Intelligent photovoltaic grid-connected adaptive system based on multi-mode sensing and predictive analysis

Abstract

The invention discloses an intelligent photovoltaic grid-connected adaptive system based on multi-modal sensing and predictive analysis, and the system comprises a multi-modal sensing unit which is used for collecting the multi-source information of a photovoltaic system; the predictive fault analysis module is used for predicting potential system faults based on the multi-source data; the self-adaptive phase-frequency synchronous control unit is used for dynamically adjusting the output parameters of the photovoltaic inverter, so that the output of the photovoltaic inverter is matched with the parameters of the power grid; the comprehensive decision engine is used for generating a grid-connected or off-grid instruction; the self-repairing execution unit is used for generating a specific response scheme and controlling related equipment to execute based on the fault early warning and the strategy instruction of the comprehensive decision engine; and the remote monitoring and manual intervention interface is used for providing a monitoring interface. Through cooperative work of multi-modal data perception, predictive fault analysis and adaptive control, safe, efficient and intelligent grid connection of the photovoltaic system and the power grid is realized, and the reliability, safety and economical efficiency of the photovoltaic power generation system are improved.

Description

Technical Field

[0001] The present invention relates to an intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis, and belongs to the technical field of photovoltaic power generation. Background Art

[0002] Photovoltaic power generation is a technology that converts solar energy directly into electricity through the photoelectric effect. It is currently one of the most important renewable energy generation methods. A photovoltaic power generation system primarily consists of photovoltaic modules, inverters, transformers, and control systems. Grid-connected photovoltaic systems connect generated power to the public grid, enabling efficient utilization of photovoltaic power.

[0003] Core technologies for photovoltaic grid connection include phase-frequency synchronization control, power quality management, and fault detection. Phase-frequency synchronization control ensures precise matching of photovoltaic inverter output with grid parameters, laying the foundation for safe grid connection. Power quality management focuses on indicators such as power factor and harmonic distortion to ensure that output power meets grid standards. Fault detection technology can promptly detect wiring faults, insulation degradation, and other problems, preventing accidents.

[0004] In recent years, multimodal sensing technology has been widely used in industrial monitoring. This technology integrates multiple sensor data types, such as electrical, temperature, and vibration, to provide more comprehensive information on equipment operating status. Combined with predictive analysis algorithms, it can provide early warning of potential equipment failures, avoid unexpected incidents, and reduce maintenance costs.

[0005] The photovoltaic grid-connected systems currently on the market have the following technical limitations:

[0006] First, traditional PV grid-connected technologies often rely on simple threshold judgments or preset rules, lacking comprehensive awareness and analysis of system operating status. For example, while SMA's Sunny series inverters utilize Grid Guard monitoring technology, they primarily make grid-connection decisions based on single parameters such as voltage and frequency, making them incapable of adapting to complex and volatile grid environments. Similarly, ABB's TRIO / PVS series products lack multimodal sensing capabilities, making it difficult to comprehensively monitor system status.

[0007] Second, existing technologies primarily focus on fault detection and handling, lacking the ability to predict potential failures. For example, while Huawei's FusionSolar smart photovoltaic solution has basic fault diagnosis capabilities, it still relies primarily on a reactive approach, often responding only after a fault has occurred, leading to equipment damage and grid fluctuations. Schneider Electric's Conext series of grid-connected inverters also lack effective predictive maintenance capabilities.

[0008] Third, traditional phase and frequency synchronization control technologies suffer from poor stability in the face of grid disturbances. For example, the conventional PLL (phase-locked loop) technology used in ACREL inverters suffers from reduced phase-locking accuracy when grid frequency fluctuates or voltage sags, impacting grid connection quality and efficiency. While SolarEdge's HD-Wave technology optimizes waveform quality, its synchronization control algorithm still struggles to adapt to changing grid conditions.

[0009] Finally, existing grid-connection decisions are mostly based on simple current-state judgments, lacking the ability to comprehensively analyze multi-source data and make forward-looking decisions. For example, while KACO's Blueplanet series products are equipped with basic MPPT and grid-connection protection functions, they lack an intelligent decision-making engine and are unable to automatically optimize grid-connection strategies based on comprehensive conditions. Similarly, Sungrow's SG series inverters lack an intelligent decision-making system based on deep learning.

[0010] Although the above technical solutions have been applied in the field of photovoltaic grid connection, they have failed to provide a comprehensive, intelligent and adaptive solution, especially in the collaborative work of multimodal data perception, predictive fault analysis and adaptive control. Summary of the Invention

[0011] The purpose of the present invention is to provide an intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis. The system realizes the safe, efficient and intelligent grid connection of the photovoltaic system and the power grid through the collaborative work of multimodal data perception, predictive fault analysis and adaptive control.

[0012] In order to achieve the above object, the present invention adopts the following technical solutions:

[0013] The present invention provides an intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis, comprising:

[0014] Multimodal sensing unit for collecting multi-source information of the photovoltaic system, including electrical parameters, temperature, vibration and environmental data;

[0015] a predictive fault analysis module, configured to receive multi-source data collected by the multimodal sensing unit and predict potential system faults;

[0016] Adaptive phase-frequency synchronization control unit, used to dynamically adjust the output parameters of the photovoltaic inverter to match the photovoltaic inverter output with the grid parameters;

[0017] An integrated decision engine for generating grid-connection or grid-offline instructions based on multi-source data and fault prediction results;

[0018] The self-repair execution unit is used to receive fault warnings from the predictive fault analysis module and policy instructions from the comprehensive decision-making engine, generate specific response plans, and control the execution of related equipment;

[0019] Remote monitoring and manual intervention interface, used to provide a monitoring interface.

[0020] Preferably, the multimodal sensing unit comprises:

[0021] Electrical parameter sensors for monitoring the voltage, current, power, and harmonic content of photovoltaic modules, inverters, and power grids; the electrical parameter sensors include voltage transformers, current transformers, and power analyzers;

[0022] Temperature sensors are used to monitor temperature changes at the terminals, inverters, and transformers. These temperature sensors include infrared thermal imagers and contact temperature sensors. The infrared thermal imager is used to scan temperature distribution over a large area, while the contact temperature sensors are located at each terminal, key inverter components, and transformer to monitor local temperature changes in real time.

[0023] Vibration sensors, used to monitor the vibration characteristics of inverters and transformers;

[0024] The environmental monitoring unit is used to collect environmental data on light intensity, temperature, humidity and wind speed; the environmental monitoring unit includes a light intensity sensor, an environmental temperature and humidity sensor, a wind speed and direction sensor and a rainfall sensor.

[0025] Preferably, the multimodal sensing unit adopts a distributed deployment mode, including edge sensor nodes, local data processing units and a centralized data aggregation center;

[0026] All sensor data are transmitted to a local data processing unit via a fieldbus network or industrial Ethernet; the local data processing unit is used to perform pre-processing operations on the data, including outlier filtering, data compression and feature extraction;

[0027] The local data processing unit is connected to the data aggregation center through the industrial Internet of Things gateway; the data aggregation center is used to integrate the data of the entire system for storage and query.

[0028] Preferably, the predictive failure analysis module includes:

[0029] A multimodal data input interface, used for receiving multi-source data from a multimodal sensing unit;

[0030] Data preprocessing unit, used for cleaning, normalizing and feature extraction of received raw data;

[0031] Deep learning networks are used to extract high-dimensional features from preprocessed data;

[0032] Fault mode library, used to store characteristic modes of various wiring faults and system anomalies;

[0033] A support vector machine classifier for classifying and predicting faults based on the high-dimensional features extracted by the deep learning network; the fault types include wiring faults, inverter faults, and environmental impact faults;

[0034] The deep learning network and support vector machine classifier constitute a fault prediction model;

[0035] The self-learning optimization unit is used to compare the prediction results of the fault prediction model with the actual fault and optimize the fault prediction model.

[0036] Preferably, the deep neural network layer adopts a convolutional neural network-long short-term memory network-residual network hybrid architecture.

