Optimization method based on distributed solar photovoltaic power generation control system

Abstract

The present invention relates to an optimization method based on a distributed solar photovoltaic power generation control system. The method comprises the following steps: collecting power generation data, voltage data, current data, and solar radiation data; inputting the data into a photovoltaic multimodal deep learning prediction model for processing to generate an active power regulation strategy and a reactive power regulation strategy; generating control instructions, and performing hierarchical safety constraint verification on the control instructions to generate remote control instructions and local control instructions; performing hierarchical reinforcement learning control and power optimization regulation on inverters in multiple control groups to obtain a total group output value; performing real-time monitoring to generate operating parameter data; calculating the active power qualification rate, control response time, control accuracy, and system fluctuation rate, and performing multi-objective optimization on the regulation dead zone range, control gain, and response time constant to obtain a target power distribution scheme. The present invention ensures reasonable load distribution for each inverter and improves the operational stability and reliability of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation control, and in particular to an optimization method based on a distributed solar photovoltaic power generation control system. Background Art

[0002] With the rapid development of photovoltaic power generation technology, the proportion of distributed photovoltaic power generation systems in the power system continues to increase. Traditional distributed photovoltaic power generation control systems mainly rely on independent control of single sites and lack the ability to coordinate group adjustments. This makes it difficult to meet the grid's requirements for large-scale distributed photovoltaic grid-connected operation.

[0003] Current photovoltaic power generation control systems present security risks during data collection and transmission, as well as slow system response and low control accuracy. This is particularly true when multiple photovoltaic power plants are operating across multiple regions. The lack of unified control strategies and safety constraints can easily lead to system instability, impacting the safe operation of the power grid. Summary of the Invention

[0004] The main purpose of the present invention is to provide an optimization method based on a distributed solar photovoltaic power generation control system, which ensures reasonable load distribution of each inverter and improves the operating stability and reliability of the system.

[0005] To achieve the above objectives, the present invention provides an optimization method based on a distributed solar photovoltaic power generation control system, comprising the following steps:

[0006] Real-time collection and encrypted transmission of photovoltaic power station operation data to obtain power generation data, voltage data, current data and solar radiation data;

[0007] Inputting the power generation data, voltage data, current data and solar radiation data into a photovoltaic multimodal deep learning prediction model for processing to generate an active power regulation strategy and a reactive power regulation strategy;

[0008] Generate control instructions based on the active power regulation strategy and the reactive power regulation strategy, and perform hierarchical safety constraint verification on the control instructions to generate remote control instructions and local control instructions;

[0009] According to the remote control instructions and the local control instructions, hierarchical reinforcement learning control and power optimization adjustment are performed on the inverters in the multiple control groups to obtain a total output value of the group;

[0010] The total output value of the group and the operating status of the equipment are monitored in real time to generate operating parameter data; based on the operating parameter data, the active power qualification rate, control response time, control accuracy and system fluctuation rate are calculated, and the adjustment dead zone range, control gain and response time constant are optimized by multiple objectives to obtain the target power allocation plan.

[0011] The present invention also provides an optimization device based on a distributed solar photovoltaic power generation control system, comprising:

[0012] The acquisition module is used to collect and encrypt the operation data of the photovoltaic power station in real time to obtain power generation data, voltage data, current data and solar radiation data;

[0013] a processing module, configured to input the power generation data, voltage data, current data, and solar radiation data into a photovoltaic multimodal deep learning prediction model for processing, and generate an active power regulation strategy and a reactive power regulation strategy;

[0014] A verification module, configured to generate control instructions based on the active power regulation strategy and the reactive power regulation strategy, and perform hierarchical safety constraint verification on the control instructions to generate remote control instructions and local control instructions;

[0015] an adjustment module, configured to perform hierarchical reinforcement learning control and power optimization adjustment on the inverters in the plurality of control groups according to the remote control instructions and the local control instructions, so as to obtain a total output value of the group;

[0016] A monitoring module, configured to monitor the total output value of the group and the operating status of the equipment in real time and generate operating parameter data;

[0017] The optimization module is used to calculate the active power qualification rate, control response time, control accuracy and system fluctuation rate based on the operating parameter data, and perform multi-objective optimization on the adjustment dead zone range, control gain and response time constant to obtain a target power allocation scheme.

[0018] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0019] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0020] In summary, the technical solution provided by the present invention improves the security and reliability of data transmission through multi-layer encrypted transmission and distributed storage technology, and realizes secure storage and efficient access to data; adopts a photovoltaic multimodal deep learning prediction model, combined with a multi-head attention mechanism and a bidirectional LSTM structure, to enhance the system's modeling ability and prediction accuracy of photovoltaic power generation characteristics; based on a hierarchical security constraint verification mechanism, multi-level security verification of control instructions is achieved, ensuring the safety and reliability of system operation; through hierarchical reinforcement learning control and power optimization adjustment, coordinated control of multiple groups of photovoltaic power stations is achieved, improving the overall control effect of the system; a multi-objective optimization algorithm is introduced to dynamically adjust system parameters, realizing adaptive optimization of system control parameters, and improving the control accuracy and response speed of the system; a power distribution scheme based on a weighted distribution algorithm ensures reasonable load distribution of each inverter and improves the operation stability and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 1 is a schematic diagram of steps of an optimization method based on a distributed solar photovoltaic power generation control system in one embodiment of the present invention;

[0022] Figure 2 This is a structural block diagram of an optimization device based on a distributed solar photovoltaic power generation control system in one embodiment of the present invention;

[0023] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0024] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0026] Reference Figure 1 This embodiment provides an optimization method based on a distributed solar photovoltaic power generation control system, comprising the following steps:

[0027] S1, real-time collection and encrypted transmission of photovoltaic power station operation data to obtain power generation data, voltage data, current data and solar radiation data;

[0028] The inverter's operating data is collected through telemetry. This process relies on a high-precision data acquisition module deployed in the system. This module acquires real-time operating parameters through the inverter's communication interface, generating a collection of inverter telemetry data. This data includes key parameters such as the inverter's real-time startup capacity, power change rate, and output upper and lower limits. This data is analyzed in real time by the system's internal data calculation module, specifically active power. Combined with the inverter's output upper and lower limits, this data generates the inverter's power regulation range. The inverter telemetry data is then input into the SSL security encryption module for proprietary protocol encryption. Based on the current industry standards for data security and confidentiality, the encryption process employs an advanced encryption algorithm to encrypt the data packets and generate first-level encrypted data packets. The routing management module distributes these first-level encrypted data packets to targeted servers, enabling targeted encrypted data transmission between different servers. This targeted transmission method ensures data security and improves data transmission efficiency and real-time performance, meeting the requirements for efficient data transmission in distributed photovoltaic control systems. The encrypted data is then input into the cloud platform for secure encryption. The cloud platform, serving as a centralized data storage and processing center, vertically encrypts this data, using secondary encryption technology to provide a higher level of protection for first-level encrypted data packets. Furthermore, to ensure dynamic data transmission security, a dynamic key update mechanism is deployed within the cloud platform. This mechanism increases the complexity of data encryption by periodically renewing keys, preventing external key theft and generating second-level encrypted data packets with a higher level of security. These second-level encrypted data packets are distributed. Using the system's efficient distributed storage architecture, encrypted data is synchronously stored across data nodes, ensuring high data availability and security. This data is stored in a real-time database containing encrypted data on power generation, voltage, current, and solar radiation. These records are categorized and stored in the real-time database. A data classification module manages recorded data in a refined manner based on timestamps and data type identifiers. Real-time data is categorized by time period and type, forming an organized data structure and generating a historical database backup. This historical database record is used for subsequent system functions such as data backtracking, trend analysis, and fault diagnosis. To improve the security and isolation of historical data, firewall service isolation technology is used to protect the historical database, ensuring that data is not attacked or tampered with during storage and transmission. Security checks are performed on the historical database, and verification algorithms are used to verify data integrity and consistency, ensuring that stored data is not lost or erroneous. Ultimately, stable and secure power generation, voltage, current, and solar radiation data can be generated.

