slowly, its reciprocal It also exhibits slow time-varying characteristics, and its dynamics are described using a random walk model.

[0122] The states of all charging piles are written in vector form, and the time-varying efficiency dynamic tracking model recursive formula in step S3 is used for iterative calculation to obtain the estimated conversion efficiency of each charging pile.

[0123] The specific structure and training process of the LSTM charging pile error estimation model based on embedded energy conservation physical constraints are as follows: Voltage, current, active power, and active energy collected by the electricity meter and DC charging pile at time k are used as features to establish a feature matrix. :

[0124]

[0125] in , , , and Let k be the voltage, current, active power, active energy, and power factor collected by the total energy meter at time k. , , , , … , , , , The voltage, current, active power, active energy, and power factor of 16 DC charging piles were measured.

[0126] Using the true value of the metering error of the DC charging pile at time k as a label, a label matrix is ​​established. :

[0127]

[0128] in … This represents the true value of the error at time k for 16 DC charging piles.

[0129] A two-layer LSTM neural network is established, and calculations are performed according to the forward propagation formula in step S4. The PINN method is used to incorporate the energy conservation physical model of the transformer area into the loss function of the LSTM.

[0130]

[0131] The twin-strip simulation data generation model generates various operating condition data to form a dataset. The dataset is divided into training and testing sets in a 7:3 ratio. The data is input into the error estimation algorithm and trained using an Adam optimizer with 600 training cycles and a learning rate of 0.05. The error estimation model structure and hyperparameter calibration are optimized. The trained error estimation model is then deployed to a real DC charging station area to monitor charging station errors in real time, providing reliable data support for the timely replacement and operation and maintenance of DC charging stations.

Claims

1. A mechanism-data collaborative driving method for online error estimation of DC charging piles, characterized in that, include: Step S1: Establish a DC charging pile twin substation simulation data generation model for offline training and hyperparameter calibration of the error estimation model. The twin substation simulation data generation model uses Simulink to build a simulation substation based on the real DC charging pile substation topology. It includes one main power meter simulation module and 16 charging branches. Each charging branch includes one DC charging pile simulation module and two power meter simulation modules. One power meter simulation module does not contain an error source, while the other power meter simulation module does contain an error source. The two power meter simulation modules run synchronously under the same simulation conditions and output time-domain aligned real power data and metered power data, respectively. The difference between the two constitutes the error truth label. The substation simulation model simulates the operation of the substation under different working conditions and generates diverse simulation data. Step S2: Establish a transformer substation line loss estimation model based on transient voltage compensation. Explicitly introduce the squared voltage ratio term into the line loss estimation recursive formula and use the voltage values ​​at the current time and the previous time to perform transient compensation for line loss. Step S3: Establish a dynamic tracking model for the time-varying efficiency of charging piles based on linear Kalman filtering. Set the reciprocal of the charging pile conversion efficiency as the state variable to linearize the observation equation. Use linear Kalman filtering to recursively estimate the conversion efficiency. Step S4: Construct an LSTM charging pile error estimation model with embedded energy conservation physical constraints. Use a two-layer LSTM to extract the time-series features of the power data of the substation area. Embed the energy conservation equation as a physical constraint term into the loss function of the LSTM. Use the line loss estimate output in step S2 and the conversion efficiency estimate output in step S3 as known quantities to input the loss function. After training, output the error estimate of each charging pile.

