A multi-gun charging pile charging power distribution system

Abstract

The application discloses a multi-gun charging pile charging power distribution system, and relates to the field of electric vehicle charging. The system comprises the following modules: a data cleaning and state estimation module that collects multi-source sensing data in real time, fuses a neural network and a physical model for deep filtering to output real-time state; a charging demand prediction module that predicts power demand based on historical and environmental data using a space-time fusion model; a multi-objective optimization decision module that generates power distribution instructions using a model predictive control with constraints; a power precise execution and compensation module that receives power distribution instructions to drive hardware execution and generates compensation control instructions through online learning algorithm feedforward compensation; a safety monitoring and fusing module that collects system operation parameters, calculates abnormal scores using an unsupervised isolation forest algorithm, and triggers hierarchical fusing according to a threshold. Through high-precision state prediction, multi-objective power distribution under constraints, and execution compensation and abnormal protection, the system realizes efficient, reliable and safe multi-gun charging scheduling.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric vehicle charging, in particular to a multi-gun charging pile charging power distribution system. BACKGROUND

[0002] With the rapid growth of electric vehicles, the demand for charging infrastructure construction is expanding. Currently, public charging stations are usually equipped with multi-gun DC charging piles to provide services for multiple electric vehicles at the same time. However, when multiple guns are charging at the same time, limited by the power supply capacity of the power grid and the rated power of the equipment, how to meet the charging demand of vehicles while maintaining the safe and stable operation of the system becomes an important problem for operation and dispatching.

[0003] In the prior art, charging pile power distribution mostly adopts fixed power or simple proportional distribution strategy, which is difficult to cope with different vehicle battery characteristics, real-time environmental changes and power grid load fluctuations, and is prone to cause low power utilization rate of some charging guns or increased risk of system overload. In addition, affected by factors such as sensor data noise and equipment aging, the traditional method lacks accuracy in state perception and abnormal monitoring, which may cause untimely dispatching or unstable control, thereby affecting charging efficiency and operation safety.

[0004] Therefore, the industry urgently needs a charging pile power management technology that can reasonably allocate charging power under the condition of multi-gun cooperation, real-time monitoring and ensure operation safety. SUMMARY

[0005] Based on the above-mentioned shortcomings of the prior art, the purpose of the present application is to provide a multi-gun charging pile charging power distribution system to solve the above technical problems.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a multi-gun charging pile charging power distribution system, comprising: a data cleaning and state estimation module, a charging demand prediction module, a multi-objective optimization decision module, a power precise execution and compensation module, and a safety monitoring and fusing module;

[0007] The data cleaning and state estimation module: real-time acquisition of multi-source sensor data, based on multi-source sensor data, using a deep filtering algorithm optimized by fusion neural network and physical model, outputting the real-time state of the system;

[0008] The charging demand prediction module: based on historical state data sequence and external environment data, using a spatio-temporal fusion prediction model, generating a charging power demand prediction value;

[0009] The multi-objective optimization decision module: taking the charging power demand prediction value and the real-time state of the system as input, using a model predictive control framework with constraints for multi-objective constraint optimization, generating a power distribution control instruction;

[0010] A power precise execution and compensation module: receiving a power distribution control instruction to drive hardware execution, feeding forward compensation on the power distribution control instruction through an online learning algorithm, and generating a compensated control instruction;

[0011] A safety monitoring and fusing module: collecting system real-time running parameters, calculating an edge real-time anomaly score using an unsupervised isolation forest algorithm, and triggering a hierarchical safety fusing mechanism according to a preset anomaly threshold.

[0012] The application further provides that the data cleaning and state estimation module comprises a data acquisition and preprocessing unit and a state estimation unit.

[0013] The data acquisition and preprocessing unit: receives a set power instruction of each charging gun, power grid power supply constraints and environmental working condition parameters, and performs time alignment and processing on the above parameters to construct a control input vector.

[0014] Real-time acquisition of original monitoring data from multi-source sensor monitoring, including current, voltage and temperature of each charging gun.

[0015] The original monitoring data is cleaned and time-stamped to obtain preprocessed monitoring data.

[0016] The application further provides that the state estimation unit:

[0017] The unmodeled nonlinear dynamics in the state estimation algorithm are compensated using a neural network to construct a deep filtering algorithm.

[0018] When the system is initially running, the state posterior estimation value at the last time is set as a default initial value as a filtering starting condition through a preset initialization strategy.

[0019] The preprocessed monitoring data is input into the deep filtering algorithm to perform state prediction, and based on the state posterior estimation value at the last time and the control input vector, a state prior estimation value at the current time is generated.