[0029] S2, inputs power generation data, voltage data, current data, and solar radiation data into the photovoltaic multimodal deep learning prediction model for processing to generate active power regulation strategies and reactive power regulation strategies;

[0030] Specifically, power generation, voltage, current, and solar radiation data are input into the multi-layer convolutional feature extraction layer of the photovoltaic multimodal deep learning prediction model for processing. The multi-layer convolutional feature extraction layer consists of four convolutional modules, each of which contains two convolutional layers and a maximum pooling layer. The convolutional layer extracts features from the input data using a sliding window and weight sharing mechanism, capturing the local correlations of photovoltaic power plant data in the spatial dimension. The maximum pooling layer uses dimensionality reduction to filter out key local features and obtain a feature vector for power plant operation. The power plant operation feature vector is then input into the bidirectional LSTM timing analysis layer of the photovoltaic multimodal deep learning prediction model for time series processing. The bidirectional LSTM timing analysis layer consists of two LSTM layers, each consisting of 128 hidden units. This layer is designed to effectively capture the temporal dependency characteristics of photovoltaic power plant data. LSTM (Long Short-Term Memory) networks, as a neural network suitable for time series data, can model both long-term and short-term dependencies in time series through a gating mechanism. The introduction of a bidirectional structure strengthens data correlations in both forward and backward directions, enabling the model to simultaneously consider past and future temporal information, resulting in more accurate temporal dependency features. These temporal dependency features are then fed into the multi-head attention layer of the photovoltaic multimodal deep learning prediction model for processing. This multi-head attention layer utilizes a multi-head self-attention architecture consisting of eight attention heads, each with a dimension of 64. The multi-head attention mechanism captures the importance and relevance of temporal dependency features from different perspectives by concurrently computing attention weight matrices across multiple subspaces, generating a feature attention matrix. This feature attention matrix weights the importance of different time steps and feature dimensions in the input data, ensuring that the model focuses on the data features most critical for power regulation strategy prediction, thereby improving the model's ability to understand and process complex photovoltaic data. The feature attention matrix is ​​then fed into the power prediction layer of the photovoltaic multimodal deep learning prediction model for power prediction. The power prediction layer consists of three fully connected layers with 512, 256, and 128 neurons, respectively, and uses the ReLU activation function to introduce nonlinear mapping. The fully connected layers reduce the dimensionality and abstract the feature attention matrix through layer-by-layer weight transformations to generate a power prediction vector. This power prediction vector is then processed by the constrained optimization layer within the photovoltaic multimodal deep learning prediction model. This layer utilizes a multi-objective optimization algorithm, where the grid dispatch instructions are set as hard constraints and the installed capacity change rate is set as a soft constraint. By balancing the conflicts between hard and soft constraints, a maximum power tracking parameter set that meets grid safety and stability requirements is determined, ensuring efficient and stable power output during system operation. The maximum power tracking parameter set is then fed into the inverter parameter optimization layer within the photovoltaic multimodal deep learning prediction model for optimization. This inverter parameter optimization layer consists of two fully connected layers and uses the Sigmoid activation function.This layer uses an optimization algorithm to adjust and optimize various inverter parameters, generating an optimized inverter parameter set to ensure maximum power output within safety constraints. This optimized inverter parameter set is then fed into the policy generation layer of the photovoltaic multimodal deep learning prediction model for processing. The policy generation layer employs a hierarchical reinforcement learning architecture, comprising a policy neural network and a value assessment network. Through reinforcement learning, the control strategy is dynamically adjusted based on current grid demand and inverter status, generating an initial control strategy. This hierarchical reinforcement learning architecture enables more efficient decision-making within multi-level control tasks, ensuring high accuracy and stability of the control strategy. The initial control strategy is then fed into the permission control layer of the photovoltaic multimodal deep learning prediction model for permission assignment and verification. The permission control layer includes a three-level permission verification module, which manages control policy access based on user level. Users with management permissions have full control, while users with maintenance and operation permissions are limited to adjusting or viewing certain parameters. After passing permission verification, active and reactive power regulation strategies are generated.

[0031] S3, generating control instructions based on the active power regulation strategy and the reactive power regulation strategy, performing hierarchical safety constraint verification on the control instructions, and generating remote control instructions and local control instructions;

[0032] It should be noted that the dispatch control module generates commands based on the active power regulation strategy and the reactive power regulation strategy, resulting in a set of power generation control commands containing the inverter output upper and lower limits, as well as the regulation rate parameters. Based on this set of power generation control commands, remote and local control parameters are calculated according to different control scenarios. Remote control parameters are calculated based on the automatic control targets issued by the master station and dynamically adjusted using a closed-loop control algorithm. Real-time feedback ensures that the inverter power output remains within the controllable range relative to the set target value, generating remote control benchmark commands. Local control parameters are calculated based on the power generation plan issued in advance by the master station and predictively adjusted using an open-loop control algorithm. Without relying on real-time feedback, the inverter power gradually follows the predetermined power generation plan trajectory, generating local control benchmark commands. This dual mechanism of remote closed-loop control and local open-loop control enables the system to flexibly respond to grid dispatch requirements and equipment operating conditions in different scenarios. To ensure the safety and effectiveness of the control commands, hierarchical safety constraint verification is applied to both remote and local control benchmark commands. In the first stage, device fault detection is performed. By evaluating key indicators such as the inverter's operating status, power supply status, and grounding status, the system determines whether the device is operating normally and generates a first-level safety test result. If a device fault, abnormal power supply, or grounding fault occurs, the system marks the corresponding command as risky and suspends control execution to ensure device safety. After device fault detection is complete, communication status detection is performed. The system evaluates the communication status and delay between the device and the master station to ensure data transmission reliability. Communication delay is determined based on a preset time threshold. When the delay exceeds the threshold, the system determines that the current communication status is unstable and generates a second-level safety test result. This process aims to ensure the timely transmission and feedback of control commands, preventing inverter misoperation or system instability caused by communication anomalies. After passing the second-level safety test, electrical parameter detection is performed. By real-time evaluating the inverter's voltage and power values ​​to ensure they are within the safe operating range, the system verifies the executableness of the control command and generates a third-level safety test result. Electrical parameter detection prevents device damage or grid safety hazards caused by inverter voltage or power exceeding the permitted range, ensuring the safe execution of control commands. Based on the results of the third-level safety checks, control lockout signals are generated. These signals primarily include power increase and power decrease lockout signals. These signals automatically lock out corresponding control functions in the event of safety issues such as equipment failure, communication anomalies, or electrical parameter violations, thereby preventing the spread of risks. These control lockout signals have a high safety priority and are only released manually. This ensures manual intervention and control restoration in emergency situations, ensuring the safety and stability of power plant operations.The generated control blocking vector is combined with the remote control reference command to form the remote control command, and the blocking control vector is combined with the local control reference command to form the local control command. These two types of control commands are transmitted to the corresponding inverter devices through different execution paths and are executed in an orderly manner based on the current system operating status and safety constraints, thereby achieving real-time active and reactive power regulation goals of the PV power plant.

[0033] S4, based on remote control commands and local control commands, performs hierarchical reinforcement learning control and power optimization adjustment on the inverters in multiple control groups to obtain the total output value of the group;

[0034] Specifically, remote and local control commands are input into the high-level decision module of the hierarchical reinforcement learning control network for processing. The high-level decision module includes a policy neural network and a value assessment network. The policy neural network consists of three convolutional layers and two fully connected layers, while the value assessment network consists of two convolutional layers and three fully connected layers. The policy neural network extracts and abstracts the features of the input control commands, and combined with the value assessment network to quantitatively analyze the current control state, the system generates an optimal set of group control actions in a high-dimensional space. The group control action set is encoded into a control state, encoding each control group as an independent control unit. Detailed state information is attached to each unit, including real-time parameters such as group identification, number of inverters, total installed capacity, and current power output. This encoded control state information is integrated into a group control state vector by the system's data integration module for subsequent control decision-making and state evaluation. The group control state vector is input into the execution control module, where the effectiveness of the control strategy is evaluated by defining a precise reward calculation function. During the reward calculation process, the group control reward is calculated by comprehensively considering three key evaluation metrics: power tracking deviation, regulation response time, and system stability. Power tracking deviation is assessed by calculating the difference between the group's actual power and target power. Regulation response time is measured by recording the time required for the system to complete regulation tasks. System stability is comprehensively evaluated based on the amplitude and frequency of power fluctuations during inverter regulation. This process enables the system to quantitatively evaluate the effectiveness of the current control state, providing feedback for subsequent reinforcement learning updates. Based on the group control reward, a deep Q-learning algorithm is used to optimize and update the group control state vector. This process incorporates an experience replay buffer and a target network mechanism. By storing past control experience and calculating the target Q value for the current state, the optimal control strategy is iteratively updated using a temporal difference algorithm to calculate the optimal control action for the current state. The introduction of the deep Q-learning algorithm effectively improves the system's learning efficiency and decision-making accuracy, enabling the system to continuously optimize the control strategy in a dynamic environment and achieve precise power regulation of the inverter group. The updated optimal control action is input into the power distribution module. Based on the operating status and capacity parameters of each inverter, the power is accurately distributed through a weighted distribution algorithm to calculate the target power command for each inverter. The power distribution algorithm achieves optimal power distribution through weight adjustment, taking into account the real-time operating status, installed capacity, and equipment load capacity of each inverter. This ensures that the power output of the entire group meets the target power demand without causing load overload on individual inverters. The inverter power command is executed and verified through the power regulation judgment mechanism. When the system detects that the deviation between the total power of the group and the target power exceeds the preset safety range, the regulation control mechanism is immediately triggered to generate a power regulation trigger signal.Based on the trigger signal, each inverter performs coordinated optimization of active and reactive power. This optimization algorithm ensures that active power regulation simultaneously meets the grid's reactive power requirements, maintaining voltage stability and system security. During this process, the inverter's real-time power output data is continuously monitored and weighted summed across all inverters to ultimately determine the total output value for the entire control group.