[0037] Preferably, the fault prediction model adopts a multi-scale time series analysis method to decompose the original input data into multiple time scales of milliseconds, seconds, minutes and hours, performs feature extraction and fault prediction on each time scale respectively, and then fuses the prediction results of multiple time scales to form a comprehensive fault prediction result.

[0038] Preferably, the adaptive phase-frequency synchronization control unit includes:

[0039] Grid parameter monitoring module, used to monitor the phase angle, frequency and voltage amplitude of the grid in real time;

[0040] Adaptive control module, used to dynamically adjust the output parameters of the photovoltaic inverter according to changes in grid parameters;

[0041] Phase-locked loop enhancement module for real-time phase tracking;

[0042] Closed-loop feedback control module for real-time correction of phase and frequency deviations.

[0043] Preferably, the phase-locked loop enhancement module adopts an adaptive phase detection method based on Hilbert transform, and the specific implementation method is:

[0044] Perform Hilbert transform on the real-time monitored grid voltage signal to obtain an analytical signal;

[0045] Extract instantaneous phase and instantaneous frequency from analytical signals;

[0046] Adaptive filtering algorithm is used to suppress harmonic interference and phase jump effects;

[0047] Update phase tracking parameters in real time.

[0048] Preferably, the adaptive control module dynamically adjusts the output parameters of the photovoltaic inverter according to changes in grid parameters, and is specifically implemented as follows:

[0049] Constructing a system dynamic model, the system dynamic model including an inverter output model and a grid parameter model; the inverter output model is represented using a second-order state space, wherein the state vector includes the inverter output voltage, current, and phase angle; the control input vector of the inverter output model includes the modulation ratio and phase offset; the output vector of the inverter output model includes the phase angle, frequency, and voltage amplitude of the inverter output; the grid parameter model is represented using a time series model with a disturbance term, wherein the grid parameter vector includes the grid voltage, frequency, and phase angle, and the disturbance term is the grid fluctuation;

[0050] Obtain the phase angle, frequency and voltage amplitude of the power grid, as well as the current output of the inverter in real time;

[0051] Based on the system dynamic model, the output parameters of the inverter under different control inputs in the next N time steps are predicted to form multiple sets of output trajectories; the control inputs include modulation ratio, phase offset and power;

[0052] Calculate the control objective function value corresponding to each control sequence; the control objective function is defined as the weighted sum of the phase error, frequency error, and voltage error in the prediction time domain, expressed as: , in, To control the objective function, 、 、 are the phase angle, frequency and voltage of the power grid respectively; 、 、 are the phase angle, frequency and voltage output by the inverter respectively; 、 、 is the corresponding weight coefficient; Indicates the current moment, Indicates the current to future step;

[0053] Select the control sequence that minimizes the control objective function value as the optimal control sequence;

[0054] Applying the first control input of the optimal control sequence generates control commands to the inverter.

[0055] Preferably, the comprehensive decision engine includes:

[0056] a data fusion module for fusing the multi-source data collected by the multimodal sensing unit, the fault prediction results of the predictive fault analysis module, and external environmental data; the external environmental data includes weather forecasts, power grid status, and energy prices; the fusion process uses multi-scale data integration and synchronization to ensure that data with different sampling rates can be analyzed collaboratively;

[0057] Deep reinforcement learning decision-making model for generating optimal decisions on grid connection, off-grid connection, and power adjustment;

[0058] The grid-connected / off-grid instruction generator is used to convert the optimal decision generated by the deep reinforcement learning decision model into a specific control command and send it to the relevant execution device.

[0059] Preferably, the deep reinforcement learning decision model adopts a dual deep Q network structure;

[0060] In the deep reinforcement learning decision model,

[0061] The state space includes multi-source data collected by the multimodal sensing unit, fault prediction results of the predictive fault analysis module, and external environment data;

[0062] The action space includes 7 discrete actions including grid connection, grid disconnection and 5-level power adjustment;

[0063] The reward function is designed based on energy efficiency, equipment safety and economic benefits.

[0064] Preferably, the self-repair execution unit is configured with a fault-response mapping library, which presets corresponding response strategies for faults of different types and severity;

[0065] The response strategy is divided into three levels: parameter adjustment, path reconfiguration and device isolation.

[0066] Preferably, the remote monitoring and manual intervention interface includes:

[0067] Visual display module, used to display the system operation status in real time, including electrical parameter curves, temperature heat maps and fault warning indicators;

[0068] Early warning push module, used to push key parameters and alarm information to the front end in real time;

[0069] Manual control module, used for manual intervention in emergency situations to override automatic decision-making and implement emergency control;

[0070] Parameter adjustment module, used to adjust system configuration parameters and threshold settings according to actual needs;

[0071] The historical query and analysis module supports multi-dimensional retrieval and in-depth analysis of historical data, including trend analysis, correlation analysis, and anomaly detection.

[0072] The present invention has the following significant technical effects:

[0073] (1) This invention uses a multimodal sensing unit to achieve comprehensive monitoring of the photovoltaic system's electrical parameters, temperature changes, vibration characteristics, and environmental data. Compared with traditional single-parameter monitoring, the system's status coverage is improved, and it can capture abnormal conditions that are easily overlooked by conventional monitoring methods. Experiments show that the system's detection rate for wiring anomalies is significantly improved compared to traditional methods.

[0074] (2) This invention significantly improves the accuracy and timeliness of fault prediction based on a hybrid prediction model combining a CNN-LSTM-residual network with a support vector machine. Experimental data show that the system can predict potential faults 30 minutes to several hours in advance, significantly outperforming the accuracy of traditional methods. In particular, for small sample fault types, the recognition accuracy of the hybrid model is significantly higher than that of pure deep learning methods.

[0075] (3) The adaptive phase detection technology based on Hilbert transform in this invention significantly improves grid connection stability and power quality. Even in the presence of grid frequency fluctuations (±3Hz / s), voltage drops (remaining 20%), and high harmonic conditions (THD up to 15%), the system can still maintain a small phase tracking error, which is significantly improved compared to the error of traditional PLL technology. Test results show that this technology enables the system to achieve a high grid connection success rate even under harsh grid conditions.

[0076] (4) The deep reinforcement learning decision-making model of the present invention enables intelligent grid-connection decision-making based on multi-source data. After one month of system operation, energy efficiency and economic benefits were improved while equipment failure rates were significantly reduced. Compared with traditional fixed strategies, the intelligent decision-making model can automatically optimize operating parameters under different meteorological conditions and grid states, improving overall performance.

[0077] (5) The fault evolution prediction and automatic repair function of the present invention significantly reduces the need for manual intervention and the duration of fault impact. Data shows that this function can handle most common faults, shortening the fault response time from hours required for traditional manual processing to minutes, and significantly reducing the average repair time. The automatic repair success rate is effectively improved, greatly enhancing system reliability and availability.

[0078] (6) This invention uses the collaborative work of multimodal data perception, predictive fault analysis, and adaptive control to reduce the overall lifecycle cost of photovoltaic systems, including maintenance costs and power losses caused by faults. The system's payback period can be shortened from the traditional 4-5 years to approximately 2-3 years, significantly improving the economic efficiency of photovoltaic power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 A schematic diagram of the structure of an intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis provided by an embodiment of the present invention;

[0080] Figure 2 A schematic diagram of a flow chart of an intelligent photovoltaic grid-connected adaptive control algorithm based on multimodal sensing and predictive analysis provided by an embodiment of the present invention;

[0081] Figure 3 A schematic diagram of the predictive fault analysis module structure provided by an embodiment of the present invention;

[0082] Figure 4 A schematic diagram of the structure of a comprehensive decision engine provided by an embodiment of the present invention;

[0083] Figure 5 A schematic flow chart of a model predictive control method for an adaptive phase-frequency synchronization control unit provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0084] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0085] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, the accompanying drawings only show structures and / or processing steps closely related to the solutions according to the present invention, while other details that are not closely related to the present invention are omitted.

[0086] It should be emphasized that the term "include / comprises" when used herein refers to the existence of features, elements, steps or components, but does not exclude the existence or addition of one or more other features, elements, steps or components.

