Vertical take-off and landing point new energy load intelligent scheduling method and system

Abstract

The application discloses a vertical take-off and landing point new energy load intelligent scheduling method and system, belongs to the technical field of new energy scheduling, and specifically comprises the following steps: obtaining multi-agent energy demand data through a power supply circuit device, and performing distributed storage and desensitization processing; combining the processed multi-agent energy demand data and real-time state information of a battery of an electric energy storage system, adopting an improved Shapley value algorithm to calculate dynamic energy distribution weights of each cooperation party; based on the weights and real-time task dynamics of each cooperation party, constructing a new energy scheduling rule, generating a control instruction according to the new energy scheduling rule, realizing scheduling and distribution of new energy resources, and feeding back the results to the electric energy storage system and each cooperation party; and the application effectively improves the intelligent level, fairness, system energy efficiency and self-adaptive capacity of new energy scheduling in a multi-cooperation party scene.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy dispatching, in particular to a vertical take-off and landing point new energy load intelligent dispatching method and system. BACKGROUND

[0002] As a key infrastructure of urban air transportation system, the energy management of vertical take-off and landing point faces many challenges. In the prior art, the energy dispatching of vertical take-off and landing point mostly adopts fixed quota or simple priority allocation method, lacks fine management of differentiated needs of multiple partners, and cannot adapt to dynamically changing task scenarios. At the same time, the traditional dispatching method has deficiencies in data processing security, allocation fairness, system adaptive ability, etc. SUMMARY

[0003] In view of the deficiencies of the prior art, the application provides a vertical take-off and landing point new energy load intelligent dispatching method and system, which obtains multi-agent energy demand data through a power supply circuit device, and performs distributed storage and desensitization processing; combines the processed multi-agent energy demand data and the real-time state information of the battery of the energy storage system, and uses an improved Shapley value algorithm to calculate the dynamic energy allocation weight of each partner; based on the weight and the real-time task dynamics of each partner, a new energy dispatching rule is constructed, a control instruction is generated according to the new energy dispatching rule, the dispatching and allocation of new energy resources are realized, and the results are fed back to the energy storage system and each partner; the application effectively improves the intelligent level, fairness, system energy efficiency and adaptive ability of new energy dispatching in the multi-partner scenario.

[0004] To achieve the above purpose, the application provides the following technical scheme:

[0005] The vertical take-off and landing point new energy load intelligent dispatching method comprises:

[0006] The power supply circuit device of the vertical take-off and landing point obtains multi-agent energy demand data containing historical electricity consumption data and task characteristics of different partners, and performs distributed storage and desensitization processing to obtain processed multi-agent energy demand data;

[0007] The energy storage system equipped in the vertical take-off and landing point is used to combine the processed multi-agent energy demand data and the real-time state information of the battery of the energy storage system, and an improved Shapley value algorithm is used to calculate the dynamic energy allocation weight of each partner in the dispatching period;

[0008] According to the dynamic energy allocation weight and the real-time task dynamics of each partner, a new energy dispatching rule is constructed in the control system of the power supply circuit device; the new energy dispatching rule structures the dispatching logic into a basic quota allocation layer, an elastic compensation adjustment layer and a cross-period learning optimization layer;

[0009] According to the new energy dispatching rule, the new energy resources connected to the vertical take-off and landing point are dispatched and distributed through the control instructions of the power supply circuit device, and the distribution results are fed back to the electric energy storage system and each partner;

[0010] After the end of the dispatching period, based on the dispatching execution data, online adaptive optimization is performed through the cross-period learning optimization layer.

[0011] Specifically, the distributed storage and desensitization processing specifically includes:

[0012] The multi-agent energy demand data is uploaded to the blockchain-based alliance chain network for distributed storage, and the user sensitive identification information in the multi-agent energy demand data is desensitized using zero-knowledge proof technology, generating processed multi-agent energy demand data containing only time series power load curve, task power demand curve and task priority label.

[0013] Specifically, the calculation process of the improved Shapley value algorithm includes:

[0014] From the processed multi-agent energy demand data, the task priority label and task power demand curve of each partner's planned task in the next dispatching period are extracted;

[0015] From the processed multi-agent energy demand data, the historical average power load of each partner is extracted;

[0016] Based on the historical average power load, the standardized historical average load ratio of each partner to the vertical take-off and landing point energy system is calculated as the basis for distribution;

[0017] The task priority label is mapped to a reference urgency value, the peak power and demand fluctuation rate characteristics of the task power demand curve are identified, and the concentration index of power demand is calculated, the reference urgency value and the concentration index of power demand are weighted and fused to generate the urgency adjustment coefficient of each partner;

[0018] A deep reinforcement learning agent is constructed, whose state space includes the real-time state of charge of the electric energy storage system, the predicted data of new energy power generation and the urgency adjustment coefficient of each partner, its action space is the fine-tuning vector of the Shapley value algorithm output result, and its reward function takes the overall energy efficiency of the system, the fairness of the partner satisfaction and the health degree of the energy storage system as the optimization target; the deep reinforcement learning agent is trained using the proximal policy optimization algorithm;

[0019] The improved Shapley value algorithm is used for comprehensive calculation, the improved Shapley value algorithm takes the basic allocation basis as a first correction factor, takes the emergency adjustment coefficient as a second correction factor, and combines the fine-tuning action output by the deep reinforcement learning agent and the current maximum allowable discharge power constraint of the electric energy storage system, and finally outputs a set of dynamic energy allocation weights with a sum of 1.

[0020] Specifically, constructing the basic quota allocation layer includes:

[0021] Obtain the new energy power generation prediction curve in the to-be-scheduled period from the new energy power generation monitoring system local to the vertical take-off and landing point;

[0022] By time integration of the new energy power generation prediction curve in the scheduling period, the total expected new energy in the scheduling period is calculated;

[0023] Obtain the dynamic energy allocation weight of the current scheduling period, and distribute the calculated total expected new energy in the scheduling period according to the proportion of the dynamic energy allocation weight to calculate the reference power supply quota of each partner; the reference power supply quota is equal to the product of the total expected new energy and the corresponding energy allocation weight;

[0024] Integrate the calculated reference power supply quota of all partners together with the corresponding partner identifier and scheduling period timestamp to generate a structured basic quota allocation plan, and send it to the scheduling rule library of the power supply circuit device control system to become a static reference component of the new energy scheduling rule.

[0025] Specifically, constructing the elastic compensation adjustment layer includes:

[0026] Establish a real-time data interface with the task management system of each partner in the control system of the power supply circuit device, continuously receive and analyze the real-time task dynamics reported by each partner in the scheduling period; the real-time task dynamics at least include the deviation of the actual start / end time of the task from the planned deviation, the deviation of the actual power demand during the task from the preset power demand curve, and the temporary addition or cancellation notification of the task;

[0027] Read the reference power supply quota of each partner generated and stored by the basic quota allocation layer from the scheduling rule library of the control system, at the same time, real-time obtain the current available discharge power and the current remaining available energy of the electric energy storage system, and real-time obtain the actual new energy power generation curve;

[0028] Determine whether to trigger the hierarchical elastic compensation decision in real time based on real-time task dynamics, wherein the triggering conditions at least include: the real-time power demand of any partner in any time period continuously exceeds the benchmark power supply power obtained by decomposing the benchmark power supply quota in the corresponding time period, and the exceeding amplitude and duration meet the preset threshold, or the actual value of the new energy actual power generation curve in any time period continuously is lower than the predicted value relied on when generating the benchmark power supply quota;

[0029] Start the hierarchical elastic compensation decision when the determination result meets any triggering condition, including: calling the electric energy storage system for compensation, using the current available discharge power of the electric energy storage system to directly provide the difference power compensation for the demand-exceeding partners, or using the current remaining available energy of the electric energy storage system to supplement the deficiency of new energy actual power generation;

[0030] If the electric energy storage system cannot meet all compensation demands, then start the second compensation mechanism, output a power supply quota reduction scheme according to the preset allocation rule, and dynamically allocate the reduced energy quota to the partners triggering compensation or use it to make up the system overall energy gap to generate a dynamic quota adjustment scheme; the input of the preset allocation rule includes the real-time interruptability of each partner's task, the emergency degree of the task, and the current respective benchmark power supply quota execution progress.

[0031] Specifically, the preset allocation rule in the second compensation mechanism is implemented through a graph neural network, wherein:

[0032] Each partner and his task is abstracted as a node in the graph, and the node features include the real-time interruptability of the task, the emergency degree, and the benchmark power supply quota execution progress;

[0033] The energy allocation relationship between partners is abstracted as an edge;

[0034] The graph neural network dynamically calculates the quota reduction priority score of each node through a message passing mechanism, and outputs a dynamic quota adjustment scheme that minimizes the overall system dissatisfaction.

[0035] Specifically, the cross-cycle learning optimization layer is constructed using a federated learning framework, including:

[0036] Deploy a local large model on each partner's energy management terminal to extract local performance indicators from the respective scheduling allocation result feedback report;

[0037] Deploy a global model on the central control system of the vertical take-off and landing point;

[0038] After the scheduling cycle ends, each local large model trains using the local scheduling allocation result feedback report, calculates a set of model update information containing only local knowledge increments, and encrypts the model update information;

[0039] Each partner's energy management terminal uploads the encrypted model update information to the central control system of the vertical take-off and landing point. The central control system uses a federated averaging algorithm to aggregate all the collected encrypted model update information and calculate the optimization direction for the global model parameters.

