A power quality degradation evolution prediction method suitable for a high proportion photovoltaic power distribution network

Abstract

The present application belongs to the technical field of power quality monitoring and prediction of distribution network, especially relates to a power quality degradation evolution prediction method suitable for high proportion photovoltaic distribution network. The method successively collects multi-dimensional data by layering and dynamically and constructs a time series database, extracts low-redundancy degradation features by using a weighted enhanced autoencoder combined with feature coupling coefficients, constructs a spatio-temporal fusion prediction model integrating multi-network and multi-head self-attention mechanism, corrects the prediction results by adaptive Kalman filtering and quantifies the feature contribution degree by improved Deep SHAP, and finally completes the model closed-loop optimization combined with knowledge distillation. The present application improves the accuracy and timeliness of power quality degradation prediction, realizes quantitative identification of dominant factors of degradation, and provides reliable technical support for power quality control of high proportion photovoltaic distribution network.

Description

Technical Field

[0001] This invention belongs to the field of power quality monitoring and prediction technology for distribution networks, and particularly relates to a method for predicting the evolution of power quality degradation applicable to high-proportion photovoltaic distribution networks. Background Technology

[0002] Driven by the "dual carbon" goals, the construction of new power systems is accelerating, and distributed photovoltaic (PV) power, as an important carrier of clean and low-carbon energy, is being increasingly widely used in agricultural areas. Laixi, a major agricultural city, has vigorously promoted PV projects such as agricultural-photovoltaic integration and greenhouse PV in recent years. Multiple PV projects using the "power generation on the shed, livestock farming underneath" model have been connected to the grid, forming a distribution network with a high proportion of PV access. This has become a core support for the construction of the local new power system, achieving efficient utilization of agricultural resources and PV energy, and providing significant assistance for energy transition.

[0003] However, the high proportion of photovoltaic power distribution networks in Laixi exhibits distinct "dual-high" operating characteristics, posing a severe challenge to power quality management. The photovoltaic power sources in this area are mainly greenhouse photovoltaics and agricultural-photovoltaic complementary clusters, whose output is significantly affected by meteorological factors such as sunlight and rainfall, exhibiting strong randomness and intermittency, easily leading to power quality problems such as voltage fluctuations and flicker in the distribution network. On the load side, agricultural loads such as irrigation and greenhouse temperature control are the main contributors, exhibiting obvious seasonal and temporal fluctuations due to seasonal changes and agricultural activities, further exacerbating the instability of the distribution network operation.

[0004] Deteriorating power quality severely impacts the safe and stable operation of power distribution networks. It not only causes malfunctions and shortens the lifespan of agricultural equipment, but also restricts the absorption efficiency of photovoltaic energy and hinders the construction of new power systems. Currently, existing power quality degradation prediction methods are mostly designed for conventional power distribution networks and are ill-suited to the unique operating scenarios of the high-proportion photovoltaic power distribution network in the Laixi area. Summary of the Invention

[0005] Purpose of the invention

[0006] This invention addresses the characteristics of random photovoltaic output and load fluctuations in high-proportion photovoltaic distribution networks, as well as the shortcomings of existing prediction methods in terms of adaptability. It aims to provide a method for predicting the evolution of power quality degradation, thereby ensuring the safe and stable operation of the distribution network.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for predicting power quality degradation evolution applicable to high-proportion photovoltaic distribution networks includes:

[0009] Layered dynamic data acquisition: Intelligent monitoring terminals are deployed in layers according to photovoltaic clusters, distribution network lines and agricultural load nodes to collect photovoltaic output, line electrical quantity and load data in real time; the sampling frequency is dynamically adjusted based on the photovoltaic output fluctuation rate and load change rate, and after preprocessing the collected data, a multi-dimensional time series prediction database is constructed.

[0010] Degradation feature extraction: Based on the time series prediction database, an autoencoder is used to extract weighted and enhanced multidimensional degradation feature vectors, and a feature coupling coefficient is introduced to quantify the correlation between the dimensions of the degradation feature vectors, thereby constructing a degradation feature vector set;

[0011] Constructing a spatiotemporal fusion prediction model: Based on a CNN-LSTM-GRU hybrid network, graph convolutional network, and physical information network, a multi-head self-attention mechanism is introduced to fuse spatiotemporal features and construct a spatiotemporal fusion prediction model; a composite loss function is designed to constrain the prediction results, and a power quality degradation judgment standard is formulated in combination with the characteristics of a high-proportion photovoltaic distribution network.