[0087] It should also be noted that, unless otherwise specified, the term "connection" herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0088] It should be emphasized here that the step marks mentioned below do not limit the order of the steps, but it should be understood that the steps can be executed in the order mentioned in the embodiment, or in a different order from the embodiment, or several steps can be executed simultaneously.

[0089] like Figure 1As shown, the intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis provided by an embodiment of the present invention includes six main functional modules: a multimodal sensing unit, a predictive fault analysis module, an adaptive phase-frequency synchronization control unit, a comprehensive decision engine, a self-repair execution unit, and a remote monitoring and manual intervention interface. The multimodal sensing unit, located at the input of the system, is responsible for collecting multi-source information such as electrical parameters, temperature, vibration, and environmental data. The predictive fault analysis module analyzes this collected multi-source information to predict potential faults. The adaptive phase-frequency synchronization control unit ensures that the photovoltaic inverter output accurately matches the grid parameters. The comprehensive decision engine generates grid connection or disconnection instructions based on multi-source data analysis. The self-repair execution unit is responsible for implementing fault responses and controlling the system's operating status. The remote monitoring and manual intervention interface provides system operators with an interface for monitoring and controlling the system. All modules work together to ensure the safe and efficient connection of the photovoltaic system to the grid.

[0090] In this embodiment, the multimodal sensing unit is the foundation of the system, responsible for comprehensively sensing the operating status and environmental conditions of the photovoltaic system, providing accurate and reliable data support for subsequent analysis and decision-making. The multimodal sensing unit is located at the input end of the system and is responsible for collecting multi-source information such as electrical parameters, temperature, vibration, and environmental data. The multimodal sensing unit includes an electrical parameter sensor, a temperature sensor, a vibration sensor, and an environmental monitoring unit. The electrical parameter sensor is used to monitor the voltage, current, power, and harmonic content of the photovoltaic modules, inverter, and power grid; the temperature sensor is used to monitor temperature changes at the terminal blocks, inverter, and key nodes; the vibration sensor is used to monitor the vibration characteristics of the system components; and the environmental monitoring unit is used to collect environmental data such as light intensity, temperature, humidity, and wind speed.

[0091] In this embodiment, the multimodal sensing unit adopts a distributed deployment architecture, comprising edge sensing nodes, local data processing units, and a centralized data aggregation center. Edge sensing nodes are located in key locations such as photovoltaic arrays, inverters, and grid connection points, collecting various parameters in real time. Local data processing units are located in various functional areas and are responsible for preliminary data processing and storage. The centralized data aggregation center integrates system-wide data for in-depth analysis and archiving.

[0092] For electrical parameter monitoring, high-precision voltage and current transformers are used to monitor voltage and current parameters at the PV panels, inverter, and grid sides, respectively. Voltage and current measurement accuracy reaches 0.5, and with a sampling frequency of 10kHz, they are capable of capturing high-frequency disturbances and harmonic components in the grid. The system also integrates a power analyzer to monitor active power, reactive power, power factor, and harmonic content, providing a basis for grid-connected control.

[0093] Temperature monitoring utilizes a combination of infrared temperature measurement and contact temperature sensors. Infrared thermal imagers are used to scan temperature distribution over large areas, quickly identifying hotspots in PV modules. Contact temperature sensors (such as Pt100 platinum resistance temperature sensors) are placed at various locations, including wiring terminals, key inverter components, and transformers, to monitor local temperature changes in real time. The temperature data acquisition frequency is 1Hz, enabling timely detection of temperature anomalies.

[0094] Vibration monitoring primarily targets key equipment such as inverters and transformers. Triaxial accelerometers are used to monitor vibration characteristics at a sampling frequency of 500Hz. Vibration spectrum analysis can promptly identify potential problems such as loose equipment and bearing wear.

[0095] The environmental monitoring unit includes sensors for light intensity, temperature and humidity, wind speed and direction, and rainfall. This environmental data is closely related to the output performance of the photovoltaic system and is an important basis for predicting system behavior and failures. The environmental data collection frequency is 0.1Hz, meeting the needs of monitoring environmental changes.

[0096] All sensor data is transmitted to local data processing units via fieldbus networks (such as RS-485, CAN bus) or industrial Ethernet. These units utilize edge computing technology to perform data preprocessing operations, including outlier filtering, data compression, and feature extraction. Specifically, for high-frequency sampled electrical parameters, the system calculates secondary metrics such as RMS values, spectral content, and disturbance signatures, significantly reducing data transmission volume.

[0097] Locally processed data is uploaded to the data aggregation center via the Industrial IoT gateway. The gateway supports multiple communication protocols, such as MQTT and OPC UA, ensuring reliable and real-time data transmission. The data aggregation center utilizes a distributed database architecture, supporting efficient storage and querying of time-series data, providing a data foundation for in-depth analysis and model training.

[0098] To ensure data security and system reliability, the multimodal sensing unit implements a multi-level fault-tolerance mechanism. Each sensor node has self-diagnosis capabilities to detect faults. Redundant sensors are used in key locations to ensure that single-point failures do not affect overall system functionality. Data transmission utilizes encryption technology and verification mechanisms to prevent data tampering and loss. The system also incorporates a data quality assessment module to evaluate the reliability and accuracy of sensor data in real time, providing credibility indicators for subsequent analysis.

[0099] In this embodiment, the predictive fault analysis module is used to receive multi-source data collected by the multi-modal sensing unit and predict potential system faults through machine learning algorithms and data pattern matching technology. Figure 3As shown in the figure, the predictive fault analysis module includes a multimodal data input interface, a data preprocessing unit, a deep learning network, a fault pattern library, a support vector machine classifier, and a self-learning optimization unit. The multimodal data input interface is used to receive multi-source data from the multimodal sensing unit. The data preprocessing unit cleans, normalizes, and extracts features from the raw data collected by the multimodal sensing unit. The deep learning network is used to extract high-dimensional features from the preprocessed data. The deep learning network uses a hybrid structure of a convolutional neural network, a long short-term memory network, and a residual network (CNN-LSTM-residual network). The fault pattern library stores characteristic patterns of various wiring faults and system anomalies. The support vector machine classifier is used to classify and predict faults based on the high-dimensional features extracted by the deep learning network. The self-learning optimization unit continuously optimizes the prediction model by comparing the predicted results with the actual faults, improving prediction accuracy.

[0100] In this embodiment, the data preprocessing unit first cleans the raw data collected by the multimodal sensing unit, including removing outliers, handling missing values, and normalizing the data. Feature extraction is then performed to convert different types of sensor data into feature vectors that can be used by the model. The preprocessed data is divided into two paths: one path is fed into a deep learning network for high-level feature extraction, and the other path is compared with historical patterns in a fault pattern library.

[0101] In this embodiment, the fault prediction model adopts a hybrid structure that combines a deep learning network with a support vector machine classifier. The deep neural network layer adopts a CNN-LSTM-residual network hybrid architecture, in which the convolutional neural network (CNN) is used to extract spatial features, especially the correlation between different sensors; the long short-term memory network (LSTM) is used to capture time series features and analyze the trend of parameter changes over time; the residual network is used to solve the gradient disappearance problem in deep network training and improve model performance. The support vector machine classifier is based on the high-dimensional features extracted by the deep learning network and combines the historical data in the fault mode library to perform fault classification and prediction. The classifier can effectively handle multi-category fault identification problems and provide probability distributions of various types of faults to facilitate subsequent decision-making. Especially for small sample fault types, the support vector machine has shown better recognition accuracy than pure deep learning methods.

[0102] To further improve the performance of the fault prediction model, a feature fusion module was introduced. The feature fusion module integrates the spatiotemporal features extracted by deep neural networks with the features generated by traditional feature engineering, such as spectral features and statistical features. It not only retains the representation ability of deep learning, but also enhances the interpretability and generalization ability of the model, enabling the system to more accurately identify and predict various fault conditions.

[0103] Furthermore, the fault prediction model also adopts a multi-scale time series analysis method to decompose the original sensor data into multiple time scales at the millisecond, second, minute and hour levels, perform feature extraction and fault pattern recognition at each time scale respectively, and then fuse the analysis results of multiple time scales to form a comprehensive fault prediction.