[0040] The central control system uses the aggregation results to update the global model it maintains, generating updated global model parameters;

[0041] The central control system distributes the updated global model parameters to the energy management terminals of all partners, and each energy management terminal updates its local large model with the received updated global model parameters.

[0042] Specifically, the process of scheduling and allocating new energy resources connected to the vertical take-off and landing point according to the new energy dispatch rules, through control commands from the power supply circuit device, and feeding back the allocation results to the energy storage system and various partners, includes:

[0043] The generated basic quota allocation plan is converted into a base power supply command for the power supply circuit device;

[0044] The generated dynamic quota adjustment scheme is converted into a power supply circuit device compensation control command sequence; the compensation control command sequence includes charging and discharging power commands for the converter of the energy storage system and power limit adjustment commands for the intelligent switches of the power supply lines of each partner.

[0045] The compensation control command sequence and the reference power supply command are integrated and arbitrated in the control system to form a real-time scheduling control command set;

[0046] The lower-level controller of the power supply circuit device executes a set of real-time scheduling and control instructions. During the execution of the instructions, the control system collects panoramic operation data in real time through sensors integrated in the power supply circuit device. The panoramic operation data includes the actual power generation of new energy sources, the actual charging and discharging power and real-time state of charge of the energy storage system, and the actual load power of each partner's line.

[0047] After the scheduling cycle ends, the control system performs data cleaning and analysis based on the panoramic operation data, generates a structured scheduling allocation result feedback report, and sends the scheduling allocation result feedback report to the energy storage system management platform and the energy management terminals of each partner. The scheduling allocation result feedback report includes an energy allocation traceability list, energy storage system operation log, and scheduling compliance analysis.

[0048] Specifically, the step of performing online adaptive optimization based on scheduling execution data after the scheduling cycle ends includes:

[0049] After the scheduling cycle ends, the cross-cycle learning optimization layer first collects all the panoramic operation data generated in the previous scheduling cycle from the sensors of the power supply circuit device, the energy storage system management platform, and the energy management terminals of each partner, and performs feature extraction to generate a cycle performance feature profile.

[0050] The cross-cycle learning optimization layer inputs the cycle performance feature profile into a preset multi-objective loss function, calculates the quantized loss value, and combines the quantized loss values ​​to form a cycle performance evaluation vector; the quantized loss value includes system energy efficiency loss value, partner fairness loss value, and energy storage health loss value.

[0051] The cross-period learning optimization layer uses the periodic performance evaluation vector as the optimization objective and employs the backpropagation algorithm to simultaneously adjust the policy network parameters and the internal weight parameters of the improved Shapley value algorithm of the deep reinforcement learning agent. The process of the coordinated adjustment is as follows: the periodic performance evaluation vector is used as the input of the gradient descent algorithm to calculate the parameter adjustment direction that makes the multi-objective loss function value decrease, and the policy network parameters and the weight parameters of the improved Shapley value algorithm are updated to generate the optimized model parameters.

[0052] Before deploying the optimized model parameters to the next scheduling cycle, the cross-cycle learning optimization layer first performs a simulation run on a validation set constructed from historical data. If the simulation results show that the key performance indicators are steadily improved, the optimized model parameters are officially updated to the scheduling rule base.

[0053] The intelligent dispatching system for new energy loads at vertical take-off and landing points includes: a data acquisition module, a dynamic weight calculation module, a dispatching rule construction module, a dispatching execution module, and an adaptive optimization module.

[0054] The data acquisition module is used to collect energy demand data from multiple entities and perform preprocessing.

[0055] The dynamic weight calculation module is used to combine the processed multi-entity energy demand data with the real-time battery status information of the energy storage system, and realize the dynamic calculation of the energy allocation weight of each partner through the improved Shapley value algorithm, and output the dynamic energy allocation weight that matches the task characteristics and battery status.

[0056] The scheduling rule construction module is used to construct new energy scheduling rules based on the dynamic energy allocation weights output by the dynamic weight calculation module and combined with the real-time task dynamics of each partner, and decompose the scheduling logic into three layers of scheduling rules.

[0057] The scheduling execution module is used to convert the three-layer scheduling rules formed by the scheduling rule construction module into control commands, schedule and allocate new energy resources at the vertical take-off and landing point, and feed back the allocation results to the energy storage system and each partner in real time.

[0058] The adaptive optimization module is used to perform online adaptive optimization based on the scheduling execution data fed back by the scheduling execution module after each scheduling cycle ends, through the cross-cycle learning optimization layer preset by the scheduling rule construction module, and feed the optimization results back to the scheduling rule construction module.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] 1. This invention proposes an intelligent scheduling system for new energy loads at vertical take-off and landing points, and optimizes and improves its architecture, operation steps, and processes. The system has the advantages of simple process, low investment and operating costs, and low production and working costs.

[0061] 2. This invention proposes an intelligent scheduling method for renewable energy loads at vertical take-off and landing points. By constructing a three-layer scheduling rule that includes basic quota allocation, elastic compensation adjustment, and cross-cycle learning optimization, it achieves dynamic and precise scheduling of renewable energy loads from multiple entities at vertical take-off and landing points. Its beneficial effects are that the system can fairly calculate dynamic energy allocation weights based on an improved Shapley value algorithm, and combine real-time task dynamics and energy storage system status to realize basic allocation and elastic compensation of renewable energy quotas through a hierarchical decision-making mechanism. This effectively improves the rationality of energy allocation in complex scenarios, the resilience of the system in responding to sudden demands, and the satisfaction among multiple partners.

[0062] 3. This invention proposes an intelligent scheduling method for new energy loads at vertical take-off and landing points, introducing a cross-cycle learning optimization mechanism. After each scheduling cycle, it can automatically perform strategy review and model parameter optimization based on panoramic operation data, enabling the scheduling rules to have the ability to continuously evolve. This mechanism significantly enhances the system's adaptability and intelligence level. By continuously learning from historical experience to optimize decisions, it can continuously improve the overall energy utilization efficiency and ensure the health of the energy storage system in long-term operation, ultimately achieving a spiral increase in scheduling performance. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the intelligent scheduling method for new energy loads at vertical take-off and landing points according to the present invention;

[0064] Figure 2 This is a diagram illustrating the architecture of the intelligent scheduling system for new energy loads at vertical take-off and landing points, as described in this invention. Detailed Implementation

[0065] Example 1:

[0066] Please seeFigure 1 The present invention provides an embodiment of a method for intelligent scheduling of renewable energy loads at vertical take-off and landing points, the method comprising S1 to S5, including the following steps:

[0067] S1: Obtain multi-entity energy demand data containing historical power consumption data of different partners and mission characteristics through the power supply circuit device of the vertical take-off and landing point, and perform distributed storage and de-identification processing to obtain the processed multi-entity energy demand data.

[0068] It should be explained that the cooperating parties are the individual users or units that share the energy resources of the vertical take-off and landing point, while multiple entities refer to multiple independent energy consumption units. In this application, each cooperating party is an entity.

[0069] The distributed storage and de-identification process specifically includes:

[0070] The multi-entity energy demand data is uploaded to a blockchain-based consortium blockchain network for distributed storage. Zero-knowledge proof technology is used to desensitize the user-sensitive identification information in the multi-entity energy demand data, generating processed multi-entity energy demand data that only includes time-series electricity load curves, task power demand curves, and task priority tags.

[0071] S2: Utilizing the energy storage system equipped at the vertical take-off and landing point, and combining the processed multi-entity energy demand data with the real-time battery status information of the energy storage system, the improved Shapley value algorithm is used to calculate the dynamic energy allocation weight of each partner during the scheduling cycle.

[0072] Furthermore, the real-time battery status information includes the current state of charge, health status, real-time charge and discharge power limits of the energy storage system, and the predicted power generation curve of renewable energy within the scheduling cycle.

