A standby module prediction energy-saving control method for a modular direct current fast charging station

Abstract

The present application belongs to the field of public fast charging station power module energy consumption management, and particularly relates to a standby module prediction energy-saving control method for a modular DC fast charging station, comprising the following steps: S1. reconstructing a vehicle-level charging power trajectory in the cloud; S2. collecting station-level real module requirements; S3. predicting module requirement probability distribution in a future scheduling period; S4. constructing a residual distribution between the predicted value and the real value of the module requirement; S5. determining a baseline standby power module quantity that meets service reliability constraints, and outputting a standby power module quantity control target; and S6. issuing the module quantity control target from the cloud to a station-side controller for control. The present application realizes dynamic optimization of the standby power module quantity, can reduce redundant module standby energy consumption on the premise of guaranteeing charging service reliability, improves the operation energy efficiency and energy management refinement level of the modular DC fast charging station, and is helpful for energy-saving operation, low-carbon management and sustainable operation of urban public fast charging infrastructure.

Description

Technical Field

[0001] This invention belongs to the field of power module energy consumption management for public fast charging stations, and specifically relates to a predictive energy-saving control method for standby modules of modular DC fast charging stations. Background Technology

[0002] Public DC fast charging stations typically employ a modular, split-type architecture. This architecture centrally arranges multiple DC power modules in a power cabinet and dynamically distributes power to multiple terminals through charging stacks or power distribution units. Power modules can be in active, standby, or sleep states during operation. Modules in standby state are already charged and can respond quickly to charging demands, but they still generate continuous standby energy consumption when not outputting power to vehicles.

[0003] To ensure charging service availability, operators typically prefer to keep a large number of power modules in standby mode for extended periods. For large fast-charging stations with a large number of modules, this conservative strategy results in many idle modules continuously consuming power during low-load periods. On the other hand, fast-charging demand exhibits significant intraday fluctuations and site variations; simply configuring modules based on peak demand or maintaining a fixed proportion of standby modules makes it difficult to balance energy-saving benefits with service risks.

[0004] Existing station-level scheduling methods mostly treat aggregated loads as continuous power demand, but the capacity supply of split-type DC fast charging systems is achieved by discrete power modules. The same aggregated power may be composed of charging power combinations from different vehicles, and the number of modules required by each vehicle needs to be rounded down before being aggregated. Therefore, directly estimating module demand based on aggregated power will result in a systematic underestimation. If dormant modules are reactively activated after demand arrives, they will also be affected by engineering processes such as pre-charging, soft start, controller initialization, and power reallocation, thus causing temporary service degradation when peak demand arrives early or there is a sudden surge in demand.

[0005] Therefore, there is a need for a standby power module number prediction and control scheme that can utilize vehicle-level discrete module demand inference, short-term demand probability prediction, risk constraints, and look-ahead activation mechanisms to enable fast charging stations to maintain acceptable service reliability while reducing the number of long-term powered standby modules. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a standby module predictive energy-saving control technology for modular DC fast charging stations, which includes a cloud server, a station controller, and a power distribution unit. This technology is used to solve the problems of high energy consumption of existing conservative standby strategies, easy underestimation of the demand of aggregated power estimation modules, and difficulty in dealing with peak time mismatch in the number of standby power modules.

[0007] Technical solution A predictive energy-saving control method for standby modules in modular DC fast charging stations includes the following steps: S1. The station-side data acquisition unit acquires historical charging session records, vehicle information, and station module configurations, and reconstructs the vehicle-level charging power trajectory in the cloud; S2. Based on the rated power of the power module, convert the vehicle-level power trajectory into vehicle-level discrete module requirements, and summarize them to obtain the actual station-level module requirements; S3. Using the actual module demand at the station level as the prediction object, predict the probability distribution of module demand during future scheduling periods based on historical operation data; S4. Based on historical forecast results, construct the residual distribution between the predicted and actual values ​​of module demand to characterize the forecast bias and uncertainty; S5. Combining historical forecast residuals and risk tolerance parameters, determine the baseline number of standby power modules that meet service reliability constraints, and perform forward activation correction based on future demand changes within a preset window, outputting the target number of standby power modules for the current scheduling interval. S6. The cloud sends the module quantity control target to the station controller, which then controls the standby, sleep, and activation states of the power modules in the station through the charging management system and power distribution unit.

[0008] Step S1 includes the following steps: (1) Collect historical charging session records, vehicle information and station module configuration data through the data acquisition unit, charging management system or station controller set in the modular DC fast charging station, and upload them to the cloud server through the station communication unit; The historical charging session record includes at least the charging session identifier, charging start time, charging end time, and actual delivered electricity. The vehicle information includes at least one or more of the following: vehicle model identifier, battery capacity, charging power level, initial state of charge, and final state of charge. The station module configuration data includes at least one or more of the following: rated power of the power module, number of power modules, and available status of the power modules.

[0009] (2) Based on high-resolution historical charging records (recording interval <= 1min), the cloud server extracts the empirical charging curves under typical vehicle models, battery capacities and charging power levels, and constructs a reference charging curve library.

[0010] For each reference trajectory, the vehicle's state of charge (SOC) value is used. Establish a reference charging power function with the variable as the independent variable. and cumulative charging time function It also stores the corresponding vehicle model identifier, battery capacity range, charging power level, SOC coverage range, and maximum allowable charging power.

[0011] (3) Based on the vehicle model identification, battery capacity, and charging power level of the target charging session, match the closest reference vehicle configuration from the reference curve library. When a perfectly matching reference vehicle configuration exists, select the reference charging curve corresponding to that reference vehicle configuration; when no perfectly matching configuration exists, select a reference curve of the same type of vehicle and within a similar charging power range as a substitute curve.

[0012] (4) To preserve the power change characteristics of the main charging stage and the final power reduction stage during DC fast charging, a SOC dividing point is set. As the boundary between the main charging phase and the final power reduction phase, and based on the initial state of charge of the target charging session. Termination of charging state With the SOC boundary point Based on the relationship, the target charging session is divided into stages. If If the target charging session crosses the SOC boundary point, it will be divided into... to The main charging phase and to The final power reduction stage; if If so, the entire target charging session is treated as the main charging phase; if The target charging session is then treated as a final power reduction phase.

[0013] When the target charging session crosses the SOC boundary point and only includes the initial state of charge and the final state of charge, it is assumed that the SOC changes linearly within the effective charging duration of the target charging session, and the target duration for the two stages is determined according to the following formula: in, Indicates the effective charging duration of the target charging session. Indicates the target duration of the main charging phase. This indicates the target duration of the end-of-phase power reduction phase. If the target charging session includes SOC sampling data within the phase, it is determined based on the actual SOC sampling sequence. The corresponding stage boundary time.

[0014] Within each stage, the target SOC progress is mapped to the curve segment corresponding to the SOC interval in the reference charging curve, and then... Obtain the initial reconstructed power trajectory .

[0015] (5) To avoid unreasonable power surges at the start of charging, a linear power ramp-up process is set during the initial ramp-up period. Subsequently, the power is increased based on the actual charging capacity of the target session. For the initial reconstructed power trajectory Perform overall scaling.

[0016] Specifically, the energy consistency scaling factor is first determined based on the integral value of the initial reconstructed power trajectory over the effective charging time. : The initial reconstructed power trajectory is corrected based on the energy consistency scaling factor to obtain the scaled vehicle-level charging power trajectory. When a maximum charging power constraint exists, the reconstructed power trajectory is truncated at the upper limit. If the interruption causes a loss of power, compensation will be prioritized during the period after the cutoff point and before the power limit is reached, in order to keep the total charging capacity as consistent as possible without violating the maximum power constraint.

