Fault detection method and system for distributed new energy access and medium

Abstract

The invention relates to the technical field of power system automation, new energy grid connection and fault detection, in particular to a distributed new energy access fault detection method and system and a medium. According to the method, firstly, data information in a micro-grid is collected through a sensor and uploaded to a control unit, and data waveform features are automatically extracted through a wavelet kernel function convolutional neural network constrained by a physical information neural network; and a near-end strategy optimization algorithm is adopted to realize fault type identification, position positioning and reason diagnosis in a distributed new energy access scene, so that the accuracy and convenience of power distribution network fault detection are improved. According to the invention, new energy access and data monitoring and fault type detection in the micro-grid formed by medium and low voltage power distribution equipment of a multi-grid structure are realized, and safe and effective operation of the micro-grid is realized.

Description

Technical Field

[0001] This invention relates to the fields of power system automation, new energy grid connection and fault detection technology, specifically a fault detection method, system and medium for distributed new energy access. Background Technology

[0002] The continued rapid growth of distributed renewable energy sources, mostly connected to distribution networks at medium and low voltage levels, has driven the transformation of traditional single-source distribution network topologies into multi-source microgrids. This high penetration rate of renewable energy brings numerous challenges to fault detection. Firstly, the nonlinear operating characteristics of renewable energy equipment result in low fault current amplitudes, short transient processes, and dynamically changing current directions, rendering traditional steady-state current-based judgment logic completely ineffective. Secondly, fluctuations in renewable energy output caused by natural conditions can easily be confused with transient signals generated by routine distribution network operations, leading to a high false alarm rate for traditional methods. Furthermore, the multi-branch nature of fault currents not only results in significant errors in traditional fault location techniques but also makes further fault diagnosis difficult, while the long detection delay fails to meet the millisecond-level fault isolation requirements of distribution networks. In existing technologies, the reliability of traditional relay protection drops significantly with high renewable energy penetration. Traditional machine learning methods require manual feature design and have weak generalization capabilities, and single intelligent algorithms struggle to adapt to the coupling relationship between faults and renewable energy operating states. Therefore, a technology that can automatically extract complex waveform features, dynamically adapt to operating states, and achieve integrated detection is urgently needed. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a fault detection method, system, and medium for distributed renewable energy access, aiming to solve the challenge of fault detection in distribution networks after the integration of distributed renewable energy. The invention first collects core operational data such as current and voltage using sensors deployed at key nodes of the microgrid, and uploads the fault data to the control unit for preprocessing. Then, relying on a wavelet kernel function convolutional neural network (WKFCNN) constrained by Physical Information Neural Networks (PINNs), it automatically captures complex fault features such as transient fluctuations and harmonic components contained in the data. This eliminates the need for manually designed feature parameters, avoids the limitation of incomplete feature coverage in traditional methods, and conforms to basic physical laws. Furthermore, by combining the PPO algorithm with the accurate features extracted by WKFCNN-PINNs as input, and through dynamic iterative optimization of the decision-making strategy, it simultaneously achieves accurate identification of fault types, rapid location of faults, and in-depth diagnosis of fault causes in distributed renewable energy access scenarios. Ultimately, this effectively overcomes the bottlenecks of traditional technologies in detection accuracy and response speed, significantly improving the reliability and convenience of distribution network fault detection, and providing technical support for the safe and stable operation of multi-source microgrids.

[0004] To solve the aforementioned technical problem, the technical solution adopted by the present invention is: a fault detection method for distributed renewable energy access, comprising the following steps: S01: Collect data on the power distribution network and new energy operation after faults occur through sensing devices, and form normalized model input data after preprocessing; S02: Construct a wavelet kernel function convolutional neural network (WKFCNN) to map the original waveform into a high-dimensional feature vector, and introduce physical information neural networks (PINNs) to ensure the physical rationality of the extracted feature vectors by using physical rules as constraints. S03: Transform fault detection into a Markov decision process (MDP), define the state, action, and reward function of proximal policy optimization (PPO), and realize the mapping from feature vectors to fault decisions; S04: The decision-making strategy is trained using a proximal strategy optimization pruning algorithm (PPO-Clip), and training stability is ensured through smooth updates; S05: Collect power distribution network fault data (voltage / current waveforms, new energy parameters, environmental data), input the preprocessed data into the WKFCNN-PINNs network to extract feature vectors, input the feature vectors into the trained PPO model, and output the fault detection results.

