New energy heavy truck charging demand prediction system based on big data analysis

Abstract

The application discloses a new energy heavy truck charging demand prediction system based on big data analysis, comprising the following modules: a data acquisition module for acquiring new energy heavy truck transportation operation data to generate a multi-source spatio-temporal data set; an energy consumption portrait module for dividing working conditions, calculating battery end predicted power consumption and updating vehicle individual energy consumption correction coefficients to generate a task energy consumption portrait; a gap determination module for determining energy margin and generating task energy gap and power supply demand level; a station planning module for constructing a station service capability constraint planning model and generating station service capability data; a demand prediction module for generating charging demand probability prediction results based on an improved DeepAR model; a demand correction module for generating charging demand correction results; and a closed-loop update module for updating relevant parameters according to execution results. The application can improve heavy truck charging demand prediction accuracy and station service capability matching.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle operation scheduling and charging demand forecasting technology, and in particular to a new energy heavy-duty truck charging demand forecasting system based on big data analysis. Background Technology

[0002] With the increasing application of new energy heavy-duty trucks in long-haul transportation, short-haul transportation in mining areas, port collection and distribution, and urban delivery, intelligent scheduling technologies focusing on vehicle charging demand forecasting, charging station load management, and transportation task continuity have received widespread attention. Existing technologies typically rely on historical charging volume, vehicle arrival numbers, station load curves, or time-of-use pricing data to statistically predict or forecast future charging loads, and then combine this with conditions such as station power capacity and the number of charging guns to perform simple scheduling or load shaving. However, the following problems commonly exist in actual transportation operations: The charging demand of new energy heavy trucks is strongly correlated with transportation tasks, load status, return distance, subsequent departure time and vehicle battery status. Existing prediction methods rely on historical charging load or changes in the number of vehicles, making it difficult to incorporate task energy consumption profiles, individual vehicle energy consumption deviations and task energy gaps into a unified prediction link, resulting in inaccurate identification of charging demand under high-intensity transportation tasks or concentrated return scenarios. The current assessment of the service capacity of charging stations is usually a rough estimate based on the number of charging guns, maximum power, or historical queuing situation. It is difficult to simultaneously handle vehicle service time windows, single gun occupancy, station power capacity, grid capacity margin, and charging demand constraints. This leads to a mismatch between the predicted charging demand and the actual service capacity of the station, which can easily result in queuing congestion, vehicles with rigid charging needs being unable to charge in time, and elastic demand occupying key capacity. While existing DeepAR time series prediction models can output predicted quantiles, they typically only extrapolate the probability of historical charging volume sequences and general exogenous variables. They lack task energy gap gating, station capacity constraint correction, and overcapacity probability analysis mechanisms, making it difficult to accurately determine the risk of charging demand exceeding the station's charging capacity in future time slices. Furthermore, it is difficult to further transform the prediction results into actionable charging demand correction results and closed-loop update parameters.

[0003] Therefore, how to provide a system for predicting the charging demand of new energy heavy-duty trucks based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a new energy heavy-duty truck charging demand prediction system based on big data analysis. This invention fully utilizes multi-source spatiotemporal correlation, EWMA energy consumption correction, CP-SAT constrained programming, and improved DeepAR probabilistic prediction technology to describe in detail a system for the coordinated prediction of heavy-duty truck task energy consumption, station service capacity, and charging demand. It has the advantages of high prediction accuracy, strong station matching, and closed-loop adaptive updating.

[0005] The new energy heavy-duty truck charging demand prediction system based on big data analysis according to an embodiment of the present invention includes the following modules: The data acquisition module is used to collect operational data of new energy heavy-duty truck transportation and generate multi-source spatiotemporal datasets. The energy consumption profiling module is used to classify operating conditions based on multi-source spatiotemporal datasets, calculate the expected power consumption of the battery, update the individual energy consumption correction coefficient of the vehicle using EWMA, and generate a task energy consumption profile. The gap determination module is used to determine the energy margin based on the task energy consumption profile and the vehicle's battery status, and to generate the task energy gap, the level of power replenishment demand, and the task urgency by combining time window constraints. The site planning module is used to construct a time-slice integer-based site service capacity constraint planning model based on the task energy gap and site operation constraints. Integer constraint variables are established through vehicle-gun position-time slice occupancy constraints, power capacity constraints and power replenishment demand constraints, and the CP-SAT solver is used to generate site service capacity data. The demand forecasting module is used to input historical charging volume sequences, task energy gaps, and station service capacity data into the improved DeepAR model, and generate charging demand probability forecasting results through task energy gap gating, station capacity constraint correction, and overcapacity probability analysis. The demand correction module is used to perform constraint correction based on the charging demand probability prediction results and the station service capacity data, and generate charging demand correction results. The closed-loop update module is used to calculate the prediction deviation based on the actual charging execution results and the transportation task completion results, and update the task energy consumption profile, station service capacity constraint parameters, and improved DeepAR model parameters.

[0006] Optionally, the data acquisition module specifically includes: The new energy heavy-duty truck transportation operation data includes vehicle operation sub-data, task scheduling sub-data, route operation sub-data, and charging station operation sub-data. Based on the data collection configuration table, access the transportation operation data of new energy heavy trucks, and complete the field standardization and unit unification according to the field mapping dictionary and unit conversion coefficient; A deduplication key is generated based on the data source identifier, vehicle identifier, task identifier, route identifier, station identifier, and collection time, and the record with the latest access time under the same deduplication key is retained; Time slice attribution is determined based on timestamp parsing format and watermark delay threshold, and corresponding tags are written for time-abnormal records and late records. The master data association is completed based on the vehicle-task, task-route, and route-station relationships. The processed data is then written into the spatiotemporal data table according to the unique data key template to generate a multi-source spatiotemporal dataset.

[0007] Optionally, the energy consumption profiling module specifically includes: Based on vehicle operation sub-data, task scheduling sub-data, route operation sub-data and working condition classification parameter table, the transportation process is divided into full-load driving working condition, empty return working condition, loading and unloading waiting working condition, congested low-speed working condition and station stopping working condition. Based on the vehicle parameter table, calculate the energy consumption of rolling resistance, air resistance, gradient resistance, and acceleration resistance under full-load driving conditions, empty return conditions, and congested low-speed conditions, and summarize them into wheel-end energy consumption. Based on the efficiency of the drive system, the energy consumption at the wheel end is converted into the power consumption of the battery end during driving, and the auxiliary power consumption during loading and unloading waiting conditions and station parking conditions is incorporated to obtain the expected power consumption of the battery end; The energy consumption deviation coefficient for this round is generated based on the ratio of the actual power consumption of historical missions to the estimated power consumption of historical missions. The EWMA is used to update the individual energy consumption correction coefficient of the vehicle, and the estimated power consumption of the battery is corrected based on the updated individual energy consumption correction coefficient of the vehicle to generate a mission energy consumption profile.