2. The online error estimation method for DC charging piles based on mechanism-data collaborative driving as described in claim 1, characterized in that, In step S1, the energy meter simulation module collects voltage and current signals from the circuit, extracts the modulus and phase angle, and calculates energy data such as power factor, active power and active energy. The metering error of a real electricity meter comes from voltage metering error, current metering error and phase metering error, so error sources are added to the electricity meter simulation module; The DC charging pile simulation module includes a charging output submodule and a charging control submodule. The charging output submodule consists of an AC-DC converter circuit and a BOOST boost circuit. The AC-DC converter circuit converts the AC power input to the DC charging pile into DC power for vehicle charging, and the BOOST boost circuit boosts the 380V low voltage to the 800V high voltage for vehicle charging. The charging control submodule controls the output of the charging pile, simulating the output process of a three-stage charging pile for a car lithium battery under real conditions. Based on the current voltage of the car lithium battery, the charging strategy is automatically adjusted. When the current voltage of the car lithium battery is less than threshold 1, a pre-charging mode is adopted, using a small current to prevent the car lithium battery from being damaged by excessive current charging when the voltage is too low. When the current voltage of the car lithium battery is greater than threshold 1 but less than threshold 2, a constant current charging mode is adopted, using a large current to charge the car lithium battery quickly. When the current voltage of the car lithium battery is greater than threshold 2, a constant voltage charging mode is adopted to avoid voltage drop leading to incomplete charging.

3. The online error estimation method for DC charging piles based on mechanism-data collaborative driving as described in claim 1, characterized in that, In step S1, the DC charging pile area simulation model is consistent with the real area topology. The voltage source is set as a 380V controlled AC voltage source based on normal distribution to simulate the voltage fluctuations in the real area, and line loss is set in the main road of the area.

4. The online error estimation method for DC charging piles based on mechanism-data collaborative driving as described in claim 1, characterized in that, In step S2, the calculation formula for the transformer substation line loss estimation model based on transient voltage compensation is as follows:

5. Among them Let the line loss be at time k. The line loss at time k-1, Let k be the output power of the charging gun at time k. The output power of the charging gun at time k-1. Let be the power factor of the charging gun at time k. Let be the power factor of the charging gun at time k-1. Let k be the output voltage of the charging gun at time k. Let be the output voltage of the charging gun at time k-1.

6. The online error estimation method for DC charging piles based on mechanism-data collaborative driving as described in claim 1, characterized in that, In step S3, the state variable is set as the reciprocal of the conversion efficiency, and the state equation of the charging pile time-varying efficiency dynamic tracking model based on linear Kalman filtering is:

7. Among them Let k be the state vector at time k. Here is the state transition matrix. For the state estimation at time k-1, This is the process noise matrix; The observation equations for the time-varying efficiency dynamic tracking model are as follows:

8. Among them Let k be the observation vector at time k. For the observation matrix, This is the observation noise matrix.

9. The online error estimation method for DC charging piles based on mechanism-data collaborative driving as described in claim 1, characterized in that, In step S4, the LSTM charging pile error estimation model with embedded energy conservation physical constraints uses a two-layer LSTM to extract the time-series features of the power data of the transformer area. These time-series features are then input into a fully connected layer for error estimation. PINN is used to introduce the energy conservation physical constraints into the loss function. Let the active power of the 16 DC charging piles at time k be... , ..., The error values ​​of the 16 DC charging piles at time k are respectively , ..., According to the law of conservation of energy, the relationship between the total power meter of the DC charging station area and the power of the DC charging station is as follows:

10. Among them Let be the active energy at time k of the total energy meter. Let k be the line loss of the transformer area at time k; The loss function is established based on the energy conservation equation:

11. Among them For the loss of the LSTM neural network, Let k be the true value of the error of the DC charging pile. , For proportional parameters, The error estimate of the DC charging pile at time k is the value of LSTM. The first term of the loss function is the mean square error between the error estimate and the true value, and the second term is the residual of the energy conservation equation. During the training process, the network output must simultaneously satisfy both statistical fitting and physical conservation conditions. The simulation data of the twin charging stations was used to generate a dataset of model operation data. The dataset was divided into training and testing sets in a 7:3 ratio. First, an improved dynamic model was used to solve for line loss. Then, a time-varying efficiency dynamic tracking model was used to solve for conversion efficiency. Finally, line loss and conversion efficiency were used as known quantities and input into the LSTM+PINN error estimation model. The model was trained using an Adam optimizer with 600 training cycles and a learning rate of lr0.05 to optimize the error estimation model structure and hyperparameter calibration. The trained error estimation model was then deployed to a real DC charging station area to monitor charging station errors in real time, providing reliable data support for the timely replacement and operation and maintenance of DC charging stations.

Images