[0104] In this embodiment, the self-learning optimization unit continuously adjusts prediction model parameters and fault assessment criteria by comparing prediction results with actual fault conditions. This adaptive learning mechanism enables the system to accumulate experience over time, improving prediction accuracy and reliability. By comparing real-time data with historical operating data, it can identify abnormal patterns and trends. This approach is particularly suitable for detecting slowly developing faults, such as equipment aging and performance degradation.

[0105] In practical applications, this module can predict various fault types, including wiring faults (such as loose wiring and incorrect phase sequence), inverter faults (such as insulated gate bipolar transistor (IGBT) damage and drive circuit anomalies), and environmental faults (such as hot spot effects and potential-induced degradation (PID) effects). The prediction window can be adjusted based on the fault type, typically ranging from 30 minutes to several hours, providing sufficient response time for maintenance personnel or self-repair systems.

[0106] In this embodiment, the adaptive phase-frequency synchronization control unit is responsible for ensuring that the output of the photovoltaic inverter accurately matches the grid parameters. The core function of this unit is to dynamically adjust the inverter output parameters through model predictive control methods to ensure that the power quality output by the photovoltaic system meets the grid standards. The adaptive phase-frequency synchronization control unit includes a grid parameter monitoring module, an adaptive control module, a phase-locked loop enhancement module, and a closed-loop feedback control module. The grid parameter monitoring module is used to monitor the phase angle, frequency, and voltage amplitude of the grid in real time; the adaptive control module dynamically adjusts the output parameters of the inverter according to changes in grid parameters; the phase-locked loop enhancement module improves the accuracy and stability of phase tracking; and the closed-loop feedback control module corrects phase and frequency deviations in real time.

[0107] In this embodiment, the phase-locked loop enhancement module uses an adaptive phase detection technology based on Hilbert transform to perform Hilbert transform on the grid voltage signal to obtain an analytical signal; extracts the instantaneous phase and instantaneous frequency from the analytical signal; uses an adaptive filtering algorithm to suppress harmonic interference and phase jump effects; and updates the phase tracking parameters in real time to improve tracking performance.

[0108] In this embodiment, the adaptive control module adopts a model predictive control method. The core of this method is to construct an accurate system dynamic model and perform forward prediction and optimization control based on the system dynamic model. The system dynamic model describes the relationship between the inverter output and the grid parameters, and mainly includes two parts: the inverter output model and the grid parameter model. The inverter output model is represented by a second-order state space, in which the state vector includes the output voltage, current, and phase angle, etc.; the control input vector includes control quantities such as modulation ratio and phase offset; the system output vector includes the phase angle, frequency and voltage amplitude of the inverter output. The grid parameter model is represented by a time series model with a disturbance term, in which the grid parameter vector includes the grid voltage, frequency and phase angle; the random disturbance vector represents the grid fluctuation.

[0109] See also Figure 5 ,The control process of the adaptive control module includes:

[0110] A1. Initialize the system, set initial parameters and model configuration;

[0111] In this embodiment, the system dynamic model describes the relationship between the inverter output and the grid parameters, and mainly includes two parts: an inverter output model and a grid parameter model.

[0112] The inverter output model is represented by a second-order state space:

[0113] ,

[0114] ,

[0115] in, is the system state vector, including output voltage, current and phase angle; To control the input vector, including modulation ratio, phase offset and other control variables; is the system output vector, including the phase angle, frequency and voltage amplitude of the inverter output; 、 、 、 is the system matrix, which is obtained through the system identification method.

[0116] The grid parameter model is represented by a time series model with disturbance terms:

[0117] ,

[0118] in, is the grid parameter vector, including grid voltage, frequency and phase angle; is the random disturbance vector, representing the power grid fluctuation; is a nonlinear mapping function obtained through training of historical data.

[0119] A2. Collect grid parameters and inverter output parameters.

[0120] In this step, the system acquires the grid's phase angle, frequency, and voltage amplitude, as well as the inverter's current output status, in real time. These parameters are collected by electrical parameter sensors at a sampling frequency of 10kHz, ensuring that the control system can capture rapid changes in grid parameters.

[0121] A3. Predict the output of the next N time steps based on the system dynamic model.

[0122] In this step, the constructed system dynamic model is used to predict the output parameters of the inverter under different control inputs in the next N time steps (prediction time domain, typical value N=10), forming multiple sets of possible output trajectories.

[0123] Under the model predictive control (MPC) framework, the system attempts different control action sequences at each future moment to predict the behavior of the inverter output. For photovoltaic inverters, these control inputs primarily refer to settings such as modulation ratio, phase offset, and power.

[0124] A4. Calculate the control objective function value corresponding to each control sequence.

[0125] In this step, the control objective function is defined as the weighted sum of phase error, frequency error, and voltage error in the prediction time domain, expressed as:

[0126] ,

[0127] in, 、 、 are the phase angle, frequency and voltage of the power grid respectively; 、 、 are the phase angle, frequency and voltage output by the inverter respectively; 、 、 is the corresponding weight coefficient. In this embodiment, =0.5, =0.3, =0.2, indicating that phase synchronization has the highest priority, followed by frequency matching and voltage amplitude matching. Indicates the current moment, Indicates the current to future step.

[0128] A5. Select the control sequence that minimizes the objective function as the optimal control sequence;

[0129] This step determines the control scheme that can minimize the synchronization error by comparing the objective function values ​​corresponding to each control sequence.

[0130] A6. Apply the first control input of the optimal control sequence;

[0131] In this step, only the first control input of the optimal control sequence is actually applied to the inverter, and the rest are recalculated in the next control cycle.

[0132] A7. Execute controls and observe system responses.

[0133] In this step, the system sends the calculated control instructions to the inverter and monitors the actual response of the system.

[0134] A key step in the control process is control performance evaluation. The system evaluates control effectiveness by comparing the actual response with the expected target. When control performance is good, the system completes the current control cycle and prepares for the next one. When control performance is poor, the system updates the system model parameters and weight coefficients for adaptive adjustment.

[0135] A rolling optimization strategy is used to form a closed-loop control system, enabling the system to adapt to changes in grid parameters in real time and maintain high-precision phase-frequency synchronization. The system also introduces an adaptive mechanism to dynamically adjust model parameters and weight coefficients based on the control effect. Specifically, the system updates the system model parameters by comparing the error between the predicted output and the actual output; and dynamically adjusts the weight coefficients by evaluating the achievement of each control target. In the event of grid disturbances, such as frequency fluctuations or phase jumps, the system also uses multi-scenario prediction technology to consider a variety of possible disturbance development trends and generate a robust control strategy. This approach significantly improves the system's adaptability in unstable grid environments.

[0136] As the core component of the adaptive phase-frequency synchronization control unit, the adaptive control algorithm works in conjunction with the grid parameter monitoring module, the phase-locked loop enhancement module, and the closed-loop feedback control system to ensure the precise matching of the photovoltaic inverter output with the grid parameters, creating conditions for safe and efficient grid connection.

[0137] In this embodiment, the comprehensive decision engine is the core decision-making unit of the system, responsible for generating the optimal grid connection or off-grid strategy. Figure 4As shown in the figure, the integrated decision engine consists of a data fusion module, a deep reinforcement learning decision model, and a grid-connection / off-grid command generator. The data fusion module integrates internal sensor data, fault prediction results, and external data (such as weather forecasts, grid status, and energy prices). The deep reinforcement learning decision model evaluates various possible actions using a policy network and a value network, generating the optimal decision based on a comprehensive consideration of energy efficiency, safety, and economics. The grid-connection / off-grid command generator converts decisions into specific control commands and sends them to the relevant execution devices. The model's performance evaluation results are used in a feedback loop to continuously optimize decision quality.

[0138] Specifically, the data fusion module integrates three key data types: internal sensor data (such as electrical parameters and temperature), fault prediction results (including fault type, probability, and expected occurrence time), and external data (such as weather forecasts, power grid status, and energy price information). The fusion process utilizes multi-scale data integration and synchronization technology to ensure collaborative analysis of data with different sampling rates. External data interfaces are used to obtain external information such as weather forecasts, power grid status, and energy market prices.