[0008] Optionally, the notch determination module specifically includes: The vehicle's current available battery power is calculated based on the vehicle's State of Charge (SOC), State of Health (SOH), rated battery capacity, and lower limit of safe SOC. A rigid reserve power is generated based on the current task power consumption, the expected return power consumption, and the safety redundancy power. A planned reserve power is generated based on the current task power consumption, the power consumption of subsequent tasks, the expected return power consumption, and the safety redundancy power. Based on the comparison results between the vehicle's current available power and the rigid and planned guaranteed power, an energy margin determination is performed to generate the task energy gap; The available charging time is determined based on the vehicle's estimated arrival time, latest departure time, and subsequent mission departure time. The required charging time is determined based on the mission's energy gap and the vehicle's acceptable charging power limit. The power replenishment demand level is generated based on the task energy gap, rigid guaranteed power supply, planned guaranteed power supply, available power replenishment time, and required power replenishment time. The task urgency is generated based on the proportion of the task energy gap to the planned guaranteed power supply and the available power replenishment time.

[0009] Optionally, the site planning module specifically includes: The station operation constraints are formed based on the power capacity of the station, the capacity margin of the grid, the upper limit of the power of a single charging gun, the acceptable charging power of the vehicle, and the time-of-use electricity price. The prediction time window, the vehicles waiting to be served, the charging gun, the power variable and the power variable are processed into time slice integers. A station service capacity constraint planning model is constructed, which includes vehicle-gun position-time slice occupancy variables, time-sharing power variables, time-sharing replenishment power variables, uncompleted replenishment power variables, vehicle waiting time variables, hard constraint set, and hierarchical weighted objective function; Establish vehicle service time window constraints, single vehicle occupancy constraints, single gun occupancy constraints, power linkage constraints, station power capacity constraints, grid capacity constraints, and supplementary power demand constraints in the set of hard constraints; The uncharged amount of vehicles with rigid charging level, the uncharged amount of vehicles with flexible charging level, vehicle waiting time, and time-of-use charging cost are taken as target items. A hierarchical weighted objective function is constructed according to the rule that the weight of the previous target item is greater than the sum of the weights of all subsequent target items. The CP-SAT solver is used to select the integer solution with the smallest hierarchical weighted objective function value from the integer solutions that satisfy the set of hard constraints. When there is no feasible solution, only the power supply demand constraints corresponding to the elastic power supply level vehicles are relaxed and then the solution is solved again. The station service capacity data is generated based on the integer solutions output by the CP-SAT solver. The station service capacity data includes the station's chargeable capacity, the number of vehicles that can be connected, the predicted queuing time, the capacity occupancy rate, the vehicle access status, the charging gun occupancy status, and the time-sharing replenishment power.

[0010] Optionally, the demand forecasting module specifically includes: The station's rechargeable capacity, predicted queuing time, and capacity occupancy rate are used as inputs for station capacity constraint correction, and together with the task energy gap, power replenishment demand level, and task urgency, they form a predictive constraint feature set. An improved DeepAR model is constructed, which includes a DeepAR autoregressive probability prediction structure and task energy gap gating, station capacity constraint correction and overcapacity probability analysis embedded in the DeepAR autoregressive probability prediction structure. The DeepAR autoregressive probability prediction structure includes an embedding mapping, a standard LSTM recurrent neural network structure, distributed parameter output and autoregressive recursion. The current time slice cyclic hidden state is generated based on the station static coding, time slice coding, prediction constraint feature set and the charging amount of the previous time slice. Based on the proportion of the task energy gap to the station's rechargeable capacity, the weight corresponding to the level of power replenishment demand, and the urgency of the task, a task energy gap gating weight is generated and applied to the current time slice cyclic hidden state to obtain the task constraint hidden state. The initial mean and initial scale parameters of the charging demand probability distribution are generated based on the hidden state of the task constraint. The initial mean parameter is adjusted by upper limit based on the station's rechargeable capacity, and the initial scale parameter is adjusted by amplification based on predicted queuing time and capacity occupancy rate, to obtain the demand distribution parameter after station capacity constraint correction. Based on the demand distribution parameters corrected by the station capacity constraints, the probability distribution of charging demand for each time slice within the future prediction time window is recursively generated. Perform overcapacity probability analysis, using the chargeable capacity of the charging station as the overcapacity judgment boundary, and the cumulative probability of the demand value being greater than the chargeable capacity of the charging station in the charging demand probability distribution as the overcapacity probability. The difference between the demand value and the chargeable capacity of the charging station is weighted and accumulated according to the corresponding probability to obtain the overcapacity demand amount. The charging demand probability prediction result, consisting of the charging demand probability distribution, prediction quantile, overcapacity probability, and overcapacity demand amount, is output according to the charging station and time slice.

[0011] Optionally, the demand correction module specifically includes: Based on the preset risk quantiles, the correction benchmark requirement is selected from the predicted quantiles, and the correction benchmark requirement is compared with the chargeable capacity of the station. When the calibration baseline requirement does not exceed the station's rechargeable capacity and the overcapacity probability is less than the preset overcapacity probability threshold, the calibration baseline requirement will be used as the calibration requirement for the corresponding station and time slice. When the calibration benchmark demand exceeds the station's chargeable capacity, or the overcapacity probability is not less than the preset overcapacity probability threshold, the amount of electricity to be diverted is determined based on the portion of the calibration benchmark demand that exceeds the station's chargeable capacity and the overcapacity demand amount. The order of vehicle charging allocation is determined based on the level of charging demand and the urgency of the task, and the charging capacity corresponding to vehicles with rigid charging level is reserved first within the charging capacity of the depot. The power to be diverted is sequentially transferred to the time slots with remaining rechargeable capacity at the same power station and to adjacent available power stations, generating charging demand correction results and power station diversion results.

[0012] Optionally, the closed-loop update module specifically includes: Based on the actual charging execution results and transportation task completion results, the charging demand correction results, station service capacity data, and task energy gap are matched; The prediction deviation is generated based on the difference between the actual charging amount and the charging demand correction result; the capacity deviation is generated based on the difference between the actual charging amount and the station's rechargeable amount; and the energy gap deviation is generated based on the difference between the actual power replenishment demand and the task energy gap. Update the vehicle-specific energy consumption correction coefficient in EWMA based on the energy gap deviation, and write back the task energy consumption profile. Update the station service capacity constraint parameters based on capacity deviation; Training samples for the improved DeepAR model are constructed based on prediction bias, and the parameters of the improved DeepAR model are updated with the negative log-likelihood of the actual charging amount label in the probability distribution of charging demand as the training objective. Write the updated task energy consumption profile, station service capacity constraint parameters, and improved DeepAR model parameters into the parameter table for the next prediction period.