[0017] (6) Obtaining the vehicle-level charging power trajectory after power constraint processing Then, the cumulative delivered energy is calculated based on the cumulative integral of the vehicle-level charging power trajectory over time.

[0018] Let the initial state of charge of the target charging session be... The state of termination of charge is Then at time Corresponding cumulative delivered energy for: Based on cumulative delivered energy and actual delivered electricity Reconstructing the SOC trajectory from the initial state of charge and the final state of charge. This represents the total battery capacity in the order. This ensures the SOC trajectory aligns with the actual energy input, avoiding deviations caused by simply using linear interpolation over time. The final output is the vehicle-level charging power trajectory. and corresponding SOC state This serves as the basic input for subsequent calculations of vehicle-level discrete module requirements.

[0019] Step S2 discretizes the reconstructed vehicle-level charging power trajectory according to the scheduling time interval based on the rated power of the power modules configured at the station. For each vehicle's instantaneous charging power within the same scheduling interval, the minimum number of power modules required is calculated to obtain the vehicle-level discrete module demand. Then, the discrete module demands of all charging vehicles within the same scheduling interval are aggregated to obtain the station-level actual module demand. This station-level actual module demand characterizes the power module occupancy level of the modular DC fast charging station under actual concurrent vehicle charging conditions. It avoids the systematic underestimation caused by the discretization of vehicle power demand and the smoothing of module allocation constraints when directly estimating module demand from the station-level aggregated power. The calculation formula is as follows: in, For vehicles At any moment The charging power, Rated power for a single power module For vehicle-level discrete module requirements, This refers to the number of vehicles charging at the same time. This is for real-world module requirements at the site level.

[0020] Step S3 includes the following steps: (1) Based on the station-level real module demand sequence, construct a heterogeneous input feature space for short-term module demand prediction.

[0021] Input features include historical station-level charging power, historical station-level actual module demand, current number of charging vehicles, number of charging requests entering the station, initial SOC distribution of vehicles, vehicle power level or model distribution, periodic time features such as hours and minutes, weekday and holiday types, electricity price signals, and meteorological features such as temperature, humidity, air pressure, and wind speed.

[0022] These characteristics collectively characterize the site load level, vehicle arrival status, vehicle charging differences, time cycle patterns, electricity price impact, and environmental impact.

[0023] (2) Continuously collect historical operating data of the above heterogeneous input feature space as the dynamic input of the prediction model. Let the length of the historical review window be... The number of future predicted step sizes is .

[0024] In one implementation, the history playback window length The scheduling intervals are selected from 24 to 144, preferably 72, i.e., one day; the number of future prediction steps... Choose 1 to 12 scheduling intervals, preferably 3 scheduling intervals, i.e., 1 hour. Input the historical sequence. According to the preset segment length and sliding step size Divided into several time segments, where the segment length is... The number of scheduling intervals can be 6 to 48, preferably 24, with a sliding step size. Choose 3 to 24 scheduling intervals, preferably 12. Perform a linear mapping on each time segment to obtain a segment embedding representation, and then superimpose learnable location embeddings and time embeddings; the time embeddings include at least hour embeddings, weekday embeddings, and date type embeddings. Correspondingly, the number of segments... It can be determined according to the following formula: The extra segments are obtained by padding the end values ​​to ensure that the end information of the history window can be encoded.

[0025] (3) Construct an exogenous variable vector from the site attributes and electricity price signal variables. The exogenous variable representation is obtained by projecting it onto the same latent space as the temporal features through a multilayer perceptron. .

[0026] In one implementation, the exogenous variable vector includes several key operational features at various stations and electricity price features, wherein the electricity price features can be obtained by hourly mapping. Subsequently, the correlation between historical time-series features and the exogenous variable representation is calculated using a cross-attention mechanism (CrossAttn), and the degree of correction of the exogenous variables for each time segment feature is controlled by a gated residual structure to obtain the fused predicted feature representation. The corresponding expression is: in, This represents the historical time-series features obtained by dividing historical input features into time segments, linear mapping, and temporal embedding. Representation of exogenous variables Cross-attention fusion is performed to obtain . Represents the projection function of exogenous variables. Represents the cross-attention function. Indicates the gating coefficient. This represents the Sigmoid function. This represents element-wise multiplication. Representation layer normalization.

[0027] (4) Represent the fused prediction features After flattening, it is compared with the exogenous variable vector. The data are concatenated and input into a multi-layer output network to obtain the module demand probability distribution parameters for each future prediction step. The output network includes at least one fully connected layer and a nonlinear activation layer. In one implementation, the output network sequentially comprises a 512-dimensional fully connected layer, a GELU activation layer, a dropout layer, a 256-dimensional fully connected layer, a GELU activation layer, and an output layer, wherein the output layer has a dimension of [missing information]. Each corresponds to the future The mean parameter of each prediction step size and dispersion parameter To ensure the parameters are non-negative, the Softplus function is used to transform the output parameters: in, and These represent the output layer of the prediction model for the future... The original output values ​​of the mean parameter and the original output values ​​of the dispersion parameter generated for each scheduling period are transformed by the Softplus function to obtain a non-negative mean parameter. and dispersion parameter . Indicates the future number The predicted average demand of power supply modules for each scheduling period Indicates the future number The dispersion parameter (reliability) of the forecast mean for each scheduling period is used to adjust the range of random fluctuations in demand. To prevent positive numbers with parameters equal to zero, it is preferable to... .

[0028] (5) Considering that the station-level module demand is a non-negative discrete variable, the prediction model outputs the probability distribution parameters of the module demand for each future prediction step. In one implementation, a negative binomial distribution is used to characterize the uncertainty of the module demand. The model outputs the mean parameter and the degree of dispersion parameter to obtain the probability distribution of the station-level module demand within the future scheduling period. The corresponding formula is: Equation (17) is the probability mass function form of the negative binomial distribution, used to describe discrete random variables. Quantity required for different modules The probability of occurrence corresponding to the given time. This indicates the number of possible module requirements. When the dispersion parameter changes, the negative binomial distribution can describe the excessive dispersion phenomenon, which differs from the Poisson distribution, thus making it suitable for characterizing the random fluctuation characteristics in the demand for fast charging station modules.

[0029] During model training, the actual station-level module requirements are used as supervision signals. By minimizing the negative log-likelihood of the observation module requirements under the prediction distribution and combining it with absolute error constraints, the accuracy of probabilistic prediction and the stability of point prediction are improved.

[0030] The training objective can be expressed as: in, For a moment For the future The actual module requirements at the station level for each scheduling period. Indicates the future number The predicted average demand of power supply modules for each scheduling period express Substitute the probability value obtained by the negative binomial distribution probability mass function predicted by the model, and learn the most suitable distribution shape; This represents the weight of the absolute error term, ranging from 0.1 to 10, with 3 being the preferred value. The AdamW optimizer is used during training, with a learning rate of [value missing]. to Preferred The batch size is 16 to 256, preferably 64, the number of training rounds is 5 to 100, and gradient clipping is used to limit the gradient norm, with the upper limit of the gradient norm being 0.1 to 5, preferably 1.0.

[0031] (6) After the prediction is completed, the probability distribution, predicted mean, and predicted quantile of the station-level module demand in the future scheduling period are obtained. The predicted mean or predicted distribution quantile is used as the input for subsequent standby power module scheduling, and the prediction results are constrained by the number of power modules already installed at the station to ensure that the predicted module demand does not exceed the number of available power modules at the station, thus providing a basis for determining the baseline number of standby power modules under risk constraints in step S5.

[0032] Step S4 includes the following steps: (1) Based on the prediction model described in step S3, rolling backtesting is performed on historical operation data according to the historical backtesting window length, prediction step size and scheduling interval consistent with online scheduling to obtain historical time points. For the future The module demand forecast results for each scheduling period are obtained, and the forecast results are time-aligned with the actual station-level module demand for the same site and the same scheduling period to construct historical backtesting samples.