[0073] It should be noted that traditional allocation methods, such as allocation based on historical electricity consumption ratios or fixed priorities, are either too rigid to adapt to sudden tasks or prone to leading to an unfair situation where the strong get stronger, and they ignore the real-time physical constraints of the energy system itself. The improved Shapley value algorithm used in this method is characterized by not performing a simple arithmetic average or priority ranking, but by simulating various possible alliance combinations of all cooperating parties and calculating the marginal contribution of each cooperating party when joining different alliances, ultimately obtaining an allocation weight that reflects its comprehensive value. Specifically, the improvements are reflected in the introduction of three key dimensions: The first dimension is the historical contribution benchmark. The historical average electricity load of each partner is extracted from the processed multi-entity energy demand data, and its proportion of the total load is calculated. This step recognizes and rewards partners who have been using the vertical take-off and landing point facilities for a long time and stably, and uses their historical behavior as a guarantee of basic rights, avoiding the unfair phenomenon of temporary partners competing for resources with long-term partners, and encouraging the sustainability of cooperation. The second dimension is the task urgency benchmark. By parsing task priority labels, such as mapping medical emergency to the highest urgency, and combining in-depth analysis of task power demand curves, such as identifying tasks with high peak power and drastic demand fluctuations, a dynamic urgency adjustment coefficient is generated. This coefficient is like a sensitive regulator. When a partner's task is related to safety hazards or has extremely high timeliness requirements, its value assessment in the current scheduling cycle will be increased, thereby giving it preferential treatment in weight allocation. This ensures that system resources can flow to where they are most needed, maximizing efficiency. The third dimension is the system's real-time power constraint benchmark, which is a manifestation of the invention's great engineering practical value. It involves real-time access to the battery status information of the energy storage system, especially the current maximum allowable discharge power. This hard constraint is embedded in the calculation logic of the Shapley value algorithm. When evaluating the value of any cooperative alliance, if the total power demand of the alliance exceeds the maximum power that the battery can currently safely provide, the value of the alliance will be judged as zero or extremely low. This means that the improved Shapley value algorithm automatically eliminates all physically infeasible allocation schemes, ensuring that the final calculated dynamic energy allocation weight is not only fair and reasonable, but also practically feasible, preventing the improved Shapley value algorithm from outputting a theoretically optimal but ultimately unexecutable scheduling plan.

[0074] S3: Based on the dynamic energy allocation weight and the real-time task dynamics of each partner, construct new energy scheduling rules in the control system of the power supply circuit device; the new energy scheduling rules structure the scheduling logic into a basic quota allocation layer, an elastic compensation adjustment layer, and a cross-cycle learning optimization layer;

[0075] S4: According to the new energy dispatch rules, the new energy resources connected to the vertical take-off and landing point are dispatched and allocated through the control commands of the power supply circuit device, and the allocation results are fed back to the energy storage system and each partner.

[0076] S5: After the scheduling cycle ends, online adaptive optimization is performed based on the scheduling execution data through the cross-cycle learning optimization layer.

[0077] In summary, this application primarily achieves intelligent scheduling of renewable energy loads at vertical take-off and landing points through data acquisition and processing, dynamic weight calculation, hierarchical rule construction, and scheduling execution feedback. First, historical electricity consumption and task characteristic data from multiple partners are acquired through power supply circuit devices. After anonymization processing using blockchain consortium chain distributed storage and zero-knowledge proof, and combined with the real-time status of the energy storage system, an improved Shapley value algorithm incorporating standardized historical average load ratios, task urgency adjustment coefficients, and energy storage constraints is used to calculate the dynamic energy allocation weights of each partner. Then, based on these weights and renewable energy generation forecast data, a basic quota allocation layer is constructed to generate a benchmark power supply quota. Simultaneously, an elastic compensation adjustment layer is constructed by real-time connection to partner task dynamics and by collecting real-time status data of energy storage and actual power generation, forming a two-layer scheduling rule. Finally, the rules are transformed into control commands for scheduling execution, and panoramic operation data is collected to generate a scheduling allocation result feedback report, which is synchronized to the energy storage system and each partner. For example, at a small drone logistics vertical take-off and landing point, partner A's scheduling cycle is 10:00-10:30. Calculated by the basic quota allocation layer, its baseline power supply quota for that time period corresponds to a baseline power supply power of 8kW, derived based on the predicted total energy of new energy power generation for that cycle and A's dynamic weight. During the scheduling process, A's drone suddenly carries additional supplies, and its actual power demand remains consistently at 10kW from 10:10-10:15, exceeding the baseline power supply power by 2kW. The excess of 25% and the duration of 5 minutes both meet preset thresholds, triggering a tiered elastic compensation decision. The system obtains the current status of the energy storage system in real time. With an available discharge power of 15kW and sufficient remaining energy, and no deviation between the actual power generation of new energy sources and the predicted value, energy storage compensation is directly initiated. Compensation control commands are sent to the power supply circuit device to control the converter of the energy storage system to output 2kW of differential power to replenish the energy of UAV A, ensuring the normal operation of the mission. During the compensation process, the system monitors the energy storage discharge power and the actual load power of A in real time through circuit sensors to ensure that the compensation accurately matches the requirements. It should be noted that the values ​​involved in the above examples are hypothetical numerical examples for ease of understanding and do not represent the specific data in actual testing. This application will not elaborate on them here.

[0078] The calculation process of the improved Shapley value algorithm includes:

[0079] S2.1: Extract the task priority labels and task power demand curves of each cooperating party's planned tasks in the next scheduling cycle from the processed multi-entity energy demand data;

[0080] S2.2: Extract the historical average electricity load of each partner from the processed multi-entity energy demand data;

[0081] S2.3: Based on the historical average electricity load, calculate the standardized historical average load ratio of each partner to the energy system of the vertical take-off and landing point, as the basis for allocation;

[0082] Furthermore, the calculation process for the basic allocation includes:

[0083] (1) The system extracts the time-series electricity load curve data of the corresponding partner from the desensitized multi-entity energy demand data through the partner's unique identifier based on the preset historical statistical period and time granularity. The time-series electricity load curve data is an ordered data set composed of timestamps and load values. For example, the system sets the historical statistical period to the past thirty natural days and the time granularity to one data point every fifteen minutes.

[0084] (2) The extracted time-series electricity load curve data is subjected to quality verification. The verification logic includes checking whether the load value is within a reasonable physical range, whether there are continuous zero or empty value segments caused by communication interruption, and for the identified abnormal data points or temporary missing values, the system uses linear interpolation of normal data at adjacent times to repair and fill them. For large unreliable data intervals, the data in that period is marked as invalid and removed from the current statistical sample to form a complete and reliable effective historical load dataset.

[0085] (3) Perform integration on the effective historical load dataset to obtain the total power consumption. Then divide the total power consumption by the total duration of the statistical period to calculate the original historical average power load of the partner. Subsequently, add up the original historical average power loads of all partners to obtain the sum. Divide the original historical average power load of each partner by the sum to obtain the standardized historical average load ratio of the partner.

[0086] (4) The standardized historical average load ratio of all partners calculated is combined and output as the basis for allocation of the improved Shapley value algorithm.

[0087] S2.4: Map task priority labels to baseline urgency values, identify peak power and demand volatility characteristics of task power demand curves, calculate power demand concentration index, and weight and fuse baseline urgency values ​​with power demand concentration index to generate urgency adjustment coefficients for each partner.

[0088] Furthermore, the specific steps in S2.4 include:

[0089] (1) Read the task priority tags marked for each partner's planned tasks in the next scheduling cycle;

[0090] (2) Convert the task priority labels into baseline urgency values ​​using preset mapping rules;

[0091] It should be noted that the system has a pre-set mapping rule library that maps each task priority label to a specific, mathematically calculable baseline urgency value. For example, a critical task may be mapped to a value of 1, an important task to 0.8, and a regular task to 0.5. This mapping process transforms the qualitative task level into a preliminary baseline urgency value based on task attributes.

[0092] (3) The system analyzes the task power demand curve, extracts the peak power and demand fluctuation rate, and weights and fuses the two to generate the power characteristic value of a single task;

[0093] The task power demand curve describes the detailed planning of the task's power demand during its execution period. The system identifies and extracts two key features from the task power demand curve: the first feature is peak power, which is the highest power point of the curve throughout the entire task cycle, reflecting the maximum instantaneous load that may occur during task execution; the second feature is demand volatility. The system quantifies the degree of instability of the curve by calculating the standard deviation of the curve's power values ​​or the rate of change within a specific band. The higher the demand volatility, the more unstable the power demand of the task, and the greater the scheduling pressure on the energy system.

[0094] (4) For each partner, the system superimposes the task power demand curves of all planned tasks of the partner in the scheduling period on the time axis to form the total power demand curve of the partner, and calculates the concentration index of its power demand based on the total power demand curve.

[0095] Furthermore, the power demand concentration index aims to measure the cumulative effect of all planned tasks of a partner on power demand over time. During calculation, the system superimposes the power demand curves of all planned tasks of the partner on the time axis to generate a total demand curve. Then, it analyzes whether there are significant power demand peak periods in this total demand curve during the scheduling period. The concentration index is quantified by calculating the proportion of demand during peak periods to total demand, or the duration of peak periods. A higher concentration index means that the partner's energy demand is highly concentrated in time and its scheduling urgency is higher.

[0096] (5) The system will weight and fuse the generated baseline urgency value with the concentration index of power demand to finally generate the urgency adjustment coefficient of each partner.

[0097] S2.5: Construct a deep reinforcement learning agent whose state space includes the real-time state of charge of the energy storage system, the predicted power of new energy generation, and the urgency adjustment coefficients of each partner; its action space is a fine-tuning vector of the output of the Shapley value algorithm; and its reward function optimizes the overall energy efficiency of the system, the fairness of partner satisfaction, and the health of the energy storage system. The deep reinforcement learning agent is trained using a proximal policy optimization algorithm.

[0098] S2.6: The improved Shapley value algorithm is used for comprehensive calculation. The improved Shapley value algorithm uses the basic allocation basis as the first correction factor, the urgency adjustment coefficient as the second correction factor, and combines the fine-tuning action output by the deep reinforcement learning agent and the current maximum allowable discharge power constraint of the energy storage system to finally output a set of dynamic energy allocation weights with a total sum of 1.

