5g private network scheduling communication and data transmission control method for distributed photovoltaic

Abstract

The application relates to the technical field of power system dispatching communication, and discloses a 5G private network dispatching communication and data transmission control method for distributed photovoltaics, which comprises the following steps: acquiring photovoltaic station distribution information and dispatching service requirements, designing a dynamic slice arrangement scheme based on power grid topology perception, establishing a private network direct-sampling direct-control communication architecture, deploying intelligent fusion terminal equipment, establishing a three-layer data model, performing hierarchical classification, monitoring and evaluating network quality and constructing a network digital twin model, performing encryption and identity authentication, acquiring operation data and meteorological data, training a hierarchical prediction model and an abnormal diagnosis model, establishing a dispatching optimization model, solving a dispatching scheme, performing multi-agent game collaborative control by the fusion terminal, collecting control process data and feeding back the control process data to a dispatching master station, and generating control execution records and effect evaluation reports; the application realizes second-level data collection and millisecond-level control response of the distributed photovoltaics, and improves the real-time performance and reliability of power grid dispatching.

Description

Technical Field

[0001] This invention relates to the field of power system dispatch communication technology, and more specifically, to a 5G private network dispatch communication and data transmission control method for distributed photovoltaic systems. Background Technology

[0002] With the rapid growth of distributed photovoltaic (PV) installations, it has become an important component of the power system. However, distributed PV is characterized by its large number, geographical dispersion, and strong output fluctuations, posing significant challenges to grid dispatching. Traditional dispatch communication architectures employ a multi-level aggregation model, which suffers from problems such as prolonged communication time, slow data acquisition cycles, and untimely control responses, making it difficult to meet the real-time dispatching requirements of large-scale distributed PV.

[0003] The power grid dispatching master station needs to achieve second-level data acquisition and millisecond-level control response for distributed photovoltaic power generation to cope with rapid fluctuations in photovoltaic output and the power balance requirements of the power grid. Meanwhile, different dispatching service scenarios have significantly different requirements for network performance; emergency control scenarios require ultra-low latency and high reliability, while data acquisition scenarios require high bandwidth support. How to provide differentiated service guarantees for different services within a unified communication network is a pressing technical challenge that needs to be addressed.

[0004] Furthermore, distributed photovoltaic (PV) sites are electrically coupled, requiring coordinated control of multiple sites to achieve global optimization when grid power fluctuations occur. Traditional centralized control methods suffer from high computational complexity, large communication overhead, and single-point-of-failure risks, making them unsuitable for the scheduling and control needs of large-scale distributed PV systems. Therefore, a novel scheduling communication and data transmission control method is urgently needed to improve the observability, measurability, adjustability, and controllability of distributed PV systems. Summary of the Invention

[0005] This invention provides a 5G private network scheduling communication and data transmission control method for distributed photovoltaics, which solves the technical problems of high computational complexity, large communication overhead and single point of failure risk in related technologies.

[0006] This invention provides a 5G private network scheduling communication and data transmission control method for distributed photovoltaic systems, including: To obtain photovoltaic site distribution information and scheduling business requirements, design a dynamic slicing and orchestration scheme based on grid topology awareness, and establish a private network direct acquisition and control communication architecture; Based on a dedicated network direct acquisition and control communication architecture, deploy intelligent converged terminal equipment, establish a three-layer data model, and obtain standardized photovoltaic operation data; Based on standardized photovoltaic operation data, the data is classified and categorized to monitor and evaluate network quality and construct a digital twin model of the network. Encryption and identity authentication are then performed to obtain the photovoltaic data upload channel. Based on the photovoltaic data upload channel, operational data and meteorological data are acquired, hierarchical prediction models and anomaly diagnosis models are trained, and power generation prediction results and equipment anomaly diagnosis information are obtained. Based on the power generation prediction results and equipment anomaly diagnosis information, scheduling optimization is performed to obtain a scheduling scheme and generate a scheduling instruction set. Based on the scheduling instruction set, multi-agent game-theoretic collaborative control is executed on the fusion terminal, and control process data is collected and fed back to the scheduling master station to generate control execution records and effect evaluation reports.

[0007] In a preferred embodiment, the design based on a grid topology-aware dynamic slicing orchestration scheme includes: Extract grid topology information, calculate the electrical distance between photovoltaic sites, and establish an electrical coupling matrix; Based on the electrical coupling matrix, a clustering algorithm is used to divide photovoltaic sites into electrical coupling clusters, and each electrical coupling cluster corresponds to a dedicated network sub-slice; The design incorporates a hierarchical network slicing architecture. The first layer is the service type slicing layer, which includes ultra-reliable low-latency communication slices, enhanced mobile broadband slices, and ordinary bandwidth slices. The second layer is the electrical coupling sub-slicing layer, where each main slice is further divided into sub-slices based on electrical coupling clusters. A dynamic slicing and scheduling mechanism based on the power grid operating status is established. A sliding time window statistical method is used to predict the flow demand in future periods. The bandwidth quota of each sub-slice is dynamically adjusted according to the flow load of the electrical coupling cluster and the power grid power deficit. Resource borrowing follows the principle of priority based on electrical coupling degree.

[0008] In a preferred embodiment, establishing the three-layer data model includes: The first layer is the device's native data layer, which retains the original data structure of each protocol; The second layer is the unified data layer at the station, which maps data from different protocols to a unified data object. The mapping is achieved by configuring a mapping table. The third layer is the dispatch system data layer, which uses a data publishing service to convert data objects from the unified data layer at the station into messages that conform to the dispatch protocol format and upload them to the dispatch master station via a dedicated network.

[0009] In a preferred embodiment, acquiring standardized photovoltaic operation data further includes performing edge-side data preprocessing and multi-level anomaly detection: Data preprocessing includes data cleaning, data compression, and data aggregation. Data cleaning uses a sliding window midpoint filtering algorithm. Anomaly detection employs a multi-level detection mechanism. The first level of detection is threshold detection, which marks a suspected anomaly when the measured value exceeds the preset threshold range. The second-level detection is a time sequence consistency detection, which uses a sliding time window method. When more than a preset proportion of data points in the window exceed a preset threshold, it is judged as a continuous anomaly. Level 3 detection is spatial correlation detection. The fusion terminal obtains the operation data of adjacent sites through the private network and calculates the correlation coefficient using Pearson correlation coefficient. When the correlation coefficient is greater than the preset correlation threshold, it is determined that the trend is consistent. Only anomalies that pass the three-level detection can generate anomaly detection results and be reported to the dispatch master station.

[0010] In a preferred embodiment, constructing the network digital twin model includes: The construction of a digital twin network model includes three layers: physical layer twin, network layer twin, and service layer twin. The physical layer twin establishes digital models of base stations, transmission links, and core network equipment; the network layer twin establishes digital models of network topology, routing strategies, and resource allocation status; and the service layer twin establishes behavioral and demand models for scheduling services. Based on a three-layer twin model, a predictive transmission strategy is designed using model predictive control. The twin model simulates the network evolution process in the future time domain and calculates the network risk index. When the network risk index falls below a preset risk threshold, a predictive transmission strategy adjustment is triggered.

[0011] In a preferred embodiment, the predictive transmission strategy adjustment includes: The priority of business semantic awareness is dynamically adjusted. An importance score is calculated based on the scheduling instructions. The importance score is weighted and fused with the basic priority of the business type to obtain the dynamic priority. Transmission path pre-optimization: When the twin model predicts that a transmission path will soon become congested, data with a priority exceeding a preset priority threshold is switched to an alternative path or detour path in advance. A predictive data pre-transmission mechanism allows data with importance scores exceeding a preset importance threshold to be transmitted in advance while the network quality is still good, when the twin model predicts a decline in network quality at a certain time in the future. An adaptive redundancy coding and error correction mechanism dynamically adjusts the redundancy coding strength of data packets based on the channel quality and packet loss probability predicted by the Siamese model.

[0012] In a preferred embodiment, the training of the hierarchical prediction model includes: The training includes a hierarchical prediction model consisting of a global base model and a site correction model. The global base model is trained using a long short-term memory network algorithm to capture the common patterns of photovoltaic sites. The site correction model takes the prediction residuals of the global base model as input and outputs the correction values. The photovoltaic power output prediction is performed based on a hierarchical prediction model. The operation data of the photovoltaic site and the weather forecast data are used to construct the prediction feature vector, which is then input into the global basic model to obtain the basic prediction value. The data is then input into the site correction model to obtain the correction value. The two are added together to obtain the final prediction value.

[0013] In a preferred embodiment, the execution of multi-agent game cooperative control includes: Each photovoltaic site's integrated terminal is modeled as an intelligent agent. The agent's utility function comprehensively considers three objectives: global scheduling target deviation penalty, self-control cost, and equipment health protection. The utility function adopts a weighted summation form, and the weight coefficients are dynamically adjusted according to the grid operation status. The communication topology adopts a dynamic construction strategy, which establishes communication connections and forms communication neighbor relationships for agents whose electrical coupling degree exceeds a preset coupling degree threshold based on the power grid topology sensing model. A distributed Nash equilibrium solution algorithm is used to perform cooperative control. Each agent receives policy information from its communication neighbors, calculates the optimal policy based on the neighbor's policy and its own utility function, and broadcasts the updated policy to its communication neighbors. The process is iterated until the policy change is less than a preset convergence threshold. After convergence, each agent calculates the inverter's power setpoint according to the equalization strategy and issues a power adjustment command.

[0014] In a preferred embodiment, the execution of multi-agent game-theoretic cooperative control further includes executing distributed reactive power cooperative control based on a voltage sensitivity matrix: The voltage sensitivity matrix is ​​established by combining online identification and offline calculation. In the offline calculation stage, the power flow calculation method is used to calculate the Jacobian matrix of the power grid and extract the sensitivity sub-matrix. In the online identification stage, the reactive power disturbance applied periodically by the terminal is integrated and the voltage change is collected. The recursive least squares method is used to update the voltage sensitivity matrix online. A distributed reactive power collaborative optimization model is established based on the voltage sensitivity matrix. The objective function of the optimization model includes a voltage deviation minimization term, a reactive power regulation cost minimization term, and a reactive power resource utilization fairness term. The distributed reactive power cooperative optimization problem is solved by the alternating direction multiplier method. Each agent solves the local subproblems in parallel and iterates until the original residual is less than the preset original residual threshold and the dual residual is less than the preset dual residual threshold. After convergence, each agent issues adjustment commands based on the reactive power adjustment amount.