[0033] (2) Calculate the module demand prediction residuals in the historical backtesting samples, and extract the positive residuals where the actual demand is higher than the predicted demand for service risks caused by demand underestimation in standby power module scheduling, and construct the empirical distribution of the underestimation residuals. The corresponding formula is: in, Representing historical moments For the future Module demand forecast results for each scheduling period This indicates the corresponding actual site-level module requirements. Indicates the predicted residual. This indicates that demand underestimates the residual. Indicates the hour to which the target scheduling period belongs. Indicates the first The set of underestimated residual samples corresponding to the hour. Represents the empirical cumulative distribution function. Indicates the first The empirical distribution of the underestimated residuals corresponding to the hour.

[0034] Step S5 includes the following steps: (1) At the current scheduling time Read the future output of step S3 The probability distribution of module demand forecast for the first scheduling period. For the future first scheduling period... For each scheduling period, let its predicted cumulative distribution function be denoted as . ,in Indicates the possible number of modules required. According to the preset risk tolerance parameters The predicted demand quantiles are extracted from the predicted cumulative distribution function and used as the upper bound of module demand under the uncertainty within the predicted distribution. in, Indicates the future number Each scheduling period is within the risk tolerance parameter The forecast module below requires quantiles. Represents the set of non-negative integers.

[0035] (2) Read the hourly-level underestimated residual empirical distribution output in step S4. Assume the future... Dispatch period The hour to which it belongs is ,in From the estimated residual empirical cumulative distribution function corresponding to that hour. In the process of extracting risk tolerance parameters Corresponding residual correction amount: in, Indicates the future number The hourly-level underestimated residual correction amount corresponding to each scheduling period. This represents the underestimated residual empirical cumulative distribution function for the hour to which the target scheduling period belongs. This correction is used to compensate for the risk of insufficient standby modules caused by historically lower predicted module demand than actual module demand within the same hour.

[0036] (3) Add the predicted demand quantiles to the hourly residual corrections to obtain the future demand quantiles. The reliability correction module requirements for each scheduling period. In one implementation, it is assumed that the short-term prediction error and the historically underestimated residual samples are independent of each other, and the corrected module requirement distribution is obtained through random sampling or discrete convolution.

[0037] Considering the limited number of available power modules at a site, the number of power modules installed or available at the site will be used as a basis. By applying an upper limit constraint, the baseline number of standby power modules is obtained: in, Indicates the future number The number of baseline standby power modules for each scheduling period. This indicates the number of power modules that have been installed or are available at the site. This step transforms the random uncertainty in the predicted probability distribution and the systematic bias in the historically underestimated residuals into a module configuration objective under reliability constraints.

[0038] (4) Considering the response time when the power module switches from sleep mode to standby mode, controlling only according to the baseline target of the current scheduling period may not be able to respond promptly to the rapidly increasing charging demand in a short period of time. Therefore, the length is set to... Forward activation window, in the future The maximum number of baseline standby power modules searched within a single scheduling period: in, , Indicates the current scheduling time The maximum baseline number of standby power modules within the next forward-looking window. This value characterizes the highest standby module demand that may occur in the near future.

[0039] (5) Based on the prospective activation coefficient The baseline target for the current scheduling period is revised in advance to obtain the current scheduling target after forward correction: in, This indicates the current scheduling objective after forward-looking correction. This represents the look-ahead activation coefficient. When... When no look-ahead activation correction is performed, the current scheduling objective is equal to the baseline objective for the current scheduling period; when When the current scheduling target is raised to the maximum baseline target within the look-ahead window, the current scheduling target is directly promoted; when At that time, the current scheduling target is smoothly increased between the current baseline target and the maximum target of the forward window.

[0040] (6) Since the number of power modules is a non-negative integer and cannot exceed the number of available power modules at the site. By rounding up and applying an upper limit constraint to the current scheduling target after look-ahead correction, the final output standby power module number control target for the current scheduling interval is obtained: in, This indicates the target number of standby power modules to be output in the current scheduling interval. This indicates rounding up. The control target... Used to send commands to the station controller to guide the switching of standby, hibernation, and early activation states of the power modules within the station.

[0041] Step S6 includes the following steps: (1) The cloud server controls the target number of standby power modules for the current scheduling interval obtained in step S5. The control is issued to the station-side controller of the corresponding modular DC fast charging station. The control target includes at least the station number, scheduling time, control effective time, and the number of target standby power modules. One or more of the following: control effective duration and abnormal protection flags.

[0042] (2) After receiving the control target, the station-side controller reads the current operating status of the power modules in the station and the real-time demand of the charging vehicles, and determines the current module status and control target accordingly. The difference is used to generate module state switching instructions. Through the charging management system and power distribution unit, the standby, hibernation, and early activation states of power modules within the station are controlled. During execution, power modules currently participating in charging output remain operational; power modules experiencing faults, maintenance, or communication abnormalities are not included in the set of schedulable modules. The station-side controller feeds back the actual number of standby modules, module state switching results, and abnormal states to the cloud server for subsequent prediction and control updates.

[0043] Beneficial effects Compared with existing technologies, the method of the present invention has the following advantages: Reduced standby energy consumption: This invention dynamically determines the number of power modules that need to remain energized within each scheduling interval through module demand probability prediction, historical residual correction, and risk-constrained scheduling, thereby reducing standby losses caused by redundant modules remaining energized for extended periods during low-load periods. Taking Shanghai's public fast-charging stations as an example, it can reduce standby energy consumption by approximately 79%, and for megawatt-level stations, the reference value is approximately 5.41 MWh / station / year.

[0044] Ensuring service reliability: This invention introduces risk tolerance parameters and hourly underestimated residual distribution. The number of baseline standby power modules is determined based on the predicted demand quantile and the historical underestimated residual correction amount. During execution, the station-side controller combines the real-time demand of vehicles being charged and the module availability status for protective control. This reduces standby energy consumption while controlling the risk of insufficient module quantity, ensuring that the scheduling results meet the requirements of public fast charging stations for service continuity and responsiveness.

[0045] Enhanced peak response capability: This invention sets up a forward activation window to adjust the current number of charged modules in advance based on future short-term module demand changes, thereby reducing the risk of insufficient response caused by sudden demand surges or predicted peak delays.

[0046] Facilitates engineering deployment: This invention outputs a target number of standby power modules for the current scheduling interval. This target can be generated by a cloud server and sent to the station controller. The station controller, combining the current status of power modules within the station and the real-time charging needs of vehicles, executes the switching of power modules between energized standby, sleep, and active states through the charging management system and power allocation unit. Therefore, this invention can be integrated with existing modular DC fast charging station control systems without relying on additional hardware, facilitating deployment at existing sites. Attached Figure Description

[0047] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the charging power curve reconstruction process according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the module demand uncertainty induced by aggregated charging demand in an embodiment of the present invention; Figure 4 This is a schematic diagram of the standby module quantity prediction control framework according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the prediction algorithm in an embodiment of the present invention; Figure 6 This is a schematic diagram of the residual prediction process in an embodiment of the present invention; Figure 7 This is a schematic diagram of the 24-hour prediction residual distribution in an embodiment of the present invention; Figure 8This is a schematic diagram of module redundancy configuration based on quantile control in an embodiment of the present invention; Figure 9 This is a schematic diagram of the module early wake-up control based on look-ahead activation buffer in an embodiment of the present invention. Detailed Implementation

[0048] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0049] like Figure 1 As shown, a predictive energy-saving control method for standby modules in modular DC fast charging stations includes the following steps: S1. The station-side data acquisition unit acquires historical charging session records, vehicle information, and station module configurations, and reconstructs the vehicle-level charging power trajectory in the cloud; S2. Based on the rated power of the power module, convert the vehicle-level power trajectory into vehicle-level discrete module requirements, and summarize them to obtain the actual station-level module requirements; S3. Using the actual module demand at the station level as the prediction object, predict the probability distribution of module demand during future scheduling periods based on historical operation data; S4. Based on historical forecast results, construct the residual distribution between the predicted and actual values ​​of module demand to characterize the forecast bias and uncertainty; S5. Combining historical forecast residuals and risk tolerance parameters, determine the baseline number of standby power modules that meet service reliability constraints, and perform forward activation correction based on future demand changes within a preset window, outputting the target number of standby power modules for the current scheduling interval. S6. The cloud sends the module quantity control target to the station controller, which then controls the standby, hibernation, and activation status of the power modules in the station through the charging management system and power distribution unit.