[0020] The monitoring data at the current time and the state prior estimation value are compared, and the state prior estimation value is corrected according to the filtering gain to obtain a state posterior estimation value at the current time.

[0021] The output state posterior estimation value is output as the system real-time state, including the current, voltage and temperature estimation values of each charging gun.

[0022] The application further provides that the charging demand prediction module comprises a data preparation unit and a prediction execution unit.

[0023] The data preparation unit:

[0024] collecting a historical state data sequence, the historical state data sequence comprising current, voltage and temperature parameters of each charging gun of the charging pile within a preset historical period;

[0025] collecting a historical environment data sequence, the historical environment data sequence comprising environmental temperature and humidity of an environment in which the charging station is located within the preset historical period;

[0026] forming a static feature sequence by taking the power allocation instruction value of the adjacent charging gun at the current time fed back by the multi-objective optimization decision module as a spatial correlation feature and extending the spatial correlation feature within the preset historical period;

[0027] obtaining a time-of-use electricity price signal within a future optimization period and reservation charging traffic information to construct a future known sequence.

[0028] The application further provides that the prediction execution unit comprises:

[0029] by introducing the static feature sequence representing the power setting value of the adjacent charging gun in the input encoding stage of the time sequence fusion model, and jointly fusing the static feature weight and the attention weight of the time dimension in the multi-head attention calculation, a time sequence fusion model enhanced by spatial correlation is constructed;

[0030] the historical state data sequence, the historical environment data sequence, the static feature sequence and the future known sequence are input into the time sequence fusion model enhanced by spatial correlation as input items;

[0031] a charging power demand prediction sequence of each charging gun of the charging pile within the future optimization period is output.

[0032] The application further provides that the multi-objective optimization decision module comprises an optimization calculation unit and an instruction execution unit.

[0033] The optimization calculation unit receives the system real-time state output by the data cleaning and state estimation module and the charging power demand prediction sequence output by the charging demand prediction module;

[0034] obtaining the upper and lower limits of the allowable power of each charging gun, the total power capacity of the charging pile and the maximum power change rate of each charging gun;

[0035] obtaining the target state of charge information provided by the battery management system of each vehicle and the environmental temperature safety threshold, and constructing a weighted multi-objective function based on a model predictive control framework;

[0036] by imposing system constraints including the upper and lower limits of the charging gun power, the total power upper limit of the system and the charging gun power change rate limit on the multi-objective function;

[0037] The optimal power distribution instruction sequence is obtained by using a hybrid strategy of a multi-objective genetic algorithm and a particle swarm optimization to solve the multi-objective function.

[0038] The application further provides that the instruction execution unit is used for:

[0039] extracting a first time step instruction to be executed at the current time from the optimal power distribution instruction sequence;

[0040] outputting the first time step instruction as a power distribution control instruction to the power accurate execution and compensation module;

[0041] feeding the power distribution instruction value at the current time as a space correlation feature to the charging demand prediction module.

[0042] The application further provides that the power accurate execution and compensation module comprises an execution control unit and a compensation learning unit.

[0043] The execution control unit:

[0044] receiving the power distribution control instruction output by the multi-objective optimization decision module;

[0045] driving the power hardware to execute the power distribution control instruction at the current time by using a closed-loop controller;

[0046] synchronously collecting the actual output power and system state parameters, wherein the system state parameters comprise the current output current, voltage and temperature.

[0047] The application further provides that the compensation learning unit:

[0048] calculating an execution deviation between the power distribution control instruction and the actual output power;

[0049] establishing an execution deviation prediction model by using an online learning algorithm based on the execution deviation and the system state parameters;

[0050] performing online adaptive update on the parameters of the execution deviation prediction model by using a recursive least square method;

[0051] performing feedforward compensation on the power distribution control instruction to be executed by using the updated execution deviation prediction model to generate a compensated control instruction;

[0052] delivering the compensated control instruction to a device execution interface to drive the power hardware to execute.

[0053] The application further provides that the safety monitoring and fusing module comprises:

[0054] collecting system real-time running parameters, wherein the running parameters comprise power instruction values, output currents, output voltages and device temperatures;

[0055] An unsupervised isolation forest algorithm is used to perform real-time calculation on the real-time running parameters of the system to generate edge real-time anomaly scores;

[0056] A hierarchical safety fusing mechanism is triggered according to a preset anomaly threshold, and the safety fusing mechanism comprises:

[0057] When the edge real-time anomaly score is between the first anomaly threshold and the second anomaly threshold, it is determined to be a mild anomaly, a power reduction operation is triggered, and alarm information is generated;

[0058] When the edge real-time anomaly score is higher than the second anomaly threshold, it is determined to be a serious anomaly, an emergency disconnect operation is triggered, the power output is stopped, and alarm information is generated.

