Cloud-edge collaborative microgrid power management system

Abstract

The application relates to the technical field of new energy and smart grid, and discloses a cloud-edge collaborative microgrid electric energy management system, which comprises a field sensing layer, an edge computing and control layer, a cloud collaborative layer and a user interaction layer. The edge side is provided with multi-modal data access, artificial intelligence local inference, monitoring control and communication management functions, synchronizes the global strategy of the cloud side in the networked state, continuously forms a control decision according to the synchronized local strategy and real-time field data when the cloud communication is abnormal, and executes the control decision by the monitoring control function; meanwhile, the local coordination is carried out on photovoltaic, energy storage, charging load and input-output link. The system can maintain the field data closed loop and control continuity when the cloud-edge communication state changes, and can also consider multi-modal sensing and microgrid energy scheduling. The cloud collaborative layer forms a global strategy of a long time scale and synchronizes the global strategy to the edge side, and the edge side forms a control decision in combination with real-time field data under the constraint of the local strategy formed by the edge side.

Description

Technical Field

[0001] This invention relates to the field of new energy and smart grid control technology, specifically to a cloud-edge collaborative microgrid power management system that can maintain data and control continuity between cloud collaboration and edge local control. Background Technology

[0002] After distributed photovoltaic (PV) systems, energy storage devices, and electric vehicle charging facilities are connected to the distribution side, a single power station simultaneously faces power generation, energy storage, charging loads, and energy exchange needs with the external power grid or adjacent power stations. Existing power management systems typically utilize cloud platforms to aggregate data from multiple sites and calculate global strategies, while simultaneously configuring monitoring and control equipment at the power station to complete data acquisition and command execution. This architecture facilitates unified operation and maintenance and cross-site optimization, but on-site control is highly sensitive to the quality of wide area networks.

[0003] For example, Chinese invention patent application CN113872323A discloses a microgrid cloud-edge collaborative control system based on edge intelligence. This system includes a cloud layer, a microgrid distributed controller, and distributed sensing units. The microgrid distributed controller collects operational status data from wind power generation, distributed photovoltaics, loads, energy storage, and controllable micro-power sources and uploads it to the cloud layer. Based on this data, the cloud layer formulates coordinated control strategies and energy dispatch strategies, which are then identified, invoked, and executed by the microgrid distributed controller. This solution can utilize the cloud layer to store, analyze, and coordinate the management of data from multiple microgrids. However, its main focus is on the cloud layer formulating strategies and the microgrid distributed controller invoking and executing them. It does not address how on-site operational data continues to enter the local control link after cloud-edge communication anomalies, or how the edge side continuously maintains the existing control loop; these aspects are not addressed in the same way as in this application.

[0004] For example, Chinese invention patent application CN113704976A discloses a smart microgrid energy management method based on edge computing. This method establishes a microgrid scheduling optimization model, partitions a polyhedron formed by multiple parameter variables, calculates the mapping relationship between the optimal scheduling output and parameters within each partition, and quickly obtains the output and electricity purchase / sale results of each microgrid unit based on input parameter samples to meet the real-time computing needs of the edge side. This scheme focuses on transforming the scheduling optimization results into a mapping relationship that can be quickly solved at the edge. However, it does not disclose the same specific collaborative processes as this application regarding the state switching between cloud collaboration and edge autonomy, the participation of multimodal unstructured data in field control, and the continuous operation of the field control link during communication anomalies.

[0005] For microgrids with computing capabilities at both the cloud and edge, their technical roles are not simply redundant. The cloud can perform long-term model training and global optimization by combining cross-site status, historical operating data, and operational constraints. However, the power, state of charge, and availability of field equipment may change within short control cycles. If each equipment-level adjustment relies solely on uploading field data to the cloud and waiting for the cloud's calculation results, the round-trip communication will affect the response to rapid changes in field status. The edge, deployed at the energy consumption site, can directly read real-time field data and form current control decisions within the constraints of the cloud's global strategy. However, its on-site computing power and available data range are not suitable to completely replace the cloud in handling cross-site, long-term simulations. Therefore, it is necessary to clarify the hierarchical relationship and data connection methods between the cloud's long-term global strategy and the edge's real-time on-site decisions.

[0006] For microgrids requiring continuous operation of photovoltaic-storage-charging systems, when cloud-edge communication fails, the site still needs to continuously obtain the status of devices such as smart meters, inverters, energy storage converters, charging piles, and environmental sensors, and maintain necessary power control. On the other hand, communication failures may be accompanied by message retransmissions, protocol session reconstruction, and changes in the timing of multi-source data. If the data access path, control permissions, and local policy takeover relationships are not clearly constrained, control link interruptions, inconsistent data timing, or the same device being controlled repeatedly by different policies can easily occur.

[0007] Furthermore, microgrids contain both structured data such as voltage, current, power, and state of charge, and potentially unstructured data such as on-site video footage used to identify photovoltaic module shading and equipment malfunctions. When different data are involved in real-time control, their sampling frequencies, computation paths, and data dimensions vary. The scheduling of output and input links also involves parameters such as target type, unit power gain, line losses, acceptable power, and equipment adjustment resolution. Therefore, a feasible and unified design is needed for the data paths during changes in cloud-edge communication status, local policy execution, the processing relationships of data across different dimensions, and the microgrid energy scheduling process. Summary of the Invention

[0008] The purpose of this invention is to provide a cloud-edge collaborative microgrid power management system. This system enables the field edge to synchronize strategies and execute collaborative control with the cloud when cloud-edge communication is normal. Even when cloud-edge communication is abnormal, it can still maintain a control loop by relying on real-time field data and synchronized local strategies. Simultaneously, it establishes clear data processing and execution relationships for multimodal data, photovoltaic maximum power point tracking, and input / output link scheduling. Specifically, the cloud-edge collaboration layer is used to form a global optimization strategy over a longer timescale and across multiple sites. The edge computing and control layer is used to form a first optimization decision based on the constraints of the local strategy formed by the global optimization strategy and real-time field data, thus enabling a layered collaboration between global strategy updates and real-time field control.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A cloud-edge collaborative microgrid power management system includes a field sensing layer, an edge computing and control layer, a cloud-to-edge collaboration layer, and a user interaction layer. The edge computing and control layer includes an intelligent edge energy control device deployed at the energy consumption site. This device includes an AI computing power localization and real-time strategy inference module, a multimodal data access module, and a SCADA server. The multimodal data access module accesses and analyzes the multimodal operation data of the microgrid collected by the field sensing layer to obtain real-time data. The AI ​​computing power localization and real-time strategy inference module generates a first optimization decision based on a local strategy and the real-time data. The SCADA server executes the first optimization decision. In the event of a cloud-edge communication anomaly, the AI ​​computing power localization and real-time strategy inference module continues to generate the first optimization decision based on the strategy updated locally before the communication anomaly and the real-time field data, to maintain the local control closed loop of the microgrid. The local strategy is synchronized and updated by the global optimization strategy formed by the cloud-to-edge collaboration layer, used to define the control boundary of the first optimization decision. The first optimization decision is formed within the control boundary in conjunction with the real-time field data.