[0050] This invention takes a large-scale DC public fast charging station (configured with 250 modules) as an example. Based on the standby power module predictive energy-saving control technology for modular DC fast charging stations proposed in this invention, an energy-saving predictive control scheme suitable for the power modules of this station is generated. The specific implementation steps are as follows: S1. Acquire basic data for the target charging session through the station-side data acquisition unit, charging management system, or station-side controller, and upload it to the cloud server for trajectory reconstruction. For the target charging session, the cloud server matches a reference trajectory in the reference charging curve library based on the vehicle model identifier, battery capacity, and charging power level. If a reference trajectory with the same vehicle model, battery capacity range, and charging power level exists, that reference trajectory is directly adopted; if no completely identical reference trajectory exists, an alternative reference trajectory is selected according to the rules of prioritizing vehicle model category, similar battery capacity, and similar charging power level.

[0051] In one implementation, the target charging session log shows that a vehicle was... SOC charging to SOC, actual delivered power is The cloud server matches a charging curve from a reference charging curve library based on the vehicle's model identifier, battery capacity, and charging power level. SOC to The reference trajectory of State of Charge (SOC). To preserve the power change characteristics of the main charging phase and the final power reduction phase during DC fast charging, a SOC cutoff point is set. In one implementation, (Right now SOC). At this time. If the target session crosses the SOC boundary, it will be divided into... to The main charging phase and to The final power reduction phase, and the target duration of the two phases are determined according to the following formula: in, Indicates the effective charging duration of the target charging session. Indicates the target duration of the main charging phase. This indicates the target duration of the end-of-phase power reduction phase. If the target session includes SOC sampling data within the phase, it is determined based on the actual SOC sampling sequence. The corresponding stage boundary time. For the example above, , , The target durations for the main charging phase and the final power reduction phase are: That is, the target session is divided into SOC to The main charging phase of the SOC, and SOC to The final power reduction phase of SOC.

[0052] Within each stage, the cloud server generates a target SOC sequence according to the time progression of that stage, and maps the target SOC sequence to the power curve segment of the corresponding SOC interval in the reference trajectory, based on the reference charging power function. Obtain the initial reconfigured power trajectory of the target charging session. .

[0053] To avoid unreasonable power surges at the start of charging, a linear power ramp-up process is implemented during the initial ramp-up period after the target charging session begins. In one embodiment, the initial ramp-up period is... to The charging power increases linearly from 0 to the target power level corresponding to the initial reconstructed power trajectory.

[0054] After completing the ramp-up correction, calculate the integral charge of the initial reconfigured power trajectory within the effective charging duration of the target session, and then calculate the actual charge delivered during the target session. Determine the energy consistency scaling factor: The initial reconstructed power trajectory is corrected based on this scaling factor to obtain the scaled vehicle-level charging power trajectory: For the above example, the actual delivered electricity was Therefore, the cloud server will initially reconstruct the power trajectory. The reconstructed charge obtained by integrating within the effective charging duration of the target session and Compare the two and determine the scaling factor based on their ratio. , so that the scaled power trajectory The credited power is consistent with the actual power delivered in the target session.

[0055] When the target charging session has a maximum charging power constraint At that time, the scaled power trajectory is truncated at an upper limit to ensure that the reconstructed power at any given moment does not exceed the maximum power allowed by the vehicle or charging equipment. If the upper limit truncation results in the cumulative charging capacity being lower than the actual delivered capacity, compensation will be made during the period before the maximum charging power constraint is reached, prioritizing the SOC (State of Charge) threshold. Compensation is then provided during periods when the power limit has not been reached, in order to keep the total charging capacity as consistent as possible without violating the maximum power constraint.

[0056] The vehicle-level charging power trajectory after power constraint processing is obtained. Then, the cloud server calculates the cumulative delivered energy in chronological order: Let the initial state of charge of the target charging session be... The state of termination of charge is The actual delivered electricity volume was Then the target session at time The SOC trajectory is reconstructed according to the following formula: For the above examples, , , The cloud server calculates the cumulative delivered energy based on the scaled and constrained power trajectory. Then, according to the above formula, the cumulative energy ratio is mapped to the SOC increment to obtain the result from... to The SOC trajectory is thus determined. This ensures that the SOC trajectory is consistent with the actual energy input, avoiding deviations caused by simply interpolating along a linear time path.

[0057] Finally, the cloud server outputs the vehicle-level charging power trajectory corresponding to the target charging session. and SOC trajectory The output results include at least a timestamp, vehicle-level charging power, cumulative delivered power, and SOC status, and serve as the basic input for calculating the vehicle-level discrete module requirements in subsequent step S2. The implementation process is as follows: Figure 2 As shown.

[0058] S2. Modular DC fast charging stations are typically equipped with several power modules of the same rated power. Let the rated power of a single power module be... For any scheduling time Read vehicle Instantaneous charging power The minimum number of power modules required for the vehicle is calculated based on the rated power of a single power module. Since power modules can only be used in integer quantities, the vehicle-level discrete module requirements are determined by rounding up. For example, when the rated power of a single power module... At that time, if a vehicle is at a certain time If one vehicle has a charging power of 42 kW, then that vehicle will require at least 2 power modules; if another vehicle has a charging power of 65 kW at the same time, then that vehicle will require at least 3 power modules. This allows us to convert continuous vehicle-level charging power requirements into integer vehicle-level module requirements.

[0059] At the same scheduling time Count the number of all vehicles currently charging. The discrete module requirements of each vehicle are then summed to obtain the actual module requirements at the station level. The method of summing after discretization at the vehicle level avoids systematic underestimation that occurs when directly estimating module requirements from station-level aggregated power. (See...) Figure 3 For example, when the charging power of the three vehicles at the same time is 42 kW, 61 kW and 18 kW respectively, and At that time, the vehicle-level discrete module requirements are 2, 3, and 1 respectively, therefore the actual station-level module requirement is 6. If we first aggregate the power of the three vehicles to 121 kW, then divide by 30 kW and round up, we only get 5 power modules, and this error will be amplified during peak hours. The error formula is shown below: This step calculates the actual module requirements at the station level by using a "vehicle-level discretization and station-level aggregation" approach. This approach can more accurately preserve the discretization of vehicle power requirements and module allocation constraints, providing a foundation for subsequent prediction of the number of backup energized modules, redundancy control, and early activation control.

[0060] S3. Construction and prediction of short-term demand forecasting model for station-level modules, specifically including the following steps: (1) Construct a heterogeneous input feature space. For each scheduling time... Construct an input feature vector consisting of station operation status, vehicle arrival status, vehicle charging characteristics, time characteristics, electricity price characteristics, and meteorological characteristics. In this embodiment, the input features may include the contents shown in Table 1.