[0059] The present application provides a multi-gun charging pile charging power distribution system, which comprises a data cleaning and state estimation module, a charging demand prediction module, a multi-objective optimization decision module, a power accurate execution and compensation module, and a safety monitoring and fusing module.

[0060] High-precision state perception and prediction: through the deep filtering algorithm optimized by the fusion neural network and the physical model, the real-time high-precision state estimation of multi-source sensing data is realized, and the future power demand of the multi-gun is accurately predicted by combining the time series fusion model enhanced by spatial correlation, which significantly improves the foresight and reliability of multi-gun charging scheduling.

[0061] Multi-objective adaptive optimization decision: the model predictive control framework with constraints is adopted, and the mixed strategy of multi-objective genetic algorithm and particle swarm optimization is introduced, so as to dynamically generate the optimal power distribution instruction under the constraints of the upper and lower limits of the charging gun power, the total power of the system and the power change rate, and to consider the charging efficiency, the power grid load balance and the safe operation of the equipment.

[0062] Online compensation of execution deviation and security protection: an execution deviation prediction model is established and updated in real time through an online learning algorithm, the power distribution instruction is fed forwardly compensated to improve the execution accuracy of the instruction; meanwhile, the hierarchical security fusing mechanism of the unsupervised isolation forest algorithm is combined to realize the instant identification and hierarchical protection of abnormal working conditions, and the security and stability of the system in complex operating environment are enhanced.

[0063] The above description is only a summary of the technical solutions of the present application. In order to enable the technical means of the present application to be more clearly understood, and to be implemented according to the content of the description, and in order to enable the above and other purposes, characteristics and advantages of the present application to be more apparent and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on these drawings. In the drawings:

[0065] Figure 1 The structure diagram of a multi-gun charging pile charging power distribution system is shown for an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0066] The embodiments of the present application will be described below with reference to the drawings and preferred embodiments, and those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the present description. The present application can also be implemented or applied by different specific embodiments, and the details in the present description can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be understood that the preferred embodiments are only for illustrating the present application, and are not intended to limit the protection scope of the present application.

[0067] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and only the components related to the present application are shown in the diagrams, not the number, shape and size of the components when actually implemented. The actual implementation of each component may be a random change, and the component layout form may be more complex.

[0068] In the following description, numerous specific details are discussed in order to provide a thorough explanation of embodiments of the application. It will be apparent, however, to one skilled in the art, that the embodiments of the application can be practiced without these specific details. In other instances, well-known structures and devices are not described in detail in order to avoid obscuring embodiments of the application. Embodiments

[0069] A multi-gun charging pile charging power distribution system, as shown in Figure 1 comprises:

[0070] Data cleaning and state estimation module: real-time acquisition of multi-source sensing data, based on multi-source sensing data, using fusion neural network and deep filtering algorithm optimized by physical model, outputting real-time state of system;

[0071] Charging demand prediction module: based on historical state data sequence and external environment data, using spatio-temporal fusion prediction model, generating charging power demand prediction value;

[0072] Multi-objective optimization decision module: taking the charging power demand prediction value and the real-time state of the system as input, using the model predictive control framework with constraints for multi-objective constraint optimization, generating power allocation control instruction;

[0073] Power accurate execution and compensation module: receiving power allocation control instruction to drive hardware execution, using online learning algorithm to feed forward compensate power allocation control instruction, generating compensated control instruction;

[0074] Safety monitoring and fusing module: collecting real-time running parameters of the system, using unsupervised isolation forest algorithm to calculate edge real-time anomaly score, triggering hierarchical safety fusing mechanism according to preset anomaly threshold.

[0075] The application further provides that the data cleaning and state estimation module comprises a data acquisition and preprocessing unit and a state estimation unit.

[0076] The data acquisition and preprocessing unit: receives parameters including the set power instruction of each charging gun, the power supply constraint of the power grid and the environmental working condition parameter, and performs time alignment and processing on the above parameters to construct a control input vector;

[0077] Real-time acquisition of original monitoring data from multi-source sensor monitoring, the original monitoring data including current, voltage and temperature of each charging gun;