[0035] The optimal control action is transmitted as input data to the power distribution module's data processing unit, where the system performs a strategic analysis of the optimal control action. This analysis derives the overall group power target vector, reflecting the target power level that each PV group needs to achieve at a specific moment. The group power target vector is then divided into groups. Based on the inverter's geographic location and electrical connections, the inverters are appropriately grouped and calibrated through distributed grid topology analysis and inter-site network communication configuration, generating group configuration data. This process ensures clear grouping information for the inverters' spatial layout and physical electrical connections, enabling precise power regulation within the group while reducing inter-group power coupling and unnecessary system interference. Based on the group configuration data, the installed capacity of each inverter is normalized. By normalizing the actual installed capacity of each inverter, a capacity normalization coefficient is generated. To adapt to dynamic changes in the inverter's operating status, the capacity normalization coefficient is dynamically adjusted. This dynamic adjustment coefficient is calculated by real-time weighting based on the inverter's communication latency and device status indicators. Communication latency reflects the delay in data transmission between the inverter and the control system. Device status indicators include parameters such as the inverter's real-time operating temperature, output efficiency, and fault status. By comprehensively evaluating these indicators, different inverters are assigned weights appropriate to their current operating status, making power allocation more accurate and efficient. A matrix operation is performed on the dynamic adjustment coefficient and the group power target vector. This matrix calculation distributes the overall power target to each inverter within the group, generating a set of initial power allocation values. To ensure safe and stable inverter operation, the initial power allocation values ​​are clipped. By incorporating the inverter's power adjustment range (including upper and lower output limits) into the constraint calculation, each inverter's output power remains within a safe range. This results in a set of power limit control values. Based on the power limit control values, power change rate optimization is performed. The control algorithm is adjusted to ensure that the inverter's power adjustment rate meets the dynamic requirements of grid operation, avoiding grid oscillation or system instability caused by excessively rapid power changes. This optimization process generates a set of optimized control values ​​by smoothing the power adjustment rate and limiting the power change slope. On this basis, the optimized control value is converted into an instruction according to the communication protocol supported by each inverter, and the standardized control data is converted into an instruction format that meets the communication interface requirements of each inverter, ensuring that the instruction can be accurately recognized and executed by the inverter to obtain the inverter power instruction.

[0036] S5, monitor the total output value of the group and the operating status of the equipment in real time and generate operating parameter data;

[0037] Data acquisition modules are deployed at each inverter and group level. These modules connect to the inverter's control interface via a high-speed communication network and acquire real-time inverter operating parameters, including key indicators such as power output, voltage, current, temperature, device fault status, and communication status. This data is transmitted to a centralized monitoring server via fieldbus or Ethernet. Security measures such as SSL encryption and proprietary protocols are employed during transmission to ensure data integrity and confidentiality. A high-performance data processing unit is deployed on the centralized monitoring server to parse and store the received real-time data. This data processing unit pre-processes the raw data, including data format conversion, outlier filtering, and time synchronization, to ensure data accuracy and consistency. The pre-processed data is stored in a real-time database. As needed, some key data is stored in an in-memory database to improve data access speed and system responsiveness. To support historical data query and analysis, real-time data is regularly archived in a historical database, compressed, and backed up to ensure long-term data availability and security. The total group output and device operating status are calculated and monitored in real time. By weightedly summing the real-time power output of each inverter, the total output value for each group is calculated. This value is then compared with the preset target power to calculate power deviations and adjustment requirements. The system monitors the operating status of each inverter, including its operating mode (e.g., standby, operating, fault), communication status (online, offline, latency), and key operating parameters (e.g., temperature, voltage, current, etc.). The monitoring interface uses visualization technology to graphically display the total group output value, device operating status, and key operating parameters. For example, a real-time graph can be used to display the changing trend of the total group output value, a bar chart or pie chart can be used to display the power contribution ratio of each inverter, and color coding or icons can be used to indicate device operating status and fault information. The interface supports alarms and notifications. When a device fault occurs or an operating parameter exceeds a preset threshold, the system promptly sends an alarm to operations and maintenance personnel, helping them quickly locate and resolve the problem. To improve the system's real-time performance and reliability, the monitoring system requires high concurrency and fault tolerance. At the software level, a distributed architecture and multi-threading technology are used to ensure the system can simultaneously handle large amounts of device data and user requests. At the hardware level, high-performance servers and network equipment should be deployed, and load balancing and redundant backup technologies should be used to prevent single points of failure from impacting the system. Furthermore, the system should be scalable, capable of flexibly increasing computing and storage resources as the PV plant scales and the number of devices increases, maintaining stable system performance. To ensure data security and compliance, the system must adhere to relevant national standards and industry regulations. Encryption and authentication mechanisms should be implemented during data transmission and storage to prevent unauthorized access and data leakage.In terms of data storage and processing, the principle of data minimization is followed, only necessary operating parameter data is collected and stored, and unnecessary data is regularly cleared. In terms of logging and auditing, user operating behaviors and system operation status are recorded to support post-audit and accountability tracing. To meet the needs of different users, the system needs to support multi-level permission management and personalized customization. Through the permission control module, the system grants different access rights and operating permissions based on the user's role and responsibilities. For example, managers can view all operating parameter data and historical records and perform system configuration and maintenance; operation and maintenance personnel can view the equipment status and alarm information in their area of ​​responsibility and perform remote control operations on the equipment; ordinary users can only view public operating data and basic information.

[0038] S6, based on the operating parameter data, calculates the active power qualification rate, control response time, control accuracy and system fluctuation rate, and performs multi-objective optimization on the adjustment dead zone range, control gain and response time constant to obtain the target power allocation plan.

[0039] Specifically, the system utilizes real-time operating parameter data collected, including the inverter's real-time output power, device operating status, control command execution feedback, and communication data with the grid dispatching system. This data enters the data processing module via a high-speed data transmission channel. Power deviation calculation is then performed. The real-time power value is compared with the control target value, and the difference between the two is calculated point by point to generate power tracking error data. Based on this power tracking error data, a statistical analysis of the active power qualification rate is performed. By screening and evaluating the time series of the power error data, the proportion of time the active power remains within the allowable deviation range is calculated to obtain the active power qualification rate. Simultaneously, a time series response analysis is performed on the operating parameter data, specifically the time difference between the issuance of a control command and the completion of its execution by the device. By monitoring the timestamps of the control command and the feedback data, the control response time is accurately calculated. These two metrics together reflect the system's control accuracy and execution efficiency. The active power qualification rate and control response time are input into a multi-objective genetic optimizer. The genetic optimizer, a global search algorithm, optimizes the parameters using a non-dominated sorting genetic algorithm. During the iterative process of the genetic algorithm, through steps such as population initialization, fitness calculation, non-dominated sorting, crossover mutation, and selection, the Pareto optimal solution boundary is continuously approached. A set of Pareto optimal solutions that meet multiple objective constraints is generated. These solutions contain optimal combinations of different deadband ranges, control gains, and response time constants, fully balancing the interactions between control accuracy, response time, and system stability. The deadband range is optimized based on the Pareto optimal solution set. By analyzing the system power fluctuation rate and regulation stability corresponding to different solutions, an optimal deadband vector is determined that best suits the current system operating state. This optimized vector effectively reduces the control system's ineffective regulation range, thereby improving control sensitivity and regulation accuracy. The resulting deadband vector is then used to optimize the PID controller parameters. The PID controller optimization process includes calculating the optimal control gain and response time constant. By integrating and differentiating the error signal and combining it with the genetic algorithm, the PID parameters (proportional gain P, integral time constant I, and differential time constant D) are adjusted to achieve the optimal balance between power response time and stability, resulting in a set of optimized controller parameter vectors. Based on the capacity ratio of each inverter, the optimized controller parameter vector is allocated, and each control parameter is weighted according to the installed capacity, current operating status and response capability of the inverter. This enables each inverter to output the most appropriate power adjustment command based on its own characteristics, thereby generating the target power allocation plan for the entire system.