[0061] Table 1. Model Input Features (List Only) The above features together constitute the moment. Input feature vector: in, Indicates the station-level charging power. This represents the actual module requirements at the station level. This indicates the number of vehicles currently charging. This indicates the number of charging requests entering the station. This indicates the initial SOC distribution characteristics of the vehicle. This indicates the vehicle power rating or model distribution characteristics. Indicates time characteristics, Indicates electricity price signal, Indicates meteorological characteristics, This indicates site attributes or module configuration characteristics. Electricity price and site attributes are not considered time-varying features in the scheme and are not used as time-series inputs.

[0062] (2) Constructing model training samples. In this embodiment, the scheduling time interval of the modular DC fast charging station is set to be... For example, it is advisable For each scheduling moment , based on the continuous time before that moment Historical operational data from each scheduling interval is used as model input, with future... The actual module demand at the station level for each scheduling interval is used as the prediction target. The input samples and output labels can be represented as: , .in, Indicates time The corresponding historical input sequence, Indicates time eigenvectors, Indicates the future predicted label sequence, Indicates the future number The actual module requirements at the station level for each scheduling period. See the detailed predictive control framework. Figure 4 .

[0063] For example, when the scheduling interval is 20 minutes, the history review window At that time, the model uses the running data from the past 24 hours as input. When the prediction step size... At that time, the model outputs the module demand prediction results for one scheduling point every 20 minutes within the next 2 hours.

[0064] Historical operational features are extracted using a time series prediction model. In one implementation, the prediction model employs a time series Transformer model based on time segment encoding. The model uses a time series encoding of length... Historical input sequences (charging information) are sorted by fragment length. and sliding step size Divide the time into several time segments, and map each time segment to a linear layer. A three-dimensional latent space. Each fragment embedding is superimposed with learnable positional and temporal embeddings, where the temporal embeddings include hour, day of the week, and date type embeddings. The number of fragments can be represented as: The insufficient portion at the end can be filled using the feature value of the last time point. For each time segment embedding, learnable location embeddings and time embeddings are superimposed, where the time embeddings include hour embeddings, day-of-week embeddings, and date type embeddings. The segment representation input to the Transformer encoder can be represented as: in, Represents a segment linear mapping function, This indicates a learnable positional embedding. Indicates time embedding, This represents the temporal feature representation obtained by encoding the historical execution sequence. This represents the Transformer time series encoding function.

[0065] In one implementation, the fragment length The number of scheduling intervals can range from 6 to 48, with 24 scheduling intervals being preferred; sliding step size The number of scheduling intervals can range from 3 to 24, with 12 scheduling intervals being preferred; implicit dimension The number of attention heads can be 64 to 512, preferably 256; the number of encoder layers can be 2 to 16, preferably 8; the number of feedforward network dimensions can be 1 to 8, preferably 4; the feedforward network dimension can be 64 to 1024, preferably 256; the dropout coefficient can be 0 to 0.5, preferably 0.3; and the activation function is GELU.

[0066] (3) Obtaining the temporal characteristics of the historical operation sequence Then, exogenous variables such as site attributes and electricity price signals are used to construct an exogenous variable vector. The exogenous variable representation is obtained by projecting it onto the same latent space as the temporal feature representation through a multilayer perceptron: in, This represents the projection function of exogenous variables. Subsequently, a cross-attention and gated residual structure are used to fuse the exogenous variable representation with historical operational characteristics to characterize the relationship between module requirements and external operating conditions. This fusion process can be expressed as: in, Represents the cross-attention function. Indicates the gating coefficient. This represents the Sigmoid function. This represents element-wise multiplication. Representation layer normalization.

[0067] (4) Considering that the station-level module requirements are non-negative integer variables, the prediction model outputs the probability distribution parameters of the module requirements for each future prediction step. In one implementation, the fused prediction features are represented as... After flattening, it is compared with the vector of exogenous variables. The layers are concatenated and fed into a multi-layer output network. The output network consists of at least one fully connected layer and a non-linear activation layer.

[0068] In one implementation, the output network sequentially comprises a 512-dimensional fully connected layer, a GELU activation layer, a dropout layer, a 256-dimensional fully connected layer, a GELU activation layer, and an output layer. The output layer has the following dimensions: Each corresponds to the future The prediction mean parameter and dispersion parameter are calculated for each prediction step size. To ensure the parameters are non-negative, the Softplus function is used for transformation. in, and These represent the output layer of the prediction model for the future... The original output values ​​of the mean parameter and the original output values ​​of the dispersion parameter generated for each scheduling period are transformed by the Softplus function to obtain a non-negative mean parameter. and dispersion parameter . Indicates the future number Average demand forecast for each scheduling period Indicates the future number The dispersion parameter for each scheduling period, To prevent the dispersion parameter from being a positive number of zero, in one implementation, we take... .

[0069] (5) Considering that the station-level module demand has non-negative discreteness and volatility, a negative binomial distribution is used to describe the random variable of future module demand: in, Indicates the future number The station-level module requirements for each scheduling period are random variables. This represents the predicted mean. This represents the degree of dispersion. The corresponding probability mass function can be written as: in, Indicates the possible number of modules required. This represents the gamma function. During model training, the actual station-level module requirements obtained in step S2 are used. As a supervisory signal, by minimizing the negative log-likelihood of the observation module requirement under the prediction distribution, and combining this with absolute error constraints, the accuracy of probabilistic prediction and the stability of point prediction are improved. The training objective can be expressed as: in, This represents the model training loss function. Indicates the number of prediction steps. Indicates the weight of the absolute error term. The value can range from 0.1 to 10, with 3 being the preferred value.

[0070] In one implementation, the model is trained using the AdamW optimizer, with a learning rate of [value missing]. to Preferred The batch size can be 16 to 256, preferably 64; the number of training rounds can be 5 to 100, preferably 20; the upper limit of the gradient norm can be 0.1 to 5, preferably 1.0.

[0071] (6) After the model training is completed, for the current scheduling time The model outputs the future. The predicted mean for each scheduling period Discreteness parameter And predict the cumulative distribution function: in, Indicates the future number The cumulative distribution function of the station-level module demand forecast for each scheduling period is used to extract the predicted demand quantile under risk constraints in step S5.

[0072] Simultaneously, point prediction results can be generated based on the predicted mean: in, Indicates the future number Demand point prediction results for station-level modules in each scheduling period This indicates the number of power modules installed or available at the site. The point prediction results are used in step S4 to perform historical rolling backtesting and construct an hourly-level underestimated residual distribution.

[0073] To illustrate the prediction output format, Table 2 provides an example of point prediction results generated from the prediction mean. In the probabilistic prediction implementation, the model simultaneously outputs the prediction distribution parameters corresponding to each future scheduling time. See [link to specific prediction results] for details. Figure 5 .

[0074] Table 2. Demand Forecast Sequence for Station-Level Modules (Partial List) S4. Construction of module demand forecast residual distribution and calculation of reliability correction amount, specifically including the following steps: (1) After the short-term prediction model for station-level module demand described in step S3 has been trained, a period of historical operating data (more than 30 days) of the modular DC fast charging station is selected as backtesting data. The cloud server performs rolling backtesting according to the same scheduling time interval, historical review window length and prediction step size as the online scheduling.

[0075] For each historical scheduling moment The model is based on Previous consecutive Using historical operation data of the first scheduling interval as input, the output will be the next scheduling interval. Station-level module demand prediction results for each scheduling period Subsequently, the predicted results were compared with the actual station-level module requirements for the same site and the same target scheduling period. Perform time alignment to form a historical backtesting sample set: in, Represents the historical backtesting sample set. This represents the set of historical scheduling moments that participated in the backtesting. Indicating a historical moment For the future Predicted module demand points for each scheduling period. This represents the actual station-level module requirements for the corresponding target scheduling period. Indicates the hour to which the target scheduling period belongs.