[0078] The original monitoring data is subjected to data cleaning and timestamp alignment processing to obtain preprocessed monitoring data. Specifically, the set power instruction of each charging gun under the upper power distribution strategy, the real-time power supply constraint provided by the power grid dispatching center and the environmental condition parameters provided by the weather module are also synchronized into the platform, and the environmental condition parameters include environmental temperature, humidity, wind speed, and other environmental parameters can also be expanded according to actual needs. All data sources are synchronized with the central clock server before entering the processing flow. The data processing program is intervalled with a preset time window, and the specific time window can be adjusted according to the actual deployment environment, and the default setting is one second, and the time length can also be appropriately increased in remote areas with small user demand, such as setting to one minute; the asynchronous data from different sources are timestamped. For delayed or missing data, interpolation algorithms such as linear interpolation or nearest neighbor strategy can be used to fill in. After time alignment, the system splices the various parameters into a control input vector in a predetermined order. At the same time, the current, voltage and temperature sensors installed on each charging gun continuously output high-frequency raw monitoring data, which can be transmitted to the local data acquisition controller through the field bus such as CAN bus, and pushed to the central data platform through the message queue such as MQTT protocol. Subsequently, the raw monitoring data is subjected to quality inspection. Abnormal values can be identified and removed by an anomaly detection algorithm such as the Isolation Forest algorithm. For missing data, time series prediction algorithms such as Kalman filter-based methods can be used to complete the data. To suppress noise, a filtering model such as an LSTM-based time series filtering network can be embedded in the data stream to ensure that the output data is smooth and meets the physical characteristics. Finally, the processed detection data and the control input vector are transmitted to the next unit or module as unit output values.

[0079] The application further provides that the state estimation unit comprises:

[0080] The neural network is used to compensate for unmodeled nonlinear dynamics in the state estimation algorithm, and a deep filtering algorithm is constructed.

[0081] When the system is initially operated, the state posterior estimation value at the previous time is set as the default initial value as the filtering starting condition through a preset initialization strategy;

[0082] The preprocessed monitoring data is input into the deep filtering algorithm to perform state prediction operation, and based on the state posterior estimation value at the previous time and the control input vector, the state prior estimation value at the current time is generated;

[0083] The monitoring data at the current time and the state prior estimation value are compared, the state prior estimation value is corrected according to the filtering gain, and the state posterior estimation value at the current time is obtained;

[0084] The output state posteriori estimation value is output as the system real-time state, which includes the current, voltage and temperature estimation values of each charging gun. Specifically, at the initial stage of system startup, the default initial value of each charging gun is configured as the first state posteriori estimation value according to the rated parameters of the charging pile and the on-site environmental conditions, so as to ensure that the filter can be started stably without historical operation data. Subsequently, the control input vector and real-time monitoring data output from the data acquisition and preprocessing unit are received and synchronously sent to the state estimation unit. In the prediction stage, the deep filtering algorithm can be based on the prediction equation of Kalman filtering, and other similar solutions capable of achieving the technology can be adopted, and the current time preliminary state priori estimation value is output according to the state posteriori estimation value of the last time and the current control input vector. At the same time, a neural network module is introduced, which uses the trained model to compensate for the unmodeled system dynamics caused by nonlinear power grid fluctuations, environmental temperature changes and other factors during the charging process, thereby generating a corrected state priori estimation value. In the correction stage, the algorithm compares the current, voltage and temperature data, i.e. the monitoring data, with the predicted state priori estimation value item by item, calculates the measurement residual, and then the algorithm calculates the Kalman gain matrix and applies the gain to the measurement residual to optimally correct the state priori estimation value and obtain the high-precision state posteriori estimation value at the current time. The posteriori state estimation value represents the optimal estimation of the current time current, voltage and temperature of each charging gun of the charging pile, and is output as the system real-time state to provide accurate input basis for the downstream module. The execution period of the entire estimation process can be configured according to system requirements.

[0085] The application further provides that the charging demand prediction module comprises a data preparation unit and a prediction execution unit.

[0086] The data preparation unit comprises:

[0087] The historical state data sequence comprises the current, voltage and temperature parameters of each charging gun of the charging pile within the preset historical period.

[0088] The historical environmental data sequence comprises the environmental temperature and humidity of the environment where the charging station is located within the preset historical period.

[0089] The current time power allocation instruction value of the adjacent charging gun fed back by the multi-objective optimization decision module is taken as a spatial correlation feature, and a static feature sequence is formed by extending it within the preset historical period.