[0010] Optionally, the field sensing layer collects multimodal operation data of the microgrid; the edge computing and control layer processes the multimodal operation data locally and sends the processed operation data to the cloud collaboration layer; the cloud collaboration layer generates a second optimization decision based on the received operation data and sends the second optimization decision to the edge computing and control layer for execution; the user interaction layer is used to display the microgrid's operating status, charging information, and the execution results of the first and / or second optimization decisions. The second optimization decision can be used to form or update the local strategy, or it can serve as a control target at the edge; the edge determines the first optimization decision based on real-time field data and the current status of the equipment before actual execution.

[0011] Furthermore, the second optimization decision being sent to the edge computing and control layer for execution means that after the edge computing and control layer completes the reception, version, and device applicability verification, it uses the verified second optimization decision to form or update the local policy and / or as a control target; the second optimization decision does not bypass the first optimization decision and directly writes the control quantity to the field device, and the actual control of the field device is formed by the edge computing and control layer in combination with the real-time field data and then executed by the SCADA server.

[0012] Optionally, the cloud collaboration layer includes an EMS cloud platform and a charging operation and maintenance platform; the EMS cloud platform is used for model training and generating the second optimization decision, and the charging operation and maintenance platform is used for charging operation management, user interaction, system operation and maintenance, and alarm management. The model training and optimization decisions performed by the EMS cloud platform are oriented towards a longer time scale and cross-site operating status, and are used to provide global optimization strategies to the edge side.

[0013] Furthermore, the EMS cloud platform trains the model using historical operating data and corresponding strategy execution results to form model parameters or rule parameters for cross-timescale prediction or candidate strategy evaluation; when generating the second optimization decision or global optimization strategy, the model parameters or rule parameters are used to obtain the prediction results of the operating status in subsequent time periods and / or the evaluation results of candidate strategies, and the strategy parameters, constraints and / or control objectives are determined in combination with equipment parameter ranges, station capacity and operational constraints.

[0014] Optionally, the user interaction layer includes a mobile terminal and / or a web terminal.

[0015] Optionally, the intelligent edge energy control device further includes a firewall module, a data preprocessing module, a communication management module, and a network switching module; the communication management module is used to access the operating data collected by the field perception layer, the data preprocessing module is used to preprocess the accessed operating data, and the SCADA server sends the processed operating data to the cloud collaboration layer through the network switching module.

[0016] Optionally, at the moment of cloud-edge communication failure and within a preset time thereafter, the operational data accessed by the communication management module is transferred to the multimodal data access module to execute the local network outage self-governance process; after the preset time, the multimodal data access module directly accesses the microgrid multimodal operational data collected by the field sensing layer.

[0017] Furthermore, after cloud-edge communication is restored, the edge side first uploads key events and current device status during the network outage, and then verifies the version, device identifier and parameter range of the cloud return policy. After the verification is passed, the cloud-edge collaboration mode is restored.

[0018] Optionally, the AI ​​computing power localization and real-time strategy reasoning module extracts on-site video data of photovoltaic modules from the real-time data, and identifies photovoltaic module shading or foreign object coverage in the on-site video data of photovoltaic modules; when photovoltaic module shading or foreign object coverage is identified, the maximum power point tracking strategy of the corresponding inverter is switched from the incremental conductance method to the intelligent hybrid algorithm.

[0019] Furthermore, the on-site video data of the photovoltaic module is associated with the electrical quantities of the same photovoltaic branch to which the corresponding inverter belongs, according to the equipment identification; the video recognition result is used to trigger the maximum power point tracking strategy switching only when the time difference between the two does not exceed the allowable time window and the shading or foreign object coverage reaches the configuration trigger condition.

[0020] Optionally, the intelligent hybrid algorithm includes: collecting the light intensity, ambient temperature, output voltage, and output current of the photovoltaic module; forming a four-dimensional input vector through A / D conversion; performing sliding window filtering and normalization on the four-dimensional input vector; determining the step size adjustment coefficient and performing coarse adjustment based on the processed data using a fuzzy logic algorithm; searching for the operating point using a particle swarm optimization algorithm; and finally using a neural network to dynamically correct the operating point.

[0021] Furthermore, when the rate of change of occlusion area in adjacent visual analysis cycles is higher than the configured threshold and the particle swarm fitness variance is lower than the configured threshold, zero-mean random perturbation is added to some particle positions and the inertia weight is temporarily increased to trigger a re-search for other power peaks.

[0022] Optionally, the microgrid includes smart meters, photovoltaic modules and inverters, energy storage units, charging piles, environmental sensors, output links, and input links.

[0023] Optionally, the AI ​​computing power localization and real-time strategy reasoning module analyzes the real-time data to determine the power generation, the remaining storable power of the energy storage unit, and the power consumption of the charging pile, and determines the power status accordingly. When the power status is a pending output status, the output link connection order corresponding to the local strategy is retrieved and the output link is connected according to the output link connection order. The output link connection order determines the first priority value, the second priority value, the third priority value, and the fourth priority value according to the power demand target type, the equity value after the target response, the link loss parameter, and the storable power of each output link, and sorts them from largest to smallest according to the sum of the first priority value, the second priority value, the third priority value, and the fourth priority value.

[0024] Furthermore, the equity value after the target response is the dimensionless equity score obtained by mapping the unit external transmission power revenue fed back after the corresponding output link completes the target response through the preset quantization table of the current strategy version; before executing the output link connection, a physical feasibility check is performed based on the link availability status and acceptance limit. If a high-priority output link refuses to access or is temporarily unavailable, the next output link will be checked.

[0025] The present invention has achieved the following beneficial effects: This invention forms a control closed loop at the field edge, seamlessly integrating multimodal data access, local policy inference, and SCADA execution. This ensures that even when cloud-edge communication is abnormal, necessary control can still be maintained based on locally updated policies and real-time field data. By clearly defining the data transfer relationship during the initial network outage and the subsequent direct access relationship for field data, the data required for local control during communication status changes has a definite source. Simultaneously, the invention utilizes photovoltaic field video recognition results to trigger maximum power point tracking (MPPT) policy switching and determines the output link connection order based on preset quantization rules, which improves the adaptability of power point tracking under partial shading conditions and the continuity of power transmission scheduling from the power plant. Furthermore, by layering the global policy formation on a longer timescale in the cloud with local decision-making within the field control cycle on the edge, the cloud can update policies using cross-site data and over longer time periods, while the edge directly forms execution quantities based on the real-time field status under predetermined policy constraints. This reduces the sensitivity of rapid changes in the field to cloud-to-cloud communication latency and complements the cloud's global optimization capabilities with the edge's real-time response capabilities. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall structure of the cloud-edge collaborative microgrid power management system in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the types of operational data collected by the field sensing layer in an embodiment of the present invention; Figure 3 This is a schematic diagram of the first topology of the intelligent edge energy control device and the cloud collaboration layer in an embodiment of the present invention; Figure 4 This is a schematic diagram of the second topology of the intelligent edge energy control device, field equipment, and cloud platform in an embodiment of the present invention. Detailed Implementation

[0027] The embodiments of the present invention will now be described with reference to the accompanying drawings. The following embodiments are presented sequentially according to the progressive relationship of the technical solutions of this application, and the same term has the same meaning in different embodiments. The parameters, interfaces, and software functions described in each embodiment can be configured according to the scale of the site and the equipment model; unless otherwise specified, safety protection, equipment rated power, state of charge boundary, and grid connection point restrictions are always considered as hard constraints higher than economic dispatch.