[0012] Degradation prediction: The degradation feature vector set is input into the spatiotemporal fusion prediction model, which outputs the degradation evolution trend, level and occurrence time, and the prediction results are corrected by adaptive Kalman filtering; the improved Deep SHAP algorithm is used to quantify the contribution of each degradation feature to power quality degradation, so as to realize the quantitative identification and importance ranking of the dominant factors.

[0013] Model closed-loop optimization: Based on the corrected prediction results and the identification results of dominant factors, generate hierarchical early warning information and differentiated prevention and control plans and issue them for execution; collect actual operation data and operation and maintenance feedback data of the distribution network, dynamically adjust the parameters of the spatiotemporal fusion prediction model, and combine knowledge distillation to realize lightweight model updates and complete closed-loop iterative optimization.

[0014] Furthermore, the tiered dynamic data acquisition specifically includes:

[0015] Intelligent monitoring terminals are deployed across three levels: photovoltaic clusters, distribution network lines, and agricultural load nodes. First, raw power quality data from multiple dimensions, including photovoltaic output, line electrical quantities, and agricultural load power, are collected with an initial step size. Then, by calculating the photovoltaic output fluctuation rate and agricultural load change rate, and combining this with preset high and low frequency thresholds, the sampling step size is dynamically adjusted to balance the capture of key fluctuation information with the reduction of data redundancy. Subsequently, outliers are removed and corrected using the 3σ criterion, noise is removed by moving average filtering, and data dimension differences are eliminated through normalization, thus completing the cleaning and standardization of the raw data. Finally, the preprocessed multi-feature data is integrated according to timestamps to construct a multi-dimensional time-series prediction database.

[0016] Furthermore, the extraction of degradation features specifically includes:

[0017] Using a time-series database as input, a weighted augmented autoencoder is constructed. By assigning differentiated weights to different input features, the expression of features related to power quality degradation is strengthened and interference from irrelevant features is suppressed. The original electrical data is compressed into a degradation feature vector by the encoder, and the autoencoder parameters are optimized using reconstruction loss. Subsequently, the linear correlation between each degradation feature dimension is quantified by the feature coupling coefficient to identify redundant features. Finally, strongly correlated redundant features are eliminated by using the coupling coefficient threshold to simplify the feature dimensions and form a low-redundancy, highly representative degradation feature vector set.

[0018] Furthermore, the construction of the spatiotemporal fusion prediction model specifically includes:

[0019] Three basic networks—a CNN-LSTM-GRU hybrid network, a graph convolutional network (GCN), and a physical information network (PINN)—were constructed to extract temporal and local spatial correlation features of degradation characteristics and global spatial correlation features driven by the distribution network topology, respectively. Physical constraint losses were formed by embedding power balance and voltage constraints into the power grid's physical equations to ensure that the prediction results conform to the power grid's operating rules. A multi-head self-attention mechanism was then used to fuse local and global spatiotemporal features to complete the spatiotemporal fusion prediction model. Subsequently, a composite loss function was constructed that integrates prediction error, physical constraints, and regularization, and the model was optimized according to preset weight coefficients to improve prediction accuracy while suppressing overfitting. Finally, combining the characteristics of a high-proportion photovoltaic distribution network and agricultural loads, a four-level power quality degradation judgment standard was formulated based on voltage deviation and total harmonic distortion rate, achieving accurate prediction and level discrimination of power quality degradation evolution.

[0020] Furthermore, the degradation prediction specifically includes:

[0021] The degradation feature vector set is input into the spatiotemporal fusion prediction model, which outputs a predicted sequence of power quality indicators for a future period. Combined with the preset four-level degradation judgment criteria, the degradation evolution trend, degradation level at each moment, and the time of first degradation and severe degradation are clarified. To address the model prediction error, an adaptive Kalman filter is used to correct the prediction results and improve prediction accuracy by dynamically adjusting the covariance of process noise and observation noise. Finally, an improved Deep SHAP algorithm with feature weight coefficients is used to quantify the contribution of each degradation feature to power quality degradation. By ranking the absolute values ​​of contribution, the dominant factors affecting degradation are identified, thus completing the accurate prediction, error correction, and location of dominant factors of power quality degradation.

[0022] Furthermore, the model closed-loop optimization specifically includes:

[0023] Based on the corrected degradation prediction results and the identification results of the dominant factors, graded early warning and differentiated prevention and control plans are generated according to different degradation levels, and targeted prevention and control measures such as photovoltaic output adjustment, reactive power compensation, and load shedding are taken. Subsequently, actual operation data and operation and maintenance feedback data of the distribution network are collected to calculate the prediction error. If the error exceeds the preset threshold, the parameters of the spatiotemporal fusion prediction model are dynamically adjusted using the stochastic gradient descent algorithm. In order to solve the deployment problem caused by too many model parameters, knowledge distillation technology is used to transfer prediction knowledge to achieve lightweight model update by using the original model as the teacher model and the lightweight model as the student model. Finally, the updated lightweight model is used as the new prediction model, and the whole process of power quality degradation evolution prediction is repeated to form a closed loop iteration.