[0139] In this embodiment, the deep reinforcement learning decision model is the core component of the comprehensive decision engine. It uses a dual-network architecture called the Dual Deep Q Network (DDQN). This architecture improves the accuracy and stability of decisions by addressing the overestimation problem in traditional Q learning.

[0140] The deep reinforcement learning decision-making model represents the system state as a multidimensional state space (42-dimensional in this example), consisting of internal state variables (such as electrical parameters and temperature), external environmental variables (such as weather conditions and grid status), and fault prediction results (including fault type, probability, and expected occurrence time). The action space includes seven discrete actions, including grid connection, grid disconnection, and five-level power adjustment. The reward function is designed based on energy efficiency (40% weight), equipment safety (30% weight), and economic benefits (30% weight). In the dual-network architecture, the policy network selects the optimal action based on the current state, while the value network evaluates the long-term value of state-action pairs, forming a complete decision-making model. This separation helps address overestimation in reinforcement learning and improves learning stability. The model uses an experience replay mechanism to store historical decision experiences and randomly sample them for learning, breaking correlation between samples, accelerating convergence, and improving generalization.

[0141] In this embodiment, the deep reinforcement learning decision model also uses a layered reinforcement learning architecture, including a strategic layer, a tactical layer, and an inter-layer coordination mechanism. The strategic layer is responsible for long-term decision planning and uses an optimization algorithm based on policy gradients; the tactical layer is responsible for short-term real-time control and uses an optimization algorithm based on value functions. The inter-layer coordination mechanism ensures consistency between strategic goals and tactical execution.

[0142] In this embodiment, the on-grid / off-grid instruction generator converts the output of the deep reinforcement learning decision model into specific control commands, such as power setpoints, grid connection timing, and energy storage system scheduling instructions. These commands are sent to relevant execution devices via the communication network to implement system control.

[0143] A key advantage of this integrated decision-making engine is its ability to adapt to different operational strategy objectives. For example, during periods of high electricity prices, the system may prioritize economic benefits and maximize grid electricity sales; during periods of grid instability, it may prioritize system stability, even opting for off-grid operation; and during equipment anomalies, the system prioritizes equipment safety and reduces operating load. This multi-objective optimization capability enables the system to achieve optimal operational results in different scenarios.

[0144] In this embodiment, the learning process of the integrated decision engine is divided into two phases: offline training and online tuning. Offline training uses historical operating data and a simulated environment for pre-training to build an initial decision model. Online tuning continuously optimizes the decision strategy based on actual operating data, enabling the system to adapt to changes and uncertainties in the actual operating environment.

[0145] In this embodiment, the self-repair execution unit is the system's execution layer, responsible for implementing fault responses and controlling the system's operating status. The self-repair execution unit receives fault warnings from the predictive fault analysis module and policy instructions from the integrated decision-making engine, generates specific response plans, and controls the execution of relevant equipment. This unit has two core functions: automatic fault response generation and automatic execution control. It can take preventive measures before a fault occurs, minimizing its impact. The self-repair execution unit maintains a fault-response mapping library with preset handling strategies for different fault types and severities. Response plans are divided into three levels: parameter adjustment (such as adjusting inverter control parameters and maximum power point tracking (MPPT) parameters), path reconfiguration (such as switching to backup transmission lines and adjusting power allocation), and device isolation (such as disconnecting faulty components and activating backup equipment). The system employs a closed-loop control strategy, monitoring response effectiveness in real time and dynamically adjusting control parameters based on feedback to ensure effective fault handling. If the initial response is ineffective, the system automatically upgrades the response level or generates a new response plan. The unit also includes a fault evolution prediction function, which is used to simulate the development trend of faults under different response plans, evaluate the risks and effects of each response plan, select the optimal response plan and dynamically adjust the execution parameters.

[0146] In this embodiment, the remote monitoring and manual intervention interface provides comprehensive system monitoring and control capabilities for operations and maintenance personnel. This interface provides a visual display of system operating status; real-time push notifications of key parameters and warning information; a manual control function that allows operators to override automatic decisions when necessary; a system configuration and parameter adjustment interface; and historical query and analysis of operating data, enabling managers to gain real-time insights into system status and perform manual intervention when necessary. This interface utilizes a browser / server (B / S) architecture and supports access via web browsers and mobile applications, ensuring that managers can monitor system status from anywhere. The visualization function utilizes Web Graphics Library (WebGL) technology to provide real-time two-dimensional / three-dimensional (2D / 3D) displays of system operating status, including electrical parameter curves, temperature heat maps, and fault warning indicators. The warning push function pushes key parameters and alarm information to the front-end in real time via WebSockets and supports multiple notification methods, including email, SMS, and application (app) notifications, ensuring that important information reaches relevant personnel promptly. The manual control function allows authorized operators to override automatic decisions and implement emergency controls when necessary. The system is designed with strict permission management and authentication mechanisms. All important operations require two-factor authentication to prevent misoperation and security risks. The parameter adjustment function allows operations and maintenance personnel to adjust system parameters and threshold settings according to actual needs, such as modifying fault judgment criteria and adjusting control parameters. The system records all configuration changes to facilitate tracking of changes in system behavior. The historical query and analysis function supports multi-dimensional retrieval and in-depth analysis of historical data, including trend analysis, correlation analysis, and anomaly detection. These functions not only facilitate fault diagnosis and system optimization, but also provide data support for long-term planning and decision-making.

[0147] Through the collaborative work of the self-repairing execution unit and the remote monitoring and manual intervention interface, the system can retain the necessary manual intervention capabilities while operating automatically, achieving an effective combination of technology and experience, and ensuring the safe, reliable, efficient and flexible operation of the system.

[0148] Based on the above-mentioned intelligent photovoltaic grid-connected adaptive system, its workflow can be divided into five main steps, such as Figure 2 As shown, including:

[0149] S1: Collect electrical parameters, temperature, vibration, and environmental data from the photovoltaic system using a multimodal sensing unit. In this step, the system uses various sensors distributed at key nodes to obtain real-time system operating data, providing basic information for subsequent analysis. Electrical parameter collection is performed at a frequency of 10kHz, temperature data at 1Hz, vibration data at 500Hz, and environmental data at 0.1Hz. All data is transmitted to the data processing center via an industrial IoT gateway.

[0150] S2: The predictive fault analysis module processes multi-source data and predicts potential faults. In this step, the system first preprocesses the raw data, including noise removal, outlier detection, and data normalization. A deep learning model then analyzes patterns and trends in the data to identify anomalies that could lead to failures. The predictive model, which utilizes a CNN-LSTM-residual network architecture and is trained on historical data, can identify and predict problems such as wiring faults and inverter anomalies, with warning times ranging from 30 minutes to several hours.

[0151] S3: The adaptive phase-frequency synchronization control unit adjusts the inverter output parameters to synchronize with the grid. In this step, the system dynamically adjusts the inverter output phase and frequency based on the real-time grid status to ensure precise matching with the grid. The system uses closed-loop feedback control to continuously monitor synchronization status and make fine adjustments to maintain a phase error within 0.5°, ensuring grid connection stability and power quality.

[0152] S4: A comprehensive decision engine generates on-grid / off-grid decisions based on multi-source data analysis. In this step, the system integrates internal sensor data, fault prediction results, and external data (such as weather, grid status, and electricity prices). Using a deep reinforcement learning model, the system evaluates various possible operational strategies to generate the optimal on-grid / off-grid decision. This decision-making process considers energy efficiency (40% weight), equipment safety (30% weight), and economic benefits (30% weight), ensuring optimal system operation under all conditions.

[0153] S5: The self-healing execution unit implements fault response and continuously monitors system operation status. In this step, the system automatically generates and executes a response plan based on the results of predictive fault analysis. This may include parameter adjustments, path reconfiguration, or device isolation. Simultaneously, the system continuously monitors the effectiveness of the response and adjusts the strategy or upgrades the response level as necessary to ensure safe and reliable system operation.