[0076] The point prediction result can be obtained by rounding up the prediction mean output in step S3 and applying the upper limit constraint on the number of site modules: (2) For each historical backtesting sample, calculate the module's required prediction residual: in, Representing historical moments For the future The module demand residual of the forecast results for each scheduling period. When When the predicted module demand is lower than the actual module demand, there is a risk of insufficient standby power modules; when When the forecast module's demand is not lower than the actual demand, it does not directly create a risk of underestimating demand.

[0077] To characterize the module insufficiency risk caused by prediction underestimation, a non-negative underestimation residual is further constructed: in, This indicates that demand has underestimated the residual. When the predicted value is lower than the actual value, This equals the number of modules underestimated in the prediction; when the predicted value is not lower than the actual value... The historical backtesting samples are shown in Table 3.

[0078] Table 3 Historical backtesting samples (partial list) (3) After obtaining the underestimated residuals, the cloud server groups the data according to the hour of the target scheduling period, rather than according to the hour of the prediction initiation time. For each hour... By summarizing the underestimated residuals of all target scheduling periods belonging to that hour, we obtain an hourly set of underestimated residual samples: Then, the empirical cumulative distribution function of the underestimated residual corresponding to that hour is constructed to obtain the empirical distribution of the underestimated residual at the 24-hour level.

[0079] This step characterizes the risk of insufficient standby power modules due to forecast underestimation at different intraday time periods. Because vehicle arrival intensity, site load levels, and forecast error characteristics vary across different hours, this embodiment does not use a uniform all-time residual distribution; instead, it constructs empirical distributions of underestimation residuals for each hour. The backtesting process is detailed below. Figure 6 In one instance, the effect of the 24-hour residual distribution is shown in [the image / document]. Figure 7 The titles of each subplot in the figure correspond to the hour of the target scheduling period. The horizontal axis represents the magnitude of the prediction residual, in units of module number; the vertical axis represents the frequency. For ease of comparison, all subplots use a uniform scale range. It can be seen that the residual distribution differs significantly across different time periods; therefore, this paper constructs empirical residual distributions for each hour. It should be noted that the prediction model is a multi-step prediction model, and theoretically, the residual distribution may also be related to the prediction lead time. To simplify modeling and ensure sufficient sample size, this paper does not currently model different lead times separately, but recommends limiting the multi-step prediction range to within 1 hour, i.e., no more than 3 steps within a 20-minute scheduling interval.

[0080] The cloud server uses the hourly-level underestimated residual empirical distribution as the output of step S4 and sends it to step S5 for use. Step S5 extracts the quantiles from the underestimated residual empirical distribution of the corresponding hour as the residual correction amount, based on the hour to which the future prediction period belongs and the preset risk tolerance parameter.

[0081] S5. Determining the number of baseline standby power modules and correcting for look-ahead activation, specifically including the following steps: (1) Obtain the predicted demand distribution and the hourly underestimated residual distribution. For the current scheduling time The cloud server reads the future output of step S3. Demand forecast results for station-level modules in the first scheduling period. For the future... For each scheduling period, the prediction result includes the prediction mean. Discreteness parameter And predict the cumulative distribution function. Meanwhile, the cloud server determines the timing based on future scheduling periods. Belonging Hour Call the hourly-level underestimated residual empirical distribution generated in step S4. Through the above processing, step S5 simultaneously obtains two types of information for each future forecast period: one is the uncertainty of module demand forecast given in step S3, and the other is the historical underestimation error characteristics of that hour given in step S4.

[0082] (2) Determine the baseline standby power module quantity under risk constraints. To simultaneously consider both predicted demand fluctuations and historical underestimation errors, the cloud server will predict the random variable of module demand. The corresponding hourly underestimation of the residual random variable Combining these, we obtain the corrected random variable for module requirements: in, Indicates the future number The module requirement random variable after residual correction for each scheduling period This represents the underestimated residual random variable characterized by the empirical distribution of the underestimated residual corresponding to the hour in step S4.

[0083] To obtain the cumulative distribution function of module requirements Module demand samples can be extracted or enumerated from the predicted demand distribution in step S3, and underestimated residual samples can be extracted or enumerated from the corresponding hourly underestimated residual empirical distribution in step S4. The two are added together to obtain the corrected module demand sample set, and then the corrected empirical distribution is constructed from this sample set. In one implementation, it is assumed that the short-term prediction error and the historically underestimated residual samples are independent of each other, and the corrected module demand distribution is obtained by random sampling or discrete convolution.

[0084] Let the risk tolerance parameter be... The service reliability level is For the future... Number of baseline standby power modules per scheduling period It should meet the following requirements: Therefore, the baseline number of standby power modules is taken as the smallest integer that satisfies the above constraints, and is also subject to the number of power modules already installed or available at the site. constraint: in, Indicates the current scheduling time Next, regarding the future... The number of baseline standby power modules determined for each scheduling period. This indicates the number of power modules installed or available at the site. See details below. Figure 7 .

[0085] For example, the number of power modules installed at a certain site Risk tolerance parameter When performing rolling scheduling at 10:00, if it is necessary to determine the baseline standby power module quantity for the next 10:20, the cloud server reads the 10:20 module demand prediction distribution output in step S3 and calls the underestimated residual empirical distribution corresponding to the 10:00 hourly segment in step S4. Combining the two yields the corrected module demand distribution. If the 95th quantile of this corrected distribution is 10 modules, then the baseline standby power module quantity for the next 10:20 is 10. If this quantile exceeds... Then take This serves as the baseline for the number of standby power modules.

[0086] (3) Adjust the current scheduling target based on the look-ahead activation window. Considering the response time when power modules switch from dormant to energized standby state, if control is only based on the baseline number of standby power modules in the current scheduling period, module activation lag may occur when demand rises rapidly. Therefore, at the current scheduling time... Set the length to The lookup activation window is then used to search for the maximum baseline standby power module number within that window: in, Indicates the length of the lookup activation window. Indicates the future The maximum number of baseline standby power modules within a scheduling interval. If the scheduling interval is... ,and The forward activation window then covers module demand changes within the next hour starting from the current moment.

[0087] After obtaining the maximum baseline standby power module number within the look-ahead window, a look-ahead activation coefficient is introduced. The current scheduling target is then modified. in, This indicates the number of baseline standby power modules corresponding to the current scheduling interval. This indicates the current scheduling objective after forward-looking correction. This represents the look-ahead activation coefficient.

[0088] when When no look-ahead activation correction is performed, the current scheduling target is equal to the current baseline number of standby power modules; when When the current scheduling target is increased to the maximum baseline standby power module number within the future window; when At that time, the current scheduling target will proportionally move towards the maximum demand level within the future window. See details for specific effects. Figure 8 .

[0089] For example, if scheduling is performed at 22:00, and the baseline standby power module counts for the next four scheduling periods are 18, 20, 21, and 19 respectively, then... If the number of baseline standby power modules corresponding to the current scheduling interval is... And the prospective activation coefficient is taken as... ,but This result indicates that it is not necessary to immediately increase the number of standby power modules to 21, but some energized standby modules should be added in advance to bring the module state configuration closer to future demand peaks, thereby alleviating the module activation lag problem when short-term demand rises.

[0090] (4) Output the target number of standby power modules in the current scheduling interval.

[0091] Since the number of power modules must be a non-negative integer and cannot exceed the number of power modules already installed or available at the site, the cloud server rounds up and applies an upper limit constraint to the forward-corrected scheduling target to obtain the target number of power modules that should ultimately remain in a powered standby state within the current scheduling interval: in, This indicates the target number of standby power modules output during the current scheduling interval. This indicates rounding up to the nearest integer.

