A charging load prediction method and device based on dynamic game and space-time state evolution

Abstract

The application relates to the technical field of charging station load prediction, and discloses a charging load prediction method and device based on dynamic game and space-time state evolution. The prediction time range and the physical state of a driver are discretized and space-time aligned, and influence factors such as electricity price, order, driver behavior and surrounding stations are parameterized. A simultaneous time recursion relationship of total electricity equivalent value, the number of drivers in a driving state and the number of drivers in a charging state is established under a discrete space-time framework, iteration is carried out along a discrete time axis until a preset stable condition is reached, and a prediction curve of the total power of a target charging station changing with time is formed according to the stable state. The application has a continuous and traceable calculation path between the change of driver behavior and the charging load of the station. The predicted power is simultaneously subjected to the constraints of driver number conservation, group electricity boundary and the upper limit of the available charging power of the station determined by the available capacity of a transformer, the rated power of charging equipment and the grid connection point limit.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle charging station load forecasting and power load data processing technology, specifically to a charging load forecasting method and device based on dynamic game theory and spatiotemporal state evolution. Background Technology

[0002] The daily load of electric vehicle charging stations is affected not only by the number of vehicles arriving at the station and the capacity of the charging equipment, but also by factors such as the listed electricity price, order demand from ride-hailing or commercial vehicles, driver work schedules, remaining battery power of vehicles, prices of surrounding charging stations, and congestion levels. When the price changes during a certain period, some vehicles that can delay charging will shift their charging demand to adjacent periods. Therefore, the charging load during the same period is not a simple one-to-one correspondence with the price during that period and subsequent periods.

[0003] For example, Chinese invention patent application CN109359389A discloses a method for urban electric vehicle charging decision-making based on dynamic game theory of typical loads. It obtains typical user load types and ideal charging load curves based on short-term grid load curves, and uses charging demand, charging time, number of electric vehicles charging, grid load, and distribution network capacity as boundary conditions. The error between the charging load curve and the ideal charging load curve, as well as user electricity costs, are used as optimization objectives. The method determines the electric vehicle charging strategy through dynamic game theory. While this document focuses on combining grid load constraints and user costs to form charging decisions, this application further focuses on the temporal state evolution of the target charging station driver group under the combined influence of factors such as price, orders, vehicle battery level, and willingness to rest, and the resulting station load forecast.

[0004] For example, Chinese invention patent application CN114725965A discloses a charging load prediction method based on active distribution network dynamic reconfiguration and active / reactive power joint optimization model. This method establishes the energy consumption per unit mile and state of charge of electric vehicles, considers the owner's subjective intentions to establish a spatiotemporal distribution prediction model for electric vehicle charging load, and constructs an active distribution network dynamic reconfiguration and active / reactive power joint optimization model based on this model. While this document already considers subjective factors of charging behavior and the spatiotemporal distribution of charging load, its focus is on the spatiotemporal prediction of charging load and the joint optimization of the distribution network. The technical path adopted in this application—discrete driver state, equivalent total energy value, simultaneous recursion of driving and charging states, and iteration along the time axis to a stable state—is different from this method.

[0005] Existing load forecasting can employ historical curve fitting, price elasticity coefficients, or machine learning models. For scenarios requiring the evaluation of new listed price curves, relying solely on static fitting of historical loads is insufficient to clearly describe the transition processes between driver states such as driving, charging, and resting. Furthermore, if only abstract behavioral factors or related functions are provided without explanations of factor quantification, state recursion, and the solution methods for self-referencing variables, it is difficult to guarantee that the model can be repeatedly implemented based on actual station data.

[0006] Especially when the entire day's price curve is used in the forecast, the willingness to charge in the current period may be simultaneously affected by electricity prices in subsequent periods, drivers' desire to rest, and the availability of surrounding charging stations. There is also a continuous dependency between different time units caused by changes in driver status and vehicle battery level. Therefore, it is necessary to establish a charging load forecasting method with a clear time discretization, state division, parameterization of influencing factors, state recursion, and parameter calibration process, so that changes in driver behavior can be transmitted along the time axis and ultimately correspond to the total charging power curve of the target charging station. Summary of the Invention

[0007] The purpose of this invention is to provide a charging load prediction method and device based on dynamic game theory and spatiotemporal state evolution. By parameterizing price, order, driver behavior and surrounding station status within a unified discrete spatiotemporal framework, and establishing a time-recursive relationship between the total power equivalent value, driving status and charging status, the charging demand transfer process of the driver group can be continuously calculated along the time axis and a charging load prediction curve of the target charging station can be formed.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A charging load prediction method based on dynamic game theory and spatiotemporal state evolution includes: dividing a continuous prediction time range into discrete time units; obtaining discrete spatial state units based on continuous physical state variables; spatiotemporally aligning the discrete time units and discrete spatial state units to obtain a discrete spatiotemporal framework; explicitly parameterizing the factors affecting charging decisions to obtain multi-source influence factor parameters; establishing a joint time recursive relationship based on the multi-source influence factor parameters within the discrete spatiotemporal framework, wherein the joint time recursive relationship describes the joint evolution process of at least two different types of driver states within the discrete spatiotemporal framework; iteratively calculating the joint time recursive relationship on the discrete time axis until a preset stability condition is reached; extracting the curve of the total charging power of the target charging station changing with time from the stable state to obtain the charging load prediction result.

[0009] Furthermore, the dynamic game referred to in this application uses driver groups in different discrete spatial states as participants, and driving, charging, and resting as mutually exclusive state strategies. The payoff of the driving strategy is represented by the response value of the nonlinear response function of the driving state; the payoff of the charging strategy is represented by the charging state response quantity jointly determined by the comprehensive charging response coefficient and the remaining number of drivers who can participate in charging; and the payoff of the resting strategy is represented by the remaining number of drivers after satisfying the conservation of the number of drivers. Within the same time unit, the self-referenced fixed point of the number of drivers in the charging state is taken as the strategy response equilibrium of this time unit; state feedback is formed between adjacent time units through the equivalent value of total electricity and spatial state migration, and the cross-time stable equilibrium is reached when the preset stability condition is met for the entire prediction cycle. The above-mentioned participants, strategies, payoff representations, and equilibria are all determined by the existing state quantities and influencing factors in the joint time recursive relationship, without setting up independent game solution chains that are detached from the aforementioned recursive relationship.

[0010] Furthermore, the combined time-based recursive relationship is constrained by the conservation of the number of active drivers, the boundary of the total energy consumption of the group, and the upper limit of the available charging power of the target charging station. The aggregated charging power is obtained based on the number of drivers in each time unit's charging status and the average charging power per vehicle, and is further constrained by the upper limit of the station's available charging power, determined jointly by the station's transformer capacity, the rated power of available charging equipment, and grid connection point limitations. Therefore, price, order, and driver status parameters only serve as inputs in the recursive calculation of the station's physical load, and the final predicted quantity is the active charging load constrained by equipment capacity and energy consumption conservation.

[0011] Furthermore, the driver's desire to rest is defined as a dimensionless parameter based on observable business status statistics such as the offline rate, rest rate, or platform online rate within the same historical time unit; charging time preference is defined as a status label based on the distribution of historical effective charging times or effective charging volume in peak and off-peak price ranges. These parameters and status labels are used to distinguish the load response of different driver groups, and psychological assessment results are not used as model input.

[0012] Furthermore, for gradient descent dynamic calibration based on historical data, the passenger order quantity and the number of active drivers in the surrounding area during the prediction phase are used as time-series input data and are not directly used as training parameters. Dynamic calibration of the passenger order quantity curve and the total number of active drivers in the surrounding area refers to calibrating the statistical prediction parameters or correction coefficients used to form their prediction inputs. The actual updated training parameters include at least one of the following: driver rest desire baseline value, driving status nonlinear response parameter, active driver correction coefficient, rest desire index parameter, price influence parameter, and related function weights.