[0090] The time-of-use price signal in the future optimization period and the reservation charging traffic information are acquired to construct a future known sequence. Specifically, in actual operation, the data preparation unit first extracts the historical state data of the past 24 hours from the monitoring system of each charging pile through the industrial Internet of Things communication interface, including the current, voltage and temperature of each charging gun. At the same time, through the environmental monitoring sensors deployed in the charging station, the environmental temperature and humidity in the past 24 hours are synchronously collected, and all the data are stored in a time series database. In order to ensure the continuity and quality of the data, the system uses a streaming data processing framework to perform automatic data cleaning at the access end, performs time interpolation on missing values, and identifies and removes abnormal points through threshold or isolation forest algorithm. Subsequently, the data preparation unit receives the current time adjacent charging gun power distribution instruction value output by the multi-objective optimization decision module. The instruction value is directly expanded along the time index of the preset historical period, and the distribution value of the current time is copied at each time step, so as to generate a static feature sequence that is constant in time, which is used to describe the power interaction relationship between the target charging gun and the adjacent charging gun. This step can be completed by using the Pandas data framework of Python, and the existing technologies such as time index filling and data broadcasting are used. Finally, the data preparation unit calls the time-of-use price prediction data of the future 2 hours from the power grid dispatching platform, and reads the vehicle flow prediction value of the future 2 hours from the reservation charging management system. The system combines these future external information in time sequence to form a future known sequence, which is used for subsequent prediction model input.

[0091] The application is further provided that the prediction execution unit:

[0092] The static feature sequence representing the adjacent charging gun power setting value is introduced in the input encoding stage of the time series fusion model, and the static feature weight is combined and fused with the attention weight of the time dimension in the multi-head attention calculation, so as to construct a time series fusion model enhanced by spatial correlation;

[0093] The historical state data sequence, the historical environment data sequence, the static feature sequence and the future known sequence are input into the time series fusion model enhanced by spatial correlation;

[0094] The output charging gun of each charging pile has a charging power demand prediction sequence in the future optimization time domain. Specifically, in actual operation, the prediction execution unit first receives four types of inputs provided by the data preparation unit: the cleaned and time-aligned historical state data sequence, the historical environment data sequence, the static feature sequence, and the future known sequence. In order to fully express the spatial interaction information in the prediction model, the system introduces the static feature sequence in the input encoding stage of the time series fusion model. Specifically, after aligning the static feature sequence with the historical state and environment sequence in the time dimension, it is mapped to a unified feature dimension through an embedding layer, and a learnable position encoding is added for each time step in the embedding stage; then, the static feature sequence is converted into a modulation vector through a light feedforward network, which is used as an additional weight to input the multi-head attention mechanism. In the multi-head attention calculation process, the model models the dependence of each time step while jointly fusing the attention weight using the static feature modulation vector, so that the attention distribution in the time dimension calculation reflects the spatial correlation brought by the adjacent charging gun power setting value. This process can be realized through the existing Transformer framework, and the invention does not limit the methods and tools used. Common engineering solutions include customizing the multi-head attention layer in the TensorFlow or PyTorch environment, using conditional modulation such as FiLM or attention weight reweighting methods. In the encoder part, the historical state data, environment data and fused static features form a spatio-temporal integrated deep representation; the decoder part introduces the future known sequence as a conditional input on this basis, and captures the influence of future external information on power demand through cross-attention mechanism. The decoder outputs the charging power demand prediction sequence of each charging gun in the future optimization time domain in parallel.

[0095] The multi-objective optimization decision module further includes an optimization calculation unit and an instruction execution unit.

[0096] The optimization calculation unit receives the system real-time state output by the data cleaning and state estimation module and the charging power demand prediction sequence output by the charging demand prediction module.

[0097] The allowed power upper and lower limits of each charging gun, the total power capacity of the charging pile, and the maximum power change rate of each charging gun are obtained.

[0098] The target state of charge information provided by each vehicle battery management system and the environmental temperature safety threshold are obtained, and a weighted multi-objective function is constructed based on a model predictive control framework.

[0099] System constraints including charging gun power upper and lower limits, system total power upper limit, and charging gun power change rate limit are applied to the multi-objective function.

[0100] The optimal power distribution instruction sequence is obtained by using a hybrid strategy of multi-objective genetic algorithm and particle swarm optimization to solve the multi-objective function. Specifically, first, the system receives the real-time state information output by the data cleaning and state estimation module, such as the current current, voltage and temperature estimation value of each charging gun, and receives the future charging power demand prediction sequence in the optimization period output by the charging demand prediction module. The system also synchronously acquires the upper and lower limits of the power allowed by each charging gun, the total power capacity that the charging pile can provide as a whole, and the maximum change rate of the power allowed by each charging gun, and reads the target state of charge information from the vehicle battery management system and the safety threshold of the environmental temperature. When constructing the optimization target, the system converts the multi-objective demand into a weighted multi-objective function based on the model predictive control framework, which comprehensively considers multiple objectives such as meeting the target state of charge at the vehicle end, reducing the peak load of the power grid, reducing the impact of power changes, and maintaining an appropriate temperature. To ensure the feasibility of the calculation results, the system imposes constraint conditions on the multi-objective function, including the upper and lower limits of the power of each charging gun, the total power limit of the charging pile, and the power change rate limit of a single charging gun at adjacent time points. To efficiently obtain an approximately global optimal solution under complex constraints, the system uses a hybrid strategy of multi-objective genetic algorithm and particle swarm optimization. The specific approach is as follows: the individual crossover and mutation mechanism of genetic algorithm is used to ensure the diversity of the solution space, and the group speed and position updating strategy of particle swarm optimization is used to accelerate the convergence of high-quality solutions; in the iteration process, the multi-objective function value is calculated as the fitness index, and the candidate solutions are sorted and selected, and the balance between genetic operation and particle update is adjusted through adaptive weight adjustment, gradually approaching the optimal power distribution scheme.