[0024] Furthermore, the raw power quality data includes:

[0025] Photovoltaic power output data ;

[0026] Line electrical quantity data: Voltage Current Power factor Total harmonic distortion ;

[0027] Agricultural load data: Active load reactive load .

[0028] Furthermore, the characteristic coupling coefficient is calculated as follows:

[0029] Introducing characteristic coupling coefficients Quantization of degraded feature vectors The correlation between the i-th and j-th dimensions, to avoid feature redundancy, is expressed by the formula:

[0030]

[0031] in, Let be the covariance between the i-th and j-th features. , Let be the variances of the i-th and j-th features, respectively. The formula is:

[0032]

[0033]

[0034] in:

[0035] : The coupling coefficient between the i-th and j-th dimension degradation features, with a value range of [-1, 1]. The larger the value, the higher the correlation.

[0036] Degraded eigenvectors at time t The i-th dimension component;

[0037] : The mean of the i-th dimension of degradation feature, ;

[0038] Covariance function: measures the degree of linear correlation between two features;

[0039] Variance function: measures the degree of dispersion of a single feature.

[0040] Furthermore, the improved Deep SHAP algorithm is specifically as follows:

[0041] Introducing feature weight coefficients optimizes the accuracy of contribution calculation. The core formula is:

[0042]

[0043] in:

[0044] The contribution of the i-th degradation feature. This indicates that the feature promotes degradation. This indicates the suppression of degradation; the larger the absolute value, the greater the impact.

[0045] : Set of deterioration features;

[0046] Feature subset This represents the set that does not contain the i-th feature;

[0047] Total number of features;

[0048] Feature subset Predicted values ​​after inputting into the model;

[0049] : The weighting coefficient of the i-th feature.

[0050] A system for implementing the power quality degradation evolution prediction method applicable to high-proportion photovoltaic distribution networks, comprising:

[0051] Layered data acquisition module: used to deploy intelligent monitoring terminals in layers according to photovoltaic clusters, distribution network lines and agricultural load nodes, collect multi-dimensional operation data in real time, dynamically adjust the sampling frequency, preprocess the data and build a time series prediction database;

[0052] Degradation Feature Extraction Module: Based on the time-series prediction database, it extracts multi-dimensional degradation feature vectors through a weighted augmented autoencoder, calculates feature coupling coefficients, removes redundant features, and constructs a set of degradation feature vectors.

[0053] Spatiotemporal fusion prediction module: used to construct a spatiotemporal fusion prediction model based on CNN-LSTM-GRU hybrid network, graph convolutional network and physical information network, introduce multi-head self-attention mechanism to fuse spatiotemporal features, design composite loss function to constrain model training, and formulate power quality degradation judgment criteria.

[0054] Degradation prediction and correction module: It is used to input the degradation feature vector set into the spatiotemporal fusion prediction model, output the degradation evolution trend, level and occurrence time, correct the prediction results through adaptive Kalman filtering, and use the improved Deep SHAP algorithm to quantify the contribution of degradation features, and complete the identification and ranking of dominant factors.

[0055] Closed-loop optimization module: Based on the corrected prediction results and the identification results of the dominant factors, it generates graded early warning information and differentiated prevention and control plans and issues them for execution, collects actual operation and maintenance feedback data, dynamically adjusts model parameters, realizes lightweight model updates through knowledge distillation, and completes closed-loop iterative optimization;

[0056] Main control module: Electrically connected to each of the above modules, used to control the coordinated work of each module, schedule data transmission and processing flow, ensure the stable operation of the entire system and realize the function of predicting the evolution of power quality degradation.

[0057] The present invention has the following beneficial effects:

[0058] I. Improve the accuracy and adaptability of data collection

[0059] Monitoring terminals are deployed hierarchically by photovoltaic cluster, power line, and agricultural load node to collect multi-dimensional data in real time, avoiding the limitations of data collected from a single node. The sampling frequency is dynamically adjusted based on photovoltaic output fluctuation rate and load change rate, balancing data timeliness and redundancy control to reduce storage and processing costs. The pre-processed time-series forecast database is adapted to the photovoltaic and agricultural load characteristics of Laixi area, providing reliable data support for subsequent forecasting work.

[0060] II. Accurately extract degradation features and quantify their correlation.