[0013] Furthermore, the comprehensive product term in the rest desire index factor is used to represent the ratio-type relative activity index of the driver's activity level in the current time period relative to the geometric mean of the driver's activity level in other time periods; the name of the comprehensive product term is used to maintain the terminology connection with the multiplicative combination relationship of multiple factors, without changing the calculation meaning of the index as a ratio-type relative quantity in terms of numerical value.

[0014] Optionally, establishing the joint time recursive relationship includes: based on the multi-source influence factor parameters, establishing a total energy equivalent value recursive relationship, a driver quantity relationship in the driving state, and a driver quantity relationship in the charging state within the discrete spatiotemporal framework; adding the driver quantity relationship in the driving state and the driver quantity relationship in the charging state to the total energy equivalent value recursive relationship to obtain the joint time recursive relationship.

[0015] Optionally, the total energy equivalent value recursive relationship is updated based on the total energy equivalent value at the initial time and the driver status data at the initial time. First, the energy consumption corresponding to the driver in the driving state is subtracted, and then the energy replenishment corresponding to the driver in the charging state is added to obtain the total energy equivalent value at the next time. Among them, the energy status of the driver in the driving state is transformed by the energy status expected conversion rate coefficient.

[0016] Optionally, the number of drivers in driving status is represented by subtracting the drivers' desire to rest in the current time period from 1 to indicate the driver's activity level. The target parameter is formed by combining the current total battery equivalent value and the number of passenger orders. The target parameter is transformed by a nonlinear response function with a positive first derivative, a negative second derivative, and an asymptote to obtain the number of drivers in driving status.

[0017] Optionally, the relationship between the number of drivers in the charging state is obtained by establishing a rest desire index factor based on 1 minus the current value of the drivers' rest desire in the current time period, a function term related to the difference between the total number of active drivers in the surrounding area and the number of drivers charging at present, and a correlation function corresponding to multiple factors affecting the charging amount in the current time period, and multiplying the rest desire index factor, the function term and the correlation function together.

[0018] Optionally, the parameters of the driver's desire to rest, the curve of the number of passenger orders, the total number of active drivers in the vicinity, and the internal parameters of the function terms and related functions are all obtained by dynamic calibration through gradient descent based on historical data.

[0019] Optionally, the plurality of factors affecting the charging amount in the current time period include a first factor and a second factor; the first factor is formed based on the fact that the charging amount in the current time period is negatively correlated with the electricity price level in the current time period and positively correlated with the charging electricity price at multiple times adjacent to the current time period; the second factor is formed based on the number of drivers with insufficient battery power as the maximum possible value of the number of charging drivers.

[0020] Optionally, the establishment of the rest desire index factor includes: obtaining 1 minus the current value of the driver's rest desire in the current time period and using it as the numerator; obtaining 1 minus the reference values ​​of the driver's rest desire at all other times; taking the geometric mean of the reference values ​​and using it as the denominator to obtain a comprehensive product term; performing an exponential operation on the comprehensive product term to obtain the rest desire index factor; wherein, the exponent is a trainable parameter used to control the influence of the comprehensive product term on the charging amount.

[0021] Optionally, the continuous physical state variables include the driver's battery status and charging time preference; the driver's battery status is divided into low battery status, medium battery status and high battery status, and the charging time preference is divided into peak preference and off-peak preference, and the combination of the driver's battery status and charging time preference forms six types of discrete spatial state units.

[0022] This invention also provides a charging load prediction device based on dynamic game theory and spatiotemporal state evolution, used to execute the charging load prediction method described in any of the preceding claims, comprising: a spatiotemporal discretization module for establishing the discrete spatiotemporal framework; a parameter definition module for explicitly parameterizing the factors affecting charging decisions to obtain the multi-source influence factor parameters; a recursive equation establishment module for establishing the simultaneous time recursive relationship based on the multi-source influence factor parameters within the discrete spatiotemporal framework; and an iterative prediction module for iteratively calculating the simultaneous time recursive relationship on the discrete time axis until a preset stability condition is reached, and extracting the curve of the total charging power of the target charging station changing with time from the stable state to obtain the charging load prediction result.

[0023] The present invention has achieved the following beneficial effects: This invention establishes a unified spatiotemporal framework by discretizing time and spatially discretizing driver battery status and charging time preferences. It incorporates price, orders, driver rest desire, and surrounding station status into an interrelated recursive process of driver behavior, enabling the driver status and group battery level of the current time unit to serve as the basis for calculations in subsequent time units. By dynamically calibrating behavioral parameters using historical data and generating a station charging total power curve based on the number of charging drivers under stable conditions, the invention allows the transmission relationship between electricity price changes and driver behavior across different time periods to correspond to specific load forecasting results.

[0024] Furthermore, by taking the conservation of driver number, changes in group power consumption, and the upper limit of available charging power at the depot as the physical boundaries of the behavioral response results, and by converting the number of charging drivers in each spatial state into the active power of the depot through the average single-vehicle charging power, this invention forms a closed technical chain between influencing factor input, state response, energy recursion, equipment capacity constraints, and load output. Attached Figure Description

[0025] Figure 1 This is a flowchart of the charging load prediction method based on dynamic game theory and spatiotemporal state evolution in an embodiment of the present invention. Figure 2 This is a schematic diagram of the charging load prediction device based on dynamic game theory and spatiotemporal state evolution in an embodiment of the present invention. Detailed Implementation

[0026] The embodiments of the present invention will now be described with reference to the accompanying drawings. The following embodiments are presented sequentially according to the progressive relationship of the technical solutions defined in the claims, and the same term has the same meaning in different embodiments. In this text, ride-hailing or commercial electric vehicle drivers are used as the subjects of charging demand; for other types of electric vehicles, the driver can be replaced with the vehicle or charging user, and the same data acquisition and status recursion method can be used. Unless otherwise specified, the prediction object is the active charging load of the target charging station within a prediction period.

[0027] Example 1: As Figure 1 As shown, this embodiment provides a charging load prediction method based on dynamic game theory and spatiotemporal state evolution, including establishing a discrete spatiotemporal framework, defining multi-source influencing factor parameters, establishing a simultaneous time recursive relationship, and iterating along the discrete time axis to obtain the charging load prediction result.

[0028] First, the continuous prediction time range is discretized. In this embodiment, a 24-hour prediction period is used, and it is divided into 48 discrete time units of 30 minutes each. The time index is denoted as... The time granularity is consistent with the power station's listed electricity price release cycle, order statistics cycle, or load sampling cycle; when the power station uses a 15-minute granularity, a forecast cycle can also be divided into 96 discrete time units. All input data are aggregated or resampled according to a unified time zone, unified time boundary, and unified time granularity.

[0029] Secondly, the continuous physical state is spatially discretized. The physical state variables used to form discrete spatial state units include at least the driver's or vehicle's battery status and charging time preference. Battery status reflects the vehicle's remaining battery power or SOC range, while charging time preference reflects whether the driver prefers to charge during peak or off-peak hours in historical driving patterns. Discrete spatial states are not geographical coordinates, but rather used to describe groups of drivers with different charging behavior characteristics. Each discrete time unit corresponds to a set of spatial state distributions, thus forming a discrete spatiotemporal framework indexed by both time units and spatial state units.

[0030] The prediction input must include at least the historical measured charging power of the target charging station and the listed electricity price for the period to be predicted. Depending on the data available at the charging station, it may also include passenger order volume or order density, number of online drivers on the platform, driver rest or offline ratio, vehicle remaining battery power, listed prices of surrounding charging stations, charging equipment utilization saturation of surrounding charging stations, weather, traffic, and holiday information. Each data record retains the data source and collection time, ensuring the same field definitions are used in both the training and prediction phases.

[0031] Data that may contain bicycle or driver identifiers, such as platform trajectories, online positioning, and historical charging records, may only be used if legally obtained and processed with the appropriate authority. Before being fed into the prediction model, data should be aggregated according to service area, discrete time unit, and discrete spatial state unit. When it is necessary to associate the same vehicle across cycles, irreversible hashing or randomized internal identifiers can be used instead of direct identification identifiers. Training parameters and prediction results should not retain direct identification information unrelated to load prediction, such as names and phone numbers; only data fields required for spatiotemporal alignment, state statistics, and power aggregation should be retained.