[0101] The application further provides that the instruction execution unit is configured to:

[0102] extract a first time step instruction to be executed at the current time from the optimal power distribution instruction sequence;

[0103] output the first time step instruction as a power distribution control instruction to the power accurate execution and compensation module;

[0104] The power allocation instruction value of the current time is fed back to the charging demand prediction module as a space-related feature. Specifically, first, the system receives the optimal power allocation instruction sequence output by the optimization calculation unit, which includes the power allocation scheme for a plurality of consecutive time steps from the current time. The system scheduling module extracts the first time step instruction corresponding to the current time from the sequence in chronological order, and updates the sequence after extraction so that the next time step instruction is placed at the head of the queue. Subsequently, the instruction execution unit packages the first time step instruction into a control message and sends it to the power accurate execution and compensation module through the internal control bus of the charging pile. The module adjusts the output power of each charging gun immediately after receiving the message to ensure that the real-time power allocation reaches the optimization goal. At the same time, the system collects the power allocation instruction value issued to each charging gun at the current time, organizes and packages the space-related features describing the power distribution state of the system through features, and feeds back the features to the charging demand prediction module.

[0105] The application further provides that the power accurate execution and compensation module comprises an execution control unit and a compensation learning unit.

[0106] The execution control unit comprises a closed-loop controller.

[0107] The execution control unit receives the power allocation control instruction output by the multi-objective optimization decision module.

[0108] The closed-loop controller drives the power hardware to execute the power allocation control instruction of the current time.

[0109] The actual output power and system state parameters are synchronously collected, and the system state parameters include the current output current, voltage and temperature. Specifically, first, the execution control unit receives the power allocation control instruction transmitted by the multi-objective optimization decision module through the standardized data interface. The instruction clearly specifies the target power value that each charging gun should output at the current time. After receiving the instruction, the closed-loop controller immediately compares the target value with the real-time output value of the power hardware, continuously calculates the deviation between the two values using closed-loop control strategies such as PID control, and adjusts the drive signal of the transformer accordingly to drive the power hardware to gradually approach the target value. At the same time, the execution control unit synchronously calls the built-in sensor network to obtain the actual output power data, measures the output current through the current transformer, detects the output voltage through the voltage sensor, and uses the thermistor to sense the temperature of the power module or the environment. All collected system state parameters are transmitted to the local controller after high-speed analog-to-digital conversion and stored in the system database for real-time monitoring and subsequent analysis.

[0110] The application further provides that the compensation learning unit comprises a learning database and a learning algorithm.

[0111] The compensation learning unit calculates the execution deviation between the power allocation control instruction and the actual output power.

[0112] establishing an execution deviation prediction model through an online learning algorithm based on the execution deviation and the system state parameters;

[0113] updating the parameters of the execution deviation prediction model through recursive least squares in an online adaptive manner;

[0114] performing feedforward compensation on the power allocation control instruction to be executed by using the updated execution deviation prediction model to generate a compensated control instruction;

[0115] delivering the compensated control instruction to a device execution interface to drive the power hardware to execute. Specifically, first, the compensation learning unit receives the power allocation control instruction issued by the multi-objective optimization decision module in each scheduling period, and obtains the current actual output power from the execution control unit. By comparing the two, the execution deviation of each charging gun at that time is calculated. Next, the execution deviation and the synchronously collected system state parameters are input into the online learning algorithm framework to establish an execution deviation prediction model. The model can be initially constructed based on a linear prediction model (such as an incremental linear regression method) or a nonlinear prediction model (such as a long short-term memory network LSTM), and continuously receives new deviation and state data during operation. The compensation learning unit calls the recursive least squares method to correct the model parameters in real time, and quickly updates the model coefficients every time a new data set is received, so as to ensure that the prediction ability always matches the dynamic changes of the system. Before the next scheduling period arrives, the compensation learning unit inputs the power allocation control instruction to be executed into the updated execution deviation prediction model, predicts the possible execution deviation in the future, and accordingly performs feedforward compensation on the original instruction, such as increasing or decreasing a certain power margin in advance. Finally, the generated compensated control instruction is delivered to the execution control unit through the device execution interface to drive the power hardware to execute.