[0005] As a further technical solution of the present invention, data is collected and preliminarily processed in S01: S11: Collect three-phase voltage and current waveforms for multiple power frequency cycles before and after a fault using current transformers and voltage transformers to form a time-series dataset of line electrical quantities; collect new energy parameters such as the operating parameters of photovoltaic inverters (e.g., output power, equipment status); collect environmental data such as sunlight and wind speed, as well as topology information such as line length and switch status; S12: For data processing, statistical methods are used to remove outliers caused by sensor noise, and missing data is filled in by interpolation. Finally, the electrical quantity waveforms are mapped to a preset range to eliminate the influence of dimensions.

[0006] As a further technical solution of the present invention, mapping the original waveform to a high-dimensional feature vector in S02 includes: S21: Construct a WKFCNN network to automatically extract deep temporal features of fault waveforms, while introducing physical constraints of PINNs to achieve the mapping from the original waveform to physically compliant high-dimensional feature vectors. The network includes an input layer, wavelet convolutional layer, pooling layer, and fully connected layer: the input layer receives the original time-series waveform data; the wavelet convolutional layer extracts local temporal features through wavelet kernel functions; the pooling layer compresses the feature dimension and retains key information. S22: The wavelet convolutional layer performs local convolution calculations with the feature sequence of the previous layer using a wavelet kernel function, adaptively extracting local temporal features at different scales in the waveform, and employing the ReLU activation function to enhance nonlinear expressive power; its calculation formula is as follows: , , For the first l Layer m Each feature map in time t The original output of the convolution. For the previous layer n Each feature map is output. For scale Translation wavelet kernel function, K For kernel width, This represents the number of feature maps in the previous layer. For bias terms, For the first l Layer m Each feature map in time t The activated feature output, To modify the activation function of the linear unit; S23: Max pooling is used to reduce the dimensionality of the feature map output by the wavelet convolutional layer, preserving key time-frequency features. The formula is as follows: , Among them, The time interval covered by the pooling window. The pooled feature map represents the temporal position, and pooling reduces computational complexity and avoids overfitting. S24: Flatten the pooled feature map in the fully connected layer and input it into the fully connected layer to map it into a high-dimensional feature vector: , in, For the final pooling output, This is the weight matrix. For flattening operation, As the bias term, the output feature vector is... ( D (for feature dimensions) S25: Introduce Physical Information Neural Networks (PINNs) to combine physical constraint loss with task loss. The total loss function is: , in, The total loss function of the model, The feature discriminative loss in WKFCNN aims to make features of similar faults close to each other in the vector space, while features of dissimilar faults are far apart. The loss function used is... ,in, The number of training samples, For the first The feature vector of each sample To and Feature vectors of random samples of the same fault type To and Feature vectors of random samples of different fault types. This is the physical loss weighting coefficient. To address the physical constraint loss, a physical consistency loss term is constructed using physical rules such as power conservation as constraints. The task loss of WKFCNN is trained together to ensure the physical validity of the extracted feature vectors. T' The feature length after pooling. M This represents the number of feature maps in the pooling layer. The power value of the feature map. The power value measured and calculated by the sensor; S26: An optimizer is used to minimize both task loss and physical loss. The optimization is iteratively trained until the total loss of the validation set converges, and the final output is a high-dimensional fault feature vector that has both high identification ability and physical rationality.

[0007] As a further technical solution of the present invention, the Markov process for decision-making defined in S03 includes: S31: Transform fault detection into a Markov decision process, define the state, action, and reward function of the PPO, and realize the mapping from feature vectors to fault decisions: S32: Integrate the high-dimensional feature vector output by WKFCNN-PINNs with power grid operating parameters, environmental data, and topology information to form a multi-dimensional state vector. It comprehensively reflects the fault operation status and fault-related information of the distribution network.