[0032] Multi-source influencing factor parameters are used to incorporate the aforementioned external incentives, competitive environment, and driver decision-making factors into subsequent state recursion. External incentive parameters may include passenger order density; competitive environment parameters may include the price difference between the target station and surrounding stations and the utilization saturation of surrounding stations; and driver decision-making parameters may include price sensitivity and desire to rest. Specific quantification methods for these parameters will be further explained in subsequent embodiments.

[0033] The desire to rest and charging time preference are both derived from observable operational data. The desire to rest can be mapped from the historical offline rate, rest rate, or platform online rate within the same time unit to a dimensionless state quantity ranging from zero to one. Charging time preference is determined by the number of effective charges or the percentage of effective charging volume for the same driver or similar drivers within peak and off-peak price ranges. These quantities describe the operational status of a repeatable group and do not require the collection of subjective psychological evaluations from drivers.

[0034] After establishing a simultaneous time recursive relationship within a discrete-time framework, the driver state distribution and the equivalent total energy consumption of the first time unit of the prediction cycle are used as initial conditions. The driver behavior state and total energy consumption changes are then calculated sequentially for each time unit. The calculation result of the current time unit serves as the input for the next time unit, continuing until a prediction cycle is completed. When using periodic steady-state iteration, the state at the end of the previous cycle is used as the initial reference for the next cycle calculation, repeating the calculation until the preset stability condition is met. Once the stability condition is reached, the number of drivers in charging mode in each time unit is extracted and converted into the total charging power of the target charging station, forming a charging load prediction curve in chronological order.

[0035] From the perspective of dynamic game theory, driver groups in different spatial states constitute participating units, and driving, charging, and resting constitute mutually exclusive state strategies. The nonlinear response function of the driving state gives the driving state response quantity, the comprehensive charging response coefficient and the remaining number of drivers available for charging give the charging state response quantity, and the resting state is supplemented by the conservation of the number of drivers. The charging state within the same time unit refers to a fixed point and constitutes the stable response of that time unit. The total power and spatial state migration of adjacent time units constitute cross-time feedback, and the complete prediction cycle iterates to the stable condition to constitute a stable result on a periodic scale.

[0036] The technical advantage of this embodiment is that it unifies continuous time, driver physical state, and influencing factors such as price, orders, and surrounding stations into a discrete spatiotemporal framework that can perform numerical recursion, and enables the current state to be continuously transmitted along the time axis. This solves the technical problem that it is difficult to describe the cross-time period migration of driver behavior as electricity price, orders, and vehicle power change when static fitting is performed based solely on historical load curves.

[0037] Example 2: Building upon Example 1, this example further illustrates the combined time recursive relationships. These combined time recursive relationships include at least the total energy equivalent value recursive relationship, the number of drivers in the driving state, and the number of drivers in the charging state. These three relationships are connected according to the input-output order within the same time unit, enabling changes in driver behavior to be fed back to the group energy status in the next time unit.

[0038] Set time unit The initial total number of active drivers is The number of drivers in driving status is The number of drivers in the charging state is The number of drivers in rest mode is The three states are counted using a mutually exclusive approach, with each driver only counted in one state at a time. To ensure driver conservation, after calculating the number of drivers in driving and charging states, the number of drivers in resting state is determined using the following formula: ; in, Representing time units The number of drivers in a resting state. (When calculated independently) Exceed At that time, priority should be given to retaining facilities that are already subject to charging capacity constraints at the charging stations. Then Cut off to This ensures that the number of drivers is always kept constant.

[0039] The total energy equivalent value is used to represent the total remaining energy of the driver group participating in the prediction at the boundary of the time unit. Driving causes a decrease in the total energy of the group, while charging causes an increase. Therefore, the total energy equivalent value of the subsequent time unit is jointly determined by the total energy of the previous time unit, the energy consumed during driving, and the energy replenished during charging. Its specific dimensions and parameter determination methods are explained in Example 3.

[0040] The driver quantity relationship in the driving state is used to convert driver activity level, current total battery level, and passenger order demand into the actual number of drivers in driving mode. The driver quantity relationship in the charging state is used to convert factors such as driver rest desire, the current number of available drivers for charging, prices at target stations and surrounding stations, and environmental factors into the actual number of drivers charging. These two behavioral states are not isolated from each other: drivers entering the charging state reduce the number of available drivers who can continue driving or participate in subsequent charging decisions, while the driving state, in turn, changes the charging demand in the next time unit through vehicle battery consumption.

[0041] During the calculation of a time unit, the current number of active drivers and the equivalent total battery power can be obtained first based on the state of the previous time unit. Then, the number of drivers in the charging state is calculated. Subsequently, the number of drivers in the driving state is calculated within the remaining driver range, and finally, the number of drivers in the resting state is obtained. Alternatively, the driving and charging relationships can be jointly updated while satisfying the quantity conservation constraint. Regardless of the calculation order used, the output is given at the end of the time unit. , , And the updated total energy equivalent value, which is used as input for the next time unit.

[0042] For different discrete spatial states, the number of driver states in each state unit can be calculated separately first, and then aggregated to obtain the result. , and Alternatively, a group average parameter can be formed first based on the proportion of people in each state unit, and then aggregated and recursively calculated. Both methods are based on the same discrete spatiotemporal framework and ultimately output the aggregated load of the target charging station.

[0043] The technical advantage of this embodiment is that it establishes a clear data transmission chain through three interrelated recursive relationships: the total power equivalent value, the driving status, and the charging status, and uses the rest status to supplement the conservation of the number of drivers. This solves the technical problem that when only independent behavior judgment rules are set, the number of drivers and the power of the group cannot be continuously transmitted, and the number of driving and charging may exceed the total number of actual active drivers.

[0044] Example 3: Based on Example 2, this example further illustrates the recursive relationship of the total energy equivalent value and the physical meaning of the expected energy state conversion rate coefficient. Let... Time unit The initial equivalent total battery capacity of all vehicles with all active drivers, in kWh. The number of drivers in the driving state for this time unit; Number of drivers in the charging state for this time unit; The average power consumption of a single driver in driving mode within a time unit, expressed in kWh / driver / time unit; For a single driver in charging state in time unit The average effective charging power, in kW; The time unit length is expressed in hours (h). This is the equivalent efficiency of the electrical energy output from the charging station to the effective electrical energy input to the vehicle. The total equivalent energy value can then be derived using the following formula: ; in, Representing time units The maximum equivalent available power for the corresponding group of mobile drivers. This indicates that the result is restricted to a given lower and upper bound. In this formula, and All have the dimension of kWh and can be compared with Simply add or subtract.

[0045] The desired conversion rate coefficient for the state of charge can be adopted. This indicates that it is not a dimensionless arbitrary coefficient, but rather an equivalent coefficient that converts the number of drivers in a driving state into the group's power consumption within a time unit. For scenarios where mileage data is available, the average power consumption per unit mileage is first calculated based on historical valid mileage. The unit is kWh / km. Then calculate the driver's average mileage in driving mode within a time unit. Then you can set: ; in, The statistics can be compiled separately for each vehicle type, and then weighted according to the proportion of each vehicle type among the driver groups. This can be obtained from ride-hailing platform tracks, odometer data, or the average mileage per unit time on similar historical days. When real-time data is unavailable... When this is the case, the historical sliding window average of the same week type and the same time unit can be used as the predicted value.