[0013] The beneficial effects of this invention are: This invention constructs a task energy consumption profile using multi-source spatiotemporal datasets and EWMA vehicle individual energy consumption correction coefficients, addressing the issue of large deviations in task power consumption estimation for new energy heavy-duty trucks due to differences in load, route, and vehicle type. This makes the task energy gap more closely reflect actual transportation needs. Through a time-slice integerized depot service capacity constraint programming model and the CP-SAT solver, vehicle-pump-time slice occupancy, power capacity, grid capacity, and replenishment demand are unified into solvable integer constraints. This solves the supply-demand mismatch problem caused by estimating depot capacity solely based on the number of pumps or historical load, yielding depot service capacity data usable for prediction. Furthermore, through task energy gap gating, depot capacity constraint correction, and overcapacity probability analysis in the improved DeepAR model, this invention addresses the problem that traditional probabilistic prediction only extrapolates historical charging volumes and fails to reflect task urgency and depot capacity boundaries, outputting correctable probabilistic prediction results for charging demand. These technical solutions improve the matching degree between charging demand prediction and transportation tasks and depot capacity, providing an executable basis for heavy-duty truck fleet replenishment arrangements, depot capacity scheduling, and grid load coordination. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a diagram illustrating the overall framework of the new energy heavy-duty truck charging demand prediction system based on big data analysis proposed in this invention. Figure 2 This is a schematic diagram of the new energy heavy-duty truck charging demand prediction system based on big data analysis proposed in this invention. Figure 3This is a framework diagram of the improved DeepAR model in the big data analysis-based new energy heavy-duty truck charging demand prediction system proposed in this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figures 1-3 The new energy heavy-duty truck charging demand prediction system based on big data analysis includes the following modules: The data acquisition module is used to collect operational data of new energy heavy-duty truck transportation and generate multi-source spatiotemporal datasets. The energy consumption profiling module is used to classify operating conditions based on multi-source spatiotemporal datasets, calculate the expected power consumption of the battery, update the individual energy consumption correction coefficient of the vehicle using EWMA, and generate a task energy consumption profile. The gap determination module is used to determine the energy margin based on the task energy consumption profile and the vehicle's battery status, and to generate the task energy gap, the level of power replenishment demand, and the task urgency by combining time window constraints. The site planning module is used to construct a time-slice integer-based site service capacity constraint planning model based on the task energy gap and site operation constraints. Integer constraint variables are established through vehicle-gun position-time slice occupancy constraints, power capacity constraints and power replenishment demand constraints, and the CP-SAT solver is used to generate site service capacity data. The demand forecasting module is used to input historical charging volume sequences, task energy gaps, and station service capacity data into the improved DeepAR model, and generate charging demand probability forecasting results through task energy gap gating, station capacity constraint correction, and overcapacity probability analysis. The demand correction module is used to perform constraint correction based on the charging demand probability prediction results and the station service capacity data, and generate charging demand correction results. The closed-loop update module is used to calculate the prediction deviation based on the actual charging execution results and the transportation task completion results, and update the task energy consumption profile, station service capacity constraint parameters, and improved DeepAR model parameters.

[0017] In this embodiment, the data acquisition module specifically includes: The operation data for new energy heavy-duty truck transportation includes vehicle operation sub-data, task scheduling sub-data, route operation sub-data, and charging station operation sub-data. The data acquisition configuration table is read to obtain the data source address, field mapping dictionary, unit conversion factor, acquisition cycle, timestamp parsing format, watermark delay threshold, and unique data key template. The watermark delay threshold is the maximum waiting time allowed for delayed data to be included in the time slice corresponding to the original acquisition time. The unique data key template is a field combination rule composed of the data source identifier, vehicle identifier, task identifier, route identifier, station identifier, and acquisition time. Access the operation data of new energy heavy truck transportation according to the collection cycle, and write the data source identifier, access batch number and access time into the access data; replace the source field name with the standard field name according to the field mapping dictionary, convert the numerical field to a unified unit according to the unit conversion factor, and write a missing mark for records with empty standard fields. A deduplication key is generated based on the data source identifier, vehicle identifier, task identifier, route identifier, station identifier, and collection time. Records with the same deduplication key are sorted by access time, and the record with the latest access time is retained. The collection time is parsed according to the timestamp parsing format and converted into a unified time format. Records with empty collection time, mismatched format, or earlier than the corresponding task creation time are marked with a time anomaly. The time slice to which a record belongs is determined based on the collection time. When the arrival time of a record does not exceed the watermark delay threshold, the record is written into the time slice corresponding to the collection time. When the arrival time of a record exceeds the watermark delay threshold, a late data mark is written into the record. Based on vehicle identifiers and task identifiers, generate vehicle-task association records; based on task identifiers and route identifiers, generate task-route association records; based on route identifiers and depot identifiers, generate route-depot association records, and attach the data to which the completion time slice belongs to the corresponding association records; generate unique data keys based on unique data key templates, check if the unique data key exists before writing to the spatiotemporal data table, insert if the unique data key does not exist, update the original record based on the access batch number and access time if the unique data key already exists, and generate a multi-source spatiotemporal dataset.