[0028] Example 1: As Figure 1 As shown, this embodiment provides a cloud-edge collaborative microgrid power management system, including a field perception layer, an edge computing and control layer, a cloud-to-edge collaboration layer, and a user interaction layer. The field perception layer is located at the microgrid site and is used to continuously acquire operational data from power generation, energy storage, loads, and environmental objects. The edge computing and control layer is deployed at the energy consumption site, with an intelligent edge energy control device at its core. The cloud-to-edge collaboration layer is used for long-term model training, cross-site analysis, and optimization decision-making. The user interaction layer is used to display the operating status and decision execution results to maintenance personnel or users.

[0029] In this application, the cloud collaboration layer and the edge computing and control layer undertake interconnected control tasks according to different time scales. The cloud collaboration layer forms a second optimization decision or global optimization strategy based on data from a longer time scale and cross-site status. This second optimization decision or global optimization strategy provides applicable strategy parameters, constraints, and / or control objectives for subsequent multiple field control cycles. The edge computing and control layer forms or updates this into a local strategy and generates a first optimization decision under the constraints of the local strategy, combined with current real-time field data. Therefore, the first optimization decision is used to implement the global strategy formed in the cloud into a field control decision that matches the current site and current equipment status. These two layers represent collaborative decisions at different levels but with continuous data and strategy relationships.

[0030] In one optional data configuration method, when the cloud side forms or updates the global optimization strategy, it can use pre-obtainable or predictable site operation and management information, including one or more of the following: time-of-use electricity price for the current and subsequent periods, number of charging orders or demand saturation, site load, site capacity or external scheduling constraints, allocable power of charging piles, and state of charge or remaining storable capacity of energy storage units. The cloud side determines the strategy boundary based on the above information and then distributes it to the edge side. The edge side does not need to repeat complete cross-site or long-term scale simulations in each field control cycle; instead, it reads real-time data such as current voltage, current, power, SOC, equipment connection status, and protection status, and determines the actual execution value of the equipment within the strategy boundary. The above list of inputs or constraints is used to illustrate one data configuration method for the global optimization strategy and is not considered a necessary technical feature of the cloud-edge hierarchical collaborative relationship limited by this application.

[0031] The intelligent edge energy control device includes at least an AI computing power localization and real-time strategy inference module, a multimodal data access module, and a SCADA server. The multimodal data access module receives multimodal operational data collected by the field perception layer, saves the data source, acquisition time, and device identifier, and forms real-time data that can be used by local strategies. The AI ​​computing power localization and real-time strategy inference module reads the most recently validated local strategy and the current real-time data to generate a first optimization decision. The first optimization decision may include one or more of the following: inverter power target, energy storage charging and discharging power, charging pile power limit, or energy link connection command. The SCADA server converts the first optimization decision into a control quantity executable by the corresponding device within the device's safety boundary and issues it. Here, the local strategy is used to store the strategy parameters, constraints, and control objectives that have been issued and adopted by the edge side in the cloud-based global optimization strategy, while the first optimization decision determines the specific execution quantity based on the real-time state of the current field control cycle.

[0032] In network-connected mode, the edge device continues to execute local rapid control while simultaneously synchronizing locally processed operational data to the cloud collaboration layer. The cloud, based on cross-site status, historical data, and operational constraints, formulates a second optimization decision or updated policy parameters and distributes them to the edge device. Upon receiving the cloud policy, the edge device first verifies the version, device identifier, and parameter range. Only policies that pass verification are added to the local valid policy set, thus establishing a definite handover relationship between cloud decisions and on-site rapid control. For second optimization decisions issued in the form of framework policy parameters or constraints, the edge device re-determines the first optimization decision within the corresponding policy parameters and security constraints, combined with current on-site data. For cases where the cloud has provided specific control objectives, the edge device still confirms the actual execution value based on the actual on-site status and device safety boundaries. Therefore, the cloud device is responsible for formulating policies over longer timescales and across sites, while the edge device is responsible for translating these policies into on-site control quantities for the current site.

[0033] To further unify the execution levels of the second and first optimization decisions, in this application, the second optimization decision is sent to the edge computing and control layer for execution. This means that the edge computing and control layer receives the second optimization decision, verifies its version and device applicability, and uses the verified content as local policy update content or control target to enter the subsequent field decision chain. The second optimization decision does not bypass the first optimization decision and directly writes control quantities to the inverter, energy storage converter, charging pile, or link switch. The actual control quantities of the field devices are still formed by the edge side in combination with the current real-time field data to form the first optimization decision, which is then converted and distributed by the SCADA server.

[0034] When cloud-edge communication fails, the edge device does not wait for the cloud to re-establish the connection. Instead, it continues to generate an initial optimization decision using valid policies and real-time field data that were updated locally before the communication failure. The SCADA server then executes this initial optimization decision, maintaining a closed loop between field sensing, edge analysis, policy reasoning, and control execution. After communication is restored, the edge device first uploads key events and the current device status during the network outage, then verifies the policy returned by the cloud, and finally resumes the cloud-edge collaborative mode, avoiding the direct overwriting of the actual field status with outdated cloud status upon recovery.

[0035] The technical advantage of this embodiment lies in deploying real-time data access, local policy inference, and SCADA execution within a closed-loop edge control link in the field. The cloud primarily undertakes global optimization over longer timescales and across sites, allowing it to issue control objectives while the edge handles real-time decision-making and execution in the field. This solves the technical problem of control link interruptions and the inability of the microgrid to maintain basic autonomous operation due to continuous reliance on cloud commands during cloud-edge communication anomalies. Simultaneously, the cloud maintains its long-term, cross-site optimization capabilities, while the edge utilizes its proximity to field data sources and execution devices to complete short-cycle control decisions, avoiding the need for cloud communication round trips as a necessary step for rapid field adjustments. This improves the timeliness of response to changes in field conditions while maintaining global policy consistency.

[0036] Example 2: Building upon Example 1, this example further illustrates the data flow between the field perception layer, edge computing and control layer, cloud collaboration layer, and user interaction layer. The multimodal operational data collected by the field perception layer is first processed at the edge side, including device identification association, timestamp organization, and validity verification. The edge side directly uses this data to form a first optimization decision, while simultaneously sending data suitable for cloud processing to the cloud collaboration layer. The cloud collaboration layer generates a second optimization decision based on the received data and returns it to the edge side. The edge side confirms whether the second optimization decision meets the execution conditions based on the policy version and the current device status. In this data flow, the second optimization decision can be represented by a control objective given by the cloud, or by policy parameters or constraints used to update the local policy. Regardless of the form, the edge side determines the final first optimization decision to be executed in this control cycle based on real-time field data and the current device status.