[0040] In one example, photovoltaic power station operation data is collected and encrypted for transmission in real time to obtain power generation data, voltage data, current data, and solar radiation data, including:

[0041] Perform telemetry collection on the inverter operating data to obtain the inverter telemetry data set, which includes the inverter's real-time startup capacity value, power change rate value, output upper limit value, and output lower limit value. Active power is calculated to obtain the inverter power adjustment range.

[0042] Input the inverter telemetry data set into the SSL security encryption module for private protocol encryption to obtain the first-level encrypted data packet, and perform directional server routing on the first-level encrypted data packet to obtain encrypted transmission data;

[0043] The encrypted transmission data is input into the cloud platform for vertical encryption and dynamic key update to obtain the second-level encrypted data packet, and the second-level encrypted data packet is distributed deployed to obtain real-time database storage records;

[0044] The real-time database storage records are classified and stored according to data type and timestamp to obtain historical database backup records, and the historical database backup records are firewall-isolated and security-verified to obtain power generation data, voltage data, current data and solar radiation data.

[0045] In this example, a real-time data acquisition module is deployed at the inverter level to collect the inverter's operating data set, including the real-time startup capacity C, through remote terminal units or smart sensors. inv , power change rate R power , output upper limit P max And the output lower limit P min For the calculation of active power, the real-time voltage V output by the inverter is used. inv , output current I inv And the power factor cos(θ) is calculated in real time. The specific formula is expressed as:

[0046] P act =V inv I inv cos(θ);

[0047] Among them, P act Indicates real-time active power in kW, V inv is the output voltage (V), I inv is the output current (A), and cos(θ) is the power factor, which is usually close to 1 when the inverter is operating normally. act , the system calculates the power regulation range P of the inverter range , whose range is expressed by the following formula:

[0048] P range =[P min ,P max ] Among them Padj =P max -P act ;

[0049] Among them, P adj Is the remaining adjustable power, which is used to indicate the current output margin of the inverter. The collected inverter operation data set D inv ={C inv ,R power ,P act ,P max ,P min} is transmitted as input to the SSL security encryption module and encrypted using the Advanced Encryption Standard. The specific encryption process is described as follows:

[0050] E1=Encrypt(D inv ,K priv );

[0051] Among them, E1 is the first level encrypted data packet, D inv is the inverter telemetry data set, K priv It is a static encryption cryptographic protocol of a private protocol. The generated E1 data packets are transmitted through a directional server routing distribution technology. This process ensures that data is distributed along the optimal path based on the network topology and the location of the target server, thereby achieving secure data transmission. The transmission path is defined as:

[0052] R path =Route(E1,S target );

[0053] where R path is the transmission path of the data packet, S target is the target server address. The encrypted data E1 after routing is uploaded to the cloud platform. In the cloud platform, the data is vertically encrypted. By introducing a dynamic key update mechanism, the key is ensured not to be reused. The vertical encryption formula is:

[0054] E2=Encrypt(E1,K dyn );

[0055] Among them, E2 is the second level encrypted data packet, K dyn It is a dynamic key that is automatically updated every period T. The cloud platform stores the encrypted data packet E2 in a real-time database through a distributed deployment mechanism to achieve high data availability and redundant backup. After the data is stored in the real-time database, it is classified and stored according to data type and timestamp. The specific classification formula is as follows:

[0056] D class ={D t |t∈[t0,t n],Type(D)};

[0057] Among them D class It is a collection of classified and stored data. t It represents the data record at a specific time t, and Type (D) indicates the type of data, such as power generation, voltage, current, solar radiation, etc. The classified data is archived to the historical database and subjected to security verification. The verification process is implemented through a hash function, and the formula is as follows:

[0058] H=Hash(D class );

[0059] Where H is the data hash value generated by the checksum, which is used to verify the integrity and tamper-proof of the data. During the storage process, the historical database is protected by the firewall business isolation mechanism to ensure that the data will not be subject to unauthorized access or malicious attacks. gen , voltage data V inv , current data I inv and solar radiation data S rad Historical database records.

[0060] In one example, power generation data, voltage data, current data, and solar radiation data are input into a photovoltaic multimodal deep learning prediction model for processing to generate active power regulation strategies and reactive power regulation strategies, including:

[0061] The power generation data, voltage data, current data, and solar radiation data are input into the multi-layer convolutional feature extraction layer of the photovoltaic multimodal deep learning prediction model for processing. The multi-layer convolutional feature extraction layer contains four convolutional modules, each of which consists of two convolutional layers and a maximum pooling layer, to obtain the power station operation feature vector;

[0062] The power plant operation feature vector is input into the bidirectional LSTM time series analysis layer in the photovoltaic multimodal deep learning prediction model for processing. The bidirectional LSTM time series analysis layer contains two LSTM layers, each with 128 hidden units, to obtain the time series dependency features.

[0063] The time-dependent features are input into the multi-head attention layer of the photovoltaic multimodal deep learning prediction model for processing. The multi-head attention layer adopts a multi-head self-attention structure, contains 8 attention heads, and the dimension of each attention head is 64, to obtain the feature attention matrix;

[0064] The feature attention matrix is ​​input into the power prediction layer of the photovoltaic multimodal deep learning prediction model for processing. The power prediction layer contains three fully connected layers with 512, 256, and 128 neurons respectively. The ReLU activation function is used to obtain the power prediction vector.

[0065] The power prediction vector is input into the constrained optimization layer of the photovoltaic multimodal deep learning prediction model for processing. The constrained optimization layer uses a multi-objective optimization algorithm, sets the grid dispatch instruction as a hard constraint condition, and the installed capacity change rate as a soft constraint, to obtain the maximum power tracking parameter set.

[0066] The maximum power tracking parameter set is input into the inverter parameter optimization layer in the photovoltaic multimodal deep learning prediction model for processing. The parameter optimization layer contains two fully connected layers and uses the Sigmoid activation function to obtain the inverter optimized parameter set.

[0067] The inverter optimization parameter set is input into the strategy generation layer of the photovoltaic multimodal deep learning prediction model for processing. The strategy generation layer adopts a hierarchical reinforcement learning structure, including a policy neural network and a value evaluation network, to obtain the initial control strategy.

[0068] The initial control strategy is input into the permission control layer of the photovoltaic multimodal deep learning prediction model for processing. The permission control layer contains a three-level permission verification module, which allocates control permissions according to user levels to obtain active power regulation strategies and reactive power regulation strategies.

[0069] In this example, an end-to-end photovoltaic prediction system is constructed, in which the data input layer includes a set of photovoltaic power station operation data collected in real time, including the power generation P gen , voltage V inv , current I inv and solar radiation data S rad These data are normalized and mapped to the standardized range required for model training. The normalization formula is:

[0070]

[0071] where X norm is the normalized data, X is the original input data, X min and X max The normalized input data is used as the input vector X. input ={P gen ,V inv ,I inv ,S rad The input data is processed by the multi-layer convolutional feature extraction layer in the photovoltaic multimodal deep learning prediction model. In the feature extraction stage, the model is equipped with four convolutional modules, each consisting of two convolutional layers and a maximum pooling layer. For the i-th convolutional layer, the input data is subjected to f convolution kernels for feature extraction, and the output feature map is calculated as follows:

[0072]

[0073] in is the output feature map of layer l, is the weight matrix of the i-th convolution kernel, * represents the convolution operation, is the bias term and σ is the activation function (e.g. ReLU). After the convolution operation, the dimension is reduced by the maximum pooling operation. The formula for maximum pooling is:

[0074]

[0075] Among them, P i represents the feature map after pooling, k×k is the pooling window size. After layer-by-layer processing by four convolutional modules, the high-dimensional feature vector F of the power station operation is obtained. conv 。 conv The input is processed into the bidirectional LSTM time series analysis layer in the photovoltaic multimodal deep learning prediction model. The bidirectional LSTM consists of two LSTM layers, each containing 128 hidden units. For each time step t, the LSTM is processed by the input vector x t and hidden state h t-1 Calculate the current hidden state h t and output o t , the specific calculation is as follows:

[0076] f t =σ(W f ·[h t-1 ,x t ]+b f ),i t =σ(W i ·[h t-1 ,x t ]+b i );

[0077]

[0078] o t =σ(W o ·[h t-1 ,x t ]+b o ),h t =o t ⊙tanh(C t );