[0099] Furthermore, the specific steps in S2.6 include:

[0100] (1) All partners at the vertical take-off and landing point are regarded as a complete energy consumption alliance. The Shapley value algorithm framework is used for calculation: by traversing all possible combinations of partner subsets, the marginal contribution of each partner to the whole system after joining the energy consumption alliance is evaluated. After calculating the marginal contribution of all combinations, an initial set of Shapley value weights is obtained by taking the average value. This set of weights is purely based on the historical interaction relationship between partners, which reflects the theoretical fairness. However, it does not consider the specific task characteristics and real-time state of the system in the current cycle. The Shapley value algorithm is the existing technology in this field and is not an inventive solution of this application. It will not be elaborated here.

[0101] (2) The initial Shapley value weights are used as the basis and integrated into the business logic: the standardized historical average load ratio extracted from historical data, i.e. the basic allocation basis, is used as the first correction factor. At the same time, the urgency adjustment coefficient calculated from real-time task data is used as the second correction factor. The system has preset the fusion rules and weights of these two factors. The initial Shapley value weights are weighted and superimposed with these two factors respectively. After correction, a set of transitional weights that initially reflect the historical contribution and the current urgency are obtained.

[0102] (3) Introduce the decision-making wisdom of deep reinforcement learning agents. The agent outputs a fine-tuning vector based on the real-time state of the current system, including the state of charge of the energy storage system, the short-term forecast of new energy power generation, and the urgency adjustment coefficient of each cooperating party. This fine-tuning vector contains subtle adjustment suggestions for the weights of each cooperating party. Its goal is to pursue the overall optimal operation of the system in the long term. The system fuses this fine-tuning vector with the obtained transition weights by vector addition to generate a set of weights optimized by intelligent strategy.

[0103] (4) Introduce the current maximum allowable discharge power constraint as the final hard boundary condition, and check whether the total power demand corresponding to the weights optimized by the intelligent strategy exceeds the maximum power supply capacity of the system. If it exceeds, the system will compress the weights of each cooperating party proportionally to ensure that the total demand falls within the feasible region. Finally, the system normalizes the weights after all corrections and constraint adjustments to ensure that their sum is strictly equal to 1, thereby outputting the final set of dynamic energy allocation weights.

[0104] The construction of the basic quota allocation layer includes:

[0105] A1: Obtain the predicted power generation curve of new energy during the waiting period from the local new energy power generation monitoring system at the vertical take-off and landing point;

[0106] A2: The total energy of new energy expected during the scheduling period is calculated by integrating the power generation prediction curve of new energy sources over time within the scheduling period. The integration process is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0107] A3: Obtain the dynamic energy allocation weight for the current scheduling period, allocate the total energy of new energy expected within the calculated scheduling period according to the proportion of the dynamic energy allocation weight, and calculate the base power supply quota for each partner; the base power supply quota is equal to the product of the total energy of new energy and its corresponding energy allocation weight.

[0108] A4: The calculated baseline power supply quotas of all partners, along with their corresponding partner identifiers and scheduling cycle timestamps, are integrated to generate a structured basic quota allocation plan, which is then sent to the scheduling rule library of the power supply circuit device control system to become a static baseline component of the new energy scheduling rules.

[0109] Constructing a resilient compensation adjustment layer includes:

[0110] B1: Establish a real-time data interface with the task management system of each partner in the control system of the power supply circuit device, continuously receive and parse the real-time task dynamics reported by each partner within the scheduling cycle; the real-time task dynamics include at least the deviation between the actual start / end time of the task and the plan, the deviation between the actual power demand during the task and the preset power demand curve, and the temporary addition or cancellation notification of the task.

[0111] B2: Read the baseline power supply quotas of each partner that have been generated and stored by the basic quota allocation layer from the scheduling rule base of the control system. At the same time, obtain the current available discharge power and current remaining available energy of the energy storage system in real time, and obtain the actual power generation curve of new energy in real time.

[0112] Furthermore, the system acquires in real-time the current available discharge power and current remaining available energy of the energy storage system, as well as the real-time curve of the actual power generation of new energy sources, including:

[0113] (1) Establish a communication bridge between the control system and the physical equipment. The communication drive unit of the control system establishes a stable data connection with the battery management system of the energy storage system. At the same time, it establishes a communication session with the monitoring equipment such as the photovoltaic inverter or wind turbine controller of the new energy power generation unit. The connection adopts a standardized industrial communication protocol to ensure that the data can be transmitted from the equipment end to the control system in a low-latency and high-reliability manner.

[0114] (2) After the communication channel is established, the control system sends data request instructions to the battery management system and the monitoring equipment of the new energy power generation unit at the same time. The request instruction sent to the battery management system clearly requests to obtain the two core state parameters of its current available discharge power and current remaining available energy. The request instruction sent to the monitoring equipment of the new energy power generation unit requests to obtain the historical sequence of actual power generation from the current moment back a short period of time. This historical sequence of actual power generation constitutes the real-time segment of the actual power generation curve of the new energy.

[0115] (3) After receiving the request, the monitoring equipment of the battery management system and the new energy unit will send back the original state data in binary encoding format stored in the internal storage through the established communication link. The data parsing unit of the control system receives the original state data stream in real time. According to the preset communication protocol, the parsing unit unpacks, decodes and parses the original state data stream, converts it into standard data formats such as floating-point numbers and integers that can be recognized and processed by the control system, starts the verification logic, checks whether the parsed data value is within a reasonable physical range, and judges the validity of the data to obtain the parsed and verified state data.

[0116] (4) The state data after parsing and verification is timestamped based on the internal high-precision clock of the control system. After the alignment is completed, the current available discharge power and the current remaining available energy of the energy storage system are encapsulated into a real-time energy storage state data packet. At the same time, the received historical sequence of actual power generation of new energy is combined with the current instantaneous power value to generate the latest actual power generation curve of new energy.

[0117] (5) The encapsulated real-time energy storage status data packet and the actual power generation curve of new energy will be published to the internal data bus of the control system or written to a specific shared memory area, and at the same time trigger a status update event to notify the elastic compensation adjustment layer that the latest system status information is ready.

[0118] B3: Based on real-time task dynamics, determine in real time whether to trigger the hierarchical elastic compensation decision. The triggering conditions include at least the following: the real-time power demand of any partner in any time period continuously exceeds the benchmark power supply obtained by decomposing the benchmark power supply quota in the corresponding time period, and the magnitude and duration of the excess meet the preset threshold, or the actual value of the new energy actual power generation curve in any time period is continuously lower than the predicted value on which the benchmark power supply quota was based.

[0119] B4: When the judgment result meets any triggering condition, the hierarchical elastic compensation decision is initiated, including: firstly, attempting to call the energy storage system for compensation, using the current available discharge power of the energy storage system to directly provide the difference power compensation to the partner with excess demand, or using the current remaining available energy of the energy storage system to supplement the actual power generation of new energy.

[0120] B5: If the energy storage system cannot meet all compensation needs, the second compensation mechanism will be activated. According to the preset adjustment rules, a temporary power supply quota reduction plan for some partners will be output, and the reduced energy quota will be dynamically adjusted to the partner that triggered the compensation or used to make up for the overall energy gap of the system, generating a dynamic quota adjustment plan. The preset adjustment rules input include the real-time interruption flag of each partner's task, the urgency of the task, and the current execution progress of their respective baseline power supply quota.

[0121] Furthermore, the specific steps of B5 include:

[0122] (1) When the system confirms that the energy storage system cannot fully meet the current compensation demand, the second compensation mechanism is automatically activated, including: the system first accurately quantifies the specific value of the unmet compensation demand gap. This gap may be due to the excess power demand of any partner, or it may be due to the overall energy shortage of the system caused by insufficient new energy power generation. At the same time, the system clearly identifies the target object that needs to be compensated, i.e. which partner or the entire system needs this additional energy quota.

[0123] (2) Based on the preset adjustment rules, potential reduction targets are selected, including: traversing all partners and checking the real-time interruptibility of their tasks. Those tasks marked as interruptible or delayed will be given priority to the partners in the potential reduction list. In addition, the system assesses the urgency of each partner's tasks. Those tasks with lower urgency will also be considered by their partners.

[0124] (3) For each potential partner that is included in the potential reduction list, the system calculates the current energy reduction limit. The calculation of the energy limit is based on the current execution progress of their respective baseline power supply quota. The system will assess how much quota the partner has consumed and how much quota is left at the current moment, and calculate a safe and reasonable reduction limit that can both alleviate the system crisis and minimize the impact on the partner's tasks, taking into account the urgency of its subsequent tasks.

[0125] (4) The system comprehensively considers the real-time interruptibility of each partner's task and the urgency of the task, and calculates a reduction priority score for each potential reduction target. Among them, the task with high interruptibility and low urgency has a higher priority score, which means it is more likely to be reduced. Based on this score, the system allocates the reduction amount in sequence from high priority until the cumulative reduction amount is sufficient to fill the quantitative compensation demand gap, thereby generating a specific temporary power supply quota reduction plan for some partners.