[0018] In this embodiment, the energy consumption profiling module specifically includes: Read vehicle operation sub-data, task scheduling sub-data, and route operation sub-data from the multi-source spatiotemporal dataset, and read the working condition division parameter table, which includes loading and unloading waiting threshold, congestion speed threshold, congestion judgment threshold, and depot stopping threshold. The vehicle trajectory between the start and end time of the task is used as the transportation process to be divided. The trajectory segment where the vehicle is marked as fully loaded and is in motion is divided into the fully loaded driving condition. The trajectory segment where the vehicle is marked as empty and is on the return route is divided into the empty return condition. The trajectory segment where the vehicle speed is zero, the parking position is within the electronic fence of the loading and unloading location, and the parking time reaches the loading and unloading waiting threshold is divided into the loading and unloading waiting condition. The trajectory segment where the vehicle speed is lower than the congestion speed threshold and the continuous driving time reaches the congestion judgment threshold is divided into the congested low speed condition. The trajectory segment where the vehicle position is within the electronic fence of the charging station and the parking time reaches the station parking threshold is divided into the station parking condition. Read the rolling resistance coefficient, air density, wind resistance coefficient, frontal area, drive system efficiency, and auxiliary power corresponding to the vehicle model from the vehicle parameter table; extract the driving distance, average speed, total vehicle mass, and altitude change for fully loaded driving conditions, unloaded return conditions, and congested low-speed conditions, respectively; and add the energy consumption of rolling resistance, air resistance, gradient resistance, and acceleration resistance to obtain the wheel end energy consumption. Rolling resistance energy consumption is obtained by multiplying the total vehicle mass, rolling resistance coefficient, and travel distance; air resistance energy consumption is obtained by multiplying the air density, drag coefficient, frontal area, average speed squared, and travel distance; gradient resistance energy consumption is obtained by multiplying the total vehicle mass, gravitational acceleration, and altitude change; and acceleration resistance energy consumption is obtained by multiplying the total vehicle mass and the square difference of the positive velocity increment. Divide the energy consumption at the wheel end by the efficiency of the drive system to obtain the battery-side driving power consumption. Multiply the duration of loading and unloading waiting conditions and station parking conditions by the auxiliary power consumption to obtain the auxiliary power consumption. Add the battery-side driving power consumption and the auxiliary power consumption to obtain the battery-side expected power consumption. Read the vehicle individual energy consumption correction coefficient, historical task actual power consumption and historical task estimated power consumption of the same vehicle in the previous round, divide the historical task actual power consumption by the historical task estimated power consumption to obtain the energy consumption deviation coefficient of the current round, multiply the current round energy consumption deviation coefficient by the EWMA update weight to obtain the first correction term, multiply the vehicle individual energy consumption correction coefficient of the previous round by the value after subtracting the EWMA update weight to obtain the second correction term, and add the first correction term and the second correction term to obtain the updated vehicle individual energy consumption correction coefficient. The revised estimated battery power consumption is obtained by multiplying the estimated power consumption of the battery by the updated vehicle-specific energy consumption correction factor, and then written into the task energy consumption profile according to the vehicle, task, and route.

[0019] In this embodiment, the gap determination module specifically includes: Read the corrected battery power consumption and vehicle power status from the task energy consumption profile. The vehicle power status includes SOC, SOH, battery rated capacity and safe SOC lower limit. SOC is the state of charge of the power battery, SOH is the state of health of the power battery, the battery rated capacity is the nominal capacity of the power battery, and the safe SOC lower limit is the preset minimum state of charge threshold. The current available capacity is obtained by subtracting the rated capacity of the battery from the SOH. The current available capacity is obtained by multiplying the difference between the current SOC and the lower limit of the safe SOC by the current available capacity. When the current SOC is not greater than the lower limit of the safe SOC, the current available capacity of the vehicle is recorded as zero. The power consumption corresponding to the return operation condition in the task energy consumption profile is read as the expected power consumption for return. The power consumption corresponding to the lower limit of safe SOC is added to the preset reserve power to obtain the safe redundancy power. The upper limit of the vehicle's acceptable charging power is determined according to the vehicle's power status and vehicle charging parameters. The rigid guarantee power is obtained by adding the corrected battery power consumption, the return power consumption, and the safety redundancy power corresponding to the current task; the planned guarantee power is obtained by adding the corrected battery power consumption, the corrected battery power consumption, the return power consumption, and the safety redundancy power corresponding to the current task. The energy margin determination process involves comparing the vehicle's current available power with the guaranteed power and the planned power to determine the task energy gap. When the vehicle's current available power is less than the guaranteed power, the difference between the guaranteed power and the vehicle's current available power is taken as the task energy gap. When the vehicle's current available power is not less than the guaranteed power and is less than the planned power, the difference between the planned power and the vehicle's current available power is taken as the task energy gap. When the vehicle's current available power is not less than the planned power, the task energy gap is recorded as zero. A time window constraint is established based on the vehicle's estimated arrival time, latest departure time, and subsequent task departure time. The duration between the estimated arrival time and the latest departure time is compared with the duration between the estimated arrival time and the subsequent task departure time. The shorter duration is taken as the available charging time. The required charging time is obtained by dividing the task energy gap by the vehicle's acceptable charging power limit. When the task energy gap is zero, the power replenishment demand level is marked as no gap level. When the vehicle's current available power is less than the rigid guaranteed power, the power replenishment demand level is marked as rigid power replenishment level. When the vehicle's current available power is not less than the rigid guaranteed power and less than the planned guaranteed power, and the available power replenishment time is less than the required power replenishment time, the power replenishment demand level is marked as rigid power replenishment level. When the vehicle's current available power is not less than the rigid guaranteed power and less than the planned guaranteed power, and the available power replenishment time is not less than the required power replenishment time, the power replenishment demand level is marked as flexible power replenishment level. The task urgency is generated based on the ratio of the task energy gap to the planned guaranteed power and the available power replenishment time, and the task energy gap, power replenishment demand level, and task urgency are written into the records of the corresponding vehicle, task, and time slice.

[0020] In this embodiment, the station planning module specifically includes: Read the station operation constraints, which include station power capacity, grid capacity margin, single-gun power limit, vehicle acceptable charging power, and time-of-use electricity price; divide the prediction time window into continuous time slices according to preset time granularity, assign vehicles to be served to vehicle sets, assign charging guns to gun position sets, convert station power capacity, grid capacity margin, single-gun power limit, and vehicle acceptable charging power into integer power units according to preset power granularity, and convert task energy gap and time-of-use supplementary power into integer power units according to preset energy granularity; A station service capacity constraint planning model is constructed, which includes vehicle-gun position-time slice occupancy variables, time-sharing power variables, time-sharing replenishment power variables, uncompleted replenishment power variables, vehicle waiting time variables, hard constraint sets, and hierarchical weighted objective functions. Among them, when the vehicle-gun position-time slice occupancy variable is 1, it means that the target vehicle occupies the target gun position in the target time slice, and when it is 0, it means that the target vehicle does not occupy the target gun position in the target time slice. The time-sharing replenishment power variable is obtained by multiplying the time-sharing power variable by the time slice length and rounding it according to the preset power granularity. In the set of hard constraints, constraints are established for vehicle service time window, single vehicle occupancy, single gun occupancy, power linkage, power station capacity, grid capacity, and charging demand. Among them, the vehicle service time window constraint sets the occupancy variable corresponding to the vehicle's expected arrival time and the latest departure time to 0. The single vehicle occupancy constraint limits the sum of all occupancy variables of the same vehicle in the same time slice to no more than 1. The single gun occupancy constraint limits the sum of all occupancy variables of the same gun position in the same time slice to no more than 1. The power linkage constraint limits the time-sharing power variable to no more than the single gun power limit corresponding to the occupancy variable and the vehicle's acceptable charging power. The power station capacity constraint limits the sum of all time-sharing power variables in the same time slice to no more than the power station capacity. The grid capacity constraint limits the sum of all time-sharing power variables in the same time slice to no more than the grid capacity margin. The charging demand constraint limits the sum of the time-sharing charging power of the same vehicle and the uncompleted charging power to equal the vehicle's task energy gap. A hierarchical weighted objective function is constructed, with the unrecharged amount of vehicles with rigid charging level, the unrecharged amount of vehicles with flexible charging level, vehicle waiting time, and time-sharing charging cost as objective items. The vehicle waiting time is accumulated based on the time slice difference between the vehicle's first access time and the vehicle's expected arrival time. The time-sharing charging cost is accumulated based on the product of the time-sharing charging amount and the corresponding time slice electricity price. Integer weights are configured according to the rule that the weight of the previous objective item is greater than the sum of the weights of all subsequent objective items. The set of hard constraints and the hierarchical weighted objective function are input into the CP-SAT solver. The CP-SAT solver searches for feasible solutions that satisfy the set of hard constraints in the integer domain and selects the integer solution with the smallest hierarchical weighted objective function value from the feasible solutions. When there are no feasible solutions in the set of hard constraints, the vehicle service time window constraints, single vehicle occupancy constraints, single gun occupancy constraints, power linkage constraints, station power capacity constraints, and grid capacity constraints remain unchanged. Only the power demand constraints corresponding to vehicles with flexible power supply levels are relaxed variables and then solved again. Based on the integer solutions output by the CP-SAT solver, the vehicle access status, charging gun occupancy status, time-sharing replenishment capacity, unfinished replenishment capacity, and vehicle waiting time are determined. The time-sharing replenishment capacity, the number of vehicles that can access the station, and the capacity occupancy rate are summarized according to the station and time slice to generate station service capacity data. The station service capacity data includes the station's available charging capacity, the number of vehicles that can access the station, the predicted queuing time, the capacity occupancy rate, the vehicle access status, the charging gun occupancy status, and the time-sharing replenishment capacity.