[0092] During actual station operation, step S5 is executed on a rolling basis according to the scheduling interval. At each new scheduling time, the cloud server rereads the future module demand prediction distribution output in step S3 and calls the corresponding hourly underestimated residual empirical distribution generated in step S4 to recalculate the baseline standby power module quantity, forward correction target, and current scheduling control target within the future prediction window. This forms a standby power module quantity control sequence that is dynamically updated according to vehicle arrival, charging power changes, prediction results, and historical underestimated residual characteristics.

[0093] The control objective Used in step S6 to send to the station controller to guide the power module to switch between working, powered standby and sleep states.

[0094] S6. The target for controlling the number of standby power modules is issued and executed at the station, specifically including the following steps: (1) Issue the control target for the current scheduling interval.

[0095] The cloud server reads the current scheduling interval standby power module quantity control target output in step S5. The control objectives are then sent to the station-side controller of the corresponding modular DC fast charging station. The control objectives include at least the station number, scheduling time, control activation time, control duration, and the number of standby power modules. One or more of the following: risk tolerance parameter identifier, forward activation window identifier, and anomaly protection identifier.

[0096] in, This indicates the number of power modules that should remain in a powered standby state within the current scheduling interval. The standby power modules are those that are in standby mode, not currently outputting charging power to the vehicles, but can quickly participate in power allocation after receiving scheduling instructions.

[0097] (2) Read the real-time status of the power modules in the station.

[0098] After receiving the control target, the station-side controller reads the current operating status of the power modules within the station and the real-time charging demand of the vehicles. The operating status of the power modules includes at least one or more of the following: working status, powered standby status, sleep status, fault status, maintenance status, and communication abnormal status.

[0099] The station-side controller calculates the current set of schedulable power modules based on the operating status, and determines the number of working power modules currently participating in charging output, the number of standby power modules, and the number of dormant power modules. Power modules currently participating in charging output are given priority to remain in working status; power modules in fault, maintenance, or communication abnormality states are not included in the set of schedulable power modules.

[0100] (3) Generate and execute power module state switching instructions.

[0101] The station-side controller will compare the current number of standby power modules with the control target issued in step S5. The system compares the current vehicle charging requirements with the number of power modules currently in operation to generate corresponding power module state switching instructions.

[0102] The current number of standby power modules is lower than the control target. At this time, the station controller selects a schedulable module from the dormant power modules and generates an early activation command to switch the corresponding power module from dormant state to standby state.

[0103] The current number of standby power modules is higher than the control target. When there is no increase in vehicle charging demand or abnormal protection conditions, the station controller can select some idle standby power modules to enter a sleep state in order to reduce standby power consumption.

[0104] When the real-time charging demand of a vehicle exceeds the supply capacity of the currently operating power module, the station controller prioritizes the use of standby power modules for charging output, switching them from standby to operating state. The state switching command can be executed through the charging management system, the power distribution unit, or a combination of both to control the power module's switching between operating, standby, and dormant states.

[0105] (4) Feedback on execution results and update them on a rolling basis.

[0106] After executing the state switching command, the station-side controller feeds back the actual execution results to the cloud server. The feedback results include one or more of the following: the actual number of standby power modules, the number of working power modules, the number of dormant power modules, the module state switching results, the number of abnormal modules, the reason for not executing the command, and the vehicle's real-time charging requirements.

[0107] After receiving the feedback result, the cloud server uses it as input data for subsequent scheduling time prediction and control updates. If the feedback result shows that the actual number of standby power modules matches the control target... If there are deviations, or abnormal situations such as communication anomalies, module failures, or a sudden increase in vehicle arrivals, the cloud server or station controller can trigger anomaly protection strategies to ensure that the number of power modules in the station is no less than the number required for the current vehicle charging demand.

[0108] During actual station operation, step S6 is executed on a rolling basis according to the scheduling time interval. At each new scheduling time, the cloud server issues new control commands based on the standby power module quantity control target regenerated in step S5. The station-side controller performs power module state switching based on the latest power module status in the station and the real-time charging demand of vehicles, thereby realizing dynamic control of the number of standby power modules in the modular DC fast charging station.

[0109] Terminology Explanation (1) Power module: The basic power unit used in modular DC fast charging stations to provide DC charging power; (2) Standby power module: A power module that is in a powered standby state, does not output power to the vehicle, but can participate in the charging power distribution under the dispatch command; (3) Working power module: The power module that is outputting power to the charging vehicle; (4) Sleeping power module: A power module that is in a low power consumption or unpowered state and needs to go through an activation process before it can participate in power distribution; (5) Scheduling interval: The time unit for updating the number of standby power modules control target by the cloud server or station controller, such as 10 minutes, 15 minutes, 20 minutes; (6) Baseline standby power module count: The target number of standby power modules calculated based on the distribution of future module demand forecasts, the distribution of historically underestimated residuals, and risk tolerance parameters before forward activation correction; (7) Proactive activation: Based on the predicted changes in module demand within a preset window, the control process of switching some dormant power modules to a powered standby state in advance before the peak demand is reached; (8) Risk tolerance parameter: The upper limit of the probability that module demand exceeds the control target, used to determine the quantile of forecast demand and the amount of residual correction.

Claims

1. A predictive energy-saving control method for standby modules in modular DC fast charging stations, characterized in that, Includes the following steps: S1. The station-side data acquisition unit acquires historical charging session records, vehicle information, and station module configurations, and reconstructs the vehicle-level charging power trajectory in the cloud; S2. Based on the rated power of the power module, convert the vehicle-level power trajectory into vehicle-level discrete module requirements, and summarize them to obtain the actual station-level module requirements; S3. Using the actual module demand at the station level as the prediction object, predict the probability distribution of module demand during future scheduling periods based on historical operation data; S4. Based on historical forecast results, construct the residual distribution between the predicted and actual values ​​of module demand to characterize the forecast bias and uncertainty; S5. Combining historical forecast residuals and risk tolerance parameters, determine the baseline number of standby power modules that meet service reliability constraints, and perform forward activation correction based on future demand changes within a preset window, outputting the target number of standby power modules for the current scheduling interval. S6. The cloud sends the module quantity control target to the station controller, which then controls the standby, hibernation, and activation status of the power modules in the station through the charging management system and power distribution unit.

2. The method according to claim 1, characterized in that, Step S1 includes the following steps: (1) Collect historical charging session records, vehicle information and station module configuration data through the data acquisition unit, charging management system or station controller set in the modular DC fast charging station, and upload them to the cloud server through the station communication unit; (2) The cloud server extracts the experience charging curves for typical vehicle models, battery capacities and charging power levels based on high-resolution historical charging records, and builds a reference charging curve library; For each reference trajectory, a reference charging power function is established with the vehicle's state of charge (SOC) as the independent variable. and cumulative charging time function It also stores the corresponding vehicle model identifier, battery capacity range, charging power level, SOC coverage range, and maximum allowable charging power; (3) Match the closest reference vehicle configuration from the reference curve library based on the vehicle model identifier, battery capacity and charging power level of the target charging session; (4) Set SOC boundary point As the boundary between the main charging phase and the final power reduction phase, and based on the initial state of charge of the target charging session. Termination of charging state With the SOC boundary point Based on the relationship, the target charging session is divided into stages; if If the target charging session crosses the SOC boundary point, it will be divided into... to The main charging phase and to The final power reduction stage; if If so, the entire target charging session is treated as the main charging phase; if Then the entire target charging session is treated as the final power reduction stage; When the target charging session crosses the SOC boundary point and only includes the initial state of charge and the final state of charge, it is assumed that the SOC changes linearly within the effective charging duration of the target charging session, and the target duration of the two stages is determined. If the target charging session includes SOC sampling data within a phase, then the actual SOC sampling sequence will be used to determine... The corresponding stage boundary time; Within each stage, the target SOC progress is mapped to the curve segment corresponding to the SOC interval in the reference charging curve, and then... Obtain the initial reconstructed power trajectory ; (5) Set a linear power ramp process during the initial ramp time; subsequently, adjust the power ramp based on the actual charging capacity of the target session. For the initial reconstructed power trajectory Scaling the whole thing up yields the following results: ; (6) Calculate the cumulative delivered energy based on the cumulative integral of the vehicle-level charging power trajectory over time. Based on cumulative delivered energy and actual delivered electricity Reconstructing the SOC trajectory from the initial state of charge and the final state of charge. The final output is the vehicle-level charging power trajectory. and corresponding SOC state This serves as the basic input for subsequent calculations of vehicle-level discrete module requirements.