[0079] where f t 、i t 、o t are the weights of the forget gate, input gate, and output gate, respectively. t is the cell state, h tis the hidden state, W and b are the weight matrix and bias vector respectively. Through the bidirectional LSTM structure, the forward and reverse dependencies of the time series data are captured to generate the time series dependency feature vector F lstm 。 lstm Input to the multi-head attention layer. The multi-head attention mechanism contains 8 attention heads, each with a dimension of 64. The attention weight matrix A of the input features is calculated through the self-attention mechanism. The formula is:

[0080]

[0081] Where Q, K, and V are query, key, and value matrices respectively, d k is the dimension of the key vector, A is the attention score matrix, and the feature attention matrix F is obtained by weighted calculation attn 。 attn The power prediction layer is input to the power prediction layer, which contains three fully connected layers with 512, 256 and 128 neurons respectively. The ReLU activation function is used for nonlinear transformation and the power prediction vector P is output. pred For the jth layer, the calculation formula of the fully connected layer is:

[0082] H j =ReLU(W j ·H j-1 +b j );

[0083] Where W j is the weight matrix, b j is the bias vector, H j Is the output vector of the current layer. The power prediction vector P pred Input to the constraint optimization layer, the constraint optimization layer adopts multi-objective optimization algorithm to calculate the power grid dispatch instruction P grid As a hard constraint, the inverter installed capacity change rate R inv is a soft constraint and is solved by the following objective function:

[0084] minJ=∥P pred -P grid ∥ 2 +λ·R inv ;

[0085] Where λ is the trade-off coefficient, J is the objective function, and the optimized maximum power tracking parameter set P mppt As output. mppt Input the inverter parameter optimization layer, which contains two fully connected layers and uses the Sigmoid activation function for nonlinear mapping to obtain the inverter optimization parameter set P opt Then P optInput strategy generation layer, strategy generation layer adopts hierarchical reinforcement learning structure, in which the strategy neural network generates control strategy, the value evaluation network evaluates the strategy, and generates the initial control strategy π through the optimal value function Q(s,a) init π init Input authority control layer, which includes three-level authority verification module, assigns control authority by verifying user level, and outputs the final active power regulation strategy P ctrl and reactive power regulation strategy Q ctrl .

[0086] In one example, control instructions are generated based on the active power regulation strategy and the reactive power regulation strategy, and hierarchical safety constraint verification is performed on the control instructions to generate remote control instructions and local control instructions, including:

[0087] Based on the active power regulation strategy and the reactive power regulation strategy, instruction generation processing is performed to obtain a power generation control instruction set, which includes an inverter output upper limit value, an output lower limit value, and a regulation rate parameter;

[0088] Calculate remote control parameters for the power generation control instruction set, and use a closed-loop control algorithm based on the automatic control target issued by the master station to obtain the remote control benchmark instruction;

[0089] Calculate local control parameters for the power generation control instruction set, and use an open-loop control algorithm based on the power generation plan issued by the master station to obtain local control benchmark instructions;

[0090] Perform equipment fault detection on remote control benchmark commands and local control benchmark commands, and obtain first-level safety detection results by evaluating the inverter operating status, power supply status and grounding status;

[0091] Perform communication status detection on the first-level security detection results, and obtain the second-level security detection results by evaluating the device communication status and communication delay status. The communication delay status is determined by the preset time threshold;

[0092] Conduct electrical parameter testing on the second-level safety test results to assess whether the inverter voltage and power values ​​are within the safe range to obtain the third-level safety test results.

[0093] Calculate the control blocking signal based on the third-level safety detection result. The control blocking signal includes a power increase blocking signal and a power decrease blocking signal. The control blocking signal can only be released by manual operation to obtain a blocking control vector.

[0094] The blocking control vector is combined with the remote control reference instruction to generate the remote control instruction, and the blocking control vector is combined with the local control reference instruction to generate the local control instruction.

[0095] In this example, the generated active power regulation strategy P is generated according to the real-time dispatch requirements of the power grid and the operating status of the photovoltaic power station. ctrl and reactive power regulation strategy Q ctrl Decompose to form a power generation control instruction set, which includes the inverter output upper limit value P max , output lower limit P min and the adjustment rate parameter R adj The output upper and lower limits are calculated using the following formula:

[0096] P max =min(C inv ,P grid ),P min =max(0,P act -R adj T);

[0097] Among them, C inv is the current installed capacity of the inverter, P grid is the maximum power allowed by the grid dispatch, P act is the current active power, R adj Represents the power regulation rate, and T is the time interval. Through the above formula, the system ensures that the inverter power output meets the grid dispatching requirements and the equipment safe operation conditions. The remote control parameter calculation of the power generation control instruction set is carried out. This stage uses a closed-loop control algorithm to process the target. The goal is to make the inverter power output follow the automatic control target P issued by the master station in real time. target The closed-loop control is achieved through the PID controller, and the control error is calculated as follows:

[0098] e(t)=P target -P act ;

[0099] Where e(t) is the power error at the current moment. The output u(t) of the PID closed-loop controller is calculated using the following formula:

[0100]

[0101] where K p , K i and K d are proportional, integral and differential gain coefficients respectively. The system generates remote control reference instruction P through the adjustment output of PID controller. ref,remote , thus achieving real-time closed-loop tracking of active power. At the same time, the local control parameter calculation of the power generation control instruction set is performed. This process is based on the power generation plan P issued by the master station. plan, using open-loop control algorithm for processing. Open-loop control does not require real-time feedback, and directly calculates the target power according to the power generation plan curve, and controls the benchmark instruction P ref,local Expressed as:

[0102] P ref,local =P plan (t);

[0103] Among them, P plan (t) is the target value of the power generation plan at time t. Through closed-loop and open-loop control algorithms, remote control benchmark instructions and local control benchmark instructions are generated. In order to ensure the safety and reliability of the control instructions, equipment fault detection is performed on the remote control benchmark instructions and local control benchmark instructions. By real-time monitoring of the inverter's operating status S inv , power supply status S power and ground state S ground , to determine whether the equipment is in normal operating conditions. The first-level safety test results are defined by the following logical conditions:

[0104] S safe,1 =(S inv ∧S power ∧S ground );

[0105] If S safe,1 = True, the device fault detection is passed, otherwise the device is marked abnormal. After passing the device fault detection, the communication status detection is performed. comm and communication delay T delay To judge whether the communication is normal, the judgment formula of communication delay is:

[0106] S safe,2 =(S comm ∧T delay <T thresh );

[0107] Where T thresh is the communication delay threshold. If the communication delay exceeds the threshold, the system will mark the communication status as abnormal. After the communication status is detected, the electrical parameters are tested to evaluate the inverter voltage V inv and power P act Whether it is within the safe range, the detection conditions are:

[0108] S safe,3 =(V min ≤V inv ≤V max )∧(P min ≤P act ≤P nax );

[0109] Where V min and V max After equipment fault detection, communication status detection and electrical parameter detection, a control blocking signal is generated, including a power increase blocking signal S lock,+ and power reduction lock signal S lock,- These signals can only be released through manual operation, and eventually form the locking control vector S lock . The locking control vector and the remote control reference instruction P ref,remote Combine to generate the final remote control command P final,remote :

[0110] P final,remote =P ref,remote ·S lock ;

[0111] At the same time, the locking control vector and the local control reference instruction P ref,local Combine to generate the final local control instruction P final,local :

[0112] P final,local =P ref,local ·S lock .

[0113] In one example, hierarchical reinforcement learning control and power optimization adjustment are performed on inverters in multiple control groups based on remote control commands and local control commands to obtain the total output value of the group, including:

[0114] The remote control commands and local control commands are input into the high-level decision module of the hierarchical reinforcement learning control network for processing. The high-level decision module includes a policy neural network and a value evaluation network. The policy neural network contains 3 convolutional layers and 2 fully connected layers, and the value evaluation network contains 2 convolutional layers and 3 fully connected layers. The group control action set is obtained.

[0115] Perform control state encoding on the group control action set, encode each control group into a control unit, including group identification, number of inverters, total installed capacity, and current power output status information, and obtain the group control state vector;

[0116] The group control state vector is input into the execution control module for reward calculation. The reward calculation includes three evaluation indicators: power tracking deviation, regulation response time, and system stability, and the group control reward value is obtained.

[0117] Based on the group control reward value, the group control state vector is updated through deep Q learning. The target Q value is calculated through the temporal difference algorithm using the experience replay buffer and target network mechanism to obtain the optimal control action.

[0118] The optimal control action is input into the power distribution module for group power regulation. Based on the operating status and capacity parameters of each inverter, a weighted distribution algorithm is used to calculate the target power of each inverter to obtain the inverter power command;

[0119] Perform power regulation judgment on the inverter power command. When the deviation between the total power of the group and the target power exceeds the preset range, the regulation control is triggered to obtain a power regulation trigger signal.