[0021] In this embodiment, the demand forecasting module specifically includes: Historical charging volume sequences are read according to the station and time slice. The historical charging volume sequence is the actual charging volume sequence of the same station arranged in consecutive time slices. The station service capacity data is read, and the station's available charging capacity, predicted queuing time, and capacity occupancy rate are used as station capacity constraint correction inputs. Together with the task energy gap, replenishment demand level, and task urgency, they form a predictive constraint feature set. An improved DeepAR model is constructed, which includes a DeepAR autoregressive probability prediction structure and task energy gap gating, station capacity constraint correction, and overcapacity probability analysis embedded in the DeepAR autoregressive probability prediction structure. The DeepAR autoregressive probabilistic prediction structure includes an embedding map, a standard LSTM recurrent neural network structure, distributed parameter output, and autoregressive recursion. The standard LSTM recurrent neural network structure is a recurrent neural network structure that takes the current time slice input vector, the previous time slice cyclic hidden state, and the previous time slice memory state as input, and outputs the current time slice cyclic hidden state and the current time slice memory state. The station static coding is a category code generated by the station identifier and the station capacity level. The time slice coding is a time feature code generated by the time slice sequence number, date type, and electricity price period. The station static coding, time slice coding and prediction constraint feature set are embedded to generate an input embedding vector. The charging amount of the previous time slice is concatenated with the input embedding vector to form the current time slice input vector, and then input into a standard LSTM recurrent neural network structure to obtain the current time slice recurrent hidden state. After the standard LSTM recurrent neural network structure outputs the current time slice cyclic hidden state, task energy gap gating is performed. The task energy gap is divided by the station's rechargeable capacity to obtain the gap ratio. When the station's rechargeable capacity is zero, the gap ratio is recorded as 1. The power replenishment demand level is mapped to the level weight. The gap ratio, level weight, and task urgency are multiplied and then normalized by Sigmoid to obtain the task energy gap gating weight. The current time slice cyclic hidden state is multiplied by the coefficient of the task energy gap gating weight to obtain the task constraint hidden state. The process of inputting the hidden state of the task constraint into the output distribution parameters generates the initial mean parameter and initial scale parameter of the charging demand probability distribution. The initial scale parameter is processed by Softplus nonnegative mapping. After the distribution parameter is output, the station capacity constraint is corrected. The station's rechargeable capacity is used as the upper boundary of the initial mean parameter. When the initial mean parameter is greater than the station's rechargeable capacity, the initial mean parameter is corrected to the station's rechargeable capacity. When the initial mean parameter is not greater than the station's rechargeable capacity, the initial mean parameter is retained. The predicted queuing time and capacity occupancy rate are used as scale correction factors to amplify and correct the initial scale parameter, resulting in the demand distribution parameter after capacity constraint correction. Based on the demand distribution parameter after capacity constraint correction, the charging demand probability distribution of the current time slice is generated. The mean of the charging demand probability distribution of the current time slice is used as the input of the autoregressive recursion of the next time slice. Combined with the input embedding vector corresponding to the next time slice, the current time slice's cyclic hidden state, and the current time slice's memory state, the charging demand probability distribution of each time slice within the future prediction time window is recursively generated. After the charging demand probability distribution is generated, the overcapacity probability analysis is performed. The chargeable amount of the station is used as the overcapacity judgment boundary. The cumulative probability of the demand value being greater than the chargeable amount of the station in the charging demand probability distribution is used as the overcapacity probability. The difference between the demand value and the chargeable amount of the station is weighted and accumulated according to the corresponding probability to obtain the overcapacity demand amount. The charging demand probability prediction result, consisting of the charging demand probability distribution, prediction quantile, overcapacity probability, and overcapacity demand amount, is output according to the station and time slice. In this embodiment, the improved DeepAR model adds a task energy gap gating to the original autoregressive probability prediction structure, and incorporates the task energy gap, the level of power replenishment demand, and the urgency of the task into the cyclic hidden state, so that the model can more accurately distinguish between rigid power replenishment demand and elastic power replenishment demand when predicting future charging demand. Meanwhile, the actual service capacity of a power station does not only depend on historical load, but is also constrained by gun occupancy, power capacity, queuing status and grid capacity. The improved DeepAR model uses the station's chargeable capacity, predicted queuing time and capacity occupancy rate as inputs to correct the station's capacity constraints, and corrects the mean and scale parameters of the demand distribution, so that the prediction results are no longer simply extrapolated from historical trends, but are constrained by the actual service boundary of the station. Building upon this foundation, the model further performs overcapacity probability analysis, using the charging capacity of the charging stations as the decision boundary to calculate the probability that charging demand will exceed the station's capacity within a future time slice, as well as the amount of excess capacity demand. The resulting charging demand probability prediction not only reflects the scale of demand but also identifies capacity risks in advance, providing a more direct basis for subsequent charging demand correction, peak-shifting, and updates to station service capacity parameters.