[0015] In a preferred embodiment, the generation of control execution records and effect evaluation reports includes: High-frequency data acquisition technology is used to record key parameters of the control process in real time. The data acquisition includes control input data, game negotiation process data, control execution data and control output data. A multi-dimensional evaluation method was adopted to quantitatively analyze the control effect, and the evaluation dimensions included control accuracy, response speed, stability and energy efficiency. A real-time feedback mechanism is used to upload control results to the dispatch master station. The feedback information includes immediate feedback and periodic feedback. After receiving feedback information, the scheduling master station evaluates the execution effect of the scheduling plan. When the command success rate is lower than the preset success rate threshold, an alarm is generated, and an adaptive adjustment mechanism is used to optimize the scheduling strategy and control parameters.

[0016] The beneficial effects of this invention are as follows: By establishing a direct acquisition and control communication architecture for the dispatch master station based on a 5G private network, the intermediate aggregation link in the traditional architecture was eliminated, enabling end-to-end direct communication between the dispatch master station and the photovoltaic sites. This reduced communication latency and improved the real-time performance of data acquisition and control response. Through the deployment of intelligent converged terminal equipment and the establishment of a three-layer data model, automatic access to heterogeneous devices and standardized data conversion were achieved, resolving the issues of significant differences in equipment configuration and inconsistent communication protocols among distributed photovoltaic sites.

[0017] This invention innovatively proposes a dynamic slicing orchestration scheme based on grid topology awareness and a 5G network digital twin model, introducing the electrical topology of the power grid into network resource allocation decisions, achieving deep coupling between network resources and grid dispatching needs. Through a predictive transmission strategy, the system can identify network quality degradation trends and potential congestion risks in advance, proactively adjusting transmission strategies to ensure the communication quality of critical dispatching services. Simultaneously, by employing multi-agent game-theoretic collaborative control and a distributed reactive power collaborative control method based on a voltage sensitivity matrix, it overcomes the computational bottlenecks and communication dependencies of traditional centralized control, achieving autonomous negotiation and global optimization of distributed photovoltaic sites, improving the system's scalability and robustness, and providing strong technical support for the safe and stable operation of new power systems. Attached Figure Description

[0018] Figure 1 This is the main flowchart of the 5G private network scheduling communication and data transmission control method for distributed photovoltaic power generation according to the present invention; Figure 2 This is a detailed flowchart of the 5G private network scheduling communication and data transmission control method for distributed photovoltaic power generation according to the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a 5G private network scheduling communication and data transmission control method for distributed photovoltaic power generation, such as... Figures 1 to 2 As shown, it includes: Step 1: Obtain photovoltaic site distribution information and scheduling business requirements, design a dynamic slicing and orchestration scheme based on grid topology awareness, and establish a private network direct acquisition and control communication architecture; Specifically, the following steps are included: Step 1.1: Collect the geographical location information and installed capacity information of the photovoltaic site; The photovoltaic power station management system based on the power grid dispatch master station uses data interface reading technology to obtain basic information datasets of all distributed photovoltaic sites within its jurisdiction. The basic information datasets include fields such as the geographical coordinates, administrative division, installed capacity, number of inverters, grid connection voltage level, and commissioning time of each photovoltaic site.

[0021] For sites with missing or inaccurate geographic coordinate data, administrative divisions are used instead of precise coordinates to assign each site to the geographic center of its county, district, or township, resulting in a complete site location dataset. Based on installed capacity information, photovoltaic sites are divided into four levels: micro sites, small sites, medium sites, and large sites. Micro sites have an installed capacity of less than 50 kilowatts, small sites have an installed capacity between 50 and 500 kilowatts, medium sites have an installed capacity between 500 and 2000 kilowatts, and large sites have an installed capacity greater than 2000 kilowatts. The communication needs and priorities of different site levels differ.

[0022] Step 1.2: Establish a library of typical scheduling service scenarios and quantify network performance indicators; Based on actual operational experience in power dispatching, and using expert interviews and demand analysis, five typical dispatching scenarios were identified: normal monitoring scenario, periodically collecting photovoltaic site operation data; emergency control scenario, rapidly issuing active power adjustment commands; voltage regulation scenario, issuing reactive power adjustment commands to maintain voltage stability; large-scale data acquisition scenario, collecting historical curve data or operation logs; and parameter configuration and firmware upgrade scenario, which are non-real-time services.

[0023] For each scenario, a benchmark table of network performance metrics is established. The benchmark values ​​for normal monitoring scenarios are: latency less than 1000 milliseconds, bandwidth greater than 100 kilobits per second, and packet loss rate less than 0.01%; for emergency control scenarios, the benchmark values ​​are: latency less than 20 milliseconds, bandwidth greater than 50 kilobits per second, and packet loss rate less than 0.1%; for voltage regulation scenarios, the benchmark values ​​are: latency less than 50 milliseconds, bandwidth greater than 50 kilobits per second, and packet loss rate less than 0.01%; for large-scale data acquisition scenarios, the benchmark values ​​are: latency less than 5000 milliseconds, bandwidth greater than 10 megabits per second, and packet loss rate less than 0.1%; and for parameter configuration scenarios, the benchmark values ​​are: latency less than 10000 milliseconds, bandwidth greater than 1 megabit per second, and packet loss rate less than 0.1%.

[0024] Step 1.3: Plan the 5G base station deployment scheme using the administrative region clustering method; Based on the site location dataset, using Geographic Information System (GIS) tools, photovoltaic (PV) sites are spatially clustered according to their respective cities and counties to obtain a PV site distribution density map. For each county or township, the number of PV sites and total installed capacity within the region are counted, and regional communication demand indicators are calculated. These indicators include the number of sites within the region, total installed capacity, maximum concurrent data traffic, and regional priority weights. Based on these regional communication demand indicators and the coverage area of ​​5G base stations, a coverage optimization algorithm is used to determine the deployment locations of 5G base stations, aiming to minimize the number of base stations while maximizing site coverage.

[0025] The coverage radius of 5G base stations is determined based on terrain and frequency band selection. Base station capacity is reserved with appropriate margins to handle peak traffic, based on 5G equipment specifications and the number of photovoltaic (PV) sites. The algorithm starts with the area with the densest distribution of PV sites, selects locations that can cover the most sites, and marks the covered sites as served. This process is repeated for the remaining unserved sites until all sites are covered, resulting in a list of 5G base station deployment locations. For remote mountainous areas or areas with scattered sites where 5G signal coverage is weak, miniaturized 5G micro base stations are deployed or satellite communication is used as a supplementary coverage method to ensure full communication network coverage.

[0026] Step 1.4: Design a dynamic slicing and orchestration scheme based on power grid topology awareness; Based on 5G private network network slicing technology and power grid topology information, this paper adopts an innovative power grid topology-aware slicing orchestration method to create dedicated network slices for different types of scheduling services, achieving logical isolation between services and differentiated service guarantees through power grid topology adaptation. The core innovation of this method lies in incorporating the electrical topology relationship of the power grid into the resource allocation decision of 5G network slicing, dynamically adjusting slice resources according to the electrical coupling strength of photovoltaic sites in the power grid.

[0027] Establish a mapping model from power grid topology to communication topology. Extract power grid topology information from the power grid model database of the dispatch master station, including substation locations, line connection relationships, and grid connection point information of distributed photovoltaics; calculate the electrical distance between any two photovoltaic sites and establish the electrical coupling matrix of the photovoltaic sites.

[0028] Based on the electrical coupling matrix, a clustering algorithm is used to divide photovoltaic (PV) sites into several electrical coupling clusters. Sites within an electrical coupling cluster are electrically close, have strong scheduling correlations, and are highly likely to require coordinated control. The number of clusters is determined based on the grid size and topology complexity. Each electrical coupling cluster corresponds to a dedicated network sub-slice; sites within the sub-slice share slice resources but are isolated from other clusters.

[0029] Based on network performance metrics and the characteristics of electrically coupled clusters for five typical scheduling service scenarios, a layered network slicing architecture is designed. The first layer is the service type slicing layer, comprising three main slices: ultra-reliable low-latency communication (uRLLC), enhanced mobile broadband (eMBB), and ordinary bandwidth slices. The uRLLC slice serves emergency control and voltage regulation scenarios, with the highest priority; the eMBB slice serves normal monitoring and large-scale data acquisition scenarios; and the ordinary bandwidth slice serves non-real-time services such as parameter configuration and firmware upgrades. The second layer is the electrically coupled sub-slicing layer, where each main slice is further divided into multiple sub-slices based on the electrically coupled cluster. Resource quotas for sub-slices are dynamically allocated based on factors such as the number of sites within the cluster, installed capacity, and historical control frequency.

[0030] Establish a dynamic slicing and orchestration mechanism based on the power grid's operating status. Deploy a network traffic monitoring system and a power grid status monitoring system to collect real-time traffic statistics and operating parameters such as power balance, voltage level, and frequency deviation of each slice.

[0031] A sliding time window statistical method is used to predict future traffic demand. The statistical window length is set to 1 hour, and the average, peak, and standard deviation of traffic flow for each slice within the window are calculated. Based on the statistical characteristics of the same time period over the past 7 days, a weighted average method is used to predict traffic flow for the next 15 minutes to 1 hour. The weight decreases with time, with the most recent day having a weight of 0.4, the previous day 0.3, the day before that 0.2, and the remaining days totaling 0.1. The predicted traffic flow value is obtained by multiplying the traffic flow values ​​for each historical time period by their corresponding weight coefficients and then summing the results; the sum of the weight coefficients equals 1. Based on forecast results and grid operating status, a rule-based dynamic resource orchestration strategy is adopted. The orchestration rules are as follows: When the predicted traffic of a uRLLC subslice of an electrically coupled cluster will exceed 80% of the current quota, 5% bandwidth is borrowed from the eMBB subslice of that cluster; when the grid power deficit is greater than 10% of the total installed capacity, the bandwidth quota of all uRLLC subslices is increased by 5% globally, and resources are borrowed from ordinary bandwidth slices; when the bandwidth utilization of a slice is below 30% for 5 consecutive minutes, the bandwidth quota of that slice is reduced by 5%, and the released resources are allocated to high-load slices; the minimum unit for resource adjustment is 3% of the total bandwidth, and after a single adjustment, the sum of the quotas of all slices remains 100%, with each slice's quota not less than 5% to ensure basic service; resource borrowing follows the principle of electrical coupling priority, prioritizing borrowing resources from clusters with low electrical coupling to avoid affecting communication at sites with strong grid correlation, and the adjustment delay is less than 5 seconds. Resource orchestration decisions are executed every 5 minutes, checking the above rules sequentially based on the current network status and traffic forecast results, and executing the corresponding resource adjustments if a rule is triggered.