[0154] The entire workflow forms a closed-loop control system. The data flow and feedback mechanism between modules ensure that the system can adapt to complex and changing operating environments and achieve safe and efficient grid connection of the photovoltaic system.

[0155] This embodiment innovatively applies a hybrid algorithm structure combining multimodal sensing, deep learning, and support vector machines to photovoltaic grid-connected systems through a multimodal data fusion and predictive analysis architecture, achieving accurate prediction of potential system faults and shifting from passive response to active prevention. The CNN-LSTM-residual network hybrid deep learning architecture is a deep learning structure designed specifically for photovoltaic grid-connected systems. It captures spatial features through convolutional neural networks, analyzes time series data through long and short-term memory networks, and solves the gradient vanishing problem through residual connections, achieving fault feature extraction in both time and space dimensions. The adaptive phase detection technology based on Hilbert transform improves the traditional phase-locked loop, enabling the system to be more robust in the grid. The system can maintain a small phase tracking error even in the presence of frequency mutations, voltage drops, and high harmonics. The adaptive phase-frequency synchronization algorithm of the model predictive control method constructs a dynamic model of the inverter output and grid parameters, achieving multi-step prediction and optimization control to ensure the stability and power quality of the grid connection process. The hierarchical decision-making architecture of deep reinforcement learning innovatively applies hierarchical reinforcement learning to grid connection decisions, separating strategic planning from tactical execution, coordinating long-term benefits with short-term control, and improving the adaptability and optimization of system decisions. The multi-scale time series analysis method and self-repair execution mechanism and fault evolution prediction improve the comprehensiveness and accuracy of fault prediction and automatically generate and execute response plans based on the fault prediction results. This embodiment achieves safe, efficient, and intelligent grid connection between the photovoltaic system and the power grid through the coordinated work of multimodal data perception, predictive fault analysis, and adaptive control, improving the reliability, safety, and economy of the photovoltaic power generation system, which is of great significance to promoting the large-scale application of photovoltaic power generation and the transformation of the energy structure.

[0156] It should be noted that in one embodiment, the predictive failure analysis module and the comprehensive decision-making engine can be integrated into a unified intelligent analysis and decision-making module to simplify the system architecture, reduce data transmission links, and improve system response speed. This integration solution is particularly suitable for application scenarios with limited computing resources.

[0157] It should be noted that in one embodiment, the system can be designed as a fully distributed architecture, where each functional module can be independently deployed on a different hardware platform and communicate through standardized interfaces. This architecture enhances the system's scalability and fault isolation capabilities, making it suitable for large-scale photovoltaic power plant applications.

[0158] It should be noted that for small photovoltaic systems, the system can be simplified into three core modules: data acquisition and analysis module, control execution module and user interface module, reducing system complexity and implementation cost.

[0159] It should be noted that, in one embodiment, it is possible to consider executing data-intensive and computing-intensive tasks (such as deep learning model training) on ​​the cloud, and executing real-time response tasks (such as data acquisition and control execution) on the edge, balancing response speed and computing power.

[0160] It should be noted that, in one embodiment, the three steps S2, S3, and S4 may be designed to be executed in parallel to reduce processing delays and improve the real-time response capability of the system.

[0161] It should be noted that, in one embodiment, the fixed sequential process can be changed to an event-driven mechanism, where the system triggers corresponding processing processes based on specific events (such as fault warnings, power grid fluctuations, etc.) to improve system flexibility.

[0162] It should be noted that, in one embodiment, different priorities may be assigned to different types of tasks to ensure that critical tasks (such as fault response) can be processed promptly, while non-critical tasks (such as data archiving) can be processed later.

[0163] It should be noted that for the predictive fault analysis module, in addition to the CNN-LSTM-residual network hybrid structure, the following alternatives can also be used for feature extraction:

[0164] Transformer Network: It uses the self-attention mechanism to capture long-distance dependencies in time series data. It does not require a recursive structure, can be computed in parallel, and is suitable for large-scale time series data analysis.

[0165] Graph Neural Network (GNN): When there are complex spatial relationships between sensors in the system, graph neural networks can be used to model the relationships between nodes to better capture the overall state of the system.

[0166] Autoencoder: Used in unsupervised learning scenarios, especially when fault samples are scarce, it can identify abnormal patterns by learning the distribution of normal operating data.

[0167] Classifier selection: In addition to support vector machines, you can also consider:

[0168] Random Forest: When the data dimension is high and the sample size is moderate, the random forest classifier usually performs well and has high computational efficiency.

[0169] XGBoost: An ensemble learning method that generally performs well on structured data, especially when there are complex nonlinear relationships between features.

[0170] Bayesian networks: Bayesian networks are a suitable choice when it is necessary to clearly express the causal relationship between variables and reflect uncertainty.

[0171] The fault prediction model training strategy can be as follows:

[0172] Transfer learning: Use models pre-trained on large-scale power plant data and transfer them to new photovoltaic systems to solve the small sample learning problem.

[0173] Active learning: The system actively selects samples with the most information for labeling and learning, improving data utilization efficiency.

[0174] Adversarial training: Generate boundary samples that are difficult to classify through generative adversarial networks (GANs) to enhance the robustness of the model.

[0175] Multiscale Time Series Analysis Alternatives:

[0176] Wavelet transform: Instead of multi-scale time series analysis, it decomposes the signal into different frequency components through wavelet decomposition to capture the characteristics of different time scales.

[0177] Empirical Mode Decomposition (EMD): Adaptively decomposes the signal into different intrinsic mode functions, suitable for the analysis of nonlinear and non-stationary signals.

[0178] Time-frequency analysis: Combined with methods such as short-time Fourier transform (STFT), signal characteristics are analyzed in both time and frequency domains.

[0179] It should be noted that there are the following alternatives for the adaptive phase-frequency synchronization control unit:

[0180] Control algorithm alternatives:

[0181] Adaptive PID control: Compared with model predictive control, adaptive PID control requires less computation and is simpler to implement. It is suitable for scenarios with limited computing resources, but it has limited capabilities in handling complex constraints and multivariable coupling control.

[0182] Robust H∞ control: When there is large uncertainty in the system model, robust H∞ control can provide more stable control performance and is suitable for harsh environments with large fluctuations in grid parameters.

[0183] Fuzzy logic control: When it is difficult to accurately model the dynamic characteristics of a system, fuzzy logic control can construct control rules based on expert experience and is suitable for scenarios where control accuracy requirements are not particularly high.

[0184] Active disturbance rejection control (ADRC): When external disturbances are large, ADRC estimates and compensates for disturbances through an extended state observer, thereby improving the system's ability to resist disturbances.

[0185] Phase detection technology alternatives:

[0186] Discrete Fourier Transform (DFT): This replaces the Hilbert transform and uses a windowed DFT to extract the phase and amplitude of the fundamental component. It is computationally simple, but its accuracy may be limited in harmonic environments.

[0187] Second-order generalized integrator (SOGI): Based on phase detection technology for quadrature signal generation, it has good filtering effect on harmonics and noise, is simple to implement, and is suitable for embedded systems.

[0188] Improved phase-locked loop (PLL) technologies, such as the phase-shifted phase-locked loop (SPLL) and decoupled dual synchronous reference frame PLL (DDSRF-PLL), perform better under unbalanced grid voltage conditions.

[0189] System Dynamics Model Alternatives:

[0190] Data-driven model: replaces the physics-based state-space model and directly learns the system dynamic characteristics from historical data through methods such as neural networks. It is suitable for complex nonlinear systems.

[0191] Hybrid model: combines the advantages of physical models and data-driven models. The physical model provides prior knowledge, and the data-driven part supplements the unmodeled dynamics.

[0192] Simplified model: For scenarios with limited computing resources, a first-order approximate model can be used instead of a second-order state-space model to reduce computational complexity.

[0193] Optimization target alternatives:

[0194] Multi-objective optimization: The control objective function is expanded into a multi-objective optimization problem, and the Pareto optimization method is used to find a balance between phase synchronization, frequency matching, voltage matching and energy efficiency.