[0022] In this embodiment, the demand correction module specifically includes: Read the probability prediction results of charging demand and the service capacity data of the station, select the correction benchmark demand from the predicted quantiles according to the preset risk quantiles, and compare the correction benchmark demand with the charging capacity of the station. When the calibration baseline requirement does not exceed the station's rechargeable capacity and the overcapacity probability is less than the preset overcapacity probability threshold, the calibration baseline requirement will be used as the calibration requirement for the corresponding station and time slice. When the calibration benchmark demand exceeds the station's chargeable capacity, or the overcapacity probability is not less than the preset overcapacity probability threshold, the excess portion between the calibration benchmark demand and the station's chargeable capacity is taken as the first excess amount, the overcapacity demand amount is taken as the second excess amount, and the larger value between the first excess amount and the second excess amount is taken as the amount to be diverted. The order of vehicle charging is determined based on the charging demand level and the urgency of the task. Vehicles with rigid charging level are arranged before vehicles with flexible charging level. Under the same charging demand level, they are arranged from high to low according to the urgency of the task. Within the charging capacity of the station, the charging capacity corresponding to vehicles with rigid charging level is reserved first. When performing capacity-constrained diversion of the power to be diverted, the power to be diverted is first migrated to the time slot with remaining rechargeable capacity in the subsequent time slot of the same station. Then, based on the line-station association record, vehicle time window constraints and remaining rechargeable capacity, adjacent available stations are determined, and the remaining power that cannot be received by the same station is migrated to adjacent available stations. The vehicle charging attribution results are updated according to the relocated depots and time slots. The charging volume after retention, migration and diversion in each depot and time slot is summarized as the charging demand correction result. The relocated electricity volume, the target depot, the target time slot and the corresponding vehicle are recorded as the depot diversion result.

[0023] In this embodiment, the closed-loop update module specifically includes: Read the actual charging execution results and the transportation task completion results. The actual charging execution results include the actual charging amount and the actual departure SOC. The transportation task completion results include the task completion status and the actual return status. Match the actual charging execution results with the charging demand correction results according to vehicle, task, station and time slice; match the transportation task completion results with the task energy gap according to vehicle and task. The difference between the actual charging amount and the charging demand correction result within the same station and time slice is calculated to generate the prediction deviation. The difference between the actual charging amount and the station's chargeable amount in the station service capacity data is calculated to generate the capacity deviation. The actual charging demand is determined based on the actual charging amount and the actual off-site SOC. The difference between the actual charging demand and the energy gap of the same vehicle and task is calculated to generate the energy gap deviation. The vehicle individual energy consumption correction coefficient in EWMA is updated based on the energy gap deviation, and the updated vehicle individual energy consumption correction coefficient is written back to the task energy consumption profile; the station service capacity constraint parameters are updated based on the capacity deviation, including the station power capacity, the vehicle acceptable charging power, the predicted queuing time, and the correction parameters corresponding to the capacity occupancy rate. The training samples for the improved DeepAR model are constructed based on the prediction bias. The training samples include historical charging volume sequences, prediction constraint feature sets, and actual charging volume labels. The actual charging volume labels are the actual charging volume of the corresponding station and time slice. The negative log-likelihood of the actual charging volume labels in the probability distribution of charging demand is used as the training objective to update the parameters of the improved DeepAR model. Write the updated task energy consumption profile, station service capacity constraint parameters, and improved DeepAR model parameters into the parameter table for the next prediction period.

[0024] Example 1: To verify the feasibility of this invention in practice, it was applied to a new energy heavy-duty truck transportation business scenario. The company has 80 new energy heavy-duty trucks, undertaking transportation tasks between Hefei and Wuhu, Hefei and Nanjing, and short-haul transportation within Hefei. Two charging stations are set up, with a total of 18 charging guns. The maximum power of a single charging gun is 180kW, and the power capacities of the two stations are set to 1800kW and 1200kW respectively. The prediction time window is set to 24 hours, the time slice granularity is set to 15 minutes, the lower limit of safe SOC is set to 20%, the EWMA update weight is set to 0.25, the watermark delay threshold is set to 5 minutes, the power integer granularity is set to 10kW, the energy integer granularity is set to 5kWh, the preset risk quantile is P90, and the preset overcapacity probability threshold is set to 0.35.

[0025] In this scenario, the data acquisition module accesses vehicle operation sub-data, task scheduling sub-data, route operation sub-data, and charging station operation sub-data, and generates a multi-source spatiotemporal dataset based on a field mapping dictionary, unit conversion coefficient, timestamp parsing format, and watermark delay threshold. The energy consumption profiling module categorizes the transportation process into full-load driving conditions, empty return conditions, loading / unloading waiting conditions, congested low-speed conditions, and charging station parking conditions based on vehicle load status, driving status, speed, parking location, and parking duration, and calculates the estimated battery power consumption using the vehicle parameter table. For historical tasks of the same vehicle, EWMA is used to update the individual vehicle energy consumption correction coefficient based on the deviation between the actual power consumption and the estimated power consumption of historical tasks, generating a task energy consumption profile.

[0026] The energy gap assessment module determines the energy margin based on the task energy consumption profile and vehicle battery status. The safe SOC lower limit for vehicle battery status is set at 20%. When the vehicle's current available battery power is insufficient to cover the rigid or planned guaranteed battery power, a task energy gap is generated. This gap is then combined with the vehicle's estimated arrival time, latest departure time, and subsequent task departure time to determine the level of charging demand and the urgency of the task. The depot planning module inputs the task energy gap and depot operational constraints into the depot service capacity constraint planning model. Based on vehicle-gun position-time slice occupancy variables, time-sharing power variables, time-sharing charging power variables, uncompleted charging power variables, hard constraint sets, and a hierarchical weighted objective function, it uses the CP-SAT solver to generate depot service capacity data. The demand forecasting module inputs historical charging volume sequences, task energy gaps, and depot service capacity data into an improved DeepAR model. Through task energy gap gating, depot capacity constraint correction, and overcapacity probability analysis, it generates charging demand probability prediction results. The demand correction module then performs constraint corrections based on the station's available charging capacity and overcapacity probability, migrating some vehicles with flexible charging levels to subsequent time slots at the same station or adjacent available stations.

[0027] To verify the effectiveness of this invention, Method A was set as a prediction method based on historical averages and fixed queuing rules, Method B was set as a general DeepAR prediction method, and Method C was set as a prediction method using the CP-SAT solver but still employing general DeepAR. This invention is a scheme that simultaneously uses a station service capacity constraint programming model and an improved DeepAR model. The verification period was 30 consecutive days, covering 2470 transportation tasks and 1685 valid charging records.