[0037] For data with different sampling periods, the edge side sets the decision time and allowable data age for each strategy calculation. The SCADA server uses the device sampling timestamp as the primary index and selects the most recent valid value no more than the maximum data age prior to the decision time. For video recognition results, both the video frame acquisition time and the recognition completion time are recorded. When the time difference between the electrical quantity of the same branch and the video result exceeds the allowable time window, the video result is only used as a status record and does not directly trigger power control for the current cycle, thereby avoiding mismatching data from different physical times as the same field state.

[0038] The user interaction layer displays the station's operating status, charging information, and the execution results of optimization decisions. Execution results prioritize the actual execution values ​​returned by the SCADA server, rather than simply displaying algorithmic candidate values. When candidate control quantities are pruned due to equipment rated power, SOC safety boundaries, grid connection point limitations, or protection status, the SCADA server records the pruning reason and the actual execution value and sends it back to the strategy inference module and the user interaction layer, ensuring that the next control cycle is based on the actual execution status of the previous cycle.

[0039] The technical advantage of this embodiment lies in creating a data closed loop with timestamps and execution feedback between on-site data, cloud-based decisions, edge execution results, and user display. This solves the technical problems that can easily lead to misjudgments and accumulated deviations in subsequent control when the sampling times of multi-source data are inconsistent and the cloud-based candidate decisions are inconsistent with the actual on-site execution state. Therefore, the actual on-site execution results can serve as the state basis for generating the first optimization decision in the next on-site control cycle, and can also be transmitted back to the cloud as the data basis for subsequent strategy updates on longer time scales, enabling the fast-time-scale on-site feedback and the slow-time-scale global optimization to be interconnected.

[0040] Example 3: Building upon Example 2, the cloud-based collaboration layer in this example includes an EMS cloud platform and a charging operation and maintenance platform. The EMS cloud platform receives device status, historical operating data, and policy execution results from the edge side for model training, cross-timescale prediction, and the generation of second optimization decisions. Model training can be performed offline in the cloud. After training, deployable model parameters or rule parameters are generated and distributed to the edge side after version management. Local runtime does not require the edge side to perform large-scale retraining during network outages; it only performs forward computation or rule inference of the deployed model. The cloud side can perform inferences based on historical operating status and cross-site resources over a longer timeframe than a single field edge side. The edge side mainly stores the distributed models, policy parameters, and short-term inference capabilities required for field control, so that the field control cycle does not require waiting for the cloud to re-complete the global inference.

[0041] Between model training and the second optimization decision, the EMS cloud platform employs a traceable conversion link: historical equipment status, historical load and generation status, historical strategy parameters, and corresponding actual execution results are used as training or rule calibration data to obtain model parameters or rule parameters for cross-timescale prediction or candidate strategy evaluation; when generating the current second optimization decision, the current operating data is input into the model or rule to obtain the load, generation, or adjustable resource status prediction results and / or candidate strategy evaluation results for subsequent time periods, which are then combined with equipment parameter ranges, station capacity, external scheduling constraints, and operational constraints to form strategy parameters, constraint conditions, and / or control objectives. Thus, the model training results are first transformed into prediction or evaluation results, and then constrained and integrated to form the second optimization decision or global optimization strategy, rather than the training actions themselves directly generating field equipment control quantities.

[0042] The charging operation and maintenance platform is used for charging operation management, user interaction, system maintenance, and alarm management. Operational information and real-time on-site control information are logically separated: charging orders, service status, and cost information are managed by the operation platform, while the current adjustable power, protection status, and connection status of the charging equipment are confirmed by the on-site perception layer and SCADA server. When the cloud generates a second optimization decision involving charging power, the edge side still uses the actual adjustable range of the on-site equipment as the final execution boundary.

[0043] The cloud-deployed policies are managed using version identifiers, applicable device sets, parameter ranges, generation times, and verification values. Upon receiving a new version, the edge device performs syntax checks, device identifier matching, and parameter boundary checks without affecting the execution of the currently valid policy. Only if the candidate policy covers a device and the control parameters do not exceed the allowed range for that device is it marked as a new valid policy. If the verification fails, the edge device continues to use the previous valid version and generates an anomaly record for the charging operation and maintenance platform.

[0044] The technical advantage of this embodiment lies in clearly defining the functional boundaries between the EMS cloud platform and the charging operation and maintenance platform, and securely converting cloud training results into executable edge-side policies through policy version verification. This solves the technical problem of unexecutable control quantities arising from changes in equipment status or version inconsistencies when cloud models or operation policies are directly distributed to the field. This division of labor further establishes a clear hierarchy between cloud-side policy updates and edge-side real-time execution, preserving the cloud's longer prediction time domain and cross-site computing capabilities while avoiding complete reliance on wide area network round trips for rapid equipment control.

[0045] Example 4: Building upon Example 2, the user interaction layer in this example includes a mobile terminal and / or a web terminal. The mobile terminal is suitable for viewing site overviews, charging status, equipment alarms, and operational information requiring on-site confirmation; the web terminal is suitable for displaying multi-device timing curves, policy versions, network outage events, and maintenance records. Both types of terminals obtain display data through a cloud platform or authorized field interfaces, without directly bypassing the SCADA server to write control values ​​to the power devices.

[0046] When cloud-edge communication is normal, the user interaction layer prioritizes displaying the aggregated status on the cloud side, while also indicating the last actual execution result on the edge side. When cloud-edge communication is abnormal, if the mobile terminal or web terminal can still access the edge side through the local operations and maintenance network, it displays the local autonomy status, the current effective policy version, the status of critical devices, and the duration of the network outage; if access is not possible, the last confirmed status is retained on the cloud interface, and this status is distinguished from the real-time status by a communication abnormality indicator to avoid mistaking outdated data for the current status.

[0047] For policy changes requiring manual intervention, the user interface layer is only responsible for submitting change requests and displaying the review results. The actual policy writing is completed by edge-side permission verification and security boundary checks, and confirmation is returned by the SCADA server or policy inference module after writing. Thus, a clear permission isolation is maintained between the user interface and the underlying power control.

[0048] The technical advantage of this embodiment is that it provides a unified interaction method for different terminals that can distinguish between real-time status and final confirmation status, while maintaining the isolation between the human-machine interface and the field control interface; thereby solving the technical problems of interface status distortion during network outages and misoperation caused by user interaction functions directly interfering with the underlying control.

[0049] Example 5: Based on Example 1, as follows Figure 3 and Figure 4 As shown, the intelligent edge energy control device further includes a firewall module, a data preprocessing module, a communication management module, and a network switching module. The communication management module is used for field protocol adaptation and operational data access, converting raw messages into data records with device identifiers and timestamps based on device addresses, communication protocols, and point tables. The data preprocessing module performs out-of-bounds value marking, duplicate message removal, time alignment, and short-term missing value preservation on structured power data. The network switching module establishes local data channels between the SCADA server, policy inference module, multimodal data access module, and cloud interface. The firewall module restricts unauthorized network access.