[0126] (5) The energy quotas reduced from each partner’s quota according to the reduction plan are summarized and dynamically allocated to the identified partners that need compensation or to make up for the overall energy gap of the system. This allocation, together with the formulated power supply quota reduction plan, constitutes a dynamic quota adjustment plan that describes the energy flow path from the supplier to the demander. The dynamic quota adjustment plan specifies the specific reduction amount for each partner whose quota is reduced, the specific compensation amount for each partner whose quota is compensated, and the effective period of the entire adjustment.

[0127] The preset adjustment rules in the second compensation mechanism are implemented through a lightweight graph neural network, wherein:

[0128] B5.1: Abstract each partner and its tasks into nodes in the diagram. Node characteristics include the real-time interruptibility of the task, its urgency, and the progress of the baseline power supply quota execution.

[0129] B5.2: Abstract the energy allocation relationship between partners as edges;

[0130] B5.3: The graph neural network dynamically calculates the quota reduction priority score of each node through a message passing mechanism, and outputs a dynamic quota adjustment scheme that minimizes the overall system dissatisfaction.

[0131] The cross-cycle learning optimization layer is constructed using a federated learning framework, including:

[0132] C1: Deploy local large-scale models on the energy management terminals of each partner to extract local performance indicators from their respective scheduling and allocation result feedback reports. The deployment of local large-scale models is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0133] C2: Deploy a global model in the central control system of the vertical take-off and landing point. The global model is existing technology in this field and is not an inventive solution of this application. It will not be described in detail here.

[0134] C3: After the scheduling cycle ends, each local large model uses the local scheduling allocation result feedback report for training, calculates a set of model update information containing only local knowledge increments, and encrypts the model update information.

[0135] Furthermore, the specific steps of C3 include:

[0136] (1) After the scheduling cycle ends, the energy management terminal of each partner will receive a scheduling allocation result feedback report. The data processing unit in the terminal extracts key fields related to model training from the scheduling allocation result feedback report, such as the deviation between the predicted power and actual power of new energy at each time point, the difference between the planned load and actual load of the partner, and the record of the hierarchical elastic compensation decision trigger. After cleaning and formatting, a structured local training dataset is constructed.

[0137] (2) Before training begins, download the latest global model parameters from the central server and use these parameters to initialize the weights of the local large model. At the same time, configure the hyperparameters for this training, such as the number of iterations to be performed, the learning rate, and the type of optimizer.

[0138] (3) Extract a small batch of data samples from the local training dataset and input them into the local large model for forward propagation calculation to obtain the prediction value of the local large model for the current energy scheduling situation; then, compare the prediction value of the model with the actual result in the scheduling allocation result feedback report to calculate the prediction error, i.e., the loss value; then, calculate the gradient of the loss value with respect to each parameter of the local large model through the backpropagation algorithm. This gradient indicates how the model parameters should be adjusted to reduce the prediction error; finally, the optimizer updates the parameters of the local large model in one round according to the calculated gradient and the preset learning rate. This process will be repeated multiple times until the performance of the local large model on the local training dataset tends to be stable.

[0139] (4) After local training is completed, the trained local large model parameters are compared with the initial model parameters downloaded from the central server before training, and the parameter difference between the two is calculated to generate plain text model update information.

[0140] (5) The system encrypts the plaintext model update information to generate protected encrypted model update information and prepares to upload it. The encryption process uses an encryption algorithm and key negotiated with the federated learning server to convert the plaintext model update information into ciphertext. Only the central aggregation server holding the corresponding private key can decrypt and read the contents.

[0141] C4: Each partner's energy management terminal uploads the encrypted model update information to the central control system of the vertical take-off and landing point. The central control system uses a federated averaging algorithm to aggregate all the collected encrypted model update information and calculate the optimization direction for the global model parameters. The federated averaging algorithm is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0142] C5: The central control system uses the aggregation results to update the global model it maintains and generates updated global model parameters;

[0143] Furthermore, the specific steps of C5 include:

[0144] (1) The central control system calls the decryption algorithm and key corresponding to the encryption to decrypt the encrypted aggregation result obtained by the federated averaging algorithm, and obtains the global update vector of the plaintext;

[0145] (2) The global update vector is scaled according to the preset global learning rate to obtain the scaled global update vector;

[0146] Furthermore, the global learning rate is obtained through grid search, that is, experiments are conducted on a set of preset values, such as [0.0001, 0.001, 0.01, 0.1], and finally the learning rate that allows the model to converge to a higher performance the fastest and most stably is selected.

[0147] (3) The central control system loads the current version of the global model parameters and adds them to the scaled global update vector to obtain the updated global model parameters;

[0148] (4) Store the updated global model parameters as the new version of global model parameters, and archive the global model parameters of the previous version.

[0149] C6: The central control system will send the updated global model parameters to the energy management terminals of all partners, and each energy management terminal will update its local large model with the received updated global model parameters.

[0150] Furthermore, the specific steps of C6 include:

[0151] (1) The central control system encapsulates the new version of global model parameters and version metadata into a global model parameter package, and obtains the address list of all on-grid partner energy management terminals;

[0152] (2) The central control system establishes a point-to-point secure communication link with each energy management terminal according to the address list. After the link is established, the global model parameter package is sent to each energy management terminal through the secure communication link. During the transmission process, an encrypted transmission protocol is used to ensure the confidentiality and integrity of the parameter data during the transmission process and to prevent it from being stolen or tampered with.

[0153] (3) After receiving the global model parameter package, the energy management terminal of the partner performs data integrity verification. After the verification is successful, the terminal temporarily stores the received global model parameter package in a temporary storage area and notifies the local model management module that there are new version parameters to be updated.

[0154] (4) The energy management terminal reads the global model parameter package from the temporary storage area, parses out the new version of the global model parameters, and uses it to completely cover the parameters of the current local large model;

[0155] (5) The energy management terminal performs rapid verification of the local large model after parameter coverage. For example, it uses a small amount of the latest verification data stored locally to quickly check whether the updated local large model can perform forward inference normally and whether the output results are reasonable. After verification, it is activated to the ready state and an update confirmation is sent to the central control system.

[0156] According to the new energy dispatch rules, the new energy resources connected to the vertical take-off and landing point are dispatched and allocated through control commands from the power supply circuit device, and the allocation results are fed back to the energy storage system and various partners, including:

[0157] S4.1: Convert the generated basic quota allocation plan into a reference power supply command that can be executed by the power supply circuit device;

[0158] Furthermore, the specific steps of S4.1 include:

[0159] (1) The control system divides the scheduling cycle into continuous time windows according to the preset control time granularity, and converts the total energy quota of each partner in the basic quota allocation plan into the average power set point in each time window according to the total duration of the scheduling cycle. For example, for a scheduling cycle that lasts for 24 hours, if the control time granularity is one hour, the system will divide it into 24 time windows.

[0160] Furthermore, the calculation process of the average power setpoint includes: dividing the total energy quota of each partner in the scheduling cycle by the total duration of the scheduling cycle to calculate the average power that the partner needs to maintain throughout the entire cycle. This average power is the reference power setpoint of the partner in each time window.

[0161] (2) For each partner and its corresponding time window, generate a power control instruction containing the target line identifier, effective timestamp, duration and the average power setpoint;

[0162] (3) Sort all power control instructions from all partners in all time windows according to their effective time and encapsulate them into a base power supply instruction set. During the encapsulation process, add control information such as the scheduling cycle identifier, the version number of the instruction sequence, and the cyclic redundancy check code of the entire instruction sequence.

[0163] (4) Send the reference power supply command set to the lower controller of the power supply circuit device, and receive the ready confirmation signal from the lower controller to complete the deployment of the reference power supply command.

[0164] S4.2: Convert the generated dynamic quota adjustment scheme into a sequence of compensation control instructions that can be executed by the power supply circuit device; the sequence of compensation control instructions includes specific charging and discharging power instructions for the converter of the energy storage system and power limit adjustment instructions for the intelligent switches of the power supply lines of each partner.

[0165] Furthermore, the specific steps in S4.2 include:

[0166] (1) The control system receives and parses the dynamic quota adjustment scheme generated by the elastic compensation adjustment layer, and clarifies the partners that need to reduce quotas and the amount of reduction, as well as the partners that need to receive compensation and the amount of compensation.

[0167] (2) Based on the reduction amount or compensation amount, combined with the preset adjustment operation duration, the instantaneous power value required to achieve the energy adjustment is calculated. For example, if a certain amount of energy needs to be compensated to any partner within five minutes, the system will calculate a corresponding compensation power value. At the same time, the system determines the effective timestamp and duration of the compensation instruction based on the start time and duration of the dynamic quota adjustment scheme.

[0168] (3) In cases where the energy storage system needs to be called for compensation, the system generates a power control command for the converter of the energy storage system. If the scheme requires discharge to compensate the partner, the command is a discharge power command with a positive power value calculated in the second step. If the scheme requires charging to absorb excess new energy, the command is a charging power command with a negative power value. The power control command includes a power value, direction, effective time and duration. The power value is taken from the instantaneous power value, and the direction is determined by the compensation or absorption requirements.