[0046] When the historical valid mileage sample is insufficient to stably obtain the average energy consumption per unit mileage, the rated energy consumption per 100 kilometers published by the vehicle manufacturer or the industry benchmark energy consumption per 100 kilometers for the applicable model can be used as a benchmark value. This benchmark value can be combined with the driver's average mileage driven within the time unit to calculate the expected energy conversion rate coefficient. The benchmark energy consumption is in kWh / 100km, and the average mileage is in km. The benchmark calibration can be performed using the following formula:

[0047] When vehicle model differences exist, a correction factor obtained from historical measured data based on vehicle model grouping can be applied. Once the historical measured average energy consumption per unit mileage reaches a preset sample size, the historical measured calibration results are used preferentially. Both calibration routes maintain consistency in the kWh dimension between the number of drivers in driving mode and the group's energy consumption.

[0048] The amount of electricity replenished corresponding to the charging state is no longer determined by arbitrary coefficients with unclear meanings, but by the average effective charging power, duration, and efficiency. It can be obtained from the actual average active power of vehicles charging simultaneously in the historical effective charging sessions of the target station; when there are charging piles with different rated power, it is weighted according to the currently available charging gun type and its historical usage ratio. Configured according to the typical efficiency range of charging equipment and vehicle, and limited to a range greater than 0 and not greater than 1.

[0049] When some vehicles start or end charging in the middle of a time unit, the effective charging time percentage can be used for correction. When only aggregated power data for charging stations is available, the actual aggregated charging power for that time unit can be directly divided by the number of valid charging drivers to generate historical data. The empirical values ​​are used. Regardless of the data source used, the dimensions of the number of drivers and the amount of electricity are kept consistent in the calculations.

[0050] After the total energy equivalent value is updated, the migration direction of each discrete energy state is determined according to the preset SOC interval boundary. The equivalent SOC of a driver in the driving state changes towards lower energy levels, while the equivalent SOC of a driver in the charging state changes towards higher energy levels. When the updated value crosses the interval boundary, the corresponding number of drivers is moved to the adjacent energy state unit in the next time unit. Therefore, the total energy recursion and the discrete spatial state evolution have a practical connection.

[0051] The technical advantage of this embodiment is that it clarifies the physical meaning, dimension, and data calibration source of the expected conversion rate coefficient of the power state, and makes the power consumption during driving and the power replenishment during charging have a unified kWh dimension; thereby, it can solve the technical problem that the total power cannot be reproduced in engineering when the power consumption of the group is directly added or subtracted by the number of drivers or when an undefined coefficient is used.

[0052] Example 4: Based on Example 2, this example further illustrates the relationship between the number of drivers in the driving state. Let... Time unit The driver's desire to rest, with values ​​limited to ,but This indicates the driver's activity level. Let... This represents the dimensionless energy state after normalization of the current total energy equivalent value. This represents the dimensionless order status, which is the normalized quantity or density of current passenger orders. Both constitute the target parameters driving driving behavior.

[0053] right Normalization can be adopted ,in To prevent positive numbers with a denominator of zero, the order quantity can be normalized to extreme values ​​based on the upper and lower bounds of the historical sliding window. Construct the comprehensive input: ; in, Representing time units The comprehensive input of driving behavior is a dimensionless quantity; and These represent the weights of battery status and order demand, respectively. To satisfy the requirements that the first derivative is positive, the second derivative is negative, and there is an asymptotic upper bound, this embodiment uses the following formula as an explicit implementation of the nonlinear response function: ; in, The nonlinear response function representing driving behavior. This represents the dimensionless input variable of the function. For the response slope parameter. When hour, , , and when When it increases It asymptotically approaches 1. Therefore, this function can express the increased willingness to drive when orders or available electricity increase, while the marginal increment gradually decreases and the response quantity has an upper bound.

[0054] As another implementation consistent with the aforementioned monotonically saturated characteristics, an improved logic styrene function can be used. By normalizing the starting point of the logic function within the operating range defined by the non-negative synthesis input and non-negative bias parameters, we can obtain:

[0055] The response slope parameter is positive. Within the operating interval, the logic function value is not less than 0.5. Therefore, the normalized improved logic function maintains a positive first derivative and a second derivative that is not greater than 0 and asymptotically close to 1, which can be used as another explicit implementation of the nonlinear response function described in the claim; this implementation is retained alongside the aforementioned exponential saturation function.

[0056] Given the total number of currently active drivers Number of drivers in charging status Afterwards, the number of drivers in the driving state can be determined according to the following formula: ; in, This is to ensure that drivers who have entered the charging state are not counted as driving again. If the station's operational data can directly identify drivers who are resting or offline, it can also start from... After deducting the number of drivers who were not in the driving state, the remaining active drivers are then counted for driving status.

[0057] The initial value can be obtained from the offline ratio, rest ratio, or platform online rate of the same time unit in the target station's service area. The data can be directly obtained from the platform's online driver location statistics; if real-time online drivers cannot be obtained, the historical daily active driver count of the same type is used and corrected using a trainable correction coefficient. , and It can be obtained by training with historical data, or by fitting historical grouped response curves when sufficient business statistics are available.

[0058] The technical advantage of this embodiment is that it uses a nonlinear response function with a clear mathematical form and boundary properties to convert driver activity level, group power status and order demand into the number of drivers in driving status, and truncates the number by the available driver margin; thereby solving the technical problems that cannot be numerically implemented when only describing "the first derivative of the function is positive and the second derivative is negative" without giving the computable function form, and that the number of drivers in driving status exceeds the actual number of available drivers.

[0059] Example 5: Based on Example 2, this example further illustrates the relationship between the number of drivers in the charging state. The number of drivers in the charging state is affected by three types of quantities: one of which is... The factors are: first, the rest desire index; second, the surplus of drivers available for charging, formed by the difference between the total number of active drivers in the surrounding area and the current number of charging drivers; and third, the correlation function formed by factors such as price, orders, weather, traffic, holidays, and the saturation of surrounding charging stations.

[0060] This embodiment uses a consistent approach. Representing time units The number of drivers in the charging state. Let the rest desire index factor be... The comprehensive correlation function is And order: ; in, This represents a time unit composed of driver activity level, rest desire index factor, and comprehensive correlation function. The overall charging response coefficient is a dimensionless quantity. Therefore, without considering the maximum simultaneous charging capacity of the depot, the recursive relationship for the number of drivers in charging mode can be written as: ; When it is necessary to further consider the maximum number of drivers that can charge simultaneously at the station. When dealing with different spatial state assignments or other nonlinear upper limits, the above analytical solution can be used as the initial value, followed by bounded fixed-point micro-iteration. Let the first... The next iteration estimate is One possible update method is: ; in, This is the relaxation coefficient. The initial value can be the number of drivers in the charging state of the previous time unit or the analytical solution mentioned above; the absolute difference between two consecutive results should not exceed the accuracy threshold. The iteration stops when the maximum number of iterations is reached but the threshold is not met, and the result of the last boundary truncation is used. This micro-iteration is an optional solution method provided for engineering implementations with capacity limits, and does not change the basic physical meaning of the charging state relationship in this embodiment.

[0061] To ensure that the self-referencing metric for the current number of charging drivers in claim 5 has a unique and direct implementation path within the same time unit, an implicit relationship of the current quantity is preferred. The number of drivers in the previous time unit's charging state is used only as the initial value for micro-iterations and does not enter the self-referencing item within the same time unit. The preferred relationship within the same time unit is as follows:

[0062] When the maximum simultaneous charging capacity of the charging station is not considered, the above implicit relationship between time units has the following analytical solution:

[0063] When considering the maximum number of drivers that can charge simultaneously at a charging station, a bounded fixed-point micro-iteration can be used:

[0064] The initial value of the micro-iteration can be either the number of drivers in the charging state of the previous time unit or the analytical solution mentioned above. The range of the relaxation coefficient is consistent with that of the aforementioned micro-iteration. The aforementioned recursive expression, which directly incorporates the number of drivers in the charging state of the previous time unit into the driver surplus term, can be used to describe the optional discrete implementation with a response lag in one time unit. When interpreting the self-referencing relationship of the current charging driver number described in claim 5, the implicit relationship, analytical solution, and bounded micro-iteration of the same time unit described in this paragraph shall be used as the basis. In Example 6, when performing automatic differentiation, the nodes of each micro-iteration in this paragraph are retained, and the dependencies of the iteration variables in the same time unit are backpropagated.