[0116] The application further provides that the safety monitoring and fusing module comprises:

[0117] collecting real-time running parameters of the system, wherein the running parameters comprise a power instruction value, an output current, an output voltage, and a device temperature;

[0118] calculating the real-time running parameters of the system in real time by using an unsupervised isolation forest algorithm to generate an edge real-time anomaly score;

[0119] triggering a hierarchical safety fusing mechanism according to a preset anomaly threshold, wherein the safety fusing mechanism comprises:

[0120] when the edge real-time anomaly score is between the first anomaly threshold and the second anomaly threshold, determining that it is a mild anomaly, triggering a power reduction operation, and generating an alarm information;

[0121] When the edge instant anomaly score is higher than the second anomaly threshold, it is determined that there is a serious anomaly, an emergency disconnect operation is triggered, power output is stopped, and an alarm information is generated. Specifically, first, the safety monitoring and fusing module continuously collects current power instruction value, output current, output voltage, and device temperature and other operating parameters through the high-speed data channel established with the charging pile power hardware interface, and aligns the collected data with time stamps and performs simple outlier rejection processing to ensure the real-time and reliability of the input data. Subsequently, these real-time operating parameters are input into the unsupervised isolation forest algorithm. This algorithm has constructed a random partition tree structure using historical normal operation data before system deployment, and in the running process, it does not need manual annotation of abnormal samples, but measures the deviation of each set of real-time data from the normal mode by comparing the average path length in the random tree, thereby outputting the edge instant anomaly score. The module compares this score with the two-level threshold set in advance according to a large amount of historical monitoring data and risk analysis. When the score is between the first anomaly threshold and the second anomaly threshold, the system determines that there is a mild anomaly, immediately issues an alarm information and automatically reduces the charging power instruction value, reduces the output power through the power control interface to reduce the device load and reduce the risk. When the score exceeds the second anomaly threshold, the system determines that there is a serious anomaly, triggers an emergency disconnect operation, immediately interrupts all power output and closes the power drive loop, and generates an alarm information and pushes it to the operation and maintenance platform or the mobile phone and email of the on-duty personnel. After executing the protection action, the module continues to monitor the operating parameters in real time, re-enters the latest data into the isolation forest model, so that the system can gradually restore normal power output according to the preset strategy after the risk is removed. Through this closed-loop adaptive safety monitoring and fusing mechanism, the charging pile operating anomaly is quickly identified and graded response, effectively ensuring the safety of the equipment and personnel.

[0122] The above merely describes a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A multi-gun charging pile charging power distribution system, characterized in that, The application relates to a charging station real-time state estimation and power allocation control method, which comprises the following steps: a data cleaning and state estimation module: real-time collection of multi-source sensing data, output of system real-time state based on multi-source sensing data and a deep filtering algorithm optimized by a fusion neural network and a physical model, the data cleaning and state estimation module comprising a data collection and preprocessing unit and a state estimation unit; the data collection and preprocessing unit: receiving parameters including set power instructions of each charging gun, power grid power supply constraints and environmental working condition parameters, performing time alignment and processing on the parameters to construct a control input vector; real-time collection of original monitoring data from multi-source sensors, the original monitoring data comprising current, voltage and temperature of each charging gun; data cleaning and time stamp alignment processing of the original monitoring data to obtain preprocessed monitoring data; the state estimation unit: compensating for unmodeled nonlinear dynamics in a state estimation algorithm by using a neural network to construct a deep filtering algorithm; setting a state posterior estimation value of the last moment as a default initial value as a filtering starting condition by a preset initialization strategy when the system is initially operated; inputting the preprocessed monitoring data into the deep filtering algorithm to perform a state prediction operation, and generating a state prior estimation value of the current moment based on the state posterior estimation value of the last moment and the control input vector; comparing the monitoring data of the current moment with the state prior estimation value, correcting the state prior estimation value according to a filtering gain, and obtaining a state posterior estimation value of the current moment; outputting the state posterior estimation value as a system real-time state, the system real-time state comprising current, voltage and temperature estimation values of each charging gun; a charging demand prediction module: generating a charging power demand prediction value by using a space-time fusion prediction model based on a historical state data sequence and external environmental data; a multi-objective optimization decision module: taking the charging power demand prediction value and the system real-time state as inputs, performing multi-objective constraint optimization by using a model predictive control framework with constraints, and generating a power allocation control instruction; a power accurate execution and compensation module: receiving the power allocation control instruction to drive hardware execution, feeding forward compensating the power allocation control instruction by using an online learning algorithm, and generating a compensated control instruction; a safety monitoring and fusing module: collecting system real-time operation parameters, calculating an edge real-time abnormal score by using an unsupervised isolation forest algorithm, and triggering a hierarchical safety fusing mechanism according to a preset abnormal threshold.