[0032] Establish a real-time slice performance evaluation mechanism. Define slice health metrics to comprehensively evaluate slice performance parameters such as latency compliance rate, packet loss rate, bandwidth utilization, and service success rate. Slice health values ​​range from 0 to 100. When the health of a slice continuously falls below the first threshold, a resource reallocation process is triggered, including resource reallocation and routing path optimization, to ensure service continuity.

[0033] By using a dynamic slicing and orchestration method based on grid topology awareness, deep coupling between 5G network resources and grid dispatching needs is achieved. When power fluctuations occur in a certain area of ​​the grid and multiple photovoltaic sites need to be controlled in a coordinated manner, the corresponding network sub-slice automatically obtains more resource guarantees, ensuring low-latency and reliable transmission of control commands.

[0034] Step 1.5: Establish a direct communication link between the dispatch master station and the photovoltaic site; Based on 5G private network and network slicing configuration, a direct communication link is established between the dispatch master station and the photovoltaic site using point-to-point or point-to-multipoint communication modes, eliminating intermediate links such as the aggregation layer and access layer in the traditional architecture. The dispatch master station deploys 5G private network core network equipment, including user plane functional units, control plane functional units, and session management functional units, which are directly connected to the dispatch master station's dispatch automation system. The photovoltaic site deploys intelligent converged terminal equipment, which has a built-in 5G communication module and accesses the 5G base station via the 5G air interface. The 5G base station connects to the 5G core network through the backhaul network, thus establishing an end-to-end communication link.

[0035] The communication link adopts the IP network protocol stack. The network layer uses the IPv6 protocol to ensure sufficient address space. The transport layer selects either the TCP or UDP protocol according to the service type. Control commands with high real-time requirements use the UDP protocol to reduce handshake latency, while data transmission with high reliability requirements uses the TCP protocol to ensure error-free transmission. The application layer adopts a communication protocol dedicated to power dispatching to realize remote data interaction between the dispatching master station and the converged terminal.

[0036] A real-time monitoring mechanism for communication link quality is established. A network performance testing module is built into the intelligent converged terminal, periodically sending ICMP echo request messages to the scheduling master station to measure round-trip time (RTT). Simultaneously, wireless link parameters reported by the 5G communication module are collected, including reference signal received power, reference signal received quality, signal-to-noise ratio (SNR), and block error rate (BER). Based on RTT and wireless link parameters, a weighted synthesis algorithm is used to calculate a communication link quality score, ranging from 0 to 100. The weighting coefficients for RTT, reference signal received power, and BER are all set to 0.4 and 0.3 respectively. When the communication link quality score falls below a second threshold, a link quality degradation alarm is generated, triggering a transmission strategy adjustment process; when the score falls below a third threshold, a link failure alarm is generated, and a backup communication link is activated. The preferred values ​​for the second and third thresholds are set to 60 and 40 respectively, and these thresholds can be adjusted based on actual operational experience.

[0037] Through the above steps, a 5G private network direct acquisition and control communication architecture covering all photovoltaic sites was established, providing basic network support for subsequent data acquisition and control functions.

[0038] Step 2: Based on the private network direct acquisition and control communication architecture, deploy intelligent converged terminal equipment, establish a three-layer data model, and obtain standardized photovoltaic operation data; Step 2.1, design the modular intelligent converged terminal hardware architecture; Based on the differences in equipment configuration and diverse functional requirements of distributed photovoltaic sites, a modular hardware architecture is adopted to design the intelligent converged terminal, consisting of a main control board and functional expansion modules. The main control board provides core computing, storage, power management, and basic communication functions. It uses a 32-bit or 64-bit embedded processor with a main frequency of at least 1GHz, at least 1GB of RAM, and at least 8GB of flash storage, supporting the Linux operating system. The main control board integrates a 5G communication module, supporting the 5G NR protocol and the Sub-6GHz band, with a downlink speed of at least 100Mbps and an uplink speed of at least 50Mbps. The main control board provides multiple standardized expansion interfaces, including Ethernet interfaces, serial communication interfaces, digital input / output interfaces, and analog input interfaces.

[0039] The functional expansion modules adopt a pluggable design and can be selected according to the actual needs of the photovoltaic site. The remote control module realizes traditional remote control functions such as telemetry, remote signaling, remote control, and remote adjustment, and is compatible with IEC 60870-5-101 and IEC 60870-5-104 protocols; the automatic generation control module and the automatic voltage control module realize closed-loop control of active power and reactive power, with built-in PID control algorithms; the multi-protocol conversion module supports the parsing and conversion of various industrial communication protocols such as Modbus RTU, Modbus TCP, IEC 61850 MMS, and OPC UA, and the pre-built protocol parsing library can be expanded through software upgrades; the edge computing module is equipped with an independent processing unit and an encryption coprocessor. The processing unit is used for data preprocessing and anomaly detection calculation, and the encryption coprocessor is used for hardware-accelerated AES256 or ChaCha20 encryption algorithms, with an encryption throughput of no less than 100Mbps.

[0040] The main control board and functional modules are connected via an internal bus. The bus protocol is SPI or CAN, and the data transmission rate is no less than 10Mbps.

[0041] Step 2.2: Achieve automatic identification and protocol adaptation of photovoltaic equipment; Based on the multi-protocol conversion module of the intelligent fusion terminal, automatic protocol identification technology is employed to achieve automatic connection and communication establishment with station-end equipment such as photovoltaic inverters, environmental monitoring devices, and power quality monitoring devices. The fusion terminal connects to the station-end equipment via an RS485 interface or an Ethernet interface, and initiates the protocol identification process after the connection is established. The protocol identification process adopts a passive listening mode, where the fusion terminal listens to data frames on the bus and analyzes the format characteristics of the data frames.

[0042] Based on a pre-built protocol feature library, a pattern matching algorithm is used to determine the protocol type of a data frame. The protocol feature library stores the frame format characteristics of common industrial protocols. When the passive monitoring mode fails to identify the protocol type within a certain time, it switches to active detection mode. In active detection mode, the converged terminal sequentially sends standard query commands for different protocols, waits for the device's response, and determines the protocol type based on the response content.

[0043] After protocol identification, the converged terminal loads the corresponding protocol's parsing rules from the protocol parsing library and establishes normal communication with the device. For non-standard or proprietary protocols not included in the protocol parsing library, a protocol learning tool is provided. Maintenance personnel input protocol message samples through the configuration interface. These samples include hexadecimal raw data of query commands and response data, with the meaning and data type of each field clearly labeled. The converged terminal analyzes the structural features of the message samples, automatically generates protocol parsing rules, and saves them to the local protocol library. The protocol learning tool supports learning common proprietary protocols such as Modbus-like protocols and free communication protocols.

[0044] For older equipment that does not support standard communication interfaces, the converged terminal provides analog and digital input channels. It directly acquires voltage, current, or dry contact status signals from the equipment and converts them into digital signals through signal conditioning circuitry. The analog input channel uses a high-precision ADC chip with a resolution of at least 16 bits, a sampling frequency of at least 1kHz, and a measurement error not exceeding 0.5% of the measurement range. The digital input channel employs opto-isolation technology, supports voltage levels from 5V to 48V, and has a response time of no more than 10 milliseconds.

[0045] Step 2.3: Establish a three-layer data model to achieve data standardization transformation; Based on the differences in raw data formats across different protocols and devices, a three-layer data model architecture is adopted to achieve standardized data conversion. The first layer is the device native data layer, which preserves the original data structure and semantics of each protocol, ensuring that complete device information is not lost. In the device native data layer, Modbus device data is stored in the form of register addresses and register values, IEC 61850 device data is stored in the form of logical nodes and data objects, and analog input data is stored in the form of channel numbers and sample values.

[0046] The second layer is the station-level unified data layer, which defines a unified data model at the station level, mapping data from different protocols to unified data objects. The station-level unified data model includes electrical measurement data objects, equipment status data objects, alarm information data objects, and statistical information data objects. Electrical measurement data objects include attributes such as measurement point identifier, measurement type, measurement value, unit, timestamp, and quality identifier. The mapping from the device's native data layer to the station-level unified data layer is achieved through a configured mapping table. The mapping table, indexed by device type and protocol type, defines the mapping relationship between native data fields and unified data objects. The mapping table is stored in XML or JSON format and supports online editing and remote updates.

[0047] The third layer is the dispatch system data layer, which conforms to the data standards of the power dispatch automation system. Data objects in the dispatch system data layer include fields such as information body address, type identifier, information element, and timestamp, and directly interface with the data processing system of the dispatch master station. The conversion from the station-side unified data layer to the dispatch system data layer is achieved through a data publishing service. The data publishing service reads data objects from the station-side unified data model, generates messages conforming to the dispatch protocol format according to the configured publishing rules, and uploads them to the dispatch master station via the 5G private network.

[0048] The transformation process of the three-layer data model is executed in real time in the data processing unit of the converged terminal. The transformation cycle is consistent with the data acquisition cycle to ensure data real-time performance. The configuration information of the data model is stored in the configuration database of the converged terminal and supports visual editing through configuration tools.

[0049] Step 2.4: Perform edge-side data preprocessing and multi-level anomaly detection; Based on the edge computing module of the converged terminal, a lightweight algorithm is used to preprocess and detect anomalies in the collected raw data. Data preprocessing includes data cleaning, data compression, and data aggregation. Data cleaning employs statistical filtering methods, applying a sliding window midpoint filtering algorithm to the collected measurement value sequence. The window length is set to 5 to 11 sampling points. The measurement values ​​within the window are sorted by size, and the median value is taken as the filtered output to remove impulse noise and random interference.

[0050] Data compression employs lossless compression algorithms such as LZ77 or Deflate for large volumes of historical data, achieving compression ratios of 2 to 5 times. For real-time telemetry data, differential coding is used, uploading only the changes from the previous moment to reduce data volume. Data aggregation calculates statistical characteristic values ​​within a time window for high-frequency sampled data, such as maximum, minimum, average, and standard deviation, and uploads these statistical characteristic values ​​instead of the original sampling sequence.

[0051] Anomaly detection employs a multi-level detection mechanism to reduce false alarm rates. The first level of detection is threshold detection. When the measured value exceeds the fourth threshold range, it is marked as a suspected anomaly. The fourth threshold is determined based on the equipment's rated parameters and operating procedures. For example, the upper limit threshold for the inverter's AC side voltage is set to 1.1 times the rated voltage, and the lower limit threshold is set to 0.9 times the rated voltage.