[0195] Constrained optimization: Introducing hard constraints into the objective function, such as inverter output limit, switching frequency limit, etc., to ensure that the control output is within a safe range.

[0196] Economic model predictive control: Introducing economic indicators into the control objectives, such as energy loss, equipment life and other factors, to achieve comprehensive technical and economic optimization.

[0197] It should be noted that there are the following alternatives to the integrated decision engine:

[0198] Decision Model Alternatives:

[0199] Model Prediction Reinforcement Learning (MBRL): Combining the advantages of model prediction and reinforcement learning, it reduces the need for interaction with the actual environment by learning the environment model, improves sample efficiency, and is particularly suitable for scenarios with high data acquisition costs.

[0200] Multi-agent reinforcement learning: The system is decomposed into multiple agents, each responsible for a part of the decision-making, such as grid connection decision, power control, fault response, etc., and global optimization is achieved through a collaborative mechanism.

[0201] Evolutionary algorithms: For problems with sparse or delayed rewards, evolutionary algorithms (such as genetic algorithms and differential evolution algorithms) search the policy space by simulating the natural selection process and are sometimes more stable than gradient-based methods.

[0202] State space representation alternatives:

[0203] Temporal difference representation: It not only considers the current state, but also the trend of state changes, enhancing the system's ability to perceive dynamic changes.

[0204] Hierarchical state representation: The state space is processed in layers, with the high-level layer representing the overall operating mode of the system (such as normal operation, fault warning, etc.) and the low-level layer representing specific parameters, reducing the dimension of the state space.

[0205] Latent state representation: Learn low-dimensional representations of states through methods such as autoencoders, reduce the dimensionality of the state space, and improve learning efficiency.

[0206] Alternative Reward Function Designs:

[0207] Hierarchical reward structure: Different levels of goals correspond to different reward signals, such as safety, economy, efficiency, etc., and the final reward is formed through hierarchical combination.

[0208] Adaptive weight adjustment: Dynamically adjust the weights of each objective in the reward function based on the system operation status and external environment to meet the needs of different scenarios.

[0209] Inverse reinforcement learning: Learn the implicit reward function from expert operation data, avoiding the subjectivity and complexity of manually designed reward functions.

[0210] Decision-Making Architecture Alternatives:

[0211] Rule-based hybrid system: Introducing a rule-based decision-making component into the reinforcement learning framework to handle the clear decision logic, while reinforcement learning is responsible for optimizing the uncertain part.

[0212] Distributed decision-making architecture: The decision-making process is distributed to multiple physical nodes, each node is responsible for local decision-making, and global consistency is achieved through a coordination mechanism.

[0213] Hierarchical decision-making system: divides decision-making into strategic level (long-term planning), tactical level (mid-term adjustment) and operational level (short-term control), using appropriate algorithms and time scales for each level.

[0214] It should be noted that there are the following alternatives for self-repair execution units:

[0215] Case-based reasoning (CBR) fault response: This replaces the fixed fault-response mapping library. By storing historical fault cases and their successful solutions, it uses similarity matching to find the most similar historical cases and makes adaptive adjustments based on the current situation.

[0216] Multi-layer decision tree response mechanism: Using a hierarchical decision tree structure, the response strategy is gradually refined based on fault symptoms, severity, and system status, improving the accuracy and adaptability of the response.

[0217] Simulation-based response evaluation: Before executing a response, digital twin technology is used to simulate the effects of different response plans, select the optimal plan for execution, and reduce trial and error costs.

[0218] Execution Control Technology Alternatives:

[0219] Distributed control architecture: Distributes execution control functions to multiple nearby controllers to reduce communication delays and improve system response speed and fault tolerance.

[0220] Soft PLC technology: A programmable logic controller implemented using software to improve system flexibility and scalability, and facilitate function upgrades and remote maintenance.

[0221] Edge computing-based execution architecture: Deploy control logic on edge devices to reduce dependence on the central processing system and improve system response speed and reliability.

[0222] It should be noted that there are the following alternatives for remote monitoring and manual intervention interfaces:

[0223] Cloud-edge-end collaborative architecture: This replaces the pure B / S architecture and adopts a three-tier architecture consisting of cloud storage and analysis, edge real-time processing, and flexible terminal display to balance real-time performance, processing power, and user experience.

[0224] Microservice architecture: Split the monitoring system into multiple independent microservices, such as data collection services, storage services, analysis services, display services, etc., to improve system flexibility and maintainability.

[0225] Mixed reality (MR) interaction: Combining augmented reality (AR) and virtual reality (VR) technologies, it provides a more intuitive system status display and interaction method, which is particularly suitable for on-site maintenance scenarios.

[0226] Data visualization technology alternatives:

[0227] Data flow visualization: In addition to traditional chart displays, data flow diagrams are used to display the energy flow and information transmission of various parts of the system, providing a more intuitive understanding of the system operation status.

[0228] Parameter association network: Graphically displays the relationship between system parameters, helping operation and maintenance personnel understand system behavior and fault propagation paths.

[0229] Abnormal Focus Technology: Automatically identifies and highlights abnormal parameters and areas, guiding operation and maintenance personnel to pay attention to key issues and reduce information overload.

[0230] Alternatives to human-computer interaction:

[0231] Natural language interaction: Natural language processing technology is introduced to support operation and maintenance personnel to query system status and issue control instructions through natural language.

[0232] Multimodal interaction: Combining multiple interaction methods such as voice, gesture, and touch to adapt to different usage scenarios and user preferences.

[0233] Knowledge-assisted decision-making: Integrates expert knowledge base and decision support system to provide professional advice and operational guidance to operation and maintenance personnel at key decision points.

[0234] It should be noted that there are the following alternatives for the multimodal sensing unit:

[0235] Fiber optic sensing technology: Replaces traditional electrical sensors and uses fiber Bragg grating (FBG) sensors to simultaneously monitor temperature and strain, with advantages such as anti-electromagnetic interference, high sensitivity, and distributed measurement.

[0236] Wireless sensor networks (WSNs): Replace wired sensor networks and use low-power wireless communication technologies (such as ZigBee, LoRa, and NB-IoT) to build self-organizing sensor networks, reducing wiring complexity and improving deployment flexibility.

[0237] Acoustic sensing technology: It can be used as a supplement or alternative to vibration monitoring by analyzing the sound characteristics of equipment operation to detect potential faults.

[0238] MEMS sensor array: Use micro-electromechanical system sensor arrays to replace single-point sensors to improve spatial resolution and redundancy.

[0239] Data collection architecture alternatives:

[0240] Fully distributed architecture: Eliminate centralized data aggregation centers and adopt blockchain or distributed ledger technology to ensure data consistency, improve system reliability and anti-attack capabilities.

[0241] Stratified sampling strategy: Set different sampling frequencies according to the importance and change characteristics of parameters, sample important parameters at high frequency, and sample auxiliary parameters at low frequency to optimize system resource utilization.

[0242] Event-driven sampling: replaces fixed-frequency sampling and triggers data collection only when parameter changes exceed the preset threshold, reducing data redundancy.

[0243] Compressed sensing technology: A compressed sensing algorithm is used at the sensor node to reduce the number of sampling points, and a reconstruction algorithm is used at the data center to restore the complete signal and reduce the transmission bandwidth requirement.

[0244] Data processing technology alternatives:

[0245] Federated learning architecture: Under the premise of protecting data privacy, each edge node trains the model locally and only shares model parameters rather than original data. The central node integrates the local models to form a global model.

[0246] Stream processing framework: Use stream processing frameworks such as Apache Flink and Spark Streaming to process real-time data streams, supporting complex event processing and real-time analysis.

[0247] Data virtualization technology: Data in different physical locations and formats are uniformly accessed through the virtualization layer, reducing data copying and conversion costs.

[0248] Active learning data processing: The system actively selects data that needs to be processed in detail based on uncertainty and information gain, reducing the computational load.

[0249] Data storage alternatives:

[0250] Hybrid storage architecture: Based on data characteristics and access patterns, hot data is stored in the in-memory database, warm data is stored in SSDs, and cold data is stored in object storage to optimize performance and cost.