[0065] Synthesis related functions The current charging volume is determined by multiple factors influencing the charging volume during the current time period. For continuous variables such as order density, surrounding station saturation, price difference, and traffic speed, the lower bound of the historical sliding window can be used as a reference. and the Upper Realm Dimensionless transformation is performed. Positive and negative indicators are represented using: ; in, This indicates the continuous influence variable to be dimensionless in time unit. The value of , and These represent the dimensionless factor values ​​when the variable is used as a positive and negative indicator, respectively. and These represent the lower and upper bounds of the continuous influencing variable within the historical sliding window, respectively. To prevent division by zero when the upper and lower bounds are equal, for categorical variables such as weather and holidays, the average charging occurrence rate per unit of active driver under the corresponding category can be calculated in the historical training window. and use: ; in, This indicates the category index used when traversing all candidate categories within the historical training window; For category Dimensionless quantization value, It is a positive number. For newly emerging categories without sufficient historical samples, the mean of all categories is used as the initial value, and then updated after obtaining actual samples.

[0066] For qualitative or hierarchical factors with fuzzy boundaries, such as weather conditions and traffic conditions, fuzzy membership functions can be used in parallel for quantitative transformation. Specifically, triangular or trapezoidal membership functions corresponding to language states such as sunny, rainy, congested, and smooth can be pre-set. The membership degree of each language state is calculated based on real-time observations, and the dimensionless correlation factors in the range of zero to one are obtained by weighting according to the membership degree. The fuzzy membership quantification route and the aforementioned historical category statistical mapping route can both be used as inputs to form the comprehensive correlation function, and they are not mutually substituted.

[0067] Each factor is restricted to Then, to prevent any factor from taking the value of 0 from causing the product to remain permanently zero, a lower bound is introduced. ,make .in, Represents the relevant factor index. Indicates the first The relevant factors in time units The original dimensionless value, Indicates the application of a lower limit The corrected factor value. The product of multiple related factors can be explicitly expressed as: ; in, The number of factors participating in the combination. The corresponding weights are defined. This preserves the product relationships of multiple factors in the original scheme while also providing the quantification, normalization, and combination methods for each type of factor. For factors with missing real-time values, the day-ahead forecast uses the statistical forecast values ​​from historical windows of the same week type and season, and records the missing value; the same rules are applied during the training phase to avoid inconsistencies between the training and prediction inputs.

[0068] Representing time units The total number of active drivers within the target charging station's service area is a time-series input that varies over time. It is preferentially obtained from online platform location data or statistics on active drivers within the service area; if only historical baseline values ​​are available... It can be used To make systematic corrections, including This is a positive coefficient calibrated based on historical measured charging sessions and online driver statistics. Therefore, the "initial total number of active drivers" is no longer mistakenly written as a fixed constant for the entire prediction period.

[0069] The technical advantage of this embodiment lies in clarifying the analytical solution of the self-referencing relationship of the number of drivers in the charging state and the optional iterative method when there is a capacity constraint. At the same time, it discloses the variable quantization, normalization, missing value handling and multi-factor multiplication methods of the relevant function. This can solve the technical problems that the charging state equation cannot be directly calculated due to self-referencing, the relevant function is only given a name but lacks internal structure, and the physical meaning of the total number of surrounding active drivers is unclear.

[0070] Example 6: Building upon Example 5, this example further illustrates the gradient descent dynamic calibration process based on historical data. The parameters involved in the calibration do not include physical quantities that can be obtained from the device nameplate or through direct measurement, but mainly include adjustable parameters in driver behavior and correlation functions. The parameters to be trained can be grouped into a vector: ; in, This represents the baseline values ​​for drivers' desire to rest in different time units. The slope of the nonlinear response during operation is represented by the slope of the nonlinear response. and This indicates the weight of battery level and order volume in the driving response. This represents the active driver correction factor. The index parameter representing the rest desire index factor. , and This indicates the parameter that affects price. This represents the factor weights of the comprehensive correlation function. In actual training, only a subset of parameters that need to be calibrated can be selected, while the remaining parameters are fixed at historical statistical values.

[0071] It is necessary to distinguish between training inputs and parameters to be trained. Price, order quantity or order density, number of active drivers in the vicinity, environmental conditions, and measured charging power are observed inputs or supervision signals from historical training samples; gradient descent updates the parameters used to map these inputs to state responses and power outputs. For the passenger order quantity curve, the smoothing coefficient, periodicity coefficient, or correction coefficient in its day-ahead statistical prediction model can be calibrated; for the total number of active drivers in the vicinity, the active driver correction coefficient mapped from historical baseline values ​​to prediction inputs can be calibrated, rather than using the actual number of orders and the number of active drivers at the prediction time themselves as parameters to be trained.

[0072] The training samples consist of several complete historical prediction periods for the target charging station. Each period includes measured charging power aligned to the same time granularity, along with corresponding price, order, driver, and environmental inputs. Let the th period be... The predicted total charging power for each historical period is The measured total charging power was The number of historical cycles is The number of time units in each cycle is Then the mean squared error loss can be written as: ; in, Represented by parameter vector The training loss function is defined as the independent variable, and the squared L2 norm in the regularization term represents the subset of parameters. The square of the second norm; The coefficient is a non-negative positive regularity coefficient. This represents the subset of parameters for which regularization constraints need to be applied. If the original approach uses full gradient descent, then each training round uses all of these parameters. There is no need to set a separate batch size for each historical period; if the engineering implementation uses batch training, then the number of complete historical periods included in the batch is fixed in the training configuration, and the time recursion chain of a single period is not truncated within the same batch.

[0073] As the preferred training configuration, mini-batch gradient descent is adopted, using complete historical prediction periods as sample units. Each batch contains a preset number of complete prediction periods, and the time recursion chain of a single period is not truncated within the batch. When the number of historical periods is small or when it is used to verify the full loss, the aforementioned full gradient descent can continue to be used. Therefore, the mini-batch route is retained as the preferred implementation method, while the full route is retained as a parallel implementation method in small sample or verification scenarios.

[0074] During forward calculation, a historical prediction period is calculated according to... arrive Expand. rely , and , rely Battery status and orders, It also depends on the rest desire index, driver margin, and related functions; therefore, a change in one parameter may affect the predicted power in multiple subsequent time units. The gradient is backpropagated along the time-expanded computational graph. For any parameter to be trained... The mean square error part satisfies: ; in, This represents the index of the parameter to be trained; the indicator function in the formula is... If the value is 1, then the value is 0; This is obtained through a chain rule that considers output power, the number of charging drivers, a comprehensive correlation function, a rest desire index, and the preceding total battery status. Example 5 uses an analytical form. When the analytical expression is directly differentiated, a fixed number of fixed-point micro-iterations can be used. When the calculation node of each iteration is retained and automatic differentiation is performed, the self-reference relationship is avoided during the backpropagation stage.

[0075] To improve the numerical stability of fully coupled training across time, gradient truncation can be added before parameter updates. The gradient vector of the current parameter to be updated, the upper bound of the gradient norm, and a positive number used to prevent the denominator from being zero are used in the truncation calculation according to the following formula:

[0076] In this diagram, the gradient vector with a wavy line represents the truncated gradient vector, while the gradient vector without a wavy line represents the original gradient vector. The quantity with the subscript "max" represents the upper bound of the gradient norm, and the positive number with the subscript "g" is used to prevent the denominator from being zero. When the gradient norm 2 does not exceed the upper bound, the original gradient is maintained; when it exceeds the upper bound, it is scaled proportionally before projective gradient descent is performed. The gradient truncation only limits the magnitude of a single update and does not change the forward recursion relationship, self-referencing fixed points, or the definition of the loss function.