[0052] The second-level detection is a time sequence consistency detection, which uses a sliding time window method. The window length is set to 10 to 30 sampling periods. When more than 50% of the data points in the window exceed the fourth threshold, it is judged as a continuous anomaly; otherwise, it is judged as a transient disturbance and ignored.

[0053] Level 3 detection is spatial correlation detection. The fusion terminal acquires key operational data from adjacent sites via the 5G private network and compares the data change trends. When a sudden drop in power occurs at the local site but the power at adjacent sites remains normal, it is determined to be a local anomaly; when both the local and adjacent sites experience a power decrease with the same trend, it is determined to be a regional weather change. Spatial correlation detection uses the Pearson correlation coefficient for calculation; a correlation coefficient greater than the fifth threshold of 0.7 is considered to indicate a consistent trend.

[0054] Only anomalies that pass the three-level detection can generate anomaly detection results and be reported to the dispatch master station. The anomaly detection results include information such as anomaly type, anomaly measurement point, anomaly value, anomaly occurrence time, and anomaly confidence level. The anomaly confidence level is calculated based on factors such as the degree to which it exceeds the fourth threshold, duration, and spatial consistency, and the value ranges from 0 to 1. The higher the confidence level, the greater the credibility of the anomaly.

[0055] Through the above steps, automatic access and protocol adaptation of heterogeneous devices were achieved, a three-layer data model was established to complete data standardization and conversion, and edge computing technology was applied to implement data preprocessing and multi-level anomaly detection, resulting in high-quality standardized photovoltaic operation data and high-confidence anomaly detection results.

[0056] Step 3: Based on standardized photovoltaic operation data, classify and categorize the data, monitor and evaluate network quality, construct a network digital twin model, encrypt and authenticate the data, and obtain the photovoltaic data upload channel. Specifically, the following steps are included: Step 3.1: Establish a business priority classification system; Based on the importance and real-time requirements of power dispatching services, a multi-level priority classification method is adopted, dividing dispatching data and control commands into five priority levels: Priority 1 is the highest priority, including emergency control commands and major equipment fault alarms; Priority 2 is high priority, including general control commands and important alarm information; Priority 3 is medium priority, including real-time telemetry data and remote signaling data; Priority 4 is low priority, including periodic statistical data and historical curve data; and Priority 5 is the lowest priority, including non-real-time configuration data and firmware upgrade packages. Different priorities correspond to different latency and packet loss rate requirements, with higher priorities having stricter requirements.

[0057] When generating data packets, the converged terminal automatically adds priority identifiers to the data packets based on their data type. The priority identifiers are encoded using DSCP fields, with different priorities mapped to corresponding forwarding categories to ensure that high-priority data is transmitted first.

[0058] Step 3.2: Implement real-time network quality monitoring and comprehensive evaluation; Based on the network performance testing module and 5G communication module of the intelligent converged terminal, a combination of active measurement and passive monitoring methods is used to acquire 5G network quality parameters in real time. Active measurement measures round-trip latency and packet loss rate by periodically sending test messages; passive monitoring collects wireless link quality parameters reported by the 5G communication module, including key indicators such as Reference Signal Received Power (RSRP), Reference Signal Received Quality (RSRQ), Signal-to-Noise Ratio (SNR), and Physical Layer Block Error Rate (BLER).

[0059] Based on the collected network quality parameters, a weighted comprehensive algorithm is used to calculate the Network Quality Analyzer (NQI). After normalization, each parameter is weighted according to the characteristics of the scheduled services and then summed to obtain a comprehensive score from 0 to 100, with higher values ​​indicating better network quality. The NQI is periodically calculated and stored in the local database of the converged terminal for use in transmission strategy decisions.

[0060] Step 3.3: Construct a 5G network digital twin model and implement a predictive transmission strategy; Based on the Network Quality Index (NQI) and historical network operation data, a virtual mirror model of the 5G network is constructed using digital twin technology. Through real-time simulation and prediction of the twin model, the network quality degradation trend and potential congestion risk can be identified in advance, realizing a shift in transmission control paradigm from passive response to proactive prediction.

[0061] The construction of a digital twin network model comprises three layers: physical layer twin, network layer twin, and service layer twin. The physical layer twin establishes digital models of 5G base stations, transmission links, and core network equipment to simulate the wireless signal propagation process. The network layer twin establishes digital models of network topology, routing strategies, and resource allocation status to simulate data packet transmission. The service layer twin establishes behavioral and demand models for scheduling services, generating service load predictions for future periods. Each layer's twin model is periodically updated and its parameters are calibrated based on measured data to maintain synchronization with the real network.

[0062] Based on a three-layer twin model, a predictive transmission strategy is designed using model predictive control. In each control cycle, the twin model predicts the network evolution process in the future time domain based on the current network state and service load, and calculates a network risk index to comprehensively assess the network health status.

[0063] When the network risk index falls below the sixth threshold, a predictive transmission strategy adjustment is triggered, including: dynamic adjustment of priority based on business semantic awareness, pre-optimization of transmission paths, predictive data pre-transmission mechanism, and adaptive redundancy coding and error correction mechanism.

[0064] The priority of business semantic awareness is dynamically adjusted. Priorities are dynamically adjusted based on the specific content of dispatch instructions, parsing key information within the instructions; an importance score is calculated for each instruction; this importance score is weighted and fused with the basic priority of the business type to obtain a dynamic priority; the dynamic priority is updated every 5 seconds, adjusting in real time according to changes in the power grid's operating status and dispatch tasks.

[0065] Transmission path pre-optimization. When the twin model predicts that a transmission path will soon become congested, the causes of congestion are analyzed to identify key factors leading to congestion. Based on the analysis results, the transmission path is adjusted in advance. For high-priority data that is about to be transmitted through the congested path, it is switched to an alternative or detour path in advance. The path switching adopts a soft handover mechanism, disconnecting the old path only after the new path is established to ensure the continuity of data transmission.

[0066] Predictive data pre-transmission mechanism. When the twin model predicts that network quality will significantly decline at some point in the future, a pre-transmission strategy is adopted for data that is not strictly time-sensitive but is of high importance, and it is transmitted in advance while the network quality is still good.

[0067] Adaptive redundancy coding and error correction mechanism. The redundancy coding strength of data packets is dynamically adjusted based on the channel quality and packet loss probability predicted by the Siamese model. A forward error correction coding scheme is employed, with the redundancy coding strength adaptively adjusted according to the predicted packet loss rate.

[0068] Establish an effectiveness evaluation mechanism for predictive strategies, record the decision-making basis, implementation actions, and actual effects of strategy adjustments, and continuously optimize the prediction accuracy of twin models and strategy decision parameters.

[0069] Step 3.4, Implement end-to-end security mechanisms. Based on the security requirements of power dispatch data, a hardware encryption acceleration and lightweight key management scheme are adopted to achieve end-to-end encryption and two-way authentication during data transmission. The edge computing module of the converged terminal is equipped with an encryption coprocessor chip, which integrates hardware implementations of encryption algorithms such as AES256. Encryption processing is completed by dedicated hardware circuitry, meeting the needs of real-time encryption of large-scale data.

[0070] Data encryption employs a symmetric encryption algorithm. The converged terminal encrypts data packets using a session key and transmits them to the dispatch master station via the 5G private network. The dispatch master station then decrypts and restores the data using the corresponding key. The session key is updated periodically, with key updates automatically completed using a key negotiation protocol.

[0071] Identity authentication employs a digital certificate mechanism. The converged terminal is pre-installed with a root certificate public key at the factory. Upon initial access to the 5G private network, it requests a device certificate from the dispatch master station and stores it in a secure storage area. A two-way identity authentication process is established, where the converged terminal and the dispatch master station mutually verify the validity of the certificates. After successful two-way authentication, a session key is negotiated and generated for subsequent data transmission encryption.

[0072] Through the above steps, a data classification system based on business priorities was established, real-time network quality monitoring and comprehensive evaluation were implemented, a digital twin model was constructed and a predictive transmission strategy was implemented, and an end-to-end encryption and authentication mechanism was executed, resulting in an efficient, reliable and secure photovoltaic data upload channel.

[0073] Step 4: Based on the photovoltaic data upload channel, obtain operational data and meteorological data, train the hierarchical prediction model and anomaly diagnosis model, and obtain the power generation prediction results and equipment anomaly diagnosis information. Based on real-time and historical photovoltaic operation data received by the dispatch master station, combined with meteorological forecast information, a photovoltaic output prediction model and an equipment anomaly diagnosis model are established using machine learning algorithms. This provides a forward-looking basis for dispatch decisions and timely fault warnings, resulting in power generation prediction results at multiple time scales and high-confidence anomaly diagnosis information.

[0074] Step 4.1: Construct a training dataset for the photovoltaic power output prediction model; Based on the historical database of the dispatch master station, historical operation data of photovoltaic sites were extracted to construct a training dataset for the prediction model. The historical operation data includes time-series data such as power generation, irradiance, ambient temperature, module temperature, and equipment operating status, with a time span of no less than one year. Data repair techniques were used to address missing and outlier values ​​in the historical data; missing values ​​were filled using interpolation, and outliers were detected and removed using statistical analysis methods.

[0075] Historical meteorological data, including irradiance, temperature, humidity, wind speed, and cloud cover, is obtained from meteorological service interfaces. A multi-source meteorological data fusion mechanism is established, simultaneously accessing data from multiple meteorological service providers. A data quality scoring method is used to select the optimal data source, and the system automatically switches to a backup data source when the primary data source fails.

[0076] Photovoltaic operation data and meteorological data are aligned by timestamps to construct a training dataset containing time features, meteorological features, and historical power features. Time features include periodic time information such as hours, days of the week, and months, represented using one-hot encoding or sine / cosine encoding; meteorological features include irradiance, temperature, cloud cover, etc., at the current and predicted future times; historical power features include power generation values ​​at several past times.

[0077] Step 4.2: Train the hierarchical photovoltaic power output prediction model; Based on the constructed training dataset, a hierarchical modeling strategy was adopted to train a photovoltaic power output prediction model. The hierarchical model consists of two layers: a global base model and a site-specific correction model. The global base model was trained using the Long Short-Term Memory (LSTM) algorithm. The network structure includes an input layer, multiple hidden layers, and an output layer. It was trained using the backpropagation algorithm and an optimizer, with mean squared error (MSE) as the loss function. The training dataset was divided into a training set and a validation set in chronological order. During the training process, the loss value of the validation set was monitored to avoid overfitting.