[0169] (4) For operations involving the adjustment of the power supply quota of the partner, the system generates a power limit adjustment instruction for the smart switch of the power supply line of the corresponding partner. For the partner whose quota is reduced, the power limit adjustment instruction lowers the upper limit of the power of its line to a new, lower value. For the partner that receives compensation, the power limit adjustment instruction may appropriately raise its upper limit of power or maintain the original benchmark upper limit but allow its actual power to exceed the original benchmark under compensation. Each power limit adjustment instruction includes the target line identifier, the new power limit, and the effective time.

[0170] (5) Synchronize the generated power control command of the energy storage converter with the power limit adjustment command of all line smart switches according to the effective time, encapsulate them into a compensation control command sequence, and send them to the corresponding lower controller for execution.

[0171] S4.3: Integrate and arbitrate the compensation control command sequence with the reference power supply command in the control system to form the final real-time scheduling control command set that integrates the reference and flexible compensation logic;

[0172] Furthermore, the specific steps of S4.3 include:

[0173] (1) The control system decodes the sequence of reference power supply command and compensation control command, parses out the specific content of each command, including its target, effective time, duration and power set value, and establishes a unified time axis, mapping all commands in the sequence of reference power supply command and compensation control command to this unified time axis according to their effective timestamp;

[0174] (2) On the unified time axis, detect whether there is a conflict of setting values ​​from two sets of instructions for the same controlled device in the same time period. If a conflict is detected, arbitration is carried out according to the preset priority rules to determine the final setting value to be executed in the time period. Among them, the compensation control instruction usually has a higher priority than the reference power supply instruction.

[0175] (3) Based on the arbitration result, generate a final power setting curve for each controlled device on a unified time axis, which is composed of different instruction segments.

[0176] (4) Perform system-level safety verification on the final power setting curve to ensure total power balance, energy storage safety and smooth command switching. If the verification fails, backtrack and adjust the command.

[0177] Furthermore, the system-level safety verification includes, but is not limited to: ensuring that the total power demand of all partners at the same time does not exceed the sum of the maximum discharge capacity of new energy power generation and energy storage; ensuring that the charging and discharging power commands of energy storage do not cause its state of charge to exceed the safe range; checking whether the switching of commands is too frequent to avoid impacting the equipment. If a problem is found during the verification, the system will backtrack the arbitration logic and fine-tune the relevant commands until all safety verification constraints are met.

[0178] (5) Encapsulate the final power setting curves of all controlled devices that have passed the verification into a real-time scheduling and control instruction set with a unified identifier and store it in the instruction library.

[0179] S4.4: The lower-level controller of the power supply circuit device executes the real-time scheduling control command set. During the execution of the command, the control system collects panoramic operation data in real time through the sensors integrated in the power supply circuit device. The panoramic operation data includes the actual power generation of new energy, the actual charging and discharging power and real-time state of charge of the energy storage system, and the actual load power of each partner's line.

[0180] S4.5: After the scheduling cycle ends, the control system performs data cleaning and analysis based on the panoramic operation data, generates a structured scheduling allocation result feedback report, and sends the scheduling allocation result feedback report to the energy storage system management platform and the energy management terminals of each partner. The scheduling allocation result feedback report includes an energy allocation traceability list, energy storage system operation log, and scheduling compliance analysis.

[0181] Furthermore, the specific steps of S4.5 include:

[0182] (1) Once the trigger signal for the end of the scheduling cycle is generated, the system first collects all panoramic operation data collected in the period from the data storage center, starts the data cleaning process, fills the temporary data gaps by time series interpolation, and sets the threshold range based on physical laws to identify and remove abnormal data points that are obviously beyond the reasonable range, forming a cleaned operation dataset.

[0183] (2) Based on the cleaned running dataset, perform energy flow source analysis, energy storage system operation analysis and scheduling compliance analysis to generate an energy allocation source list, energy storage system operation log and scheduling compliance analysis results;

[0184] Furthermore, the energy flow traceability analysis includes: the system accurately matches the energy correspondence between new energy power generation, energy storage charging and discharging and the loads of each partner through timestamp alignment, and generates an energy distribution traceability list that records the source, destination and time of each kilowatt-hour of electricity.

[0185] Furthermore, the energy storage system operation analysis includes: the system statistically analyzing the cumulative charging amount, cumulative discharging amount, number of charge-discharge cycles, and state of charge change curves of the energy storage equipment throughout the entire cycle, forming an energy storage system operation log.

[0186] Furthermore, the scheduling compliance analysis includes: the system compares the actual load power curves of each partner with the planned power curves in the real-time scheduling control instruction set point by point, and calculates indicators such as the root mean square error and the overall plan completion rate.

[0187] (3) The system organizes the energy allocation traceability list, energy storage system operation log and scheduling compliance analysis results according to the preset standardized template and generates a structured scheduling allocation result feedback report;

[0188] Furthermore, due to the different recipients of the reports and their varying concerns, the system will adapt the content of the dispatch allocation result feedback reports accordingly. The version of the report sent to the energy storage system management platform will focus on detailed data from the energy storage system's operation logs, including battery health assessment indicators, to facilitate battery life management and maintenance decisions. The version of the report sent to the energy management terminals of each partner will focus on fragments of the energy allocation traceability list related to that partner and its own dispatch compliance analysis, providing them with transparent electricity billing basis and efficiency assessment. Finally, the system will package the adapted content into different report files.

[0189] (4) The report documents are sent to the energy storage system management platform and the energy management terminals of each partner through a secure communication link;

[0190] (5) The system obtains the receipt confirmation signal from the report recipient and archives the distribution record and content of this report.

[0191] After the scheduling cycle ends, based on the scheduling execution data, online adaptive optimization is performed through the cross-cycle learning optimization layer, including:

[0192] S5.1: At the moment the scheduling cycle ends, the cross-cycle learning optimization layer first collects all the panoramic operation data generated in the previous scheduling cycle from the sensors of the power supply circuit device, the energy storage system management platform, and the energy management terminals of each partner, and performs feature extraction to generate a cycle performance feature profile.

[0193] Furthermore, the specific steps in S5.1 include:

[0194] (1) Once the trigger signal for the end of the scheduling cycle is generated, the data acquisition module of the cross-cycle learning optimization layer will start immediately. According to the preset data source address list, it will actively send data retrieval requests to the sensor data buffer of the power supply circuit device, the historical database of the power storage system management platform, and the local storage of the energy management terminals of each partner to obtain the panoramic operation data recorded in time sequence during the previous complete scheduling cycle.

[0195] (2) The collected panoramic operation data are time-stamp aligned and fused using a unified time axis to form a multi-dimensional time series dataset;

[0196] (3) Based on the predefined performance evaluation model, calculate a number of key performance indicators from the multidimensional time series dataset, such as new energy prediction deviation, energy storage usage intensity, quota execution deviation, and frequency and total amount of flexible compensation. For example, calculate the average absolute error between the predicted power and the actual power of new energy to evaluate the prediction accuracy, count the total charge and discharge cycle depth of the energy storage system in one cycle to evaluate its usage intensity, analyze the average deviation rate between the actual electricity consumption of each partner and the planned quota to evaluate the allocation fairness, and calculate the total volume and frequency of energy adjustment due to flexible compensation to evaluate the system's ability to cope with emergencies.

[0197] Furthermore, the performance evaluation model is built based on domain knowledge and mathematical quantification methods, and is a comprehensive computational framework that integrates physical principles, business objectives, statistical methods, and weighting strategies.

[0198] (4) Normalize the calculated key performance indicators and organize them into fixed-dimensional numerical vectors according to a predetermined structure to generate a periodic performance feature profile for that period. Store the generated periodic performance feature profile along with the period identifier in the feature library and establish an index.

[0199] S5.2: The cross-cycle learning optimization layer inputs the cycle performance feature profile into a preset multi-objective loss function, calculates the quantized loss value, and combines the quantized loss values ​​to form a cycle performance evaluation vector; the quantized loss value includes system energy efficiency loss value, partner fairness loss value, and energy storage health loss value;

[0200] Furthermore, the specific steps in S5.2 include:

[0201] (1) The cross-cycle learning optimization layer loads a preset multi-objective loss function. The multi-objective loss function model includes three loss calculation sub-functions: system energy efficiency, partner fairness and energy storage health, which correspond to the three optimization objectives of system energy efficiency, partner fairness and energy storage health, respectively. The cross-cycle learning optimization layer uses the generated cycle performance feature profile as input data and prepares to send it into the multi-objective loss function for calculation. The specific feature values ​​in the cycle performance feature profile will be mapped to the corresponding input variables of each loss function.

[0202] (2) The cross-cycle learning optimization layer extracts the actual total power generation of new energy, the actual total power consumption of each partner and the total charging and discharging loss of the energy storage system from the cycle performance feature profile, and inputs them into the system energy efficiency loss calculation sub-function. The energy efficiency loss is quantified by calculating the ratio of wasted energy to total available energy. For example, if the new energy power generation is not fully utilized and forced to abandon solar or wind power, or if the energy storage charging and discharging process is too lossy, the energy efficiency loss value will increase. Finally, a specific system energy efficiency loss value is output.