[0050] Figure 3 and Figure 4The device marked AIMGC corresponds to the intelligent edge energy control device described in this application. This marking is only the device name in the engineering drawing and does not represent another independent functional entity. The above modules can run together on a single edge industrial control computer, or they can be distributed across a communication management machine, an edge server, and a SCADA server connected via a field LAN, as long as the operating data can enter the local inference and execution link on-site.

[0051] In one optional implementation, the communication management module sets up a ring buffer written in chronological order when the cloud-edge link status changes. The buffer records include device identifier, original timestamp, received timestamp, message sequence number, and payload. This ring buffer is used to absorb short-term bursts of traffic caused by protocol reconnection or message retransmission. The buffer medium can be existing memory or non-volatile storage on the edge device and does not constitute the essential technical features defined in the claims of this application. After communication is restored, necessary deduplication and sequence restoration are performed based on the timestamp and message sequence number, and then the data to be uploaded is synchronized to the cloud.

[0052] To prevent erroneous data from propagating to the control link when a local anomaly occurs in the communication module, the data preprocessing module sets effective ranges and maximum data ages for critical control variables. Data exceeding the physical allowable range or the data age is not included in the current policy calculation but is marked as abnormal for maintenance traceability. Device protection events have higher priority than link economic scheduling and manual policy updates. In the event of energy storage overvoltage, inverter protection shutdown, grid connection point exceeding limits, or charging equipment emergency stop, the SCADA server first executes the corresponding protection logic and removes the affected device from the adjustable resource set.

[0053] The technical advantage of this embodiment is that it uses communication access, data preprocessing, network switching and access isolation to jointly constitute the edge data entry point, and in optional cases, it uses a cache to absorb short-term burst messages caused by changes in communication status; thereby, it can solve the technical problems of lack of unified verification of multi-protocol field data before entering the local control link and the interference of communication jitter and message duplication on the continuity of real-time control data.

[0054] Example 6: Building upon Example 5, this example further illustrates the data path when cloud-edge communication is abnormal. The edge side can use heartbeat messages and service connection status to jointly determine whether communication is abnormal. When multiple consecutive heartbeat cycles fail to receive a valid response and the service connection reconnection reaches an abnormal condition, a communication abnormality event is generated; when a single heartbeat is lost but service messages continue to arrive, the original collaborative state can be maintained, thereby avoiding frequent mode switching caused by short-term network jitter.

[0055] During the moment of communication failure and for a preset period thereafter, the communication management module continues to access on-site operational data and directly transfers data requiring local control to the multimodal data access module, ensuring that local policy inference and SCADA execution links do not cease due to cloud link interruption. The preset time can be determined based on the site's communication reconnection mechanism, for example, within the range of 1 to 20 minutes allowed by the original engineering configuration. The purpose of the preset time is to cover the protocol reconstruction and message burst phases that may occur in the initial stage of communication failure, rather than waiting for cloud recovery before initiating local control.

[0056] After a preset time, the multimodal data access module directly accesses the multimodal operational data collected by the field sensing layer. "Direct access" here means that the data required for control no longer needs to be forwarded from the cloud; it can reach the multimodal data access module via the field LAN, RS-485, CAN, or the device's internal switching channel. The communication management module can still handle general protocol management and non-critical data access, and does not necessarily need to completely exit the system.

[0057] In one optional enhanced implementation, in addition to the preset time, the system can also determine whether a stable autonomous state has been entered by combining the cache backlog and fieldbus occupancy rate: when the number of pending messages is below the configured low water level for several consecutive detection cycles and the fieldbus occupancy rate recovers to the allowable range, a switch to the field direct connection path can be allowed ahead of schedule or confirmed. This enhanced judgment is only used to improve the smoothness of the project switchover and does not change the basic relationship of "initially transferring data through the communication management module after a network outage, and then directly accessing field data through the multimodal data access module".

[0058] If the network outage lasts for an extended period and on-site personnel are required to modify local policies, a controlled authorization window can be further configured. Let the historical forward update count be... The number of times historical regressions or protection alarms were triggered was [number]. The forward gain step size and the reverse penalty step size are respectively and ,and The staff reputation weight can be updated in a bounded manner using the following formula: ; in, For reputation weight, and These are its lower limit and upper limit, respectively. This indicates that the calculation results are restricted to a given interval, and the constant 1 in the formula represents the normalized baseline value of the reputation weight. Both the forward and reverse iterations are derived from historical update records that have already undergone online verification. Let... Based on the allowable risk value, This represents the number of authorized windows that have been fully accessed since the internet connection was disconnected. To reduce the risk of increasing window size, If credit weight is the adjustment metric, then the current permissible risk threshold can be expressed as: ; In the formula, The current permissible risk threshold, and These are the lower and upper limits of the risk threshold, respectively. , , The risk values ​​in the strategy library are measured using the same dimensionless scale. The switching risk value of the strategy to be updated reaches or exceeds [a certain threshold]. Write operations are denied when the threshold is reached, and are only permitted when the device's security boundaries are met. This authorization control is an optional security enhancement for network-disconnected autonomy and does not alter the data access switching relationship defined in the claims of this embodiment.

[0059] The switching risk value of the proposed update strategy is generated by the edge side before each write verification based on the difference between the current effective strategy and the proposed update strategy. The edge side reads the change range of control parameters involved in the proposed update strategy, the number of affected adjustable devices, and the remaining margin of the target control quantity relative to the safety boundary of the corresponding device. It normalizes the change range of each control parameter according to the allowable adjustment range of the corresponding parameter, normalizes the number of affected adjustable devices according to the current total number of adjustable devices, and converts the remaining margin into the degree of safety margin occupancy. Then, it weights and sums the above dimensionless quantities according to the non-negative weights preset in the current strategy version and limits the sum to 0 to 1 to obtain the switching risk value. The weights remain fixed within the same strategy version. Device protection status, rated power, SOC safety boundary, and grid connection point restrictions remain as independent hard constraints and are not replaced by the switching risk value. Therefore, the switching risk value has a definite data source, calculation method, and generation time.

[0060] The technical advantage of this embodiment is that it establishes a clear sequential relationship between data transfer in the initial stage of network outage and direct on-site access in the stable autonomous stage, and allows for the addition of status determination and controlled manual update mechanisms without changing the basic network outage autonomous scheme. This solves the technical problems of unclear on-site data entry after cloud-edge link interruption, control link waiting for cloud recovery, and lack of risk constraints for manual modification of strategies during long-term network outages.

[0061] Example 7: Building upon Example 1, this example utilizes on-site video data from photovoltaic modules to participate in the maximum power point tracking (MPPT) strategy switching. The multimodal data access module associates video frames with electrical quantities of the same photovoltaic branch according to device identifiers. The video recognition function uses the module or string area as the target area and calculates the proportion of the target area covered by shading or foreign objects. When shading or foreign object coverage is detected and the configuration trigger condition is met, the localization of artificial intelligence computing power and real-time strategy reasoning module switches the corresponding inverter's MPPT strategy from the incremental conductance method to the intelligent hybrid algorithm.