[0078] The trained global base model can capture the common patterns of all photovoltaic sites. For the individual characteristics of each photovoltaic site, a site-specific model is trained using Support Vector Regression (SVR) or Gradient Boosting Tree Derivative (GBDT) algorithms. The input is the prediction residual of the global base model, and the output is the correction value.

[0079] To address the lack of historical data for newly built photovoltaic (PV) sites, a method of borrowing data from similar sites is adopted. All PV sites are clustered, and newly built sites are assigned to the most similar cluster based on their characteristic attributes. Historical data from sites within this cluster are then used to train the initial model.

[0080] Establish an online model update mechanism to periodically use the latest data to incrementally train and update the model parameters. The online update adopts an incremental learning algorithm, which only uses newly added data to adjust the model parameters, thus saving computing resources.

[0081] Step 4.3: Perform multi-timescale photovoltaic power output prediction; Based on a pre-trained hierarchical prediction model, real-time photovoltaic power output prediction is performed, generating prediction results at multiple time scales. The prediction process involves obtaining the latest operational data and weather forecast data of photovoltaic sites from the dispatch master station, constructing prediction feature vectors, inputting them into the global basic model to obtain basic prediction values, then inputting them into the site correction model to obtain correction values, and finally adding the two to obtain the final prediction value.

[0082] The forecast timescales include 15-minute ultra-short-term forecasts, 1-hour short-term forecasts, 4-hour medium-term forecasts, and 24-hour day-ahead forecasts. The forecasts at different timescales use the same model structure but the input feature weights differ slightly. Ultra-short-term forecasts rely more on historical power inertia, while day-ahead forecasts rely more on meteorological forecast information.

[0083] The accuracy of the prediction results is evaluated using the Mean Absolute Percentage Error (MAPE), with corresponding MAPE target values ​​set for different time scales. A dynamic evaluation mechanism for prediction accuracy is established, calculating the prediction error and updating the statistical indicators each time new actual operating data is obtained. When the MAPE continuously exceeds the target value, an alarm is generated and the model retraining process is triggered.

[0084] Step 4.4: Establish a hierarchical equipment anomaly diagnosis model; Based on the time-series characteristics and multi-measuring-point correlations of photovoltaic operation data, a data-driven unsupervised learning method is used to establish an equipment anomaly diagnosis model. The anomaly diagnosis adopts a hierarchical framework: the first layer determines whether an anomaly exists, the second layer determines the major category of the anomaly, and the third layer determines the specific anomaly type.

[0085] The first layer of anomaly detection employs an autoencoder algorithm. The autoencoder consists of an encoder and a decoder, trained using normal operating data. The training objective is to minimize the reconstruction error between the input and output data. After training, real-time operating data is input into the autoencoder to calculate the reconstruction error. An anomaly is identified when the reconstruction error exceeds a seventh threshold.

[0086] The second layer of anomaly classification employs a multi-classification algorithm, categorizing into three main types: equipment failure, performance degradation, and external interference. Features utilize statistical and frequency domain features, and the classification algorithm is trained using historical labeled data.

[0087] The third layer, specific anomaly type judgment, establishes dedicated diagnostic rules or classification models for each major category. Equipment failure is further subdivided into inverter failure, combiner box failure, communication failure, etc.; performance degradation is subdivided into component aging, component obstruction, dust accumulation, etc.; external interference is subdivided into grid voltage fluctuations, interference from nearby equipment, etc.

[0088] For each abnormal diagnosis result, the anomaly confidence level is calculated, taking into account factors such as the extent of reconstruction error exceeding the limit, the duration of the anomaly, and the consistency of multiple measurement points. A confidence level greater than 0.8 is considered a high-confidence anomaly, and an alarm is generated immediately. A confidence level between 0.5 and 0.8 is considered a medium-confidence anomaly, and continuous observation is conducted. A confidence level less than 0.5 is considered a low-confidence anomaly, and only logs are recorded.

[0089] Step 4.5: Generate intelligent alarm information; Based on anomaly diagnosis results and confidence assessment, an alarm filtering and aggregation mechanism is employed to generate structured intelligent alarm information. Alarm filtering targets multiple similar alarms generated at the same site within a short period, using a time window aggregation method to merge alarms of the same type within a window into a single comprehensive alarm.

[0090] Alarm priority scoring comprehensively considers factors such as the severity of the anomaly, the scope of impact, and historical frequency, and is calculated using a weighted comprehensive algorithm. Alarm scores range from 0 to 100. Scores above the eighth threshold are considered urgent alarms requiring immediate attention; scores between the ninth and eighth thresholds are considered important alarms requiring prompt attention; and scores below the ninth threshold are considered general alarms that can be addressed later. The preferred value for the eighth threshold is set to 70, and the preferred value for the ninth threshold is set to 40.

[0091] Alarm information is in a structured format, including fields such as alarm identifier, site identifier, anomaly type, occurrence time, anomaly confidence level, severity, impact range, alarm priority score, diagnostic description, and handling suggestions. The diagnostic description is converted into readable text using natural language generation technology, and the handling suggestions are retrieved from a knowledge base.

[0092] Through the above steps, a hierarchical prediction model was trained, multi-timescale power output prediction was performed, a hierarchical anomaly diagnosis model was established, intelligent alarm information was generated, and high-precision photovoltaic power generation prediction results and high-confidence equipment anomaly diagnosis information were obtained.

[0093] Step 5: Based on the power generation prediction results and equipment anomaly diagnosis information, optimize the scheduling to obtain a scheduling scheme and generate a scheduling instruction set; Specifically, the following steps are included: Step 5.1: Establish a scheduling optimization model with multiple constraints; Based on the power balance requirements of the power grid and the operational constraints of photovoltaic (PV) sites, a scheduling optimization model is established using mathematical modeling methods. The objective function of the optimization model is to minimize the power grid deviation, thereby adjusting the PV output to bring the power deficit close to zero.

[0094] The constraints of the optimization model include photovoltaic (PV) site output capacity constraints, power regulation rate constraints, equipment health constraints, and voltage stability constraints. Output capacity constraints require that the target power of dispatch commands not exceed the site's maximum available power, which is determined comprehensively based on current irradiance, equipment capacity, and equipment operating status. Power regulation rate constraints limit the maximum rate of power change to prevent rapid power fluctuations from impacting equipment and the grid. Equipment health constraints reduce the dispatch participation of sites diagnosed with anomalies or exclude them from the dispatch plan. Voltage stability constraints require that the voltage at the grid connection point remain within the allowable range after regulation.

[0095] Uncertainty parameters are introduced into the optimization model to handle photovoltaic power output prediction errors. A robust optimization method is adopted, which represents the predicted power output as the predicted value plus or minus the uncertainty interval. The uncertainty interval is determined based on the statistics of historical prediction errors, ensuring that the scheduling scheme still meets the constraints under uncertainty.

[0096] Step 5.2: Solve the scheduling scheme using a hierarchical decomposition optimization algorithm; To address the high computational complexity of scheduling optimization for large-scale photovoltaic (PV) sites, a hierarchical decomposition optimization strategy is adopted. The first layer is regional optimization, where PV sites are divided into several regions based on geographical location. The optimal scheduling scheme for each region is solved independently, with the optimization objective being to meet the region's power regulation requirements. Regional optimization can be performed in parallel, with each region's optimization executed on independent computing nodes, improving computational speed. The regional optimization uses either linear programming or quadratic programming algorithms, selected based on the characteristics of the objective function.

[0097] The second layer is global coordination optimization, which aggregates the optimization schemes from all regions and performs cross-regional power balance adjustments. Global coordination optimization checks whether the sum of the power of the schemes from all regions meets the overall power grid demand. If there is a deviation, the power allocation ratio of each region is adjusted, and the regional tasks are reissued for iterative calculation until convergence.

[0098] For ultra-large-scale problems, heuristic algorithms are used instead of exact optimization algorithms. By searching for near-optimal solutions in the solution space, computational speed is improved to meet the needs of real-time applications. An adaptive weight adjustment mechanism for multi-objective optimization is established. Optimization objectives include not only power balance but also equipment health protection and economic objectives. The target weights are dynamically adjusted according to the current operating state of the power grid. When there is a large power deficit requiring emergency adjustment, the weight of the power balance target is increased; when the power grid is operating smoothly, the weights of each objective are relatively balanced.

[0099] Step 5.3: Generate a scheduling instruction set and add a unique identifier; Based on the scheduling scheme obtained from the optimization solution, a structured set of scheduling instructions is generated. Each scheduling instruction includes fields such as instruction identifier, station identifier, control type, target value, execution time, priority, and validity period. The instruction identifier is generated using a globally unique identifier (UUID) to ensure that it is unique and non-repeating throughout the entire system.

[0100] Control types are divided into two categories: active power control and reactive power control. Execution time is the expected execution time of the instruction, using UTC time format with second-level precision. Priority is set according to the urgency of the instruction, and validity period is the duration the instruction remains in effect; instructions not executed after the validity period automatically become invalid.

[0101] An instruction status tracking table is established to record the lifecycle status of each instruction, including statuses such as generated, issued, received, executing, completed, timed out, and canceled. The instruction status tracking table is stored in the database of the scheduling master station, and the update of instruction status is achieved through feedback messages from the converged terminal.

[0102] Step 5.4: Reliably send control commands through the 5G private network; uRLLC slicing based on a scheduling instruction set and a 5G private network employs a reliable transmission protocol to send control commands to the converged terminals of the corresponding photovoltaic sites. Command issuance uses a request-response mode: the scheduling master station sends instruction messages to the converged terminals, and the converged terminals return acknowledgment response messages after receiving the commands.

[0103] The transport protocol employs either TCP or a custom reliable UDP protocol. TCP provides reliable connection-oriented data transmission, suitable for transmitting critical commands. The custom reliable UDP protocol adds reliability assurance mechanisms to UDP while maintaining low latency, making it suitable for urgent commands with extremely high real-time requirements.

[0104] The instruction issuance process involves the dispatch master station encrypting the instruction message and sending it via the 5G private network, with the message priority marked as the highest priority. The converged terminal receives the message, decrypts it, and verifies the integrity and authenticity of the instruction. Upon successful verification, it returns a response message. The dispatch master station sets an instruction response timeout timer; if no response is received within the timeout period, a retransmission mechanism is triggered.

[0105] The retransmission mechanism employs a backoff strategy for multiple retransmissions. If no response is received after retransmission, a communication failure is identified, a failure alarm is generated, and an attempt is made to send the command via a backup communication link. If the backup link also fails, the command status is updated to "failed to send," the station is marked as uncontrollable, and the station is temporarily excluded from the scheduling plan.