[0203] (3) Extract the deviation data between the planned power supply quota and the actual power consumption of each partner from the periodic performance characteristic profile. Among them, the Gini coefficient is used to analyze the distribution of the quota completion of all partners. If the actual power consumption of each partner differs greatly from the planned quota, it indicates that the allocation result is unfair. The partner fairness loss calculation sub-function will output a higher partner fairness loss value.

[0204] (4) Extract the total cycle depth, high-rate discharge times and average state of charge of the energy storage system from the cycle performance characteristic profile, and input them into the energy storage health loss calculation sub-function. The energy storage health loss calculation sub-function integrates the battery life decay model to estimate the equivalent loss caused to battery life. Among them, the greater the usage intensity and the more severe the working conditions, the higher the calculated energy storage health loss value.

[0205] (5) The cross-cycle learning optimization layer normalizes the calculated system energy efficiency loss value, partner fairness loss value, and equivalent depreciation amount to obtain the normalized system energy efficiency loss value, partner fairness loss value, and energy storage health loss value, and combines them in a predetermined order to generate the periodic performance evaluation vector for the scheduling cycle.

[0206] S5.3: The cross-period learning optimization layer uses the periodic performance evaluation vector as the optimization objective and employs the backpropagation algorithm to simultaneously adjust the policy network parameters and the internal weight parameters of the improved Shapley value algorithm of the deep reinforcement learning agent. The process of the coordinated adjustment is as follows: the periodic performance evaluation vector is used as the input of the gradient descent algorithm to calculate the parameter adjustment direction that makes the multi-objective loss function value decrease, and the policy network parameters and the weight parameters of the improved Shapley value algorithm are updated to generate a set of optimized model parameters. The backpropagation algorithm and the gradient descent algorithm are existing technologies in this field and are not inventive solutions of this application, and will not be described in detail here.

[0207] S5.4: Before deploying the optimized model parameters to the next scheduling cycle, the cross-cycle learning optimization layer first performs a simulation run on a validation set constructed from historical data. If the simulation results show that the key performance indicators are steadily improved, the old parameters are archived and the optimized model parameters are officially updated to the scheduling rule base.

[0208] Example 2:

[0209] Please see Figure 2 Another embodiment of the present invention provides: a smart dispatching system for new energy loads at vertical take-off and landing points, comprising:

[0210] Data acquisition module, dynamic weight calculation module, scheduling rule construction module, scheduling execution module, adaptive optimization module;

[0211] The data acquisition module is used to collect energy demand data from multiple entities and perform preprocessing.

[0212] The dynamic weight calculation module combines the processed multi-entity energy demand data with the real-time battery status information of the energy storage system. It uses an improved Shapley value algorithm to dynamically calculate the energy allocation weights of each partner and outputs dynamic energy allocation weights that match the task characteristics and battery status, providing core decision-making basis for the scheduling rule construction module.

[0213] The scheduling rule construction module is used to construct structured and hierarchical new energy scheduling rules based on the dynamic energy allocation weights output by the dynamic weight calculation module and combined with the real-time task dynamics of each partner. It decomposes the scheduling logic into three executable scheduling rules, providing a clear execution basis for the scheduling execution module.

[0214] The scheduling and execution module is used to convert the three-layer scheduling rules formed by the scheduling rule building module into actual control instructions, to accurately schedule and allocate new energy resources at vertical take-off and landing points, and to feed the allocation results back to the energy storage system and various partners in real time.

[0215] The adaptive optimization module is used to perform online adaptive optimization based on the scheduling execution data fed back by the scheduling execution module after each scheduling cycle. It uses the cross-cycle learning optimization layer preset by the scheduling rule construction module to perform optimization. The optimization results are fed back to the scheduling rule construction module to realize the continuous iterative upgrade of scheduling rules and improve the long-term scheduling accuracy and energy utilization efficiency of the system.

[0216] The scheduling rule construction module includes: allocation layer unit, adjustment layer unit, and optimization layer unit;

[0217] The allocation layer unit is used to formulate the basic energy quota standards for each partner based on their basic electricity demand and dynamic weight ratio, to ensure the stability of energy supply for each partner's core tasks, and to form the basic execution logic of the scheduling rules.

[0218] The adjustment layer unit is used to construct a dynamic compensation energy allocation logic based on the real-time task dynamics and battery redundancy of each partner. That is, on the basis of the basic quota allocation, energy elastic supply is provided to partners with incremental task requirements and battery status allows, and energy quota is recovered for partners with reduced tasks, so as to realize the dynamic adjustment of energy resources.

[0219] The optimization layer unit is used to construct learning-iterative rule optimization logic, that is, to pre-set a cross-cycle learning index system, define learning optimization algorithms, and provide a solidified learning framework and logical support for the adaptive optimization module to ensure the continuous iteration capability of scheduling rules.

[0220] The scheduling and execution module includes: a dynamic acquisition unit, an instruction generation unit, an allocation and execution unit, and a result feedback unit;

[0221] The dynamic acquisition unit is used to collect real-time task dynamic data from each partner, such as task start / stop, energy consumption demand increment / decrease, task priority adjustment, etc., to provide real-time dynamic input to the instruction generation unit and ensure the synchronization of scheduling execution with the actual needs of the partners.

[0222] The instruction generation unit is used to combine the three-layer scheduling rules of the scheduling rule construction module with the real-time task dynamic data of the dynamic acquisition unit, and transform them into control instructions that the power supply circuit device can recognize, such as energy allocation, power supply period, charging and discharging control parameters, to ensure the accuracy and executability of the instructions.

[0223] The distribution execution unit is used to execute the control commands output by the command generation unit through the power supply circuit device of the vertical take-off and landing point, and distribute new energy resources, such as the electrical energy converted from photovoltaic and wind power and the electrical energy stored in batteries, to each partner according to the control commands, so as to realize the physical distribution and supply of energy resources.

[0224] The result feedback unit is used to collect the actual allocation results of the allocation execution unit in real time, such as the actual amount of energy obtained by each partner, the allocation response time, and the stability of energy supply. The results are fed back bidirectionally, on the one hand to the energy storage system, and on the other hand to each partner, to ensure that the partners are aware of the energy allocation situation. At the same time, it provides execution data input for the adaptive optimization module.

[0225] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A vertical take-off and landing point new energy load intelligent scheduling method, characterized in that, The application relates to a vertical take-off and landing point power supply circuit device, and a method for dynamically allocating energy resources of the vertical take-off and landing point power supply circuit device. The method comprises the following steps: obtaining multi-agent energy demand data including historical power consumption data and task characteristics of different partners through the vertical take-off and landing point power supply circuit device, and performing distributed storage and desensitization processing to obtain processed multi-agent energy demand data; using the power storage system equipped at the vertical take-off and landing point, combining the processed multi-agent energy demand data and real-time state information of the battery of the power storage system, and using an improved Shapley value algorithm to calculate the dynamic energy allocation weight of each partner in a scheduling period; constructing a new energy scheduling rule in the control system of the power supply circuit device according to the dynamic energy allocation weight and the real-time task dynamics of each partner; the new energy scheduling rule structures the scheduling logic into a basic quota allocation layer, an elastic compensation adjustment layer and a cross-period learning optimization layer; scheduling and allocating new energy resources connected to the vertical take-off and landing point through the control instruction of the power supply circuit device according to the new energy scheduling rule, and feeding back the allocation result to the power storage system and each partner; 2. The vertical take-off and landing point new energy load intelligent scheduling method of claim 1, wherein, after the end of the scheduling period, performing online adaptive optimization through the cross-period learning optimization layer based on the scheduling execution data. The distributed storage and desensitization processing specifically comprises:

3. The vertical take-off and landing point new energy load intelligent scheduling method of claim 2, wherein, uploading the multi-agent energy demand data to a blockchain-based consortium chain network for distributed storage, and using zero-knowledge proof technology to desensitize the user sensitive identification information in the multi-agent energy demand data to generate processed multi-agent energy demand data containing only time sequence power consumption load curve, task power demand curve and task priority label. The calculation process of the improved Shapley value algorithm comprises: extracting the task priority label and the task power demand curve of the planned task of each partner in the next scheduling period from the processed multi-agent energy demand data; extracting the historical average power load of each partner from the processed multi-agent energy demand data; based on the historical average power load, calculating the standardized historical average load proportion of each partner to the vertical take-off and landing point energy system as a basis for basic allocation; mapping the task priority label to a reference urgency value, identifying the peak power and demand fluctuation rate characteristics of the task power demand curve, and calculating the concentration index of power demand, and weighting and fusing the reference urgency value and the concentration index of power demand to generate an emergency adjustment coefficient of each partner; constructing a deep reinforcement learning agent, the state space of which comprises the real-time state of charge of the power storage system, the prediction data of the new energy power generation and the emergency adjustment coefficient of each partner, the action space of which is a fine-tuning vector of the output result of the Shapley value algorithm, and the reward function of which takes the overall energy efficiency of the system, the satisfaction fairness of the partners and the health degree of the energy storage system as optimization objectives; the deep reinforcement learning agent is trained by using a proximal policy optimization algorithm. The improved Shapley value algorithm is used for comprehensive calculation, the improved Shapley value algorithm takes the basic allocation basis as a first correction factor, takes the emergency adjustment coefficient as a second correction factor, and finally outputs a set of dynamic energy allocation weights with a sum of 1 in combination with the fine-tuning action output by the deep reinforcement learning agent and the current maximum allowable discharge power constraint of the energy storage system.