[0062] After entering the intelligent hybrid tracking process, the illuminance of the same photovoltaic branch is obtained. Ambient temperature Output voltage and output current In the first At each sampling time, the voltage change, output power, and power change at adjacent times are determined according to the following formula: ; in, and They represent the first Photovoltaic output voltage and output current at each sampling time. and These represent the photovoltaic output voltage and photovoltaic output power at the previous sampling time, respectively. This indicates the corresponding output power. This represents the voltage change at adjacent sampling times. This represents the power change between adjacent sampling times. The output current is not an isolated additional input, but rather, together with the output voltage, it forms the real-time power and power change, and is used for subsequent operating point evaluation.

[0063] Fuzzy logic pre-positioning stage and For input, the input quantities are dimensionless within their respective historical or calibrated ranges. Among them, This indicates the voltage change during the current sampling period. This represents the power change during the current sampling period, consistent with the meanings of the voltage and power changes with sampling time indices mentioned earlier. The membership set can be set to negative maximum, negative minimum, zero, positive minimum, and positive maximum, with the regular direction centered around the point of maximum power. Based on this, the derivative term represents the local rate of change of photovoltaic output power relative to output voltage. The output of the fuzzy logic is a perturbation step size adjustment coefficient, which is used to limit the candidate voltage range of subsequent particle swarm search and set the initial perturbation scale, rather than directly replacing the final operating voltage of the inverter.

[0064] When the video recognition result recovers to an unobstructed state and continues to meet the recovery confirmation conditions, the incremental conductance method can be switched back. If the time difference between the video result and the electrical quantity exceeds the allowable time window, the algorithm switching will not be triggered in this cycle to avoid using outdated images to affect the current electrical control. The inverter voltage, current, and power protection boundaries remain effective before and after the algorithm switching.

[0065] The technical advantage of this embodiment is that it correlates video information that reflects the spatial state of local shading with electrical quantities of the same photovoltaic branch, and uses the shading identification result as the switching condition for the maximum power point tracking strategy; thereby it can solve the technical problem that it is difficult to distinguish between the multi-peak power characteristics caused by local shading and ordinary irradiance fluctuations when relying solely on electrical quantities.

[0066] Example 8: Based on Example 7, this example further illustrates the intelligent hybrid algorithm. The collected light intensity, ambient temperature, output voltage, and output current are converted into a four-dimensional input vector after A / D conversion, and a sliding window filter is used to suppress high-frequency noise; each input is mapped to a corresponding calibration interval. Normalization only changes the numerical scale, not the origin and meaning of the four physical quantities. In the fuzzy logic stage, the coarse adjustment direction and step size are first determined, and then particle swarm optimization is performed within the corresponding operating voltage range.

[0067] For the first The particle, at the _ , in the _ ... In the next iteration, velocity and position can be updated according to the following formula: ; in, and They represent the first The particle in the first Candidate operating voltage and speed at the next iteration; and These represent the candidate working voltage and velocity of the particle after the next iteration update, respectively. This indicates the particle's historical best position. Indicates the optimal position of the group. and As a learning factor, and for Internal random number, For the first The inertia weights for each iteration. Candidate positions are always limited to the inverter's allowable operating voltage range, with the output power at the corresponding candidate voltage used as the fitness and increasing power as the preferred direction.

[0068] The number of particles, inertia weight, and maximum number of iterations can be configured according to the original edge computing power of the power station. For example, 20 to 30 particles can be set, with the inertia weight gradually decreasing from an initial value in the range of 0.8 to 1.0 to a range of 0.4 to 0.6. The iteration ends early when the relative change in optimal power is lower than the stopping threshold after several consecutive iterations; otherwise, the maximum number of iterations is not exceeded. The above values ​​are only one configuration method, and the actual values ​​are determined based on the inverter control cycle and the edge processor computing time.

[0069] In one alternative implementation, the change in the video occlusion area is used to determine whether a perturbation needs to be added to the particle swarm. Let the... The occlusion area ratio for each visual analysis cycle is The time interval between adjacent visual analysis cycles is The rate of change of the shading area is: ; in, This represents the occlusion area ratio of the previous visual analysis cycle. The absolute value sign in the formula indicates that the absolute value of the difference between the occlusion area ratios of two adjacent visual analysis cycles is taken. Indicates the first The rate of change of occlusion area corresponding to each visual analysis cycle. When When the particle swarm fitness variance is higher than the configured threshold and lower than the configured threshold, zero-mean random perturbation can be added to some particle positions and the inertia weight can be temporarily increased. This mechanism is used to improve the ability to search for other power peaks again under conditions of rapid changes in local occlusion, and is an optional enhancement to the original particle swarm search.

[0070] After candidate operating points are obtained through particle swarm optimization (PSO), a neural network is used for dynamic correction. The neural network can employ a three-layer feedforward structure, using light intensity, ambient temperature, and current output voltage as direct inputs, and outputting a reference operating voltage. The output current has already been used in the search based on real-time power and PSO fitness. Training samples consist of historical environmental quantities, electrical quantities, and stable maximum power point operating voltages. The model is trained offline in the cloud or at the edge and then synchronized locally. During network outages, only forward inference is performed. If the deviation between the PSO result and the neural network reference value exceeds a correction threshold, a limiting correction is performed within the inverter's allowable range.

[0071] The technical advantage of this embodiment is that it narrows the search range through fuzzy logic, searches for power peaks within the allowable voltage range using particle swarm optimization and performs continuous correction by a neural network, while retaining the optional re-search capability when occlusion changes rapidly. This solves the technical problems that traditional unidirectional perturbation tends to stay at local peaks when there are multiple peaks in the power curve under partial occlusion, and that tracking lags when the environment changes rapidly.

[0072] Example 9: Based on Example 1, as follows Figure 2 As shown, the microgrid in this embodiment includes smart meters, photovoltaic modules and inverters, energy storage units, charging piles, environmental sensors, output links, and input links. Smart meters are used to obtain voltage, current, power, and electrical energy at the grid connection point or key nodes; photovoltaic inverters provide DC and AC side operating data and receive active power or operating point control quantities; energy storage units provide SOC, remaining storable energy, chargeable / dischargeable power, and alarm status; charging piles provide connection status, current charging power, and adjustable range; and environmental sensors provide environmental quantities such as light intensity and temperature.