[0008] S33: Employs a discrete-continuous hybrid action space. Fault types are categorized, with typical faults such as photovoltaic inverter overcurrent, single-phase grounding, and phase-to-phase short circuits configured as discrete actions. Set the key fault parameters, including fault location and fault resistance, to continuous operation. Therefore, the action vector is represented as ; S34: Guide the PPO to learn accurate fault detection strategies through multi-objective rewards, and design the reward function as follows: The specific formula is as follows: , Definitions of each item: Accuracy Bonus :like Consistent with the actual fault type, A positive reward applies; otherwise, a negative reward applies. Precision reward The smaller the detection error of fault location and resistance, the better. The higher the value, the more the correlation between error and reward is quantified using a negative exponential function; Efficiency Rewards The shorter the detection time, The higher the time, the more negative the reward; if the time exceeds the preset threshold, a negative reward will be given. Restraint and Punishment If the test results violate the physical laws of the power grid (e.g., the short-circuit current exceeds the rated capacity of the equipment but is not detected). A positive penalty value is applied; otherwise, it is 0.

[0009] As a further technical solution of the present invention, training the decision-making strategy in S04 includes: S41: The decision strategy is trained using the PPO-Clip algorithm, and training stability is ensured through smooth updates, as detailed below: S42: Design the PPO network structure as an Actor-Critic network, where the policy network (Actor) takes the state vector as input and outputs the action probability distribution; the value network (Critic) takes the state vector as input and outputs the state value. To assess the expected benefits of actions taken in the current state; S43: Using Generalized Advantage Estimation (GAE) to balance bias and variance, the relative value of actions is quantified, as shown in the following formula: , in, For the dominant function, This is a discount factor (balancing current and future rewards). The time-series difference error is a core hyperparameter of GAE. , For the first t Instant rewards for each step Let be the state value function, representing the state value at the th . t Step in state At that time, the expected total reward that can be obtained in the future. k This is the step offset, used to indicate the number of steps from the current step. t The number of steps to push backward. T This represents the number of steps in a single training cycle. Generalized advantage estimation is used to estimate the balance bias and variance, quantifying the relative value of actions. S44: Limit the policy update magnitude by clipping the probability ratio to avoid policy abrupt changes. The formula is as follows: , The strategy parameters are updated by maximizing the above formula. This is to limit the update range. In the formula, The probability ratio of the new and old strategies , For the dominant function, This refers to the editing coefficient; S45: Update the policy network separately using the optimizer (maximize) The policy network (which minimizes the mean square error between the state value and the target value) is used, and at the beginning of each iteration, the current policy network parameters are set. Copy to old strategy parameters This is used for subsequent advantage and probability ratio calculations to ensure that the data distribution is fixed during the gradient update phase; for policy updates, the current old policy is used. It interacts with the environment to collect a batch of trajectory data and stores it in a buffer. During updates, it randomly selects small batches of data from this batch and performs multiple gradient updates.

[0010] As a further technical solution of the present invention, the fault detection result output in S05 includes: S05: Collect data during power distribution network faults (voltage / current waveforms, renewable energy parameters, environmental data), preprocess the data, and input it into the trained WKFCNN-PINNs model to output high-dimensional feature vectors. Then, fuse the feature vectors with the fault-prone operating parameters to generate the state of the PPO model. Input the PPO policy network and output the optimal action. Final analysis Output the fault detection results (fault type, location, resistance parameters).

[0011] The present invention also discloses a fault detection system for distributed renewable energy access, including a processor and a memory storing program instructions. The processor is configured to execute the fault detection method for distributed renewable energy access as described above when running the program instructions.

[0012] The present invention also discloses a storage medium storing program instructions, which, when executed, perform the fault detection method for distributed new energy access as described above.

[0013] The beneficial effects of this invention are as follows: Addressing the challenge of fault detection in distribution networks after the integration of distributed renewable energy sources, this invention collects microgrid data via sensors and uploads it to the control unit, providing a reliable data foundation for analysis and avoiding the bias problems of traditional data acquisition methods. Utilizing WKFCNN-PINNs to automatically extract data waveform features eliminates the need for manual parameter design, overcoming the pain point of poor feature adaptability in traditional methods. Combined with the PPO algorithm, fault identification, location, and diagnosis are simultaneously achieved, avoiding the limitations of single-algorithm approaches. Ultimately, this breakthrough overcomes the bottleneck of detection technology, significantly improving fault detection accuracy and convenience, and providing support for the safe and stable operation of multi-source microgrids. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method described in Example 1. Detailed Implementation

[0015] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0016] Example 1 Medium and low voltage power distribution equipment with new energy access and multi-grid structure is a comprehensive device that integrates new energy power generation and energy storage technologies such as photovoltaics. It connects to new energy power generation systems such as photovoltaics, monitors the power output of the power generation system and detects faults, thereby improving the stability and reliability of medium and low voltage power distribution equipment with new energy access and multi-grid structure.