[0028] Table 1. Comparison of different methods in predicting the charging demand of new energy heavy-duty trucks.

[0029] As shown in Table 1, Method A, relying on historical averages and fixed queuing rules, fails to reflect the energy gap and depot operational constraints of new energy heavy-duty trucks. Its predicted charging demand MAPE is 18.7%, with an average queuing time of 47.6 minutes. Method B, using a standard DeepAR model, can perform probabilistic predictions using historical charging volume sequences, reducing the predicted charging demand MAPE to 13.6%. However, due to the lack of depot service capacity data, the prediction results still easily exceed the depot's chargeable capacity, resulting in an overcapacity identification accuracy of only 71.3%. Method C, with the addition of the CP-SAT solver, reduces the depot's chargeable capacity error to 64.7 kWh, indicating that the depot service capacity constraint planning model can improve depot capacity estimation. However, the standard DeepAR model still does not deeply constrain the energy gap and capacity boundaries, and the rigid charging failure rate remains at 6.7%.

[0030] This invention updates the individual vehicle energy consumption correction coefficient through EWMA, enabling the task energy consumption profile to reflect the energy consumption differences of different vehicles under different load, route, and driving conditions. The CP-SAT solver selects the integer solution with the smallest hierarchical weighted objective function value from the integer solutions satisfying constraints such as vehicle service time window, single vehicle occupancy, single charging gun occupancy, depot power capacity, grid capacity, and recharge demand, making the depot service capacity data closer to the actual executable results. An improved DeepAR model incorporates task energy gap, recharge demand level, task urgency, depot rechargeable capacity, predicted queuing time, and capacity occupancy rate into the prediction process, reducing the MAPE of charging demand prediction to 8.4%, increasing the overcapacity identification accuracy to 88.6%, reducing the rigid recharge unmet rate to 3.9%, and reducing the average queuing time to 22.4 minutes.

[0031] As illustrated in this embodiment, the present invention can integrate the vehicle's charging demand, the service capacity of the depot, and probability prediction results for coordinated processing. This avoids prediction lag caused by relying solely on historical charging volumes and also prevents predicted demand from exceeding the actual service capacity of the depot. In scenarios involving concentrated return trips, high-load transportation, and limited charging gun resources, the system can identify overcapacity risks in advance, prioritize vehicles requiring rigid charging, and rationally divert vehicles requiring flexible charging, thereby improving the continuity of new energy heavy-duty truck transportation tasks and the utilization efficiency of depot charging resources.

[0032] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A new energy heavy-duty truck charging demand forecasting system based on big data analysis, characterized in that, Includes the following modules: The data acquisition module is used to collect operational data of new energy heavy-duty truck transportation and generate multi-source spatiotemporal datasets. The energy consumption profiling module is used to classify operating conditions based on multi-source spatiotemporal datasets, calculate the expected power consumption of the battery, update the individual energy consumption correction coefficient of the vehicle using EWMA, and generate a task energy consumption profile. The gap determination module is used to determine the energy margin based on the task energy consumption profile and the vehicle's battery status, and to generate the task energy gap, the level of power replenishment demand, and the task urgency by combining time window constraints. The site planning module is used to construct a time-slice integer-based site service capacity constraint planning model based on the task energy gap and site operation constraints. Integer constraint variables are established through vehicle-gun position-time slice occupancy constraints, power capacity constraints and power replenishment demand constraints, and the CP-SAT solver is used to generate site service capacity data. The demand forecasting module is used to input historical charging volume sequences, task energy gaps, and station service capacity data into the improved DeepAR model, and generate charging demand probability forecasting results through task energy gap gating, station capacity constraint correction, and overcapacity probability analysis. The demand correction module is used to perform constraint correction based on the charging demand probability prediction results and the station service capacity data, and generate charging demand correction results. The closed-loop update module is used to calculate the prediction deviation based on the actual charging execution results and the transportation task completion results, and update the task energy consumption profile, station service capacity constraint parameters, and improved DeepAR model parameters.

2. The new energy heavy-duty truck charging demand prediction system based on big data analysis according to claim 1, characterized in that, The data acquisition module specifically includes: The new energy heavy-duty truck transportation operation data includes vehicle operation sub-data, task scheduling sub-data, route operation sub-data, and charging station operation sub-data. Based on the data collection configuration table, access the transportation operation data of new energy heavy trucks, and complete the field standardization and unit unification according to the field mapping dictionary and unit conversion coefficient; A deduplication key is generated based on the data source identifier, vehicle identifier, task identifier, route identifier, station identifier, and collection time, and the record with the latest access time under the same deduplication key is retained; Time slice attribution is determined based on timestamp parsing format and watermark delay threshold, and corresponding tags are written for time-abnormal records and late records. The master data association is completed based on the vehicle-task, task-route, and route-station relationships. The processed data is then written into the spatiotemporal data table according to the unique data key template to generate a multi-source spatiotemporal dataset.

3. The new energy heavy-duty truck charging demand prediction system based on big data analysis according to claim 1, characterized in that, The energy consumption profiling module specifically includes: Based on vehicle operation sub-data, task scheduling sub-data, route operation sub-data and working condition classification parameter table, the transportation process is divided into full-load driving working condition, empty return working condition, loading and unloading waiting working condition, congested low-speed working condition and station stopping working condition. Based on the vehicle parameter table, calculate the energy consumption of rolling resistance, air resistance, gradient resistance, and acceleration resistance under full-load driving conditions, empty return conditions, and congested low-speed conditions, and summarize them into wheel-end energy consumption. Based on the efficiency of the drive system, the energy consumption at the wheel end is converted into the power consumption of the battery end during driving, and the auxiliary power consumption during loading and unloading waiting conditions and station parking conditions is incorporated to obtain the expected power consumption of the battery end; The energy consumption deviation coefficient for this round is generated based on the ratio of the actual power consumption of historical missions to the estimated power consumption of historical missions. The EWMA is used to update the individual energy consumption correction coefficient of the vehicle, and the estimated power consumption of the battery is corrected based on the updated individual energy consumption correction coefficient of the vehicle to generate a mission energy consumption profile.