[0061] An autoencoder is employed to extract weighted and enhanced multidimensional degraded feature vectors, effectively mining key information and avoiding redundancy and information loss. A feature coupling coefficient is introduced to quantify the correlation between features, clearly understanding the mutual influence of factors such as photovoltaic output and load changes, resulting in a highly targeted feature vector set. This solves the prediction bias caused by inaccurate feature extraction in traditional methods, laying the foundation for accurate prediction.

[0062] III. Improving forecast accuracy and timeliness

[0063] A spatiotemporal prediction model fusing CNN-LSTM-GRU, graph convolution, and physical information is constructed, incorporating a multi-head self-attention mechanism to adapt to the stochastic and fluctuating characteristics of photovoltaic power output. A composite loss function is designed and a dedicated degradation judgment criterion is formulated, combined with adaptive Kalman filtering to correct the prediction results, significantly reducing errors. It can accurately output degradation trends, levels, and occurrence times, with strong timeliness, allowing maintenance personnel sufficient time to handle the situation.

[0064] IV. Quantitatively Identify the Dominant Factors of Deterioration

[0065] An improved Deep SHAP algorithm is used to quantify the contribution of each degradation feature, accurately identify the dominant factors such as photovoltaic output fluctuations and load abrupt changes, and rank their importance. This breaks through the limitations of traditional methods that can only predict trends but cannot pinpoint the root causes, helping maintenance personnel accurately locate the source of degradation, avoid blind prevention and control, improve prevention and control efficiency, reduce costs, and adapt to the photovoltaic and load distribution characteristics of the Laixi area.

[0066] V. Achieving Closed-Loop Optimization of the Model

[0067] Based on the prediction and identification results, tiered early warning and differentiated prevention and control plans are generated and implemented. Actual operation and maintenance feedback data are collected to dynamically adjust model parameters, and knowledge distillation is used to achieve lightweight model updates. This forms a virtuous cycle of "prediction-prevention-feedback-optimization," ensuring the model adapts to dynamic changes in the distribution network, continuously improving prediction accuracy and prevention and control effectiveness, and reducing operation and maintenance costs. Attached Figure Description

[0068] Figure 1 This is a flowchart of a power quality degradation evolution prediction method for high-proportion photovoltaic distribution networks proposed in this invention. Detailed Implementation

[0069] The present invention will be further described in detail below with reference to specific embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Example 1:

[0071] A method for predicting power quality degradation evolution applicable to high-proportion photovoltaic distribution networks includes:

[0072] Layered dynamic data acquisition: Intelligent monitoring terminals are deployed in layers according to photovoltaic clusters, distribution network lines and agricultural load nodes to collect photovoltaic output, line electrical quantity and load data in real time; the sampling frequency is dynamically adjusted based on the photovoltaic output fluctuation rate and load change rate, and after preprocessing the collected data, a multi-dimensional time series prediction database is constructed.

[0073] Degradation feature extraction: Based on the time series prediction database, an autoencoder is used to extract weighted and enhanced multidimensional degradation feature vectors, and a feature coupling coefficient is introduced to quantify the correlation between the dimensions of the degradation feature vectors, thereby constructing a degradation feature vector set;

[0074] Constructing a spatiotemporal fusion prediction model: Based on a CNN-LSTM-GRU hybrid network, graph convolutional network, and physical information network, a multi-head self-attention mechanism is introduced to fuse spatiotemporal features and construct a spatiotemporal fusion prediction model; a composite loss function is designed to constrain the prediction results, and a power quality degradation judgment standard is formulated in combination with the characteristics of a high-proportion photovoltaic distribution network.

[0075] Degradation prediction: The degradation feature vector set is input into the spatiotemporal fusion prediction model, which outputs the degradation evolution trend, level and occurrence time, and the prediction results are corrected by adaptive Kalman filtering; the improved Deep SHAP algorithm is used to quantify the contribution of each degradation feature to power quality degradation, so as to realize the quantitative identification and importance ranking of the dominant factors.

[0076] Model closed-loop optimization: Based on the corrected prediction results and the identification results of dominant factors, generate hierarchical early warning information and differentiated prevention and control plans and issue them for execution; collect actual operation data and operation and maintenance feedback data of the distribution network, dynamically adjust the parameters of the spatiotemporal fusion prediction model, and combine knowledge distillation to realize lightweight model updates and complete closed-loop iterative optimization.