[0017] This invention considers a microgrid scenario consisting of multiple new energy sources and multi-grid structure medium- and low-voltage power distribution equipment. It is used to monitor and detect faults and fault types within the system at the system level, such as... Figure 1 As shown, it includes the following steps: S01 data acquisition and processing includes: S11: Collect three-phase voltage and current waveforms for multiple power frequency cycles before and after a fault using current transformers and voltage transformers to form a time-series dataset of line electrical quantities; collect new energy parameters such as the operating parameters of photovoltaic inverters (e.g., output power, equipment status); collect environmental data such as sunlight and wind speed, as well as topology information such as line length and switch status.

[0018] S012: For data processing, statistical methods are used to remove outliers caused by sensor noise, and missing data is filled in by interpolation. Finally, the electrical quantity waveform is mapped to a preset range to eliminate the influence of dimensions.

[0019] In S02, mapping the original waveform to a high-dimensional feature vector includes: S21: Construct a WKFCNN network to automatically extract deep temporal features from fault waveforms, while introducing physical constraints of PINNs to achieve mapping from the original waveform to physically compliant high-dimensional feature vectors. The network consists of an input layer, wavelet convolutional layers, pooling layers, and fully connected layers: the input layer receives the original time-series waveform data; the wavelet convolutional layer extracts local temporal features through wavelet kernel functions; the pooling layer compresses the feature dimension while retaining key information.

[0020] S22: The wavelet convolutional layer performs local convolution calculations with the feature sequence of the previous layer using a wavelet kernel function, adaptively extracting local temporal features at different scales in the waveform, and employing the ReLU activation function to enhance nonlinear expressive power. Its calculation formula is as follows: , , in, For the first l Layer m Each feature map in time t The original output of the convolution. For the previous layer n Each feature map is output. For scale Translation wavelet kernel function, K For kernel width, This represents the number of feature maps in the previous layer. For bias terms, For the first l Layer m Each feature map in time t The activated feature output, To modify the activation function of the linear unit.

[0021] S23: Max pooling is used to reduce the dimensionality of the feature map output by the wavelet convolutional layer, preserving key time-frequency features. The formula is as follows: , in The time interval covered by the pooling window. This represents the temporal position of the feature map after pooling. Pooling reduces computational complexity and avoids overfitting.

[0022] S24: Flatten the pooled feature map in the fully connected layer and input it into the fully connected layer to map it into a high-dimensional feature vector: , For the final pooling output, This is the weight matrix. For flattening operation, As the bias term, the output feature vector is... ( D (For feature dimensions).

[0023] S25: Introduce Physical Information Neural Networks (PINNs) to combine physical constraint loss with task loss. The total loss function is: , The total loss function of the model, The feature discriminative loss in WKFCNN aims to make features of similar faults close to each other in the vector space, while features of dissimilar faults are far apart. The loss function used is... , The number of training samples, For the first The feature vector of each sample To and Feature vectors of random samples of the same fault type To and Feature vectors of random samples of different fault types. This is the physical loss weighting coefficient. To address the physical constraint loss, a physical consistency loss term is constructed using physical rules such as power conservation as constraints. The task loss of WKFCNN is trained together to ensure the physical validity of the extracted feature vectors. T' The feature length after pooling. M This represents the number of feature maps in the pooling layer. The power value of the feature map. The power value measured and calculated by the sensor.

[0024] S26: An optimizer is used to minimize both task loss and physical loss. The optimization is iteratively trained until the total loss of the validation set converges, and the final output is a high-dimensional fault feature vector that has both high identification ability and physical rationality.

[0025] The Markov process for decision-making defined in S03 includes: S31: Transform fault detection into a Markov decision process, define the state, action, and reward function of the PPO, and realize the mapping from feature vectors to fault decisions: S32: Integrate the high-dimensional feature vector output by WKFCNN-PINNs with power grid operating parameters, environmental data, and topology information to form a multi-dimensional state vector. It comprehensively reflects the fault operation status and fault-related information of the distribution network.

[0026] S33: Employs a discrete-continuous hybrid action space. Fault types are categorized, with typical faults such as photovoltaic inverter overcurrent, single-phase grounding, and phase-to-phase short circuits configured as discrete actions. Set the key fault parameters, including fault location and fault resistance, to continuous operation. Therefore, the action vector is represented as .