[0120] Based on the power regulation trigger signal, the active power and reactive power of each inverter are collaboratively optimized to obtain the real-time power data of the inverter. The real-time power data of the inverter is weighted and summed to obtain the total output value of the group.

[0121] In this example, the remote control command P remote And local control instruction P local As input, it is passed into the high-level decision module of the hierarchical reinforcement learning control network. This module contains a policy neural network and a value evaluation network. The policy neural network extracts policy-related decision information from the input features. Its network structure includes 3 convolutional layers and 2 fully connected layers. The output feature map of the first convolutional layer is Calculated by the following formula:

[0122]

[0123] in is the weight matrix of the i-th convolution kernel, is the corresponding bias, * represents the convolution operation, and σ is the ReLU activation function. After extracting deep features through the stacking of multiple layers of convolution, it enters the fully connected layer and obtains the output vector through linear transformation:

[0124] Z policy =σ(W fc ·F conv +b fc );

[0125] Where W fc and b fc are the weights and biases of the fully connected layer, Z policy is the output vector of the policy network. Meanwhile, the value evaluation network is used to evaluate the value of the current policy. Its structure consists of 2 convolutional layers and 3 fully connected layers. It calculates the value function V(s) of the current state through convolution and linear transformation. The specific formula is:

[0126] V(s)=W value ·F value +b value ;

[0127] Among them F valueIs the convolution output of the value network, V(s) represents the value evaluation result under the current state s. The output of the policy network and the value network jointly determine the group control action set A group , where A group Contains power adjustment actions for the group. The control state of the group control action set is encoded, and each control group is encoded as an independent control unit. The characteristics of the control unit include the group identifier G id 、Number of inverters inv , total installed capacity C total and the current power output state P current , these features are combined into the group control state vector S group :

[0128] S group ={G id ,N inv ,C total ,P current};

[0129] S group Input execution control module to calculate reward, which includes three indicators: power tracking deviation e(t), adjustment response time T response and system stability S stability The power tracking deviation is calculated using the following formula:

[0130] e(t)=P target -P current ;

[0131] Among them, P target is the target power, P current is the current power output. The reward function R is defined as:

[0132] R=-α|e(t)|-βT response +γS stability ;

[0133] Among them, α, β, γ are weight coefficients, S stability Represents the inverse value of the system power fluctuation, which is used to measure the stability of the system. Based on the group control reward value R, the group control state vector S group Perform deep Q learning updates, store historical states and rewards through the experience replay buffer, and use the temporal difference (TD) algorithm to calculate the target Q value Q in combination with the target network mechanism target :

[0134]

[0135] Where r is the current reward, γ is the discount factor, Q(s ′ ,a ′) is the next state s ′ Next take action a ′ The Q value of target -Q(s,a)) 2 , the system updates the parameters θ of the Q network and obtains the optimal control action A opt The optimal control action is input into the power distribution module, and the system adjusts the power distribution according to the operating status S of each inverter. inv and installed capacity C inv , the target power P of each inverter is calculated using the weighted allocation algorithm inv , the formula is as follows:

[0136]

[0137] Among them C inv,i is the installed capacity of the i-th inverter, P group is the target power of the group, and N is the number of inverters. The calculated inverter power command is used to make power adjustment decisions, and the total power of the group P is determined by total With the target power P target The deviation e, when |e|>ΔP thresh When the adjustment control is triggered, the power adjustment trigger signal S is generated trigger Based on the adjustment trigger signal, the active power and reactive power of each inverter are optimized in a coordinated manner, and the power output P is calculated in real time through the optimization control algorithm. opt,i and reactive power Q opt,i , and finally the total output value of the group P is obtained by real-time summation total :

[0138]

[0139] In one example, the optimal control action is input into the power distribution module for group power regulation. Based on the operating status and capacity parameters of each inverter, a weighted distribution algorithm is used to calculate the target power of each inverter to obtain the inverter power instruction, including:

[0140] The optimal control action is input into the data processing unit of the power allocation module for strategy analysis to obtain the group power target vector;

[0141] Divide the group power target vector into groups, group and calibrate the inverters based on their geographic location data and electrical connection relationships, and obtain group configuration data;

[0142] Normalizing the installed capacity of each inverter based on the group configuration data to obtain a capacity normalization coefficient;

[0143] Dynamically correct the capacity normalization coefficient and adjust the weight according to the inverter's communication delay and device status indicators to obtain a dynamic adjustment coefficient;

[0144] Perform matrix operations on the dynamic adjustment coefficient and the group power target vector to obtain the initial power allocation value of each inverter. The initial power allocation value is then limited. Constraint calculation is performed based on the inverter power adjustment range to obtain the power limit control value.

[0145] The power change rate is optimized based on the power limit control value to ensure that the power regulation rate meets the grid operation requirements, and the optimized control value is obtained. The optimized control value is then converted into an instruction according to the communication protocol of each inverter to obtain the inverter power instruction.

[0146] In this example, the optimal control action A calculated by the upper decision network or reinforcement learning module is opt As the data processing unit of the input power distribution module, the module maps the input action through strategy analysis and converts the optimal control action A opt Decomposed into group power target vector P group , which is expressed as:

[0147] P group,i =A opt W group,i ;

[0148] Among them, P group,i is the target power of the i-th group, W group,i is the corresponding group weight coefficient, and the weight is allocated based on the group priority or total installed capacity. Through this analytical process, the power target vector P of all groups is obtained group ={P group,1 ,P group,2 ,…,P group,n}. For the group power target vector P group Divide the groups based on geographic location data L geo And electrical connection relationship R elec All inverters are logically grouped and calibrated to form group configuration data. The specific rules for group calibration are as follows:

[0149] G k ={I j ∣L geo,j ∈L k ∧R elec,j ∈R k};

[0150] Among them, G k represents the kth group, I j represents the jth inverter, L kand R k are the geographical location area and electrical connection area of ​​the kth group respectively. Through this process, the group configuration data G = {G1, G2, ..., G k Based on the group configuration data, the installed capacity C of each inverter is inv,j Perform normalization to eliminate the differences in capacity between different inverters and calculate the normalization coefficient α j , the formula is:

[0151]

[0152] Among them, α j is the normalization coefficient of the j-th inverter, C inv,i is the installed capacity of the jth inverter, and N is the number of inverters in the group. The normalization process maps the installed capacity of all inverters to a unified proportional interval to facilitate subsequent power allocation calculations. j Perform dynamic correction, combined with the inverter's communication delay T comm,j and device status indicator S inv,j Calculate the dynamic adjustment coefficient β j The dynamic correction formula is:

[0153]

[0154] Among them, ω is the delay impact factor, T comm,j is the communication delay of the j-th inverter, S inv,j is the equipment status coefficient. When the equipment is normal, S inv,j =1, when fault occurs S inv,j = 0. Dynamic adjustment coefficient β j Based on the consideration of communication delay and equipment status, the power allocation is dynamically modified. The dynamic adjustment coefficient β j and the group power target vector P group Perform matrix operations to obtain the initial power distribution value P of each inverter init,j , the calculation formula is:

[0155] P init,j =β j ·P group,k ,I j ∈G k ;

[0156] Among them, P init,j is the initial power allocation value of the j-th inverter, P group,k For group G k The initial power allocation value is limited, combined with the power adjustment range of the inverter [P min,j,P max,j ] is constrained, and the limit control value P lim,j The calculation formula is:

[0157] P lim,j =min(P max,j ,max(P min,j ,P init,j ));

[0158] Among them, P lim,j Is the power distribution value after limiting, ensuring that the distribution result meets the safe operation constraints of the inverter. Based on the power limit control value P lim,j , optimize the power change rate to ensure the power regulation rate R adj,j To meet the grid operation requirements, the calculation formula is:

[0159] P opt,j =P prev,j +sgn(P lim,j -P prev,j )·min(R adj,j ·Δt,|P lim,j -P prev,j |);

[0160] Among them, P opt,j is the optimized power distribution value, P prev,j is the power value of the previous step, Δt is the time step, sgn is the sign function, R adj,j is the maximum power regulation rate. The optimized power distribution value P opt,j The inverter power command P is generated by converting the command according to the communication protocol of each inverter. cmd,j and sends it to the corresponding inverter for execution, thereby achieving precise adjustment of the power of each inverter.

[0161] In one example, the active power qualification rate, control response time, control accuracy, and system fluctuation rate are calculated based on operating parameter data. Multi-objective optimization is then performed on the regulation deadband range, control gain, and response time constant to obtain a target power allocation solution, including:

[0162] Calculate power deviation based on operating parameter data and obtain power tracking error data by comparing real-time power value with control target value;

[0163] Based on the power tracking error data, the qualified rate statistics are performed to obtain the active power qualified rate, and the timing response analysis is performed on the operating parameter data to calculate the control instruction execution completion time and obtain the control response time;

[0164] The active power qualification rate and control response time are input into the multi-objective genetic optimizer, and the non-dominated sorting genetic algorithm is used to optimize the parameters to obtain the Pareto optimal solution set.