[0077] Layered dynamic data acquisition specifically includes:

[0078] Intelligent monitoring terminals are deployed across three levels: photovoltaic clusters, distribution network lines, and agricultural load nodes. First, raw power quality data from multiple dimensions, including photovoltaic output, line electrical quantities, and agricultural load power, are collected with an initial step size. Then, by calculating the photovoltaic output fluctuation rate and agricultural load change rate, and combining this with preset high and low frequency thresholds, the sampling step size is dynamically adjusted to balance the capture of key fluctuation information with the reduction of data redundancy. Subsequently, outliers are removed and corrected using the 3σ criterion, noise is removed by moving average filtering, and data dimension differences are eliminated through normalization, thus completing the cleaning and standardization of the raw data. Finally, the preprocessed multi-feature data is integrated according to timestamps to construct a multi-dimensional time-series prediction database.

[0079] Degradation feature extraction specifically includes:

[0080] Using a time-series database as input, a weighted augmented autoencoder is constructed. By assigning differentiated weights to different input features, the expression of features related to power quality degradation is strengthened and interference from irrelevant features is suppressed. The original electrical data is compressed into a degradation feature vector by the encoder, and the autoencoder parameters are optimized using reconstruction loss. Subsequently, the linear correlation between each degradation feature dimension is quantified by the feature coupling coefficient to identify redundant features. Finally, strongly correlated redundant features are eliminated by using the coupling coefficient threshold to simplify the feature dimensions and form a low-redundancy, highly representative degradation feature vector set.

[0081] The construction of a spatiotemporal fusion prediction model specifically includes:

[0082] Three basic networks—a CNN-LSTM-GRU hybrid network, a graph convolutional network (GCN), and a physical information network (PINN)—were constructed to extract temporal and local spatial correlation features of degradation characteristics and global spatial correlation features driven by the distribution network topology, respectively. Physical constraint losses were formed by embedding power balance and voltage constraints into the power grid's physical equations to ensure that the prediction results conform to the power grid's operating rules. A multi-head self-attention mechanism was then used to fuse local and global spatiotemporal features to complete the spatiotemporal fusion prediction model. Subsequently, a composite loss function was constructed that integrates prediction error, physical constraints, and regularization, and the model was optimized according to preset weight coefficients to improve prediction accuracy while suppressing overfitting. Finally, combining the characteristics of a high-proportion photovoltaic distribution network and agricultural loads, a four-level power quality degradation judgment standard was formulated based on voltage deviation and total harmonic distortion rate, achieving accurate prediction and level discrimination of power quality degradation evolution.

[0083] Degradation prediction specifically includes:

[0084] The degradation feature vector set is input into the spatiotemporal fusion prediction model, which outputs a predicted sequence of power quality indicators for a future period. Combined with the preset four-level degradation judgment criteria, the degradation evolution trend, degradation level at each moment, and the time of first degradation and severe degradation are clarified. To address the model prediction error, an adaptive Kalman filter is used to correct the prediction results and improve prediction accuracy by dynamically adjusting the covariance of process noise and observation noise. Finally, an improved Deep SHAP algorithm with feature weight coefficients is used to quantify the contribution of each degradation feature to power quality degradation. By ranking the absolute values ​​of contribution, the dominant factors affecting degradation are identified, thus completing the accurate prediction, error correction, and location of dominant factors of power quality degradation.

[0085] Model closed-loop optimization specifically includes:

[0086] Based on the corrected degradation prediction results and the identification results of the dominant factors, graded early warning and differentiated prevention and control plans are generated according to different degradation levels, and targeted prevention and control measures such as photovoltaic output adjustment, reactive power compensation, and load shedding are taken. Subsequently, actual operation data and operation and maintenance feedback data of the distribution network are collected to calculate the prediction error. If the error exceeds the preset threshold, the parameters of the spatiotemporal fusion prediction model are dynamically adjusted using the stochastic gradient descent algorithm. In order to solve the deployment problem caused by too many model parameters, knowledge distillation technology is used to transfer prediction knowledge to achieve lightweight model update by using the original model as the teacher model and the lightweight model as the student model. Finally, the updated lightweight model is used as the new prediction model, and the whole process of power quality degradation evolution prediction is repeated to form a closed loop iteration.

[0087] The raw power quality data includes:

[0088] Photovoltaic power output data ;

[0089] Line electrical quantity data: Voltage Current Power factor Total harmonic distortion ;

[0090] Agricultural load data: Active load reactive load .

[0091] The characteristic coupling coefficient is calculated as follows:

[0092] Introducing characteristic coupling coefficients Quantization of degraded feature vectors The correlation between the i-th and j-th dimensions, to avoid feature redundancy, is expressed by the formula:

[0093]

[0094] in, Let be the covariance between the i-th and j-th features. , Let be the variances of the i-th and j-th features, respectively. The formula is:

[0095]

[0096]

[0097] in:

[0098] : The coupling coefficient between the i-th and j-th dimension degradation features, with a value range of [-1, 1]. The larger the value, the higher the correlation.