[0073] For the input link, the local policy determines whether to allow external power access based on the power generation capacity, load capacity, and remaining storable capacity of the energy storage at the local power station. Let the current power generation capacity be... The current load power is The remaining storable electricity is To avoid When approaching zero, division by zero occurs, introducing a positive number no greater than the energy corresponding to the energy storage capacity measurement resolution. The power consumption parameter can be expressed as: ; in, For parameters related to battery level adjustment; This indicates that there is a power shortage on site. This indicates that there is a surplus of electricity generated on-site. It is only used to limit the numerical boundaries of the denominator and does not change the direction of change within the normal remaining capacity range. The upper limit of allowed access power can be determined according to... Determined by a monotonic relationship that is inversely proportional to the remaining storable proportion, one possible implementation is: ; In the formula, To limit the maximum allowed access power, As the reference access power, For the rated available capacity of energy storage, This represents the maximum access power allowed by both the input link and the site hardware. Indicates the range limit; To and A power regulation coefficient with matching dimensions enables It has the dimension of power. For the power scaling factor, make It has power dimensions. The two coefficients are calibrated by the safe access power corresponding to different power gaps and remaining energy storage capacity during the trial operation of the power station.

[0074] To ensure that the aforementioned alternative implementations satisfy the monotonic relationship, both kψ and kE are non-negative, and this formula is only used within the pre-defined access judgment working interval of the power station. The working interval at least defines the maximum allowable power gap Pdef,max and the minimum remaining storable energy ER,min to be used in the calculation. As a set of sufficient constraints that can be directly applied, Under this constraint, when clip limiting is not triggered, Pallow remains non-decreasing with increasing ER within the allowable power gap range; after triggering 0 or Phw limiting, Pallow remains constant, still exhibiting a non-decreasing relationship. Simultaneously, when ER remains constant, since kψ is not less than 0, Pallow remains non-increasing with increasing ψ. If the calibration coefficient of a certain power station cannot satisfy the above sufficient constraints, this formula is only used as a formula for calculating candidate access capacity, and is not used to claim monotonicity regarding ER. Feasibility filtering is completed by subsequent upper limits of rechargeable energy storage power, adjustable range of charging piles, and grid connection point restrictions.

[0075] When there are multiple input links, first determine the input path based on... The upper limit of the energy storage charging power, the adjustable range of the charging pile, and the grid connection point restrictions eliminate infeasible requests, and then select link combinations and access objects that meet the constraints. To convert continuous power requests into control quantities that can be executed by the actual PCS or charging module, the first step is to read... Minimum power adjustment step of each adjustable device And select a reference control resolution from the equipment that meets the control accuracy requirements. For those allowed to be assigned to the first... Power of individual devices , can be represented as: ; in, The number of complete control units that can be allocated. This indicates rounding down. This represents the remaining power, which is insufficient for a single baseline control resolution. The system first allocates power according to the complete control unit; the remaining power can be retained without allocation or used as end-point compensation when the device allows continuous fine-tuning. This discretization process is executed based on the actual step size of the device interface, without using the greatest common divisor method for real power values.

[0076] The technical advantage of this embodiment is that it integrates smart meters, photovoltaics, energy storage, charging equipment, environmental sensors, and input / output links into the field sensing and control objects, and provides a complete path from energy state judgment to power allocation that the device can perform. This solves the technical problems of lack of boundary conditions for input link access judgment and mismatch between continuous power requests and the discrete adjustment capabilities of actual power devices.

[0077] Example 10: Based on Example 9, this example further illustrates output link scheduling. The AI ​​computing power localization and real-time strategy reasoning module determines the current power generation, the remaining storable power of the energy storage unit, and the power consumption of the charging piles based on real-time data. When the total power consumption of the charging piles is less than the total power generation and the remaining storable power of the energy storage is less than or equal to a preset first threshold, it is determined that the power station is in a state of waiting for external transmission, and the output link connection order is retrieved from the local strategy.

[0078] The output link connection order can be pre-constructed and synchronized locally based on historical monitoring data of each link during network connectivity. For the first... Each candidate link is prioritized based on its electricity demand target type, its unit electricity benefit after the target response, its link loss parameters, and its capacity capacity. The sum of these four priorities is used as the ranking criterion in the original strategy. Each priority value is obtained through a pre-defined quantification table, ensuring that different physical quantities are mapped to a unified priority scale before summing, rather than directly adding the original electricity price, loss ratio, and electricity consumption.

[0079] The equity value after the target response described in this application adopts a unified and traceable acquisition method: for the corresponding output link, after the target response is completed, the unit external transmission volume revenue confirmed by the operation metering record or external settlement interface is read, and mapped to a dimensionless equity score in the range of 0 to 1 according to the equity quantification table preset in the current strategy version. The dimensionless equity score serves as the basis for the second priority value. The original unit of the unit external transmission volume revenue is yuan / kWh; the same equity quantification table and the same data source are used within the same strategy version, and different methods are not mixed in a single sorting calculation. Therefore, the equity value after the target response in the claims has a definite data source, unit conversion relationship, and comparison scale.

[0080] In an alternative equivalent implementation, the above quantization table can be further represented as dimensionless values ​​from 0 to 1. For positive indices where larger values ​​are better and negative indices where smaller values ​​are better, the following methods are used respectively: ; In the formula, For the first The metrics to be normalized for each link and These serve as the lower and upper bounds of the current candidate link set or the calibrated interval. To prevent positive numbers divided by zero from occurring when the upper and lower bounds are equal, and These are the normalized values ​​for positive and negative indicators, respectively. When the target type is a categorical variable, a dimensionless categorical score can be directly set based on the supply guarantee level, contract service level, or scheduling priority.

[0081] Let the dimensionless values ​​corresponding to the target type, equity, loss, and accommodative capacity be uniformly denoted as: ,in And when the index is 1, 2, 3, and 4, it corresponds to the four indicators of target type, equity, loss, and capacity, respectively. The priority score for each output link can be expressed as: ; in, Prioritize scoring for output links. For the first The non-negative weights of the indicators. These weights are configured by the site operation rules and remain fixed within the same strategy version. The output range of this formula is consistent with each dimensionless indicator, and is not multiplied by 100 again, thus avoiding inconsistencies between the score value range and the text definition. If a link repeatedly refuses access within a preset historical window, its ranking value can be deducted according to the rejection frequency, reducing the probability of consecutively requesting invalid links.

[0082] During power output, the system checks the availability and acceptance limits of each output link in descending order of connection priority. If a high-priority link refuses access or is temporarily unavailable, the system continues checking the next link. Ultimately, the output power is constrained by the inverter's rated power, grid connection point power limit, and link acceptance limit. The SCADA server converts the allowed output power into the corresponding device's power setting or link switching control and records the actual execution results for subsequent policy updates.

[0083] The step-by-step checks constitute the physical feasibility filtering after sorting: if a link is disconnected for maintenance, blocked for protection, communication is unconfirmed, or the acceptance limit is zero, the link is skipped from the candidate set for this cycle; if the acceptance limit is less than the current planned outgoing power, the allocable power for this link is determined according to its acceptable limit, and the remaining outgoing power is checked on the next link. If a high-priority link refuses access or is temporarily unavailable, the system does not terminate the outgoing power decision for this cycle, but continues to check subsequent links along the connection order, so that the sorting result can be implemented as actual link control.

[0084] The technical advantage of this embodiment is that it forms a definite output link connection order by taking the demand target, rights, line loss and accommodative power according to the preset quantification rules, and further uses a unified dimensionless scoring method to verify the sorting relationship in the optional implementation; thereby it can solve the technical problems of different dimensions of evaluation indicators for multiple output links, inability to directly compare priorities, and easy stagnation of field station external transmission scheduling after high priority links are rejected.