[0165] Based on the Pareto optimal solution set, the parameters of the dead zone range are optimized to obtain the dead zone optimization vector;

[0166] The PID controller is optimized for the dead zone optimization vector, the optimal control gain and response time constant are calculated, and the controller parameter vector is obtained. The controller parameter vector is then allocated according to the inverter capacity ratio to obtain the target power allocation scheme.

[0167] In this example, the operating parameter data including the real-time power value P is obtained from the real-time monitoring system of the inverter. real (t) and control target value P target (t), and the power tracking error e(t) is calculated by the difference between the two. The calculation formula of the power tracking error data is as follows:

[0168] e(t)=P target (t)-P real (t);

[0169] Where e(t) represents the power error at time t, P target (t) is the target power set by the dispatching system, P real (t) is the actual power monitored in real time. By statistically analyzing the time series data e(t), the system's regulation accuracy and real-time tracking capability are evaluated. Based on the power tracking error data e(t), the active power qualification rate is statistically analyzed. The qualification rate R pass It refers to the proportion of time that the actual power error e(t) is within the allowable deviation range [-∈,∈] during the operating cycle T. The calculation formula is as follows:

[0170]

[0171] Where 1(·) is an indicator function that takes 1 when |e(t)|≤∈ and 0 otherwise. ∈ is the allowable error threshold. By statistically evaluating the pass rate, the stability and compliance of the system in achieving performance requirements during the target power adjustment process are evaluated. Timing response analysis is performed on the operating parameter data to calculate the execution completion time of the control instruction. Timing response analysis focuses on the system from the control instruction P cmd Sent to actual power P real Time to reach the set target T resp , which is calculated based on the detection of time points:

[0172] T resp =t settle -t cmd ;

[0173] Among them, t cmd is the time when the control instruction is issued, t settle is the moment when the system is stable near the target value, that is, |P real (t settle )-P target |≤δ, where δ is the stability threshold. By evaluating the timing response, the rapidity and execution efficiency of the system control are measured. The active power qualification rate R pass and control response time T resp Input the multi-objective genetic optimizer and use the non-dominated sorting genetic algorithm to optimize the parameters and find a balance between maximizing the pass rate and minimizing the response time. The objective function of the optimization process is defined as:

[0174] minJ=[-R pass ,T resp ];

[0175] The genetic algorithm iteratively updates the population through operations such as initial population generation, fitness calculation, non-dominated sorting, crossover mutation and elite selection to obtain the Pareto optimal solution set P. Pareto , where each solution corresponds to a set of parameter combinations. The Pareto optimal solution set reflects the control parameter set under different trade-offs, providing a selection basis for subsequent controller optimization. Based on the Pareto optimal solution set, the dead zone range is optimized, and the dead zone range D is adjusted. opt It is defined as the interval in which the controller does not perform adjustment when the error is small. Specifically, the optimal dead zone vector D is obtained through optimization search. opt , the result is expressed as:

[0176] Where J(D) = -R pass +λ·T resp ;

[0177] Among them, λ is the trade-off coefficient. By optimizing the dead zone parameters, frequent adjustment actions can be effectively reduced and the stability of the system can be improved. Based on the optimized dead zone vector D opt Optimize the parameters of the PID controller. The goal of the PID controller is to use the proportional gain K p , integration time K i and differential time K d The power error is adjusted. The PID control output u(t) is expressed as follows:

[0178]

[0179] Search for the optimal K by genetic optimizer p ,K i ,K d, so that the system responds quickly and minimizes errors during the adjustment process. The final optimized PID parameter vector is:

[0180]

[0181] The optimal PID control parameter vector is allocated according to the installed capacity ratio of the inverter. The specific parameter allocation formula is:

[0182]

[0183] Among them, θ inv,j is the PID parameter assigned to the j-th inverter, C inv,j is the installed capacity of the inverter, and N is the total number of inverters. The control output of each inverter is calculated according to the allocated PID parameters and used as the target power allocation scheme P target,inv,j .

[0184] Reference Figure 2 This embodiment provides an optimization device based on a distributed solar photovoltaic power generation control system, including:

[0185] Acquisition module 1 is used to collect and encrypt the operation data of the photovoltaic power station in real time to obtain power generation data, voltage data, current data and solar radiation data;

[0186] Processing module 2 is used to input power generation data, voltage data, current data and solar radiation data into the photovoltaic multimodal deep learning prediction model for processing and generate active power regulation strategy and reactive power regulation strategy;

[0187] Verification module 3, used to generate control instructions based on the active power regulation strategy and the reactive power regulation strategy, and perform hierarchical safety constraint verification on the control instructions to generate remote control instructions and local control instructions;

[0188] Adjustment module 4 is used to perform hierarchical reinforcement learning control and power optimization adjustment on the inverters in multiple control groups according to remote control instructions and local control instructions to obtain the total output value of the group;

[0189] Monitoring module 5, used to monitor the total output value of the group and the operating status of the equipment in real time and generate operating parameter data;

[0190] Optimization module 6 is used to calculate the active power qualification rate, control response time, control accuracy and system fluctuation rate based on the operating parameter data, and perform multi-objective optimization on the adjustment dead zone range, control gain and response time constant to obtain the target power allocation plan.

[0191] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0192] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, memory, display screen, input device, network interface and database connected via a system bus. The processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the above method is implemented.

[0193] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0194] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which implements the above-described method when executed by a processor. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0195] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Among them, any reference to memory, storage, database, or other media provided by the present invention and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM.

[0196] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0197] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An optimization method based on a distributed solar photovoltaic power generation control system, characterized in that: The following steps are involved: Real-time collection and encrypted transmission of photovoltaic power station operation data to obtain power generation data, voltage data, current data and solar radiation data; Inputting the power generation data, voltage data, current data and solar radiation data into a photovoltaic multimodal deep learning prediction model for processing to generate an active power regulation strategy and a reactive power regulation strategy; Generate control instructions based on the active power regulation strategy and the reactive power regulation strategy, and perform hierarchical safety constraint verification on the control instructions to generate remote control instructions and local control instructions; According to the remote control instructions and the local control instructions, hierarchical reinforcement learning control and power optimization adjustment are performed on the inverters in the multiple control groups to obtain a total output value of the group; Monitor the total output value of the group and the operating status of the equipment in real time to generate operating parameter data; Based on the operating parameter data, the active power qualification rate, control response time, control accuracy and system fluctuation rate are calculated, and the adjustment dead zone range, control gain and response time constant are optimized by multiple objectives to obtain the target power allocation scheme.

2. The optimization method based on distributed solar photovoltaic power generation control system according to claim 1, characterized in that: The real-time collection and encrypted transmission processing of photovoltaic power station operation data to obtain power generation data, voltage data, current data and solar radiation data includes: Performing telemetry collection on inverter operation data to obtain an inverter telemetry data set, wherein the inverter telemetry data set includes the inverter's real-time startup capacity value, power change rate value, output upper limit value, and output lower limit value, and performing active power calculation to obtain the inverter power adjustment range; Inputting the inverter telemetry data set into the SSL security encryption module for private protocol encryption to obtain a first-level encrypted data packet, and performing directional server routing assignment on the first-level encrypted data packet to obtain encrypted transmission data; Input the encrypted transmission data into the cloud platform for vertical encryption and dynamic key update to obtain a second-level encrypted data packet, and distribute the second-level encrypted data packet to obtain a real-time database storage record; The real-time database storage records are classified and stored according to data type and timestamp to obtain historical database backup records, and the historical database backup records are firewall-isolated and security-verified to obtain power generation data, voltage data, current data and solar radiation data.