[0099] Degraded eigenvectors at time t The i-th dimension component;

[0100] : The mean of the i-th dimension of degradation feature, ;

[0101] Covariance function: measures the degree of linear correlation between two features;

[0102] Variance function: measures the degree of dispersion of a single feature.

[0103] The improved Deep SHAP algorithm is specifically as follows:

[0104] Introducing feature weight coefficients optimizes the accuracy of contribution calculation. The core formula is:

[0105]

[0106] in:

[0107] The contribution of the i-th degradation feature. This indicates that the feature promotes degradation. This indicates the suppression of degradation; the larger the absolute value, the greater the impact.

[0108] : Set of deterioration features;

[0109] Feature subset This represents the set that does not contain the i-th feature;

[0110] Total number of features;

[0111] Feature subset Predicted values ​​after inputting into the model;

[0112] : The weighting coefficient of the i-th feature.

[0113] Example 2:

[0114] A system for implementing the power quality degradation evolution prediction method applicable to high-proportion photovoltaic distribution networks, comprising:

[0115] Layered data acquisition module: used to deploy intelligent monitoring terminals in layers according to photovoltaic clusters, distribution network lines and agricultural load nodes, collect multi-dimensional operation data in real time, dynamically adjust the sampling frequency, preprocess the data and build a time series prediction database;

[0116] Degradation Feature Extraction Module: Based on the time-series prediction database, it extracts multi-dimensional degradation feature vectors through a weighted augmented autoencoder, calculates feature coupling coefficients, removes redundant features, and constructs a set of degradation feature vectors.

[0117] Spatiotemporal fusion prediction module: used to construct a spatiotemporal fusion prediction model based on CNN-LSTM-GRU hybrid network, graph convolutional network and physical information network, introduce multi-head self-attention mechanism to fuse spatiotemporal features, design composite loss function to constrain model training, and formulate power quality degradation judgment criteria.

[0118] Degradation prediction and correction module: It is used to input the degradation feature vector set into the spatiotemporal fusion prediction model, output the degradation evolution trend, level and occurrence time, correct the prediction results through adaptive Kalman filtering, and use the improved Deep SHAP algorithm to quantify the contribution of degradation features, and complete the identification and ranking of dominant factors.

[0119] Closed-loop optimization module: Based on the corrected prediction results and the identification results of the dominant factors, it generates graded early warning information and differentiated prevention and control plans and issues them for execution, collects actual operation and maintenance feedback data, dynamically adjusts model parameters, realizes lightweight model updates through knowledge distillation, and completes closed-loop iterative optimization;

[0120] Main control module: Electrically connected to each of the above modules, used to control the coordinated work of each module, schedule data transmission and processing flow, ensure the stable operation of the entire system and realize the function of predicting the evolution of power quality degradation.

[0121] Application example:

[0122] A method for predicting the evolution of power quality degradation in high-proportion photovoltaic distribution networks

[0123] Step 1: Layered Dynamic Data Acquisition

[0124] Intelligent monitoring terminals are deployed at three levels: photovoltaic clusters, distribution network lines, and agricultural load nodes. This enables real-time acquisition, dynamic adjustment, and preprocessing of multi-dimensional data, and the construction of a standardized time-series forecasting database, specifically including:

[0125] 1.1 Layered Data Acquisition Deployment

[0126] Monitoring terminals are deployed at the photovoltaic cluster outlet to collect photovoltaic output data. Monitoring terminals are deployed at key nodes of the main and branch lines of the distribution network to collect line electrical quantity data (voltage). Current Power factor Total harmonic distortion ); deploy monitoring terminals at the nodes where agricultural loads (irrigation pumps, temperature control equipment, etc.) are connected to collect load data. , (Active load, reactive load), the initial time step for data acquisition is set to... .

[0127] 1.2 Dynamic Sampling Frequency Adjustment

[0128] Based on photovoltaic power output volatility With load change rate Dynamically adjust sampling frequency To avoid data redundancy and missing key information, the formula is as follows:

[0129]

[0130]

[0131]

[0132] Parameter definition:

[0133] The photovoltaic power output fluctuation rate at time t reflects the degree of instantaneous change in photovoltaic power output, expressed in % (%).

[0134] : Total output of the photovoltaic cluster at time t, in kW;

[0135] The agricultural load change rate at time t reflects the instantaneous fluctuation of the load, expressed as a percentage (%).