2. The charging power distribution system of claim 1, wherein, The charging demand prediction module comprises a data preparation unit and a prediction execution unit. The data preparation unit: collects a historical state data sequence, the historical state data sequence comprising current, voltage and temperature parameters of each charging gun in a preset historical period; collects a historical environmental data sequence, the historical environmental data sequence comprising environmental temperature and humidity of an environment where the charging station is located in the preset historical period; uses power allocation instruction values of adjacent charging guns in the current moment as space correlation features, and expands the space correlation features into a static feature sequence in the preset historical period; obtains a time-of-use electricity price signal in a future optimization period and vehicle flow information of a pre-booking charging station to construct a future known sequence.

3. The charging power distribution system of claim 2, wherein, The prediction execution unit comprises: By introducing a static feature sequence representing the power setting value of the adjacent charging gun in the input encoding stage of the time fusion model, and jointly fusing the static feature weight with the attention weight of the time dimension in the multi-head attention calculation, a time fusion model enhanced by spatial correlation is constructed; The historical state data sequence, the historical environment data sequence, the static feature sequence and the future known sequence are input into the time fusion model enhanced by spatial correlation. The power demand prediction sequence of each charging pile in the future optimization time domain is output.

4. The charging power distribution system of claim 1, wherein, The multi-objective optimization decision module comprises an optimization calculation unit and an instruction execution unit. The optimization calculation unit receives the system real-time state output by the data cleaning and state estimation module and the charging power demand prediction sequence output by the charging demand prediction module. The upper and lower limits of the allowable power of each charging gun, the total power capacity of the charging pile and the maximum power change rate of each charging gun are obtained. The target state of charge information provided by each vehicle battery management system and the environmental temperature safety threshold are obtained, and a weighted multi-objective function is constructed based on a model predictive control framework. System constraints including the upper and lower limits of the charging gun power, the total power upper limit of the system and the power change rate limit of the charging gun are applied to the multi-objective function. A hybrid strategy of multi-objective genetic algorithm and particle swarm optimization is used to solve the multi-objective function, and an optimal power distribution instruction sequence is obtained.

5. The charging power distribution system of claim 4, wherein, The instruction execution unit is configured to: extract the first time step instruction to be executed at the current time from the optimal power distribution instruction sequence; output the first time step instruction as a power distribution control instruction to the power accurate execution and compensation module; feed back the power distribution instruction value at the current time to the charging demand prediction module as a spatial correlation feature.

6. The charging power distribution system for a multi-gun charging station of claim 1, wherein, The power accurate execution and compensation module comprises an execution control unit and a compensation learning unit. The execution control unit is configured to: receive the power distribution control instruction output by the multi-objective optimization decision module; drive the power hardware to execute the power distribution control instruction at the current time using a closed-loop controller; synchronously collect the actual output power and system state parameters, including the current output current, voltage and temperature.

7. The charging power distribution system of claim 6, wherein, The compensation learning unit is configured to: calculate the execution deviation between the power distribution control instruction and the actual output power; establish an execution deviation prediction model based on the execution deviation and the system state parameters through an online learning algorithm; update the parameters of the execution deviation prediction model online and adaptively using the recursive least squares method; use the updated execution deviation prediction model to feed forward compensate the power distribution control instruction to be executed soon, and generate a compensated control instruction; deliver the compensated control instruction to the device execution interface to drive the power hardware to execute.

8. The charging power distribution system of claim 1, wherein, The safety monitoring and fusing module comprises: collect system real-time running parameters, including power instruction value, output current, output voltage and device temperature; use an unsupervised isolation forest algorithm to calculate the system real-time running parameters in real time to generate an edge real-time anomaly score; trigger a hierarchical safety fusing mechanism according to a preset anomaly threshold, the safety fusing mechanism comprising: When the edge instant abnormality score is between the first abnormality threshold and the second abnormality threshold, it is determined to be a mild abnormality, a power reduction operation is triggered, and alarm information is generated; When the edge instant abnormality score is higher than the second abnormality threshold, it is determined to be a serious abnormality, an emergency shutdown operation is triggered, power output is stopped, and alarm information is generated.

Images