3. The optimization method based on distributed solar photovoltaic power generation control system according to claim 2, characterized in that: The power generation data, voltage data, current data and solar radiation data are input into the photovoltaic multimodal deep learning prediction model for processing to generate an active power regulation strategy and a reactive power regulation strategy, including: Inputting the power generation data, voltage data, current data, and solar radiation data into a multi-layer convolutional feature extraction layer in the photovoltaic multimodal deep learning prediction model for processing, wherein the multi-layer convolutional feature extraction layer includes four convolutional modules, each of which is composed of two convolutional layers and a maximum pooling layer, to obtain a power station operation feature vector; Inputting the power station operation feature vector into the bidirectional LSTM time series analysis layer in the photovoltaic multimodal deep learning prediction model for processing, the bidirectional LSTM time series analysis layer includes two LSTM layers, each LSTM layer includes 128 hidden units, to obtain time series dependency features; Inputting the time-dependent features into the multi-head attention layer in the photovoltaic multimodal deep learning prediction model for processing, the multi-head attention layer adopts a multi-head self-attention structure, includes 8 attention heads, and each attention head has a dimension of 64, to obtain a feature attention matrix; Inputting the feature attention matrix into the power prediction layer in the photovoltaic multimodal deep learning prediction model for processing, the power prediction layer comprises three fully connected layers with 512, 256, and 128 neurons, respectively, and uses a ReLU activation function to obtain a power prediction vector; The power prediction vector is input into the constraint optimization layer of the photovoltaic multimodal deep learning prediction model for processing. The constraint optimization layer adopts a multi-objective optimization algorithm, sets the grid dispatch instruction as a hard constraint condition and the installed capacity change rate as a soft constraint condition, and obtains a maximum power tracking parameter set; Inputting the maximum power tracking parameter set into the inverter parameter optimization layer in the photovoltaic multimodal deep learning prediction model for processing, wherein the parameter optimization layer includes two fully connected layers and uses a Sigmoid activation function to obtain the inverter optimization parameter set; Inputting the inverter optimization parameter set into the strategy generation layer in the photovoltaic multimodal deep learning prediction model for processing, the strategy generation layer adopts a hierarchical reinforcement learning structure, including a strategy neural network and a value evaluation network, to obtain an initial control strategy; The initial control strategy is input into the authority control layer in the photovoltaic multimodal deep learning prediction model for processing. The authority control layer includes a three-level authority verification module, which allocates control authority according to user level to obtain active power regulation strategy and reactive power regulation strategy.

4. The optimization method based on distributed solar photovoltaic power generation control system according to claim 3 is characterized in that: The generating of control instructions based on the active power regulation strategy and the reactive power regulation strategy, and performing hierarchical safety constraint verification on the control instructions to generate remote control instructions and local control instructions include: Performing instruction generation processing based on the active power regulation strategy and the reactive power regulation strategy to obtain a power generation control instruction set, wherein the power generation control instruction set includes an inverter output upper limit value, an output lower limit value, and a regulation rate parameter; Performing remote control parameter calculation on the power generation control instruction set, and performing calculation using a closed-loop control algorithm based on the automatic control target issued by the master station to obtain a remote control reference instruction; Calculating local control parameters for the power generation control instruction set, using an open-loop control algorithm based on the power generation plan issued by the master station to obtain a local control benchmark instruction; Performing equipment fault detection on the remote control reference command and the local control reference command, and obtaining a first-level safety detection result by evaluating the inverter operation state, power supply state, and grounding state; Performing a communication status test on the first-level security test result, by evaluating the device communication status and communication delay status, wherein the communication delay status is determined by a preset time threshold, to obtain a second-level security test result; Conduct electrical parameter testing on the second-level safety test results to obtain the third-level safety test results by evaluating whether the inverter voltage and power values ​​are within the safe range; Calculating a control blocking signal based on the third-level safety detection result, wherein the control blocking signal includes a power increase blocking signal and a power decrease blocking signal, and the control blocking signal is released only by manual operation to obtain a blocking control vector; The locking control vector is combined with the remote control reference instruction to generate a remote control instruction, and the locking control vector is combined with the local control reference instruction to generate a local control instruction.

5. The optimization method based on distributed solar photovoltaic power generation control system according to claim 4, characterized in that: The step of performing hierarchical reinforcement learning control and power optimization adjustment on the inverters in the plurality of control groups according to the remote control instructions and the local control instructions to obtain a total output value of the group includes: Inputting the remote control instructions and the local control instructions into the high-level decision module of the hierarchical reinforcement learning control network for processing, the high-level decision module includes a policy neural network and a value evaluation network, the policy neural network includes 3 convolutional layers and 2 fully connected layers, and the value evaluation network includes 2 convolutional layers and 3 fully connected layers, to obtain a group control action set; Performing control state encoding on the group control action set, encoding each control group into a control unit, including group identification, number of inverters, total installed capacity, and current power output status information, to obtain a group control state vector; The group control state vector is input into the execution control module to perform reward calculation, wherein the reward calculation includes three evaluation indicators: power tracking deviation, adjustment response time, and system stability, to obtain a group control reward value; Based on the group control reward value, the group control state vector is updated through deep Q learning, and the target Q value is calculated by a temporal difference algorithm using an experience replay buffer and a target network mechanism to obtain an optimal control action; The optimal control action is input into the power distribution module for group power regulation. According to the operating status and capacity parameters of each inverter, a weighted distribution algorithm is used to calculate the target power of each inverter to obtain the inverter power instruction; Performing power regulation judgment on the inverter power command, triggering regulation control when the deviation between the total power of the group and the target power exceeds a preset interval, and obtaining a power regulation trigger signal; Based on the power regulation trigger signal, active power and reactive power of each inverter are collaboratively optimized to obtain real-time power data of the inverter, and the real-time power data of the inverter are weighted and summed to obtain the total output value of the group.

6. The optimization method based on distributed solar photovoltaic power generation control system according to claim 5, characterized in that: The optimal control action is input into the power distribution module for group power regulation, and a weighted distribution algorithm is used to calculate the target power of each inverter according to the operating status and capacity parameters of each inverter to obtain the inverter power instruction, including: Inputting the optimal control action into the data processing unit of the power allocation module for strategy analysis to obtain a group power target vector; Dividing the group power target vectors into groups, grouping and calibrating the inverters based on geographic location data and electrical connection relationships, and obtaining group configuration data; performing a normalization operation on the installed capacity of each inverter based on the group configuration data to obtain a capacity normalization coefficient; Dynamically correcting the capacity normalization coefficient and adjusting the weight according to the communication delay and device status index of the inverter to obtain a dynamic adjustment coefficient; Performing a matrix operation on the dynamic adjustment coefficient and the group power target vector to obtain an initial power allocation value for each inverter, performing a limit processing on the initial power allocation value, and performing a constraint calculation based on the inverter power adjustment range to obtain a power limit control value; The power change rate is optimized based on the power limit control value to ensure that the power regulation rate meets the grid operation requirements, and the optimized control value is obtained. The optimized control value is converted into an instruction according to the communication protocol of each inverter to obtain an inverter power instruction.

7. The optimization method based on distributed solar photovoltaic power generation control system according to claim 6, characterized in that: The active power qualification rate, control response time, control accuracy and system fluctuation rate are calculated based on the operating parameter data, and multi-objective optimization is performed on the adjustment dead zone range, control gain and response time constant to obtain a target power allocation scheme, including: Calculating power deviation based on the operating parameter data, and obtaining power tracking error data by comparing the real-time power value with the control target value; Performing pass rate statistics based on the power tracking error data to obtain an active power pass rate, and performing a timing response analysis on the operating parameter data to calculate the control instruction execution completion time to obtain a control response time; Inputting the active power qualified rate and the control response time into a multi-objective genetic optimizer, and performing parameter optimization using a non-dominated sorting genetic algorithm to obtain a Pareto optimal solution set; Optimizing the parameters of the dead zone range based on the Pareto optimal solution set to obtain a dead zone optimization vector; The dead zone optimization vector is optimized by a PID controller, and the optimal control gain and response time constant are calculated to obtain a controller parameter vector. The controller parameter vector is then parameter-distributed according to the inverter capacity ratio to obtain a target power distribution scheme.

8. An optimization device based on a distributed solar photovoltaic power generation control system, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the apparatus comprises: The acquisition module is used to collect and encrypt the operation data of the photovoltaic power station in real time to obtain power generation data, voltage data, current data and solar radiation data; a processing module, configured to input the power generation data, voltage data, current data, and solar radiation data into a photovoltaic multimodal deep learning prediction model for processing, and generate an active power regulation strategy and a reactive power regulation strategy; A verification module, configured to generate control instructions based on the active power regulation strategy and the reactive power regulation strategy, and perform hierarchical safety constraint verification on the control instructions to generate remote control instructions and local control instructions; an adjustment module, configured to perform hierarchical reinforcement learning control and power optimization adjustment on the inverters in the plurality of control groups according to the remote control instructions and the local control instructions, so as to obtain a total output value of the group; A monitoring module, configured to monitor the total output value of the group and the operating status of the equipment in real time and generate operating parameter data; The optimization module is used to calculate the active power qualification rate, control response time, control accuracy and system fluctuation rate based on the operating parameter data, and perform multi-objective optimization on the adjustment dead zone range, control gain and response time constant to obtain a target power allocation scheme.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images