[0106] A command execution monitoring mechanism is established, with the scheduling master station monitoring the command issuance and execution process in real time. Monitoring information includes command sending time, response receiving time, round-trip time, retransmission count, execution start time, and execution completion time. A sliding window algorithm is used to calculate the average round-trip time and latency jitter; a communication quality alarm is generated when key indicators exceed the tenth threshold.

[0107] The design command execution rehearsal function allows for optional rehearsal evaluation before the formal issuance of control commands. After receiving the rehearsal command, the fusion terminal simulates the execution of the control process, evaluates the feasibility and expected effects of the control, and returns the rehearsal results to the dispatch master station to assist in decision-making. The rehearsal process does not involve actual operation of the equipment.

[0108] Through the above steps, a multi-constraint scheduling optimization model was established. A robust scheduling scheme was obtained by using a hierarchical decomposition optimization algorithm. A scheduling instruction set with a unique identifier was generated, and control instructions were reliably sent to the photovoltaic site through the 5G private network.

[0109] Step 6: Based on the scheduling instruction set, execute multi-agent game collaborative control on the fusion terminal, collect control process data and feed it back to the scheduling master station, and generate control execution records and effect evaluation reports; Specifically, the following steps are included: Step 6.1: Parse the scheduling instructions and perform a feasibility check; Based on the scheduling instructions received by the converged terminal, instruction parsing and feasibility verification technologies are employed to ensure the correctness and executability of the instructions. The instruction parsing process involves the converged terminal reading the instruction message from the receive buffer, extracting fields such as instruction identifier, control type, target value, execution time, and priority according to the protocol format, and storing them in the instruction queue. The instruction queue adopts a priority queue data structure, with higher-priority instructions placed at the front of the queue for priority execution.

[0110] Feasibility verification includes three aspects: target value rationality verification, equipment status verification, and time validity verification. Target value rationality verification checks whether the target power value of the instruction is within the equipment's output capacity range. The output capacity range is determined comprehensively based on the current irradiance, equipment rated capacity, and equipment operating status. A real-time power estimation method is used to calculate the maximum achievable power based on the current irradiance and module efficiency, taking into account the effect of module temperature on efficiency. When the instruction target value exceeds the reasonable range of the maximum achievable power, the target value is deemed unreasonable, and a verification failure feedback is generated.

[0111] The equipment status verification checks whether critical equipment such as inverters and combiner boxes are in normal operating condition. If any equipment malfunctions or alarms, control commands will be refused to be executed to avoid further damage to the equipment. The time validity verification checks whether the execution time of the command is within the validity period. If the current time exceeds the validity period of the command, it is determined that the command has expired and will not be executed.

[0112] Instructions that pass verification proceed to execution; those that fail verification generate a failure feedback message and are uploaded to the dispatch master station via the 5G private network. The dispatch master station adjusts the dispatch plan based on the reason for the failure, either regenerating the instruction or marking the site as uncontrollable.

[0113] Step 6.2: Establish a multi-agent game model and execute distributed active power cooperative control; Based on verified active power control commands, an innovative multi-agent game-theoretic collaborative control method is adopted to achieve autonomous negotiation and collaborative optimization among distributed photovoltaic (PV) sites. This method models the fusion terminal of each PV site as an agent with autonomous decision-making capabilities. The agent's goal is to minimize its own control costs and equipment losses while meeting the global power regulation targets issued by the master dispatch station.

[0114] The establishment of a multi-agent game model includes four elements: agent definition, policy space definition, utility function definition, and communication topology definition. In the agent definition, each photovoltaic site's fusion terminal is considered an agent, and the state of agent i includes information such as current output power, adjustable power range, equipment health status, and historical control frequency. In the policy space definition, agent i's policy is the power adjustment amount it chooses, and the policy space consists of continuous or discrete values ​​within the adjustable power range.

[0115] The definition of the utility function is a core innovation of the game theory model. The utility function of agent i comprehensively considers three objectives: global scheduling target deviation penalty, self-control cost, and equipment health protection. The global scheduling target deviation penalty term measures the deviation between the sum of power adjustments made by all agents and the global target issued by the scheduling master station; the larger the deviation, the higher the penalty. The self-control cost term measures the cost for an agent to perform power adjustments, including the power adjustment magnitude, adjustment rate, and losses due to deviation from the maximum power point. The equipment health protection term measures the impact of frequent control on equipment lifespan; based on the historical control frequency and current health status of the equipment, a higher cost penalty is imposed on agents that frequently control, achieving a long-term equilibrium allocation of the adjustment task.

[0116] The utility function uses a weighted summation, with the weight coefficients of the three objective terms dynamically adjusted according to the grid's operating status. When a significant power deficit in the grid necessitates emergency adjustments, the weight coefficient for the global dispatch objective deviation penalty increases, while the weight coefficients for self-control costs and equipment health protection decrease. When the grid operates smoothly, the weight coefficients of the three objectives are relatively balanced.

[0117] In the communication topology definition, agents interact with each other via a 5G private network. This interaction includes their own state, policy selections, and utility function values. The communication topology employs a dynamic construction strategy. Based on the power grid topology sensing model established in step 1, communication connections are established between agents corresponding to photovoltaic sites with high electrical coupling, forming communication neighbor relationships. The communication topology is updated periodically according to changes in the power grid topology and the operating status of the sites.

[0118] Based on a multi-agent game theory model, a distributed Nash equilibrium solution algorithm is used to implement cooperative control. Nash equilibrium is a core concept in game theory, representing that, given the policies of other agents, each agent chooses the optimal response that maximizes its own utility, and no agent can obtain higher utility by unilaterally changing its policy.

[0119] The distributed Nash equilibrium solution employs an iterative optimization algorithm. Initially, each agent calculates an initial policy based on the global objective issued by the scheduling master station and its own adjustability. The initial policy uses a proportional allocation method, distributing adjustment tasks according to the installed capacity of each station. The agent broadcasts the initial policy to its communication neighbors via the 5G private network.

[0120] In the k-th iteration, agent i receives policy information from its communicating neighbors. Based on the neighbors' policies and its own utility function, it calculates its optimal policy using the optimal response algorithm. The optimal response algorithm models agent i's policy optimization problem as a constrained optimization problem, with the objective function being agent i's utility function. Constraints include power adjustment range constraints, adjustment rate constraints, and equipment capacity constraints. The constrained optimization problem is solved using the Lagrange multiplier method or sequential quadratic programming to obtain agent i's optimal policy in the k-th iteration.

[0121] Agent i broadcasts its updated policy to its neighbors, triggering policy updates from the neighboring agents. The iterative process continues until a convergence condition is met. The convergence condition is that the policy changes of all agents are less than the eleventh threshold in two consecutive iterations, or the number of iterations reaches its maximum value. Theoretical analysis and simulation experiments show that, when the communication topology is connected and the utility function satisfies the convexity condition, the distributed Nash equilibrium solution algorithm can converge to the Nash equilibrium solution within a finite number of iterations.

[0122] After convergence, each agent calculates the inverter's power setpoint according to its own equilibrium strategy and sends power adjustment commands to the inverter via the local communication interface. The inverter adjusts its operating point to regulate active power output. The agent collects the inverter's output power in real time, compares it with the equilibrium strategy, and calculates the control deviation. If the control deviation exceeds the allowable range, the agent re-triggers the game-solving process and updates its strategy.

[0123] Establish a robustness guarantee mechanism for game control. In actual operation, some agents may be unable to participate in the game due to communication failures, equipment failures, or other reasons. To ensure system robustness, a virtual agent substitution mechanism is adopted. When an agent is detected to be out of contact, its communication neighbors construct a virtual agent model based on the agent's historical information to replace it in the game, ensuring the continuity of the game process.

[0124] Step 6.3: Perform distributed reactive power cooperative control based on voltage sensitivity matrix; Based on reactive power control commands or voltage control commands, an innovative distributed collaborative control method based on a voltage sensitivity matrix is ​​adopted to achieve collaborative optimization of reactive power at multiple photovoltaic sites and precise voltage regulation. The core innovation of this method lies in establishing a global sensitivity matrix model between grid voltage and photovoltaic site reactive power. A distributed optimization algorithm is then used to solve a multi-objective optimization problem considering voltage constraints, capacity constraints, and fairness constraints, achieving the dual objectives of optimal voltage control performance and optimal reactive power resource utilization.

[0125] The voltage sensitivity matrix is ​​established using a combination of online identification and offline calculation. In the offline calculation phase, based on the power grid model database of the dispatch master station, information such as the power grid topology, line parameters, transformer parameters, and load distribution is extracted. The power flow calculation method is then used to calculate the Jacobian matrix of the power grid. The Jacobian matrix describes the partial derivative relationship between node injected power and node voltage. The voltage sensitivity submatrix to reactive power is extracted from the Jacobian matrix to obtain the initial voltage sensitivity matrix.

[0126] The element S(i,j) of the voltage sensitivity matrix represents the voltage change at node i when the reactive power of photovoltaic site j changes by 1 kVvar, in volts per kVvar. The diagonal element S(i,i) of the matrix represents the impact of the reactive power of site i on its own grid connection voltage. The off-diagonal element S(i,j) of the matrix represents the cross-influence of the reactive power of site j on the grid connection voltage of site i.

[0127] During the online identification phase, the fusion terminal periodically applies small-amplitude reactive power disturbances. The terminal collects voltage changes at its own grid connection point before and after the disturbances, and simultaneously receives voltage change data reported by other sites within the electrical coupling cluster via the 5G private network. A recursive least squares method is used to update the elements of the voltage sensitivity matrix online based on the measured reactive power disturbance and voltage response. Online identification is performed during periods of stable grid operation and stable photovoltaic output to avoid interference from external disturbances.

[0128] The sensitivity matrix obtained through online identification is weighted and fused with the matrix calculated offline, with the fusion weights determined based on the confidence level of the identification data. The fused sensitivity matrix is ​​stored in the local database of the fusion terminal and synchronized to other intelligent agents within the electrical coupling cluster via a 5G private network.

[0129] Based on the voltage sensitivity matrix, a distributed reactive power collaborative optimization model is established. The objective function of the optimization model includes three terms: voltage deviation minimization, reactive power regulation cost minimization, and reactive power resource utilization fairness. The voltage deviation minimization term measures the sum of squares of voltage deviations at all grid-connected points within the electrical coupling cluster, aiming to make the voltages of all grid-connected points as close as possible to the target voltage. The reactive power regulation cost minimization term measures the cost of reactive power regulation at each site, including the regulation magnitude, inverter capacity occupancy, and sacrifice of active power output. The reactive power resource utilization fairness term measures the degree of balance in reactive power regulation across sites, avoiding excessive regulation tasks for some sites.