[0027] S34: Guide the PPO to learn accurate fault detection strategies through multi-objective rewards, and design the reward function as follows: The specific formula is as follows: , Definitions of each item: Accuracy Bonus :like Consistent with the actual fault type, A positive reward applies; otherwise, a negative reward applies. Precision reward The smaller the detection error of fault location and resistance, the better. The higher the value, the more the correlation between error and reward is quantified using a negative exponential function; Efficiency Rewards The shorter the detection time, The higher the time, the more negative the reward; if the time exceeds the preset threshold, a negative reward will be given. Restraint and Punishment If the test results violate the physical laws of the power grid (e.g., the short-circuit current exceeds the rated capacity of the equipment but is not detected). A positive penalty value is applied; otherwise, it is 0.

[0028] Training decision-making strategies in S04 includes: S41: The decision strategy is trained using the PPO-Clip algorithm, and training stability is ensured through smooth updates, as detailed below: S42: Design the PPO network structure as an Actor-Critic network, where the policy network (Actor) takes the state vector as input and outputs the action probability distribution; the value network (Critic) takes the state vector as input and outputs the state value. To assess the expected benefits of actions taken in the current state.

[0029] S43: Using Generalized Advantage Estimation (GAE) to balance bias and variance, the relative value of actions is quantified, as shown in the following formula: , in, The dominant function, time-series difference error , This is a discount factor (balancing current and future rewards). These are the core hyperparameters of GAE. For the first t Instant rewards for each step Let be the state value function, representing the state value at the th . t Step in state At that time, the expected total reward that can be obtained in the future. k This is the step offset, used to indicate the number of steps from the current step. t The number of steps to push backward. T The number of steps per training cycle. Generalized advantage estimation is used to estimate balance bias and variance, quantifying the relative value of actions.

[0030] S44: Limit the policy update magnitude by clipping the probability ratio to avoid policy abrupt changes. The formula is as follows: , The strategy parameters are updated by maximizing the above formula. This is to limit the update range. In the formula, The probability ratio of the new and old strategies , For the dominant function, This is the editing coefficient.

[0031] S45: Update the policy network separately using the optimizer (maximize) The policy network (which minimizes the mean square error between the state value and the target value) is used, and at the beginning of each iteration, the current policy network parameters are set. Copy to old strategy parameters This is used for subsequent advantage and probability ratio calculations to ensure that the data distribution is fixed during the gradient update phase; for policy updates, the current old policy is used. It interacts with the environment to collect a batch of trajectory data and stores it in a buffer. During updates, it randomly selects small batches of data from this batch and performs multiple gradient updates.

[0032] S05: Collect distribution network fault data (voltage / current waveforms, renewable energy parameters, environmental data), preprocess the data, and input it into the trained WKFCNN-PINNs model to output high-dimensional feature vectors. Then, fuse the feature vectors with the fault-prone operating parameters to generate the state of the PPO model. Input the PPO policy network and output the optimal action. Final analysis Output the fault detection results (fault type, location, resistance parameters).

[0033] Example 2: Fault Detection System for Distributed New Energy Access This disclosure provides a fault detection system for distributed renewable energy access, including a data acquisition module, a control unit, a communication module, and an operation and maintenance terminal, wherein: Data acquisition module: Composed of several current sensors and voltage sensors, deployed at distributed new energy access points and line measurement points of the microgrid, used to collect operating data and send it to the control unit; Control unit: includes processor and memory, and optionally equipped with communication interface and bus; processor, communication interface and memory realize information interaction with each other through bus, communication interface is used for information transmission with data acquisition module and operation and maintenance terminal; memory stores logic instructions (including WKFCNN-PINNs feature extraction model and PPO diagnostic algorithm program), processor can call logic instructions to execute the fault detection method in the above embodiment 1; The memory is further divided into a program storage area and a data storage area: the program storage area is used to store the operating system and application programs required for fault detection (such as data preprocessing programs); the data storage area is used to store the raw data collected by the sensor, the preprocessed data, and the fault diagnosis results; and the memory can use high-speed random access memory, while being compatible with non-volatile memory to ensure data persistence.

[0034] Furthermore, if the logical instructions in the memory are implemented in the form of software functional units and are sold or used as independent products, they can be stored in a computer-readable storage medium.