[0077] The parameters are updated using projective gradient descent. Let the learning rate be... , No. Wheel parameters are ,but: ; in, Indicates the first One parameter to be trained The set of feasible constraints This indicates that the parameter is projected onto its allowed range. Project to , , , , and To maintain non-negativity, the weights of the relevant function are projected onto a non-negative simplex with a weight sum of 1 after the update. As a configuration example, a learning rate of 0.01 can be used and no more than 500 updates can be performed, while stopping early when the validation set error no longer decreases to avoid unnecessary over-iteration.

[0078] When reliable historical group fitting initial values ​​are lacking, the parameters of the exponent to be trained, the response slope parameter, and the weights of the correlation function can be randomly initialized within their respective feasible constraint intervals. For example, initial values ​​can be generated uniformly within preset upper and lower bounds for non-negative bounded parameters, and non-negative random numbers can be generated for the correlation function weights and then normalized until the weight sum is 1; subsequently, the feasible set is still constrained by projective gradient descent. When historical group fitting results are available, the fitted values ​​can be used directly for initialization. Neither initialization route changes the training objective function or parameter constraints.

[0079] After training, the parameter versions, training data time range, input field definitions, and validation errors are saved. When a new listed electricity price curve is entered into prediction, the most recently validated parameter version is loaded for forward recursion. As new measured charging power accumulates, calibration can be re-performed at fixed intervals, but all parameters will not be unconditionally trained in every real-time control cycle.

[0080] The technical advantage of this embodiment is that it provides a set of trainable parameters, a supervision signal, a loss function, a gradient propagation path with time expansion, and parameter constraint update rules, so that historical charging power can be used to calibrate driver behavior and related function parameters. This solves the technical problem that the model training process cannot be reproduced when only "dynamic calibration through gradient descent" is proposed without explaining the training object, error source, and how gradients propagate across time dependencies.

[0081] Example 7: Based on Example 5, this example further illustrates the first and second factors affecting the charging amount in the current time period. The first factor reflects the combined effect of the electricity price in the current time period and the electricity price in subsequent adjacent time periods on the current charging behavior, and the second factor reflects the physical upper limit of the number of drivers with insufficient charging capacity on the current number of drivers who can charge.

[0082] First, the target charging station time unit Listed electricity price Normalize within the range allowed by the current pricing strategy. Let's consider the future. The future normalized average price for each adjacent time unit is: ; in, Represented by time unit The normalized average price of the next nearest neighbor calculated at the current moment. This represents the time offset index relative to the current time unit. This represents the actual number of adjacent future time units available within the forecast period. If the price strategy repeats on a daily basis, the tail of the forecast period may be insufficient. If a time unit is used, it can be supplemented from the starting period of the next cycle based on the price sequence within the same day; if cycle supplementation is not used, it is determined by the actual remaining time unit. .

[0083] One implementation of the price factor uses a logical function: ; in, Representing time units Price factors, Represents a logical function. This represents any real number input to the logic function; , This indicates the positive sensitivity of future adjacent prices to current willingness to charge in advance. This indicates the degree of negative sensitivity of the current price to the current willingness to charge. This indicates bias. Therefore, a rise in the current price will decrease... The subsequent increase in adjacent price will increase the relative willingness to charge in advance, which is consistent with the relationship in the claim that the current charging amount is negatively correlated with the current electricity price and positively correlated with the electricity price at subsequent adjacent times.

[0084] The price factor does not directly output individual driver consumption decisions, but only serves as a dimensionless input in the relationship between the number of drivers in charging status. Its effect on the number of charging drivers is simultaneously limited by the number of drivers with insufficient battery power, the current remaining number of drivers available for charging, the maximum number of drivers that can charge simultaneously at the charging station, and the upper limit of the station's available charging power. The resulting number of charging drivers continues to be incorporated into the group's total battery power recursion and the SOC state transition in the next time unit. Therefore, the electricity price relationship is used in this application to form the station load forecast constrained by vehicle battery power and charging equipment capacity.

[0085] If the prices of surrounding charging stations are also considered, the dimensionless price difference between the target station's price and the representative prices of surrounding stations can be calculated first, and then used as... Additional input. The representative price for surrounding stations can be weighted by distance, driver's historical selection ratio, or the number of available charging stations within the same service area. This expansion does not change the fundamental direction of the current price being negative and the future adjacent price being positive.

[0086] The second factor is based on the number of drivers whose battery level has not reached the preset charging completion condition. (Time unit is defined.) The number of active drivers whose batteries are not fully charged is Then the rechargeable margin factor can be expressed as: ; in, To prevent A positive number divided by zero. When the depot can obtain the SOC of a single vehicle, Count the number of drivers whose SOC is below the preset charging completion threshold; when only the aggregated state can be obtained, estimate based on the number of people in the low and medium charge state units and the historical proportion of those in the high charge state who have not yet reached the completion threshold.

[0087] The first and second factors, along with other relevant factors such as weather, orders, traffic, holidays, and the saturation of surrounding stations, are incorporated into the comprehensive correlation function in Example 5. If only the two core factors of this embodiment are used, it can also be made Each non-negative weight is set and multiplicative aggregation is performed.

[0088] The technical advantage of this embodiment is that it transforms the driver's cross-time response to the current electricity price and the future adjacent electricity price into a dimensionless price factor with a definite direction, and limits the actual population size that can generate charging demand by the number of drivers whose electricity is not fully charged; thereby solving the technical problems that cannot reflect the driver's waiting or early charging behavior when only considering the current price, and predicting that the number of charging drivers exceeds the number of drivers with actual charging demand.

[0089] Example 8: Based on Example 5, this example further illustrates the calculation method of the rest desire index factor. Assume the prediction period has a total of... Discrete time units For the driver's desire to rest in the current time unit, This represents the current driver activity level. To compare the current activity level with the overall activity level in other time units, we first calculate the activity levels in all time units except the current one. The geometric mean of the activity level of each time unit.

[0090] To avoid numerical underflow caused by directly multiplying a large number of decimals and then taking higher roots, logarithmic field calculation is used. Let... If the value is positive, then the geometric mean of the activity level in the remaining time units is: ; in, This represents the time index used to traverse other time units within the prediction period. Representing time units The driver's desire to rest Indicates time unit The geometric mean of the activity levels of all other time units. The relative activity levels of the current time unit and the remaining time units form a comprehensive product term: ; in, Representing time units The product term is a composite term representing the overall activity level relative to the remaining time units. Performing an exponential operation on this composite term yields the rest desire index factor: From an operational perspective, the comprehensive product term is a ratio-type relative activity index between the current activity level and the geometric mean of the activity levels of other time units. The name "comprehensive product term" is retained because the denominator's geometric mean is derived from the product aggregation of activity levels from multiple other time units, and the comprehensive product term subsequently enters a multiplicative combination with other relevant factors; its name does not imply that the index itself must be the result of directly multiplying multiple current quantities.

[0091] ; in, It is a trainable non-negative exponential parameter used to adjust the strength of the influence of the current relative activity level on the number of drivers in the charging state. The time index factor is always 1, indicating the difference in relative activity during periods of non-use; As the value increases, the sensitivity to the differences between the current time unit and other time units increases.

[0092] To prevent individual When the ratio approaches 1, it becomes too large. This can be addressed before the driver number relationship enters the charging state. Set the upper limit allowed by the business. Even using . The magnitude is determined based on the accuracy of the rest desire data storage and is only used to avoid zero logarithmic values; it is not used as a freely adjustable parameter to change behavioral relationships.

[0093] Updated using historical data training process from Example 6, the initial value can be set to 1 or obtained by statistical fitting of historical groups, and projected onto a preset non-negative interval after training. Since... Simultaneously dependent on other time units Training time is used to carry out any The changes will be through The exponential factors affecting other time units; the time computation graph described in Example 6 preserves this cross-time dependency during backpropagation, without considering... It is considered a constant.

[0094] The technical advantage of this embodiment is that it explicitly calculates the activity level of the current time unit relative to other time units using logarithmic geometric mean, and adjusts its influence intensity through a constrained trainable exponent. This solves the technical problems of unclear calculation path, potential numerical underflow due to direct multiplication, and lack of training and boundary definition for exponent parameters when the original geometric mean expression is coupled with multiple time units.