[0130] The constraints of the optimization model include voltage constraints, capacity constraints, and regulation rate constraints. Voltage constraints require that the voltage at all grid connection points remain within the allowable range after optimization. Capacity constraints require that the reactive power output of each site does not exceed the inverter's reactive power capacity limit, which is calculated based on the inverter's apparent power capacity and current active power output. Regulation rate constraints limit the rate of change of reactive power to prevent rapid changes from impacting the power grid.

[0131] The Alternating Direction Method of Multipliers (ADMM) is employed to solve the distributed reactive power cooperative optimization problem. ADMM is a distributed optimization algorithm that decomposes the global optimization problem into multiple local subproblems. Each agent solves the local subproblems in parallel and reaches the global optimum through iterative negotiation.

[0132] The iterative process of the ADMM algorithm is as follows: Initially, each agent calculates the initial reactive power regulation based on the voltage deviation at its own grid connection point and the diagonal elements of the voltage sensitivity matrix. The initial regulation adopts a local optimization method, considering only minimizing its own voltage deviation. The agent broadcasts the initial regulation to other agents within the electrical coupling cluster via the 5G private network.

[0133] In the k-th iteration, agent i receives reactive power adjustments from other agents and calculates the impact of other stations' reactive power on its own grid connection voltage based on the voltage sensitivity matrix. Agent i updates its local optimization problem by adding a penalty term for coordination with other agents to the objective function; this penalty term uses an augmented Lagrange form. Agent i solves the updated local optimization problem to obtain the reactive power adjustment for the k-th iteration and broadcasts the updated adjustment to other agents, while simultaneously updating the Lagrange multipliers.

[0134] The iterative process continues until convergence conditions are met. Convergence conditions include the original residual being less than the twelfth threshold and the dual residual being less than the thirteenth threshold. The original residual measures the consistency of the adjustments made by each agent, while the dual residual measures the change in the Lagrange multipliers. Theoretical analysis shows that the ADMM algorithm has global convergence in convex optimization problems.

[0135] After convergence, each agent calculates the inverter's reactive power setpoint based on the optimized reactive power regulation, and sends reactive power regulation commands to the inverter via the local communication interface. The inverter adjusts the phase angle of the output current, changing the ratio of active current to reactive current to achieve reactive power output. The agent collects the grid connection point voltage and inverter reactive power output in real time, compares them with the optimization target, and calculates the control deviation. If the voltage deviation or reactive power deviation exceeds the allowable range, the agent re-triggers the ADMM optimization process and updates the reactive power regulation.

[0136] A coordinated control mechanism for active and reactive power is established. When the dispatch master station simultaneously issues active and reactive power control commands, the fusion terminal coordinates their execution using a joint optimization strategy. Joint optimization models active and reactive power control as a unified multi-objective optimization problem, with constraints including the inverter's apparent power capacity constraint, ensuring that the vector sum of active and reactive power does not exceed the capacity limit. Joint optimization is performed iteratively using a sequential optimization method until the adjustments for both active and reactive power converge simultaneously.

[0137] An adaptive adjustment mechanism for voltage control is established. When voltage control is ineffective and voltage deviation continues to exceed the allowable range, the agent automatically adjusts the parameters of the optimization model, including increasing the weighting coefficient of the voltage deviation term, relaxing the constraint on reactive power regulation cost, and expanding the reactive power regulation range. Through adaptive adjustment, the system can cope with uncertainties such as grid topology changes, load fluctuations, and equipment failures, maintaining the robustness of voltage control.

[0138] Step 6.4: Collect control process data and generate execution records; The execution process of multi-agent game-theoretic collaborative control and distributed reactive power optimization control employs high-frequency data acquisition technology to record key parameters of the control process in real time, generating detailed control execution records. Data acquisition includes four categories: control input data, game negotiation process data, control execution data, and control output data.

[0139] Control input data includes instruction identifiers, global scheduling objectives, execution time, initial power values, initial voltage values, and initial agent states. Game negotiation process data records the iterative negotiation process between agents, including the agent's strategy selection, utility function value, communication neighbor information, and convergence metrics in each iteration. The recorded data is used to analyze performance metrics such as game convergence speed, negotiation efficiency, and communication overhead.

[0140] The control execution data recorder tracks the process by which the agent executes control based on the game equilibrium strategy or ADMM optimization results, including power deviation, voltage deviation, control quantity, inverter response time, and capacity utilization for each control cycle. Control output data includes final power value, final voltage value, settling time, overshoot, steady-state error, game convergence time, number of iterations, and collaborative control effect. The collaborative control effect evaluation assesses the performance improvement of multi-site collaborative control compared to independent control, including global goal achievement, resource utilization efficiency, and fairness metrics.

[0141] The data acquisition cycle is consistent with the control cycle. The acquired data is stored in the local time-series database of the fusion terminal, which uses an efficient compression storage algorithm. Data exceeding the retention period is automatically deleted or archived to the scheduling master station.

[0142] The control execution log adopts a structured format, including fields such as record identifier, instruction identifier, station identifier, control type, start time, end time, target value, initial value, final value, settling time, overshoot, steady-state error, and control status. Control status includes successful execution, failed execution, and partial success. For controls that fail, the reason for failure and exception information are recorded for subsequent fault analysis and control strategy optimization.

[0143] Step 6.5: Evaluate the control effectiveness and generate a multi-dimensional evaluation report; Based on control execution records and real-time operational data, a multi-dimensional evaluation method is employed to quantitatively analyze control effectiveness and generate a comprehensive evaluation report. The evaluation dimensions include four aspects: control accuracy, response speed, stability, and energy efficiency.

[0144] The control accuracy assessment uses the steady-state error index, calculating the percentage deviation between the actual power and the target power after stabilization. A deviation percentage less than 1% is rated as excellent, between 1% and 3% as good, between 3% and 5% as acceptable, and greater than 5% as unacceptable. The accuracy distribution of all control commands within a certain time window is statistically analyzed, and the proportion of each level is calculated.

[0145] Response speed is evaluated using settling time as an indicator. The time required from control initiation to power stabilization is statistically analyzed. A settling time of less than 30 seconds is considered a fast response, between 30 and 60 seconds is considered a normal response, and more than 60 seconds is considered a slow response. The average settling time and standard deviation are calculated; the standard deviation reflects the consistency of the response speed.

[0146] Stability assessment uses overshoot and oscillation frequency as indicators. Overshoot less than 5% and no oscillation is rated as excellent stability; overshoot between 5% and 10% or 1 to 2 oscillations is rated as good stability; and overshoot greater than 10% or more than 3 oscillations is rated as poor stability. The number of oscillations is obtained by analyzing the number of peaks and troughs in the power curve.

[0147] Energy efficiency assessment uses control energy consumption indicators to calculate the power loss of the inverter and the power consumption of auxiliary equipment during the control process, comparing it with the energy consumption before control. An increase in control energy consumption of less than 0.5% of the rated power is rated as excellent energy efficiency, an increase between 0.5% and 1% is rated as good energy efficiency, and an increase greater than 1% is rated as poor energy efficiency. Energy efficiency assessment is used to optimize control strategies and reduce control costs.

[0148] The comprehensive evaluation report uses a scoring system. The evaluation results of each dimension are converted into scores and weighted and summed to obtain a comprehensive score. A comprehensive score greater than 90 is rated as excellent, a score between 80 and 90 is rated as good, a score between 70 and 80 is rated as satisfactory, and a score less than 70 is rated as needing improvement.

[0149] Evaluation reports are generated regularly and include the evaluation period, total number of control commands, success rate, evaluation results for each dimension, overall score, existing problems, and improvement suggestions. The reports are uploaded to the dispatch master station via a 5G private network for dispatch personnel to review and use in decision-making.

[0150] Step 6.6: Feedback control results to the dispatch master station to form a closed loop; Based on control execution records and effect evaluation reports, a real-time feedback mechanism is adopted to upload control results to the dispatch master station, forming a complete closed loop from dispatch command issuance to control execution and result feedback. Feedback information includes two types: immediate feedback and periodic feedback.

[0151] Real-time feedback is sent immediately after the control command is executed. The feedback includes key information such as command identifier, execution status, final power value, and adjustment time. It is transmitted via uRLLC slicing on the 5G private network to ensure that the dispatch master station can promptly grasp the control results. After receiving the feedback, the dispatch master station updates the command status tracking table.

[0152] Periodic feedback uploads detailed control execution records and evaluation reports at fixed intervals. Due to the large volume of data in periodic feedback, data compression technology is used to reduce transmission bandwidth consumption. Data is transmitted via eMBB slicing on the 5G private network, and the TCP protocol is used to ensure complete data transmission.

[0153] After receiving feedback information, the scheduling master station uses data analysis technology to evaluate the execution effect of the scheduling scheme and identify problems and anomalies in the control process. The analysis includes command success rate statistics, control accuracy analysis, response time analysis, and failure cause analysis. When the command success rate falls below the fourteenth threshold, a poor scheduling effect alarm is generated, triggering the scheduling strategy optimization process.

[0154] Based on feedback and analysis results, the master dispatch station employs an adaptive adjustment mechanism to optimize dispatch strategies and control parameters. Dispatch strategy optimization includes adjusting station dispatch priorities, modifying power allocation ratios, and updating prediction model parameters. Control parameter optimization includes adjusting PID parameters, modifying control cycles, and optimizing coordinated control strategies. The optimized strategies and parameters are pushed to the fusion terminal via a configuration distribution interface. The fusion terminal updates its local configuration and applies the new strategies and parameters in subsequent control operations.

[0155] Establish a long-term tracking mechanism for control effectiveness. The dispatch master station stores historical control records and evaluation reports, creating a control effectiveness database. Analyze long-term trends in control effectiveness, identifying seasonal patterns and equipment performance degradation trends. Based on long-term data, train a control strategy optimization model. The model's inputs include site characteristics, meteorological conditions, and power grid status; the outputs are recommended control parameters and dispatch strategies, achieving continuous optimization of the dispatch system.

[0156] Through the above steps, the scheduling instructions were parsed and verified at the fusion terminal. Innovatively, distributed collaborative control of active power based on multi-agent game theory and distributed collaborative control of reactive power based on voltage sensitivity matrices were implemented. This broke through the computational bottlenecks and communication dependencies of traditional centralized control, realizing autonomous negotiation and global optimization of distributed photovoltaic sites. Detailed control process data and game negotiation process data were collected, generating a multi-dimensional control effect evaluation report. A real-time feedback mechanism from control results to the scheduling master station was established, forming a distributed closed-loop control system.