[0136] : Total active power of agricultural load at time t, in kW;

[0137] : The dynamic sampling step size at time t, in seconds;

[0138] High-frequency sampling threshold, set to 0.1 (i.e. 10%). When volatility / rate of change is ≥10%, high-frequency sampling is used to capture key volatility.

[0139] The intermediate frequency sampling threshold is set to 0.05 (i.e., 5%). When the volatility / rate of change is between 5% and 10%, intermediate frequency sampling is used.

[0140] 1.3 Data Preprocessing

[0141] The collected raw data undergoes preprocessing to eliminate noise and outliers, achieving data standardization. This includes:

[0142] (1) Outlier handling: Outliers are removed using the 3σ criterion for data sequences. ,like If a value is found to be outlier, it is replaced by a linear interpolation of two adjacent valid data points, using the following formula:

[0143]

[0144] (2) Data smoothing: Noise is eliminated by moving average filtering, formula:

[0145]

[0146] (3) Standardization: Map the data to the [0,1] interval to eliminate the influence of dimensions. Formula:

[0147]

[0148] Parameter definition:

[0149] : The original data sequence, where n is the data length;

[0150] The mean of a data series. ;

[0151] Standard deviation of the data series ;

[0152] Corrected data after outlier replacement;

[0153] The time of data collection for the i-th data point is expressed in seconds.

[0154] : Slide the window size, take an odd number, here it is set to 5;

[0155] Smoothed data;

[0156] Standardized data;

[0157] , : The maximum and minimum values ​​of a data sequence.

[0158] 1.4 Construction of Time Series Prediction Database

[0159] The preprocessed standardized data is integrated according to the format of "timestamp-PV output-line electrical quantity-load data" to construct a multi-dimensional time series forecast database. ,in T represents the total amount of data. (7 feature dimensions).

[0160] Step 2: Deterioration Feature Extraction

[0161] Based on the time-series prediction database constructed in step 1, a weighted augmented autoencoder is used to extract multidimensional degraded feature vectors. A feature coupling coefficient is introduced to quantify the correlation between features, and a degraded feature vector set is constructed, specifically including:

[0162] 2.1 Construction of Weighted Augmented Autoencoder

[0163] An autoencoder (AE) consists of an encoder and a decoder. Weighted coefficients are introduced to enhance the weights of relevant features and suppress interference from irrelevant features. Its structure is as follows:

[0164] (1) Encoder: Transforms the input 7-dimensional time series data Mapped to d-dimensional degraded feature vector ,formula:

[0165]

[0166] (2) Decoder: converts the feature vector Reconstructed into a reconstructed vector with the same dimension as the input. ,formula:

[0167]

[0168] (3) Loss function: The reconstruction loss is used to optimize the autoencoder parameters. The formula is:

[0169]

[0170] Parameter definition:

[0171] The degraded feature vector at time t. d is the feature dimension, which is set to 16 here;

[0172] Encoder weight matrix ;

[0173] Feature weighted coefficient vector For the seven input features, the higher the correlation with degradation, the greater the weight (e.g., ...). The weight is set to 0.25. The weight is set to 0.2, and the weights of the other features are all 0.11.

[0174] : Diagonal matrix of weighted coefficients ;

[0175] Encoder bias vector ;

[0176] Decoder weight matrix ;

[0177] Decoder bias vector ;

[0178] The activation function is the ReLU function. ;

[0179] : Autoencoder reconstruction loss, which measures the difference between the input and the reconstructed output;

[0180] : L2 norm.

[0181] 2.2 Calculation of Characteristic Coupling Coefficients

[0182] Introducing characteristic coupling coefficients Quantization of degraded feature vectors The correlation between each dimension (the i-th dimension and the j-th dimension), avoiding feature redundancy, is expressed by the formula:

[0183]

[0184] in, Let be the covariance between the i-th and j-th features. , Let be the variances of the i-th and j-th features, respectively. The formula is:

[0185]

[0186]

[0187] Parameter definition:

[0188] : The coupling coefficient between the i-th and j-th dimension degradation features, with a value range of [-1, 1]. The larger the value, the higher the correlation.

[0189] Degraded eigenvectors at time t The i-th dimension component;

[0190] : The mean of the i-th dimension of degradation feature, ;

[0191] Covariance function: measures the degree of linear correlation between two features;

[0192] Variance function: measures the degree of dispersion of a single feature.

[0193] 2.3 Construction of Degraded Feature Vector Set

[0194] Set coupling coefficient threshold ,like Then, one redundant feature is removed; features with coupling coefficients less than a threshold are retained to construct the final degraded feature vector set. ,in ( , here ).