[0095] Example 9: Building upon Example 1, this example further illustrates the specific composition of the discrete spatial state unit and the correspondence between the state and the aggregation recursion. The spatial state is obtained by combining two dimensions: driver battery status and charging time preference. The driver battery status is divided into three intervals: low battery, medium battery, and high battery; the charging time preference is divided into two categories: peak preference and off-peak preference, thus forming six types of discrete spatial states: low battery + peak preference, low battery + off-peak preference, medium battery + peak preference, medium battery + off-peak preference, high battery + peak preference, and high battery + off-peak preference.

[0096] The boundaries of the battery level intervals are defined based on the vehicle's State of Charge (SOC). For example, low, medium, and high battery levels can each correspond to three consecutive SOC intervals, with the specific boundaries set according to the target fleet's operating rules and charging strategy. Time preference is not temporarily assigned at the current moment but is determined by historical charging records: the effective charging frequency or effective charging volume of the same driver or similar drivers within peak and off-peak price intervals is statistically analyzed. A higher proportion in the peak interval is marked as peak preference, and a higher proportion in the off-peak interval is marked as off-peak preference. If the sample size is insufficient, initial allocation is made according to the proportion of similar driver groups at the depot.

[0097] Each discrete-time unit retains six spatial states, rather than using the approach of "one time unit can only map to one spatial state". Let the spatial state index be... Then the discrete-time frame can be expressed as: ; In each The system records the number of active drivers, the number of drivers in driving mode, the number of drivers in charging mode, the number of drivers in resting mode, and the equivalent battery level. The aggregate variables in Examples 2 to 8 can be obtained by summing the six state units, for example: ; in, , and Representing time units respectively Spatial state The number of charging drivers, driving drivers, and available drivers are considered. Therefore, the use of aggregated variables in the core recursive equation does not mean that the discrete spatial dimension is discarded; rather, it is first statistically analyzed or evolved in the spatial state, and then aggregated into the variables required for the station load.

[0098] Spatial states transition between adjacent time units based on actual energy changes. While driving, the driver reduces the equivalent State of Charge (SOC) based on the average energy consumption determined in Example 3; while charging, the driver increases the equivalent SOC based on actual replenished energy; and while resting and not charging, the driver maintains the original energy state. When the equivalent SOC crosses the boundaries of low, medium, and high energy ranges, the driver enters the adjacent energy state in the next time unit. Charging time preferences remain unchanged within a single prediction period and are periodically recalibrated based on newly added historical charging records during rolling multi-day training.

[0099] While charging time preference retains its label within a single prediction period, it is not merely a classification marker. For the two state units of peak preference and off-peak preference, the current price-sensitive parameter, the future adjacent price-sensitive parameter, or the price factor weight in the comprehensive correlation function in Example 7 can be trained respectively. Within each discrete time unit, the number of drivers in charging status is first calculated according to spatial status, and then aggregated into the number of drivers charging at the depot. Thus, peak preference and off-peak preference states within the same SOC interval can produce different charging status responses, and the differentiated charging results are then incorporated into the SOC state transition of the next time unit through the electricity replenishment in Example 3.

[0100] The power of the target charging station is still calculated by aggregating the number of charging drivers across all spatial states, without defining a separate "total station power" for each of the six preference states. A time unit is defined. The average charging power per vehicle is The maximum available charging power at the station is ,but: ; in, Time unit for target charging station The predicted total charging power. It can be obtained from the average power of historical effective charging sessions; It is determined by the available capacity of the station transformer, the rated power of the available charging equipment, and the grid connection point restrictions.

[0101] For the steady state described in Example 1, the results can be compared at the same time unit between complete prediction cycles to avoid mistaking intraday natural load changes as non-convergence. Let the first... Time unit in round cycle iteration The predicted power is Define the relative rate of change between wheels: ; in, This is a positive number corresponding to the power measurement resolution. As a clear implementation method connecting with the original scheme's requirement that "fluctuations do not exceed 4% within three consecutive time units," the power measurement resolution can be checked for three consecutive time units in each round. Are all values ​​no greater than 4%, and are all 48 time units required to fall within at least one three-unit window that meets the conditions? At the period boundary, according to... and Processing. For more stringent requirements, one can directly demand processing of all time units. All values ​​are no greater than 4%. If the condition is not met even after reaching the maximum number of iterations, the final result is output and marked as not fully converged, without generating false stability flags.

[0102] The technical advantage of this embodiment lies in clarifying the combination rules of the six types of spatial states, the correspondence between each time unit and all spatial states, the spatial state migration method, and the conversion path of spatial state data to the aggregated power of the power station. This can solve the technical problem that when the discrete spatiotemporal framework only remains at the conceptual level and the aggregated recursive equation does not have spatial subscripts, those skilled in the art cannot determine how the six types of states participate in the actual load calculation.

[0103] Example 10: As Figure 2 As shown, this embodiment provides a charging load prediction device based on dynamic game theory and spatiotemporal state evolution, used to execute the methods described in Embodiments 1 to 9. The device includes a spatiotemporal discretization module, a parameter definition module, a recursive equation establishment module, and an iterative prediction module. Each module can be deployed on a target charging station operation server, an energy management server, a cloud computing platform, or a computing device connected to the station database. The module division represents the software functional relationship and does not require that they be implemented by mutually independent physical processors.

[0104] The spatiotemporal discrete module is used to establish discrete time units according to the prediction period and time granularity, and to establish discrete spatial state units based on driver battery status and time preferences. After aligning the timestamps of data from different sources, it stores the number of drivers and battery status corresponding to each spatial state in each time unit. This module outputs a unified discrete spatiotemporal index and initial state to the parameter definition module and the recursive equation establishment module.

[0105] The parameter definition module is used to explicitly parameterize the factors influencing charging decisions. It performs range validation and dimensionless transformation on continuous variables such as electricity price, orders, and surrounding charging station saturation. Categorical variables such as weather and holidays are mapped to numerical values ​​based on historical statistics, resulting in price factors, order factors, rest desire parameters, surrounding active driver parameters, and related function internal parameters. The parameter definition module also reads the most recently calibrated parameter version using historical data to ensure that a defined parameter set is used for prediction.

[0106] The recursive equation establishment module is used to establish the recursive relationship of the total energy equivalent value, the relationship of the number of drivers in the driving state, and the relationship of the number of drivers in the charging state. This module calculates the total energy for the next time unit based on the dimensionally consistent energy update process of Example 3, calculates the number of drivers in the driving state based on the nonlinear response function of Example 4, and calculates the number of drivers in the charging state based on the driver margin, price factor, correlation function, and rest desire index factor of Examples 5, 7, and 8. When a self-referencing relationship with capacity boundaries is used, the module obtains a unique calculation result for the current time unit according to the analytical form of Example 5 or bounded micro-iteration.

[0107] The iterative prediction module calls the recursive equation building module in chronological order to complete the forward calculation of one prediction cycle, and repeats the complete prediction cycle when a periodic steady state is required. After reaching the preset stability condition, the iterative prediction module summarizes the number of charging drivers in six discrete spatial states, forms the total charging power for each time unit based on the average charging power per vehicle and the upper limit of available power at the depot, and outputs the complete charging load prediction curve according to the time index.

[0108] The device can also perform parameter calibration as described in Example 6 using historical data. Parameter calibration can be completed collaboratively by the parameter definition module and the recursive equation establishment module, or by a training program connected to the aforementioned modules; the training results are written to the parameter definition module in the form of parameter versions. This training implements the data processing of the four functional modules described above and does not change the composition of the four core modules in the device claims.

[0109] When outputting prediction results, the prediction period, time granularity, model parameter version, input data time range, and predicted power for each time unit are saved simultaneously. For time units where input data is missing, iteration has not reached a stable condition, or the upper limit of the actual available power of the power station has changed, corresponding status flags are output to help operators distinguish between normal prediction values ​​and results affected by data quality or equipment status.