[0035] Example 3: Fault Detection Media for Distributed New Energy Access This disclosure provides a fault detection medium for distributed renewable energy access. The medium is a computer-readable storage medium that stores computer-executable instructions. The computer-executable instructions are configured to execute the distributed renewable energy access fault detection method in Embodiment 1 when the computer device invokes the instructions.

[0036] The aforementioned computer-readable storage media are divided into two categories: transient and non-transient. Transient storage media can be used to temporarily store instructions to meet the needs of instruction calls in a short period of time. Non-transient storage media can store instructions for a long period of time, specifically including media that can store program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, and optical disks.

[0037] The technical solutions of this disclosure can be presented in the form of a computer software product, which is stored in the aforementioned storage medium and includes one or more instructions for causing a computer device to execute all or part of the steps of the fault detection method. It should be noted that this embodiment is merely exemplary; other embodiments may adjust the media storage format and instruction calling logic. Unless explicitly required, individual instruction modules are optional configurations, and the instruction execution order can be adjusted as needed. The media characteristics of some embodiments can be incorporated into or replace the media characteristics of other embodiments. The terminology used in this application is only used to describe embodiments and is not intended to limit the scope of protection.

Claims

1. A fault detection method for distributed renewable energy access, characterized in that: Includes the following steps: S01. Collect data on the operation of the power distribution network and new energy sources after faults through sensing devices, and form normalized model input data after preprocessing; S02: Construct a physical constraint-wavelet kernel function convolutional neural network. The physical constraint-wavelet kernel function convolutional neural network automatically extracts the time-domain deep features of the fault waveform, and at the same time introduces physical constraints to realize the mapping from the original waveform to the physical compliance high-dimensional feature vector. S03: Transform fault detection into a Markov decision process, define the state, action, and reward function of the proximal policy optimization model, and realize the mapping from feature vectors to fault decisions; S04: Employ a near-end strategy to optimize the pruning algorithm for training the decision strategy, and ensure training stability through smooth updates; S05: Collect distribution network fault data, input the preprocessed data into the physical constraint-wavelet kernel function convolutional neural network to extract feature vectors, input the feature vectors into the trained near-end strategy optimization model, and output the fault detection results.

2. The fault detection method for distributed renewable energy access according to claim 1, characterized in that: Step S01 includes: S11. Collect three-phase voltage and current waveforms for multiple power frequency cycles before and after a fault using current transformers and voltage transformers to form a time-series dataset of line electrical quantities; collect new energy operation parameters, environmental data, and topology information. S12. Data preprocessing: missing data is filled in by interpolation using statistical methods, and the electrical quantity waveforms are normalized to eliminate the influence of dimensions.

3. The fault detection method for distributed renewable energy access according to claim 2, characterized in that: The physically constrained wavelet kernel convolutional neural network consists of an input layer, a wavelet convolutional layer, a pooling layer, and a fully connected layer. The input layer receives the original time-series waveform data. The wavelet convolutional layer performs local convolution calculations with the feature sequence of the previous layer using a wavelet kernel function, adaptively extracting local time-series features at different scales in the waveform. The ReLU activation function is used to enhance the nonlinear expressive power. The calculation formula is as follows: ， in For the first l Layer m Each feature map in time t The original output of the convolution. For the first l -1st floor n Each feature map is output. For scale Translation wavelet kernel function, K For kernel width, For the first l -1 number of feature maps in layer 1 For bias terms, For the first l Layer m Each feature map in time t The activated feature output, To modify the activation function of the linear unit; The pooling layer uses max pooling to reduce the dimensionality of the feature map output by the wavelet convolutional layer, preserving key time-frequency features. The formula is as follows: ， in The time interval covered by the pooling window. The pooled feature map represents the temporal position, and pooling reduces computational complexity and avoids overfitting. The pooled feature maps, after being flattened, are input into a fully connected layer and mapped to high-dimensional feature vectors. ， in For the final pooling output, This is the weight matrix. For flattening operation, As a bias term, the fully connected layer outputs a feature vector. .