4. The vertical take-off and landing point new energy load intelligent scheduling method of claim 3, wherein, The basic quota allocation layer is constructed by: obtaining a new energy power generation prediction curve in a to-be-scheduled period from a new energy power generation monitoring system local to the vertical take-off and landing point; calculating the total new energy in the scheduling period by time integration of the new energy power generation prediction curve in the scheduling period; obtaining the dynamic energy allocation weight of the current scheduling period, distributing the calculated total new energy in the scheduling period according to the proportion of the dynamic energy allocation weight, and calculating the reference power supply quota for each partner; the reference power supply quota is equal to the product of the total new energy and the corresponding energy allocation weight; integrating the calculated reference power supply quota of all partners, together with the corresponding partner identifier and scheduling period timestamp, to generate a structured basic quota allocation plan, and sending the plan to the scheduling rule library of the power supply circuit device control system as a static reference component of the new energy scheduling rule.

5. The vertical take-off and landing point new energy load intelligent scheduling method of claim 4, wherein, The elastic compensation adjustment layer is constructed by: establishing a real-time data interface with the task management system of each partner in the control system of the power supply circuit device, continuously receiving and analyzing the real-time task dynamics reported by each partner in the scheduling period; the real-time task dynamics at least include the deviation of the actual start / end time of the task from the planned deviation, the deviation of the actual power demand during the task from the preset power demand curve, and the temporary addition or cancellation of the task; reading the reference power supply quota of each partner generated and stored by the basic quota allocation layer from the scheduling rule library of the control system, and simultaneously, obtaining the current available discharge power and the current remaining available energy of the energy storage system in real time, and obtaining the actual new energy generation power curve in real time; based on the real-time task dynamics, real-time judgment is made on whether to trigger the layered elastic compensation decision, wherein the triggering conditions at least include: the real-time power demand of any partner in any period continuously exceeds the reference power supply power obtained by decomposing the reference power supply quota in the corresponding period, and the exceeding amplitude and duration meet the preset threshold, or the actual value of the new energy actual generation power curve in any period continuously is lower than the predicted value relied on when the reference power supply quota is generated; when the judgment result meets any triggering condition, the layered elastic compensation decision is started, including: calling the energy storage system for compensation, using the current available discharge power of the energy storage system to directly provide the excess power compensation for the demand-exceeding partner, or using the current remaining available energy of the energy storage system to supplement the deficiency of the new energy actual generation; If the energy storage system cannot meet all compensation requirements, a second compensation mechanism is started, a power supply quota reduction scheme is output according to a preset allocation rule, and the reduced energy quota is dynamically allocated to the cooperative parties triggering compensation or used to make up for the overall energy gap of the system, thereby generating a dynamic quota adjustment scheme; the preset allocation rule input includes real-time interruptability identification of each cooperative party task, emergency degree of the task, and current respective baseline power supply quota execution progress.

6. The vertical take-off and landing point new energy load intelligent scheduling method of claim 5, wherein, The preset allocation rule in the second compensation mechanism is implemented through a graph neural network, wherein: Each cooperative party and its task is abstracted as a node in the graph, and the node features include real-time interruptability, emergency degree, and baseline power supply quota execution progress of the task; The energy allocation relationship between the cooperative parties is abstracted as an edge; The graph neural network dynamically calculates the quota reduction priority score of each node through a message passing mechanism, and outputs a dynamic quota adjustment scheme that minimizes the overall dissatisfaction of the system.

7. The vertical take-off and landing point new energy load intelligent scheduling method of claim 6, wherein, The cross-cycle learning optimization layer is constructed using a federated learning framework, including: A local large model is deployed on the energy management terminal of each cooperative party to extract local performance indicators from the scheduling allocation result feedback report of each party; A global model is deployed on the central control system of the vertical take-off and landing point; After the scheduling cycle ends, each local large model trains using the local scheduling allocation result feedback report, calculates a set of model update information containing only local knowledge increments, and encrypts the model update information; Each cooperative party's energy management terminal uploads the encrypted model update information to the central control system of the vertical take-off and landing point, and the central control system uses a federated averaging algorithm to aggregate all the encrypted model update information collected, and calculates the optimization direction of the global model parameters; The central control system updates the global model maintained by it using the aggregation result to generate updated global model parameters; The central control system distributes the updated global model parameters to the energy management terminals of all cooperative parties, and each energy management terminal updates its local large model using the received updated global model parameters. 8.The vertical take-off and landing point new energy load intelligent scheduling method of claim 7, wherein, According to the new energy scheduling rule, the new energy resources connected to the vertical take-off and landing point are scheduled and allocated through the control instructions of the power supply circuit device, and the allocation results are fed back to the energy storage system and each cooperative party, including: The generated basic quota allocation plan is converted into a power supply circuit device baseline power supply instruction; The generated dynamic quota adjustment scheme is converted into a power supply circuit device compensation control instruction sequence; the compensation control instruction sequence includes charge and discharge power instructions for the energy storage system converter and power limit adjustment instructions for each cooperative party power supply line intelligent switch; The compensation control instruction sequence and the baseline power supply instruction are integrated and arbitrated in the control system to form a real-time scheduling control instruction set; The real-time scheduling control instruction set is executed by the lower controller of the power supply circuit device, and in the instruction execution process, the control system collects panoramic operation data in real time through the sensors integrated in the power supply circuit device; the panoramic operation data includes actual power generation of new energy, actual charge and discharge power and real-time state of charge of the energy storage system, and actual load power of each cooperative party line; After the end of the scheduling period, the control system performs data cleaning and analysis based on the panoramic operation data, generates a structured scheduling allocation result feedback report, and sends the scheduling allocation result feedback report to the energy storage system management platform and the energy management terminal of each partner; the scheduling allocation result feedback report includes an energy allocation traceability list, an energy storage system operation log, and a scheduling compliance analysis.

9. The vertical take-off and landing point new energy load intelligent scheduling method of claim 8, wherein, After the end of the scheduling period, the cross-period learning optimization layer performs online adaptive optimization based on scheduling execution data, including: After the end of the scheduling period, the cross-period learning optimization layer first collects all panoramic operation data generated during the last scheduling period from the sensors of the power supply circuit device, the energy storage system management platform, and the energy management terminal of each partner, performs feature extraction, and generates a cycle performance feature portrait; The cross-period learning optimization layer inputs the cycle performance feature portrait into a preset multi-objective loss function, calculates a quantitative loss value, combines the quantitative loss value to form a cycle performance evaluation vector; the quantitative loss value includes system energy efficiency loss value, partner fairness loss value, and energy storage health loss value; The cross-period learning optimization layer uses the cycle performance evaluation vector as the optimization target and uses the back propagation algorithm to simultaneously adjust the policy network parameters of the deep reinforcement learning agent and the internal weight parameters of the improved Shapley value algorithm; the process of the collaborative adjustment is as follows: the cycle performance evaluation vector is input into the gradient descent algorithm to calculate the parameter adjustment direction that reduces the multi-objective loss function value, and the policy network parameters and the improved Shapley value algorithm weight parameters are updated to generate optimized model parameters; Before deploying the optimized model parameters to the next scheduling period, the cross-period learning optimization layer first performs simulation running on the validation set constructed by historical data, and if the simulation running result shows that the key performance indicators are steadily improved, the optimized model parameters are updated to the scheduling rule library.

10. The vertical take-off and landing point new energy load intelligent scheduling system is used for realizing the vertical take-off and landing point new energy load intelligent scheduling method in any one of claims 1-9, characterized in that, including: a data acquisition module, a dynamic weight calculation module, a scheduling rule construction module, a scheduling execution module, and an adaptive optimization module; The data acquisition module is configured to collect multi-agent energy demand data and perform preprocessing; The dynamic weight calculation module is configured to combine the processed multi-agent energy demand data, fuse the real-time state information of the battery of the energy storage system, and realize dynamic calculation of the energy allocation weight of each partner through the improved Shapley value algorithm to output dynamic energy allocation weights matched with task characteristics and battery states. The scheduling rule construction module is configured to construct new energy scheduling rules based on the dynamic energy allocation weights output by the dynamic weight calculation module and in combination with the real-time task dynamics of each partner, and to decompose the scheduling logic into three layers of scheduling rules. The scheduling execution module is configured to convert the three layers of scheduling rules formed by the scheduling rule construction module into control instructions, schedule and allocate new energy resources at the vertical take-off and landing point, and feed back the allocation results to the energy storage system and each partner in real time. The adaptive optimization module is configured to, after each scheduling period, perform online adaptive optimization based on scheduling execution data fed back by the scheduling execution module, through a cross-period learning optimization layer preset by the scheduling rule construction module, and feed back an optimization result to the scheduling rule construction module.

Images