[0110] The charging load forecast curve can also serve as a technical input for the power station energy management system to verify transformer load margin, check charging equipment capacity, and schedule day-ahead power plans or charging facility operation and maintenance. Whether to further generate and issue real-time charging control commands based on the forecast curve is determined by the specific power station control architecture and is not a necessary condition for the validity of this charging load forecasting method.

[0111] The technical advantage of this embodiment is that by using four modules—spatiotemporal discretization, parameter definition, recursive equation establishment, and iterative prediction—the calculation steps of embodiments 1 to 9 are formed into a deployable software device, while maintaining the complete path of data from the station's historical and real-time inputs to the final total charging power curve. This solves the technical problem that the solution steps exist only in the form of abstract algorithms and that the lack of input-output boundaries between functional modules in the actual system makes it difficult to deploy and reproduce.

Claims

1. A charging load prediction method based on dynamic game theory and spatiotemporal state evolution, characterized in that, include: The continuous prediction time range is divided into discrete time units, and discrete spatial state units are obtained based on continuous physical state variables. The discrete time units and discrete spatial state units are spatiotemporally aligned to obtain a discrete spatiotemporal framework. Factors affecting charging decisions are explicitly parameterized to obtain multi-source influence factor parameters. Based on the multi-source influence factor parameters, a joint time recursive relationship is established under the discrete spatiotemporal framework. The joint time recursive relationship is used to describe the joint evolution process of at least two different types of driver states under the discrete spatiotemporal framework. The joint time recursive relationship is iteratively calculated on the discrete time axis until a preset stability condition is reached. The curve of the total charging power of the target charging station changing with time is extracted from the stable state to obtain the charging load prediction result. In this dynamic game, driver groups in different discrete spatial states are the participants, driving, charging, and resting are the mutually exclusive state strategies, the state response quantity corresponding to each strategy is the payoff characterization, and the self-referenced fixed point of the charging state in the same time unit and the stable result of cross-time recursion are used as the equilibrium. The combined time recursion relationship is constrained by the conservation of the number of available drivers, the boundary of the total energy of the group, and the upper limit of the available charging power of the target charging station. The upper limit of the available charging power of the station is determined based on the available capacity of the station transformer, the rated power of the available charging equipment, and the grid connection point limit.

2. The charging load prediction method according to claim 1, characterized in that, The establishment of the joint time recursive relationship includes: based on the multi-source influence factor parameters, establishing a total energy equivalent value recursive relationship, a driver quantity relationship in the driving state, and a driver quantity relationship in the charging state within the discrete spatiotemporal framework; adding the driver quantity relationship in the driving state and the driver quantity relationship in the charging state to the total energy equivalent value recursive relationship to obtain the joint time recursive relationship.

3. The charging load prediction method according to claim 2, characterized in that, The total energy equivalent value recursive relationship is updated based on the total energy equivalent value at the initial time and the driver status data at the initial time. First, the energy consumption corresponding to the driver in the driving state is subtracted, and the energy replenishment corresponding to the driver in the charging state is added to obtain the total energy equivalent value at the next time. Among them, the energy status of the driver in the driving state is transformed by the expected energy status conversion rate coefficient.

4. The charging load prediction method according to claim 2, characterized in that, The number of drivers in the driving state is represented by subtracting the driver's desire to rest in the current time period from 1 to indicate the driver's activity level. The target parameter is formed by combining the current total battery equivalent value and the number of passenger orders. The target parameter is transformed by a nonlinear response function with a positive first derivative, a negative second derivative, and an asymptote to obtain the number of drivers in the driving state. The driver's desire to rest is a dimensionless state parameter formed by statistical analysis of the offline ratio, rest ratio, or platform online rate in the same historical time unit.

5. The charging load prediction method according to claim 2, characterized in that, The relationship between the number of drivers in the charging state is obtained by establishing a rest desire index factor based on 1 minus the current value of the driver's rest desire in the current time period, a function term related to the difference between the total number of active drivers in the surrounding area and the number of drivers charging at present, and a correlation function corresponding to multiple factors affecting the charging amount in the current time period, and then multiplying the rest desire index factor, the function term and the correlation function together.

6. The charging load prediction method according to claim 5, characterized in that, The parameters of driver rest desire, the curve of passenger order quantity, the total number of active drivers in the vicinity, and the internal parameters of the function terms and related functions are all obtained by dynamic calibration through gradient descent based on historical data. Among them, the dynamic calibration of the curve of passenger order quantity and the total number of active drivers in the vicinity is to calibrate the statistical prediction parameters or correction coefficients used to form the corresponding prediction inputs. The order quantity values ​​and the number of active drivers in the vicinity obtained in the prediction stage are used as input data rather than training parameters. The training parameters actually updated by gradient descent include at least the baseline value of driver rest desire, the nonlinear response parameter of driving status, the active driver correction coefficient, the rest desire index parameter, the price influence parameter, and the weights of related functions.

7. The charging load prediction method according to claim 5, characterized in that, The multiple factors affecting the charging amount in the current time period include a first factor and a second factor. The first factor is formed based on the fact that the charging amount in the current time period is negatively correlated with the electricity price level in the current time period, and positively correlated with the charging electricity price at multiple times adjacent to the current time period. The second factor is formed by taking the number of drivers with insufficient charging power as the maximum possible value of the number of charging drivers. The first factor is only used as the cross-time price response input for the number of drivers in the charging state, and together with the second factor, the current remaining number of drivers who can participate in charging, and the maximum number of drivers that can charge at the station at the same time, it constrains the number of drivers in the charging state. The number of drivers in the charging state is further used for updating the equivalent value of the total charging power in the next time unit.

8. The charging load prediction method according to claim 5, characterized in that, The establishment of the rest desire index factor includes: obtaining 1 minus the current value of the driver's rest desire in the current time period and using it as the numerator; obtaining 1 minus the reference values ​​of the driver's rest desire at all other times; taking the geometric mean of the reference values ​​and using it as the denominator to obtain a comprehensive product term; performing an exponential operation on the comprehensive product term to obtain the rest desire index factor; wherein, the exponent is a trainable parameter used to control the influence of the comprehensive product term on the charging amount; the comprehensive product term is a ratio-type relative activity index of the driver's activity level in the current time period relative to the geometric mean of the driver's activity level at other times.

9. The charging load prediction method according to claim 1, characterized in that, The continuous physical state variables include driver battery status and charging time preference; the driver battery status is divided into low battery status, medium battery status, and high battery status, and the charging time preference is divided into peak preference and off-peak preference. The combination of driver battery status and charging time preference forms six discrete spatial state units; the charging time preference is a status label formed by the statistics of the number of historical effective charging times or effective charging volume of the same driver or similar drivers in the peak price range and off-peak price range, and the number of drivers in the charging status is calculated in each of the six discrete spatial state units using the price response parameters corresponding to the charging time preference, and then aggregated.

10. A charging load prediction device based on dynamic game theory and spatiotemporal state evolution, used to execute the charging load prediction method as described in any one of claims 1 to 9, characterized in that, include: The spatiotemporal discrete module is used to establish the discrete spatiotemporal framework; The parameter definition module is used to explicitly parameterize the factors affecting charging decisions to obtain the parameters of the multi-source influencing factors. The recursive equation establishment module is used to establish the simultaneous time recursive relationship based on the multi-source influence factor parameters within the discrete spatiotemporal framework. The iterative prediction module is used to iteratively calculate the simultaneous time recursive relationship on the discrete time axis until a preset stability condition is reached, and extract the curve of the total charging power of the target charging station changing with time from the stable state to obtain the charging load prediction result. The iterative prediction module is also used to form the predicted total charging power based on the average charging power per vehicle and the number of drivers in each time unit, and to impose boundary constraints based on the upper limit of the available charging power of the station determined by the available capacity of the station transformer, the rated power of the available charging equipment, and the grid connection point limitation.

Images