4. The fault detection method for distributed renewable energy access according to claim 3, characterized in that: The Physical Constraint-Wavelet Kernel Convolutional Neural Network combines the physical constraint loss with the task loss, and the total loss function is: ， For the total loss function, For mission losses, , The number of training samples, For the first The feature vector of each sample To and Feature vectors of random samples of the same fault type To and Feature vectors of random samples of different fault types This is the physical loss weighting coefficient. To address the physical constraint loss, a physical consistency loss term is constructed using physical rules as constraints. and Joint training ensures the physical validity of the extracted feature vectors. T' The feature length after pooling. The power value of the feature map. The power value is measured and calculated by the sensor; during training, an optimizer is used to minimize both task loss and physical loss, and training is iterated until the verification lumped loss converges, finally outputting a high-dimensional fault feature vector that has both high identification ability and physical rationality.

5. The fault detection method for distributed renewable energy access according to claim 4, characterized in that: Step S03 includes: S31. The high-dimensional feature vector output by the physical constraint-wavelet kernel function convolutional neural network is fused with new energy operation parameters, environmental data, and topology information to comprehensively reflect the fault operation status and fault correlation information of the distribution network. S32. Classify the fault types, setting photovoltaic inverter overcurrent, single-phase grounding of the line, and phase-to-phase short circuit as discrete actions. Set the critical fault parameters to continuous operation. Therefore, the action vector is a discrete-continuous hybrid action space, represented as: ; S33, Design the reward function as follows: The specific formula is as follows: ， For accuracy rewards, if Consistent with the actual fault type, A positive reward is given; otherwise, a negative reward is given. To increase accuracy, the smaller the detection error of fault location and resistance, the better. The higher the value, the more the correlation between error and reward is quantified using a negative exponential function; As an efficiency bonus, the shorter the detection time, The higher the threshold, the more negative rewards will be given if the time taken exceeds a preset threshold. Constraints and penalties: If the test results violate the physical laws of the power grid, A positive penalty value is applied; otherwise, it is 0.

6. The fault detection method for distributed renewable energy access according to claim 5, characterized in that: Step S04 includes: S41. The near-end policy optimization model involves an Actor-Critic network. The Actor network takes a state vector as input and outputs an action probability distribution; the Critic network takes a state vector as input and outputs a state value. To assess the expected benefits of actions taken in the current state; S42. Using the generalized dominance estimation method to estimate the balance deviation and variance, the relative value of actions is quantified using the following formula: in, For the dominant function, As a discount factor, The core hyperparameter for generalized dominance estimation is the time-series difference error. , For the first t Instant rewards for each step Let be the state value function, representing the state value at the th . t Step in state At that time, the expected total reward that can be obtained in the future. k This is the step offset, used to indicate the number of steps from the current step. t The number of steps to push backward. T The number of steps per training cycle; generalized advantage estimation is used to estimate balance bias and variance, quantifying the relative value of actions; S43: Limit the policy update magnitude by clipping the probability ratio to avoid policy abrupt changes. The formula is as follows: ， The strategy parameters are updated by maximizing the above formula. This is to limit the update range; in the formula, The probability ratio of the new and old strategies , For the dominant function, This refers to the editing coefficient; S44: Update the Actor network and Critic network separately using the optimizer, and at the beginning of each iteration, change the current policy network parameters. Copy to old strategy parameters This is used for subsequent advantage and probability ratio calculations to ensure that the data distribution is fixed during the gradient update phase; for policy updates, the current old policy is used. It interacts with the environment to collect a batch of trajectory data and store it in a buffer; during updates, it randomly selects small batches of data from this batch to perform multiple gradient updates.

7. The fault detection method for distributed renewable energy access according to claim 6, characterized in that: Step S05 is as follows: Collect distribution network fault data, preprocess the data and input it into the trained physical constraint-wavelet kernel function convolutional neural network to output a high-dimensional feature vector; then fuse the feature vector at the fault with the operating parameters to generate the state of the PPO model, input it into the near-end strategy optimization model, output the optimal action, and finally analyze and output the fault detection result.

8. The fault detection method for distributed renewable energy access according to claim 7, characterized in that: Power grid fault data includes voltage / current waveforms, renewable energy parameters, and environmental data. Renewable energy parameters include power output and equipment status. Environmental data includes illumination and wind speed. Topology information includes line length and switch status. Fault detection results include fault type, location, and resistance parameters.

9. A fault detection system for distributed renewable energy access, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute the fault detection method for distributed new energy access as described in any one of claims 1 to 8 when running the program instructions.

10. A storage medium storing program instructions, characterized in that, When the program instructions are executed, they perform the fault detection method for distributed new energy access as described in any one of claims 1 to 8.

Images