[0251] Multimodal database: A specially designed hybrid storage and query engine that supports different types of data (time series data, relational data, graph data, etc.).

[0252] Tiered data lifecycle management: Automates the management of data throughout its lifecycle, from collection, processing, storage, archiving, and deletion, optimizing storage efficiency.

[0253] Data Lake Architecture: This replaces traditional data warehouses, supports storing multi-source heterogeneous data in its original format, and enables flexible data processing using schema-on-read.

[0254] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis, characterized in that: include: Multimodal sensing unit for collecting multi-source information of the photovoltaic system, including electrical parameters, temperature, vibration and environmental data; a predictive fault analysis module, configured to receive multi-source data collected by the multimodal sensing unit and predict potential system faults; Adaptive phase-frequency synchronization control unit, used to dynamically adjust the output parameters of the photovoltaic inverter to match the photovoltaic inverter output with the grid parameters; An integrated decision engine for generating grid-connection or grid-offline instructions based on multi-source data and fault prediction results; The self-repair execution unit is used to receive fault warnings from the predictive fault analysis module and policy instructions from the comprehensive decision-making engine, generate specific response plans, and control the execution of related equipment; Remote monitoring and manual intervention interface, used to provide a monitoring interface.

2. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 1, characterized in that: The multimodal sensing unit comprises: Electrical parameter sensors for monitoring the voltage, current, power, and harmonic content of photovoltaic modules, inverters, and power grids; the electrical parameter sensors include voltage transformers, current transformers, and power analyzers; Temperature sensors are used to monitor temperature changes at the terminals, inverters, and transformers. These temperature sensors include infrared thermal imagers and contact temperature sensors. The infrared thermal imager is used to scan temperature distribution over a large area, while the contact temperature sensors are located at each terminal, key inverter components, and transformer to monitor local temperature changes in real time. Vibration sensors, used to monitor the vibration characteristics of inverters and transformers; The environmental monitoring unit is used to collect environmental data on light intensity, temperature, humidity and wind speed; the environmental monitoring unit includes a light intensity sensor, an environmental temperature and humidity sensor, a wind speed and direction sensor and a rainfall sensor.

3. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 2, characterized in that: The multimodal sensing unit adopts a distributed deployment mode, including edge sensor nodes, local data processing units and a centralized data aggregation center; All sensor data are transmitted to a local data processing unit via a fieldbus network or industrial Ethernet; the local data processing unit is used to perform pre-processing operations on the data, including outlier filtering, data compression and feature extraction; The local data processing unit is connected to the data aggregation center through the industrial Internet of Things gateway; the data aggregation center is used to integrate the data of the entire system for storage and query.

4. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 2, characterized in that: The predictive failure analysis module includes: A multimodal data input interface, used for receiving multi-source data from a multimodal sensing unit; Data preprocessing unit, used for cleaning, normalizing and feature extraction of received raw data; Deep learning networks are used to extract high-dimensional features from preprocessed data; Fault mode library, used to store characteristic modes of various wiring faults and system anomalies; A support vector machine classifier for classifying and predicting faults based on the high-dimensional features extracted by the deep learning network; the fault types include wiring faults, inverter faults, and environmental impact faults; The deep learning network and support vector machine classifier constitute a fault prediction model; The self-learning optimization unit is used to compare the prediction results of the fault prediction model with the actual fault and optimize the fault prediction model.

5. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 4, characterized in that: The deep neural network layer adopts a convolutional neural network-long short-term memory network-residual network hybrid architecture.

6. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 4, characterized in that: The fault prediction model adopts a multi-scale time series analysis method to decompose the original input data into multiple time scales of milliseconds, seconds, minutes and hours, performs feature extraction and fault prediction at each time scale respectively, and then fuses the prediction results of multiple time scales to form a comprehensive fault prediction result.

7. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 4, characterized in that: The adaptive phase-frequency synchronization control unit includes: Grid parameter monitoring module, used to monitor the phase angle, frequency and voltage amplitude of the grid in real time; Adaptive control module, used to dynamically adjust the output parameters of the photovoltaic inverter according to changes in grid parameters; Phase-locked loop enhancement module for real-time phase tracking; Closed-loop feedback control module for real-time correction of phase and frequency deviations.

8. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 7, characterized in that: The phase-locked loop enhancement module adopts an adaptive phase detection method based on Hilbert transform, and the specific implementation method is as follows: Perform Hilbert transform on the real-time monitored grid voltage signal to obtain an analytical signal; Extract instantaneous phase and instantaneous frequency from analytical signals; Adaptive filtering algorithm is used to suppress harmonic interference and phase jump effects; Update phase tracking parameters in real time.

9. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 7, characterized in that: The adaptive control module dynamically adjusts the output parameters of the photovoltaic inverter according to the changes in the grid parameters. The specific implementation method is as follows: Constructing a system dynamic model, the system dynamic model including an inverter output model and a grid parameter model; the inverter output model is represented using a second-order state space, wherein the state vector includes the inverter output voltage, current, and phase angle; the control input vector of the inverter output model includes the modulation ratio and phase offset; the output vector of the inverter output model includes the phase angle, frequency, and voltage amplitude of the inverter output; the grid parameter model is represented using a time series model with a disturbance term, wherein the grid parameter vector includes the grid voltage, frequency, and phase angle, and the disturbance term is the grid fluctuation; Obtain the phase angle, frequency and voltage amplitude of the power grid, as well as the current output of the inverter in real time; Based on the system dynamic model, the output parameters of the inverter under different control inputs in the next N time steps are predicted to form multiple sets of output trajectories; the control inputs include modulation ratio, phase offset and power; Calculate the control objective function value corresponding to each control sequence; the control objective function is defined as the weighted sum of the phase error, frequency error, and voltage error in the prediction time domain, expressed as: ， in, To control the objective function, 、 、 are the phase angle, frequency and voltage of the power grid respectively; 、 、 are the phase angle, frequency and voltage output by the inverter respectively; 、 、 is the corresponding weight coefficient; Indicates the current moment, Indicates the current to future step; Select the control sequence that minimizes the control objective function value as the optimal control sequence; Applying the first control input of the optimal control sequence generates control commands to the inverter.

10. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 7, characterized in that: The comprehensive decision engine includes: a data fusion module for fusing the multi-source data collected by the multimodal sensing unit, the fault prediction results of the predictive fault analysis module, and external environmental data; the external environmental data includes weather forecasts, power grid status, and energy prices; the fusion process uses multi-scale data integration and synchronization to ensure that data with different sampling rates can be analyzed collaboratively; Deep reinforcement learning decision-making model for generating optimal decisions on grid connection, off-grid connection, and power adjustment; The grid-connected / off-grid instruction generator is used to convert the optimal decision generated by the deep reinforcement learning decision model into a specific control command and send it to the relevant execution device.

11. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 10, characterized in that: The deep reinforcement learning decision model adopts a dual deep Q network structure; In the deep reinforcement learning decision model, The state space includes multi-source data collected by the multimodal sensing unit, fault prediction results of the predictive fault analysis module, and external environment data; The action space includes 7 discrete actions including grid connection, grid disconnection and 5-level power adjustment; The reward function is designed based on energy efficiency, equipment safety and economic benefits.

12. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 10, characterized in that: The self-repair execution unit is configured with a fault-response mapping library, which presets corresponding response strategies for faults of different types and severity; The response strategy is divided into three levels: parameter adjustment, path reconfiguration and device isolation.

13. The intelligent photovoltaic grid-connected adaptive system based on multimodal sensing and predictive analysis according to claim 12, characterized in that: The remote monitoring and manual intervention interface includes: Visual display module, used to display the system operation status in real time, including electrical parameter curves, temperature heat maps and fault warning indicators; Early warning push module, used to push key parameters and alarm information to the front end in real time; Manual control module, used for manual intervention in emergency situations to override automatic decision-making and implement emergency control; Parameter adjustment module, used to adjust system configuration parameters and threshold settings according to actual needs; The historical query and analysis module supports multi-dimensional retrieval and in-depth analysis of historical data, including trend analysis, correlation analysis, and anomaly detection.

Images