4. The new energy heavy-duty truck charging demand prediction system based on big data analysis according to claim 1, characterized in that, The gap determination module specifically includes: The vehicle's current available battery power is calculated based on the vehicle's state of charge (SOC), state of equilibrium (SOH), rated battery capacity, and lower limit of safe SOC. A rigid reserve power is generated based on the current task power consumption, the expected return power consumption, and the safety redundancy power. A planned reserve power is generated based on the current task power consumption, the power consumption of subsequent tasks, the expected return power consumption, and the safety redundancy power. Based on the comparison between the vehicle's current available power and the rigid and planned guaranteed power, an energy margin determination is performed to generate a mission energy gap; The available charging time is determined based on the vehicle's estimated arrival time, latest departure time, and subsequent mission departure time. The required charging time is determined based on the mission's energy gap and the vehicle's acceptable charging power limit. The power replenishment demand level is generated based on the task energy gap, rigid guaranteed power supply, planned guaranteed power supply, available power replenishment time, and required power replenishment time. The task urgency is generated based on the proportion of the task energy gap to the planned guaranteed power supply and the available power replenishment time.

5. The new energy heavy-duty truck charging demand prediction system based on big data analysis according to claim 1, characterized in that, The site planning module specifically includes: The station operation constraints are formed based on the power capacity of the station, the capacity margin of the grid, the upper limit of the power of a single charging gun, the acceptable charging power of the vehicle, and the time-of-use electricity price. The prediction time window, the vehicles waiting to be served, the charging gun, the power variable and the power variable are processed into time slice integers. A station service capacity constraint planning model is constructed, which includes vehicle-gun position-time slice occupancy variables, time-sharing power variables, time-sharing replenishment power variables, uncompleted replenishment power variables, vehicle waiting time variables, hard constraint set, and hierarchical weighted objective function; Establish vehicle service time window constraints, single vehicle occupancy constraints, single gun occupancy constraints, power linkage constraints, station power capacity constraints, grid capacity constraints, and supplementary power demand constraints in the set of hard constraints; The uncharged amount of vehicles with rigid charging level, the uncharged amount of vehicles with flexible charging level, vehicle waiting time, and time-of-use charging cost are used as target items. A hierarchical weighted objective function is constructed according to the rule that the weight of the previous target item is greater than the sum of the weights of all subsequent target items. The CP-SAT solver is used to select the integer solution with the smallest hierarchical weighted objective function value from the integer solutions that satisfy the set of hard constraints. When there is no feasible solution, only the power supply demand constraints corresponding to the elastic power supply level vehicles are relaxed and then the solution is solved again. The station service capacity data is generated based on the integer solutions output by the CP-SAT solver. The station service capacity data includes the station's chargeable capacity, the number of vehicles that can be connected, the predicted queuing time, the capacity occupancy rate, the vehicle access status, the charging gun occupancy status, and the time-sharing replenishment power.

6. The new energy heavy-duty truck charging demand prediction system based on big data analysis according to claim 1, characterized in that, The demand forecasting module specifically includes: The station's rechargeable capacity, predicted queuing time, and capacity occupancy rate are used as inputs for station capacity constraint correction, and together with the task energy gap, power replenishment demand level, and task urgency, they form a predictive constraint feature set. An improved DeepAR model is constructed, which includes a DeepAR autoregressive probability prediction structure and task energy gap gating, station capacity constraint correction and overcapacity probability analysis embedded in the DeepAR autoregressive probability prediction structure. The DeepAR autoregressive probability prediction structure includes an embedding mapping, a standard LSTM recurrent neural network structure, distributed parameter output and autoregressive recursion. The current time slice cyclic hidden state is generated based on the station static coding, time slice coding, prediction constraint feature set and the charging amount of the previous time slice. Based on the proportion of the task energy gap to the station's rechargeable capacity, the weight corresponding to the level of power replenishment demand, and the urgency of the task, a task energy gap gating weight is generated and applied to the current time slice cyclic hidden state to obtain the task constraint hidden state. The initial mean and initial scale parameters of the charging demand probability distribution are generated based on the hidden state of the task constraint. The initial mean parameter is adjusted by upper limit based on the station's rechargeable capacity, and the initial scale parameter is adjusted by amplification based on predicted queuing time and capacity occupancy rate, to obtain the demand distribution parameter after station capacity constraint correction. Based on the demand distribution parameters corrected by the station capacity constraints, the probability distribution of charging demand for each time slice within the future prediction time window is recursively generated. Perform overcapacity probability analysis, using the chargeable capacity of the charging station as the overcapacity judgment boundary, and the cumulative value of the probability that the demand value is greater than the chargeable capacity of the charging station in the charging demand probability distribution as the overcapacity probability. The difference between the demand value and the chargeable capacity of the charging station is weighted and accumulated according to the corresponding probability to obtain the overcapacity demand electricity. The charging demand probability prediction result, consisting of the charging demand probability distribution, prediction quantile, overcapacity probability, and overcapacity demand electricity, is output according to the charging station and time slice.

7. The new energy heavy-duty truck charging demand prediction system based on big data analysis according to claim 1, characterized in that, The demand correction module specifically includes: Based on the preset risk quantiles, the correction benchmark requirement is selected from the predicted quantiles, and the correction benchmark requirement is compared with the chargeable capacity of the station. When the calibration baseline requirement does not exceed the station's rechargeable capacity and the overcapacity probability is less than the preset overcapacity probability threshold, the calibration baseline requirement will be used as the calibration requirement for the corresponding station and time slice. When the calibration benchmark demand exceeds the station's chargeable capacity, or the overcapacity probability is not less than the preset overcapacity probability threshold, the amount of electricity to be diverted is determined based on the portion of the calibration benchmark demand that exceeds the station's chargeable capacity and the overcapacity demand amount. The order of vehicle charging allocation is determined based on the level of charging demand and the urgency of the task, and the charging capacity corresponding to vehicles with rigid charging level is reserved first within the charging capacity of the depot. The power to be diverted is sequentially transferred to the time slots with remaining rechargeable capacity at the same power station and to adjacent available power stations, generating charging demand correction results and power station diversion results.

8. The new energy heavy-duty truck charging demand prediction system based on big data analysis according to claim 1, characterized in that, The closed-loop update module specifically includes: Based on the actual charging execution results and transportation task completion results, the charging demand correction results, station service capacity data, and task energy gap are matched; The prediction deviation is generated based on the difference between the actual charging amount and the charging demand correction result; the capacity deviation is generated based on the difference between the actual charging amount and the station's rechargeable amount; and the energy gap deviation is generated based on the difference between the actual power replenishment demand and the task energy gap. Update the vehicle-specific energy consumption correction coefficient in EWMA based on the energy gap deviation, and write back the task energy consumption profile. Update the station service capacity constraint parameters based on capacity deviation; Training samples for the improved DeepAR model are constructed based on prediction bias, and the negative log-likelihood of the actual charging amount label in the probability distribution of charging demand is used as the training objective to update the parameters of the improved DeepAR model. Write the updated task energy consumption profile, station service capacity constraint parameters, and improved DeepAR model parameters into the parameter table for the next prediction period.

Images