[0157] In one embodiment of the present invention, a specific example is provided: A 90-day field test was conducted in three counties of a certain city. During the test, 12 5G base stations were deployed, covering 1500 distributed photovoltaic (PV) sites with a total installed capacity of approximately 350 MW, covering a test area of ​​approximately 2000 square kilometers. Intelligent converged terminal equipment was deployed at each PV site, establishing a 5G private network direct acquisition and control communication architecture. Adaptive transmission control and collaborative control strategies were implemented, achieving second-level data acquisition and millisecond-level control response for the distributed PV sites within the test area. The following data tables illustrate some key data acquisition examples.

[0158] Table 1 shows examples of standardized photovoltaic operation data obtained through intelligent integrated terminals: Table 1: Examples of standardized photovoltaic operation data obtained through intelligent integrated terminals;

[0159] Table 2 shows examples of photovoltaic power output prediction results and equipment anomaly diagnosis information obtained from prediction models trained based on historical operating data and meteorological data. Table 2: Examples of photovoltaic power output prediction results and equipment anomaly diagnosis information obtained by the prediction model;

[0160] Through the implementation of the technical solution of this invention, the test area has achieved comprehensive monitoring and precise control of distributed photovoltaic sites. The end-to-end communication latency has been reduced from hundreds of milliseconds in the traditional architecture to less than 20 milliseconds, the data acquisition cycle has been improved from minutes to seconds, the success rate of acquiring standardized photovoltaic operation data has reached over 99.5%, the average absolute percentage error of 15-minute ultra-short-term prediction is 8.2%, and the accuracy rate of equipment anomaly diagnosis has reached over 92%. This has improved the real-time performance and reliability of power grid dispatching and provided strong support for the safe and stable operation of the new power system.

[0161] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A 5G private network scheduling communication and data transmission control method for distributed photovoltaic power generation, characterized in that, Includes the following steps: To obtain photovoltaic site distribution information and scheduling business requirements, design a dynamic slicing and orchestration scheme based on grid topology awareness, and establish a private network direct acquisition and control communication architecture; Based on a dedicated network direct acquisition and control communication architecture, deploy intelligent converged terminal equipment, establish a three-layer data model, and obtain standardized photovoltaic operation data; Based on standardized photovoltaic operation data, the data is classified and categorized to monitor and evaluate network quality and construct a digital twin model of the network. Encryption and identity authentication are then performed to obtain the photovoltaic data upload channel. Based on the photovoltaic data upload channel, operational data and meteorological data are acquired, hierarchical prediction models and anomaly diagnosis models are trained, and power generation prediction results and equipment anomaly diagnosis information are obtained. Based on the power generation prediction results and equipment anomaly diagnosis information, scheduling optimization is performed to obtain a scheduling scheme and generate a scheduling instruction set. Based on the scheduling instruction set, multi-agent game collaborative control is executed on the fusion terminal, and control process data is collected and fed back to the scheduling master station to generate control execution records and effect evaluation reports. The design, based on a grid topology-aware dynamic slicing orchestration scheme, includes: Extract grid topology information, calculate the electrical distance between photovoltaic sites, and establish an electrical coupling matrix; Based on the electrical coupling matrix, a clustering algorithm is used to divide photovoltaic sites into electrical coupling clusters, and each electrical coupling cluster corresponds to a dedicated network sub-slice; The network slicing architecture is designed with a hierarchical structure. The first layer is the service type slicing layer, which includes ultra-reliable low-latency communication slices, enhanced mobile broadband slices, and ordinary bandwidth slices. The second layer is the electrical coupling sub-slicing layer, where each main slice is further divided into sub-slices based on the electrical coupling cluster. Establish a dynamic slice scheduling mechanism based on the power grid operation status, use the sliding time window statistical method to predict the flow demand in future periods, dynamically adjust the bandwidth quota of each sub-slice according to the flow load of the electrical coupling cluster and the power grid power deficit, and follow the principle of priority of electrical coupling degree in resource borrowing; The establishment of the three-layer data model includes: The first layer is the device's native data layer, which retains the original data structure of each protocol; The second layer is the unified data layer at the station, which maps data from different protocols to a unified data object. The mapping is achieved by configuring a mapping table. The third layer is the dispatch system data layer, which converts the data objects of the unified data layer at the station into messages that conform to the dispatch protocol format through the data publishing service, and uploads them to the dispatch master station through the private network; The execution of multi-agent game-theoretic cooperative control also includes executing distributed reactive power cooperative control based on the voltage sensitivity matrix: The voltage sensitivity matrix is ​​established by combining online identification and offline calculation. In the offline calculation stage, the power flow calculation method is used to calculate the Jacobian matrix of the power grid and extract the sensitivity sub-matrix. In the online identification stage, the reactive power disturbance applied periodically by the terminal is integrated and the voltage change is collected. The recursive least squares method is used to update the voltage sensitivity matrix online. A distributed reactive power collaborative optimization model is established based on the voltage sensitivity matrix. The objective function of the optimization model includes a voltage deviation minimization term, a reactive power regulation cost minimization term, and a reactive power resource utilization fairness term. The distributed reactive power cooperative optimization problem is solved by the alternating direction multiplier method. Each agent solves the local subproblems in parallel and iterates until the original residual is less than the preset original residual threshold and the dual residual is less than the preset dual residual threshold. After convergence, each agent issues adjustment commands based on the reactive power adjustment amount.

2. The 5G private network scheduling communication and data transmission control method for distributed photovoltaic power generation according to claim 1, characterized in that, The acquisition of standardized photovoltaic operation data also includes performing edge-side data preprocessing and multi-level anomaly detection: Data preprocessing includes data cleaning, data compression, and data aggregation. Data cleaning uses a sliding window midpoint filtering algorithm. Anomaly detection employs a multi-level detection mechanism. The first level of detection is threshold detection, which marks a suspected anomaly when the measured value exceeds the preset threshold range. The second-level detection is a time sequence consistency detection, which uses a sliding time window method. When more than a preset proportion of data points in the window exceed a preset threshold, it is judged as a continuous anomaly. Level 3 detection is spatial correlation detection. The fusion terminal obtains the operation data of adjacent sites through the private network and calculates the correlation coefficient using Pearson correlation coefficient. When the correlation coefficient is greater than the preset correlation threshold, it is determined that the trend is consistent. Only anomalies that pass the three-level detection can generate anomaly detection results and be reported to the dispatch master station.

3. The 5G private network scheduling communication and data transmission control method for distributed photovoltaic power generation according to claim 1, characterized in that, The construction of the network digital twin model includes: The construction of a digital twin network model includes three layers: physical layer twin, network layer twin, and service layer twin. The physical layer twin establishes digital models of base stations, transmission links, and core network equipment; the network layer twin establishes digital models of network topology, routing strategies, and resource allocation status; and the service layer twin establishes behavioral and demand models for scheduling services. Based on a three-layer twin model, a predictive transmission strategy is designed using model predictive control. The twin model simulates the network evolution process in the future time domain and calculates the network risk index. When the network risk index falls below a preset risk threshold, a predictive transmission strategy adjustment is triggered.

4. The 5G private network scheduling communication and data transmission control method for distributed photovoltaic power generation according to claim 3, characterized in that, The predictive transmission strategy adjustment includes: The priority of business semantic awareness is dynamically adjusted. An importance score is calculated based on the scheduling instructions. The importance score is weighted and fused with the basic priority of the business type to obtain the dynamic priority. Transmission path pre-optimization: When the twin model predicts that a transmission path will soon become congested, data with a priority exceeding a preset priority threshold is switched to an alternative path or detour path in advance. A predictive data pre-transmission mechanism allows data with importance scores exceeding a preset importance threshold to be transmitted in advance while the network quality is still good, when the twin model predicts that the network quality will decline in a certain period of time in the future. An adaptive redundancy coding and error correction mechanism dynamically adjusts the redundancy coding strength of data packets based on the channel quality and packet loss probability predicted by the Siamese model.

5. The 5G private network scheduling communication and data transmission control method for distributed photovoltaic power generation according to claim 1, characterized in that, The training hierarchical prediction model includes: The training includes a hierarchical prediction model consisting of a global base model and a site correction model. The global base model is trained using a long short-term memory network algorithm to capture the common patterns of photovoltaic sites. The site correction model takes the prediction residuals of the global base model as input and outputs the correction values. The photovoltaic power output prediction is performed based on a hierarchical prediction model. The operation data of the photovoltaic site and the weather forecast data are used to construct the prediction feature vector, which is then input into the global basic model to obtain the basic prediction value. The data is then input into the site correction model to obtain the correction value. The two are added together to obtain the final prediction value.

6. The 5G private network scheduling communication and data transmission control method for distributed photovoltaic power generation according to claim 1, characterized in that, The execution of multi-agent game cooperative control includes: Each photovoltaic site's integrated terminal is modeled as an intelligent agent. The agent's utility function comprehensively considers three objectives: global scheduling target deviation penalty, self-control cost, and equipment health protection. The utility function adopts a weighted summation form, and the weight coefficients are dynamically adjusted according to the grid operation status. The communication topology adopts a dynamic construction strategy, which establishes communication connections and forms communication neighbor relationships for agents whose electrical coupling degree exceeds a preset coupling degree threshold based on the power grid topology sensing model. A distributed Nash equilibrium solution algorithm is used to perform cooperative control. Each agent receives policy information from its communication neighbors, calculates the optimal policy based on the neighbor's policy and its own utility function, and broadcasts the updated policy to its communication neighbors. The process is iterated until the policy change is less than a preset convergence threshold. After convergence, each agent calculates the inverter's power setpoint according to the equalization strategy and issues a power adjustment command.

7. The 5G private network scheduling communication and data transmission control method for distributed photovoltaic power generation according to claim 1, characterized in that, The generated control execution record and effect evaluation report include: High-frequency data acquisition technology is used to record key parameters of the control process in real time. The data acquisition includes control input data, game negotiation process data, control execution data and control output data. A multi-dimensional evaluation method was adopted to quantitatively analyze the control effect, and the evaluation dimensions included control accuracy, response speed, stability and energy efficiency. A real-time feedback mechanism is used to upload control results to the dispatch master station. The feedback information includes immediate feedback and periodic feedback. After receiving feedback information, the scheduling master station evaluates the execution effect of the scheduling plan. When the command success rate is lower than the preset success rate threshold, an alarm is generated, and an adaptive adjustment mechanism is used to optimize the scheduling strategy and control parameters.

Images