Claims

1. A cloud-edge collaborative microgrid power management system, characterized in that, The system comprises a field perception layer, an edge computing and control layer, a cloud collaboration layer, and a user interaction layer. The edge computing and control layer includes an intelligent edge energy control device deployed at the energy consumption site. This device includes an AI computing power localization and real-time strategy inference module, a multimodal data access module, and a SCADA server. The multimodal data access module accesses and analyzes the multimodal operation data of the microgrid collected by the field perception layer to obtain real-time data. The AI ​​computing power localization and real-time strategy inference module generates a first optimization decision based on the local strategy and the real-time data, and the SCADA server executes this first optimization decision. In the event of a cloud-edge communication anomaly, the AI ​​computing power localization and real-time strategy inference module continues to generate the first optimization decision based on the strategy updated locally before the communication anomaly and the real-time field data, to maintain the local control closed loop of the microgrid. The local strategy is formed or updated by a global optimization strategy synchronized by the cloud collaboration layer, and the first optimization decision is a field control decision formed based on the real-time field data under the constraints of the local strategy.

2. The cloud-edge collaborative microgrid power management system according to claim 1, characterized in that, The field sensing layer collects multimodal operation data of the microgrid; the edge computing and control layer processes the multimodal operation data locally and sends the processed operation data to the cloud collaboration layer; the cloud collaboration layer generates a second optimization decision based on the received operation data and sends the second optimization decision to the edge computing and control layer for execution; The user interaction layer is used to display the microgrid's operating status, charging information, and the execution results of the first optimization decision and / or the second optimization decision; The second optimization decision is used to form or update the local policy and / or serve as the control target of the edge computing and control layer. The edge computing and control layer determines the first optimization decision to be actually executed based on real-time field data. The second optimization decision being sent to the edge computing and control layer for execution means that after receiving and verifying the second optimization decision, the edge computing and control layer uses the verified second optimization decision to form or update the local policy and / or serve as the control target. The second optimization decision does not bypass the first optimization decision and directly writes control quantities to the field devices.

3. The cloud-edge collaborative microgrid power management system according to claim 2, characterized in that, The cloud collaboration layer includes an EMS cloud platform and a charging operation and maintenance platform. The EMS cloud platform is used for model training and generating the second optimization decision. The charging operation and maintenance platform is used for charging operation management, user interaction, system operation and maintenance, and alarm management. The EMS cloud platform is also used for model training on a longer time scale and cross-site operation status analysis to form or update the global optimization strategy sent to the edge. The EMS cloud platform trains a model using historical operating data and corresponding strategy execution results to obtain model parameters or rule parameters for cross-timescale prediction or candidate strategy evaluation. When generating the second optimization decision, the model parameters or rule parameters are used to obtain the prediction results of the operating status in subsequent time periods and / or the evaluation results of candidate strategies. The strategy parameters, constraints and / or control objectives are determined by combining the equipment parameter range, site capacity and operational constraints to form the second optimization decision or the global optimization strategy.

4. The cloud-edge collaborative microgrid power management system according to claim 2, characterized in that, The user interaction layer includes mobile terminals and / or web terminals.

5. The cloud-edge collaborative microgrid power management system according to claim 1, characterized in that, The intelligent edge energy control device also includes a firewall module, a data preprocessing module, a communication management module, and a network switching module; the communication management module is used to access the operation data collected by the field perception layer, the data preprocessing module is used to preprocess the accessed operation data, and the SCADA server sends the processed operation data to the cloud collaboration layer through the network switching module.

6. The cloud-edge collaborative microgrid power management system according to claim 5, characterized in that, During the instant of cloud-edge communication failure and within a preset time thereafter, the operational data accessed by the communication management module is transferred to the multimodal data access module to execute the local network outage self-governance process; after the preset time, the multimodal data access module directly accesses the microgrid multimodal operational data collected by the field perception layer; after cloud-edge communication is restored, key events during the network outage and the current device status are uploaded first, and the version, device identifier, and parameter range of the cloud return strategy are verified. After the verification is passed, the cloud-edge collaborative mode is restored.

7. The cloud-edge collaborative microgrid power management system according to claim 1, characterized in that, The AI ​​computing power localization and real-time strategy reasoning module extracts on-site video data of photovoltaic modules from the real-time data, and identifies photovoltaic module shading or foreign object coverage in the on-site video data of photovoltaic modules. When photovoltaic module shading or foreign object coverage is identified, the maximum power point tracking strategy of the corresponding inverter is switched from the incremental conductance method to the intelligent hybrid algorithm. The on-site video data of photovoltaic modules is associated with the electrical quantities of the same photovoltaic branch to which the corresponding inverter belongs according to the device identifier, and the maximum power point tracking strategy switching is triggered only when the time difference between the on-site video data of photovoltaic modules and the electrical quantities does not exceed the allowable time window and the shading or foreign object coverage meets the configured triggering conditions.

8. The cloud-edge collaborative microgrid power management system according to claim 7, characterized in that, The intelligent hybrid algorithm includes: collecting the illuminance, ambient temperature, output voltage, and output current of the photovoltaic module; forming a four-dimensional input vector through A / D conversion; performing sliding window filtering and normalization on the four-dimensional input vector; determining the step size adjustment coefficient and performing coarse adjustment based on the processed data using a fuzzy logic algorithm; searching for the operating point using a particle swarm optimization algorithm; and finally using a neural network to dynamically correct the operating point. When the rate of change of the occlusion area in adjacent visual analysis cycles is higher than the configured threshold and the particle swarm fitness variance is lower than the configured threshold, zero-mean random perturbation is added to some particle positions and the inertia weight is increased to trigger a re-search for other power peaks.

9. The cloud-edge collaborative microgrid power management system according to claim 1, characterized in that, The microgrid includes smart meters, photovoltaic modules and inverters, energy storage units, charging piles, environmental sensors, output links, and input links.

10. The cloud-edge collaborative microgrid power management system according to claim 9, characterized in that, The AI ​​computing power localization and real-time strategy reasoning module analyzes the real-time data to determine the power generation, the remaining storable power of the energy storage unit, and the power consumption of the charging pile, and determines the power status accordingly. When the power status is pending external transmission, the output link connection order corresponding to the local strategy is retrieved and the output link is connected according to the output link connection order. The output link connection order determines the first priority value, second priority value, third priority value, and fourth priority value according to the power demand target type, the equity value after the target response, the link loss parameter, and the storable power of each output link, and sorts them from largest to smallest according to the sum of the first priority value, second priority value, third priority value, and fourth priority value. The equity value after the target response is the dimensionless equity score obtained by mapping the unit external transmission power revenue fed back after the corresponding output link completes the target response through the preset quantization table of the current strategy version. Before establishing an output link connection, a physical feasibility check is performed based on the link's availability and acceptance limit. If a high-priority output link refuses access or is temporarily unavailable, the next output link is checked.

Images