3. The method according to claim 1, characterized in that, Step S2 discretizes the reconstructed vehicle-level charging power trajectory according to the scheduling time interval based on the rated power of the power modules configured at the station; calculates the minimum number of power modules required for each vehicle's instantaneous charging power within the same scheduling interval to obtain the vehicle-level discrete module requirements; and then summarizes the discrete module requirements of all charging vehicles within the same scheduling interval to obtain the station-level real module requirements.

4. The method according to claim 1, characterized in that, Step S3 includes the following steps: (1) Based on the station-level real module demand sequence, construct a heterogeneous input feature space for short-term module demand prediction; Input features include historical station-level charging power, historical station-level actual module demand, current number of charging vehicles, number of charging requests entering the station, initial SOC distribution of vehicles, vehicle power level or model distribution, periodic time features, weekday and holiday types, electricity price signals, and meteorological features. The above features collectively characterize the site load level, vehicle arrival status, vehicle charging differences, time cycle patterns, electricity price impact, and environmental impact. (2) Continuously collect the historical operation data of the above heterogeneous input feature space as the dynamic input of the prediction model, and divide the input historical sequence into preset segment lengths. and sliding step size Divided into several time segments, where the segment length is... Choose from 6 to 48 scheduling intervals, with a sliding step size Choose 3 to 24 scheduling intervals; perform linear mapping on each time segment to obtain the segment embedding representation, and superimpose the learnable position embedding and time embedding; (3) Construct an exogenous variable vector from the site attributes and electricity price signal variables. The exogenous variable representation is obtained by projecting it onto the same latent space as the temporal features through a multilayer perceptron. ; Subsequently, the correlation between historical time-series features and exogenous variable representations is calculated using a cross-attention mechanism, and the degree of modification of exogenous variables on the features of each time segment is controlled by a gated residual structure to obtain the fused predicted feature representation. ; (4) Represent the fused prediction features After flattening, it is compared with the exogenous variable vector. The data are concatenated and input into a multi-layer output network to obtain the probability distribution parameters of module requirements for each future prediction step. The multilayer output network includes at least one fully connected layer and a nonlinear activation layer; (5) Considering that the station-level module requirements are non-negative discrete variables, the prediction model outputs the probability distribution parameters of the module requirements for each future prediction step size; During model training, the actual station-level module requirements are used as supervision signals. By minimizing the negative log-likelihood of the observation module requirements under the prediction distribution and combining it with absolute error constraints, the accuracy of probabilistic prediction and the stability of point prediction are improved. The training objective is represented as: in, For a moment For the future The actual module requirements at the station level for each scheduling period. Indicates the future number The predicted average demand of power supply modules for each scheduling period express Substitute the probability value obtained by the negative binomial distribution probability mass function predicted by the model, and learn the most suitable distribution shape; This represents the weight of the absolute error term, ranging from 0.1 to 10. The AdamW optimizer is used during training, with a learning rate of [value missing]. to The batch size is 16 to 256, the number of training rounds is 5 to 100, and gradient clipping is used to limit the gradient norm, with the upper limit of the gradient norm being 0.1 to 5. (6) After the prediction is completed, the probability distribution, prediction mean and prediction quantile of the station-level module demand in the future scheduling period are obtained. The prediction mean or prediction distribution quantile is used as the input for subsequent standby power module scheduling. The prediction result is constrained by the upper limit according to the number of power modules installed at the station, so that the predicted module demand does not exceed the number of available power modules at the station, and provides a basis for determining the baseline number of standby power modules under risk constraints in step S5.

5. The method according to claim 4, characterized in that, The output network sequentially comprises a 512-dimensional fully connected layer, a GELU activation layer, a dropout layer, a 256-dimensional fully connected layer, a GELU activation layer, and an output layer. The output layer has a dimension of [missing value]. Each corresponds to the future The mean parameter and dispersion parameter of each prediction step are used; to ensure that the parameters are non-negative, the Softplus function is used to transform the output parameters.

6. The method according to claim 4, characterized in that, The uncertainty of module demand is represented by a negative binomial distribution. The model outputs mean parameters and dispersion parameters to obtain the probability distribution of station-level module demand during future scheduling periods.

7. The method according to claim 1, characterized in that, Step S4 includes the following steps: (1) Based on the prediction model in step S3, rolling backtesting is performed on historical operation data according to the same historical backtesting window length, prediction step size and scheduling interval as online scheduling, to obtain historical time points. For the future The module demand forecast results for each scheduling period are obtained, and the forecast results are time-aligned with the actual station-level module demand of the same site and the same scheduling period to construct historical backtesting samples. (2) Calculate the module demand prediction residuals in the historical backtest samples, and extract the positive residuals where the actual demand is higher than the predicted demand for service risks caused by the underestimation of demand in the standby power module scheduling, and construct the empirical distribution of the underestimation residuals.

8. The method according to claim 1, characterized in that, Step S5 includes the following steps: (1) At the current scheduling time Read the future output of step S3 The probability distribution of module demand forecast for the first scheduling period; for the future first... For each scheduling period, let its predicted cumulative distribution function be denoted as . ,in Indicates the possible number of modules required. According to preset risk tolerance parameters Extract the predicted demand quantile from the predicted cumulative distribution function as the upper bound of module demand under the uncertainty of the predicted distribution; (2) Read the hourly-level estimated residual empirical distribution output in step S4; extract the risk tolerance parameter from the accumulated empirical distribution function of the estimated residual corresponding to that hour. The corresponding residual correction amount; (3) Add the predicted demand quantiles to the hourly residual corrections to obtain the future demand quantiles. Reliability correction module requirements for each scheduling period ; Based on the number of power modules installed or available at the site Apply an upper limit constraint to obtain the baseline number of standby power modules; (4) Set the length to Forward activation window, in the future The maximum number of baseline standby power modules searched within a single scheduling period: in, , Indicates the current scheduling time The maximum baseline standby power number of modules within the next forward-looking window; (5) Based on the prospective activation coefficient The baseline target for the current scheduling period is revised in advance to obtain the current scheduling target after forward revision; (6) Round up and impose upper limit constraints on the current scheduling target after look-ahead correction to obtain the final output of the number of standby power modules in the current scheduling interval, which is then sent to the station controller to guide the switching of standby, hibernation and early activation states of the power modules in the station.

9. The method according to claim 1, characterized in that, Step S6 includes the following steps: (1) The cloud server sends the current scheduling interval standby power module number control target obtained in step S5 to the station-side controller of the corresponding modular DC fast charging station; the control target includes one or more of the following: station number, scheduling time, control effective time, target standby power module number, control effective duration and abnormal protection flag; (2) After receiving the control target, the station controller reads the current operating status of the power modules in the station and the real-time demand of the charging vehicles, and generates a module state switching instruction based on the difference between the current module state and the control target; through the charging management system and the power distribution unit, it controls the standby, hibernation and early activation states of the power modules in the station; the station controller feeds back the actual number of standby modules, the module state switching results and abnormal states to the cloud server for subsequent prediction and control updates.

Images