[0195] Step 3: Construct a spatiotemporal fusion prediction model

[0196] Based on a hybrid CNN-LSTM-GRU network, graph convolutional network (GCN), and physical information network (PINN), a multi-head self-attention mechanism is introduced to fuse spatiotemporal features. A composite loss function is designed to constrain the prediction results, and a power quality degradation judgment standard is established, specifically including:

[0197] 3.1 Basic Network Structures and Formulas

[0198] 3.1.1 Hybrid CNN-LSTM-GRU Network (Temporal + Local Spatial Feature Extraction)

[0199] (1) CNN module: Extracts local spatial features from the feature vector set (such as the correlation between features at adjacent time points), formula:

[0200]

[0201] (2) LSTM module: Extracts long-term time-dependent features, state update formula:

[0202]

[0203]

[0204]

[0205]

[0206]

[0207] (3) GRU module: Optimizes the extraction efficiency of long-term dependencies, state update formula:

[0208]

[0209]

[0210]

[0211]

[0212] Parameter definition:

[0213] CNN convolution kernel size, set to 3;

[0214] CNN convolution weight matrix c is the number of convolution kernels, set to 32;

[0215] CNN bias vector ;

[0216] Max pooling operation, with a pooling kernel size of 2;

[0217] CNN module outputs features, ;

[0218] LSTM's input gate, forget gate, and output gate;

[0219] : Cellular state of LSTM : The output state of the LSTM;

[0220] The input gate's weight matrix and bias are calculated similarly, as are the other gates; all have the same dimension. (h is the number of neurons in the LSTM hidden layer, set to 64);

[0221] GRU's update gate and reset gate;

[0222] : GRU candidate hidden states The final output state of the GRU. ;

[0223] Update the weight matrix and bias of the gate; the same applies to the other gates, with all dimensions being equal to 1. ;

[0224] Element-wise multiplication.

[0225] 3.1.2 Graph Convolutional Networks (GCN, Global Spatial Feature Extraction)

[0226] Based on the distribution network topology, construct a graph. V represents the set of nodes (photovoltaic clusters, line nodes, load nodes), and E represents the set of edges (electrical connections between nodes). Global spatial correlation features are extracted using GCN, as shown in the formula:

[0227]

[0228] in, The normalized adjacency matrix, The original adjacency matrix ( This indicates that nodes i and j are connected; otherwise, it is 0. For the identity matrix, the normalization formula is:

[0229]

[0230] Parameter definition:

[0231] Distribution network topology diagram;

[0232] : The degradation feature matrix of all nodes at time t, where N is the number of nodes. ;

[0233] GCN weight matrix ;

[0234] GCN bias vector ;

[0235] The global spatial feature matrix output by GCN. ;

[0236] Degree matrix , a diagonal matrix.

[0237] 3.1.3 Physical Information Network (PINN, with physical constraints introduced)

[0238] Introducing physical constraints of the distribution network (power balance, voltage constraints), and embedding physical equations into the network using PINN, the rationality of the constraint prediction results is constrained. The physical constraint loss formula is as follows:

[0239]

[0240] Among them, line loss , This is the voltage reference value (rated voltage of the distribution network, set to 10kV). This represents the line resistance.

[0241] Parameter definition:

[0242] Physical constraint losses ensure that the prediction results conform to the physical laws of power distribution network operation;

[0243] Line active power loss at time t, in kW;

[0244] Equivalent resistance of distribution network lines, in Ω;

[0245] Voltage reference value, in kV.

[0246] 3.2 Multi-head self-attention mechanism (spatiotemporal feature fusion)

[0247] Introducing a multi-head self-attention mechanism to fuse local spatiotemporal features from CNN-LSTM-GRU outputs Global spatial features output by GCN ,formula:

[0248]

[0249]

[0250] in, (Query vector) (Key vector, value vector), fused to output spatiotemporal features .

[0251] Parameter definition:

[0252] : Number of heads to focus on, set to 8;

[0253] : The output of the k-th attention head;

[0254] The query, key, and value weight matrix of the k-th attention head, with each dimension being [dimension 1]. ;

[0255] : Feature dimensions of each attention head, here ;

[0256] : Fusion weight matrix, ;

[0257] : The fused spatiotemporal feature vector ;

[0258] : Normalization function, .

[0259] 3.3 Design of Composite Loss Function

[0260] Combining prediction error, physical constraints, and regularization terms, a composite loss function is designed to constrain model training. The formula is:

[0261]

[0262] in:

[0263] (1) Predicting losses Mean squared error (MSE) is used to measure the difference between predicted and actual values. , The power quality index value predicted by the model. The actual value;

[0264] (2) Physical constraint loss Same as step 3.1.3;

[0265] (3) Regularization term L2 regularization is used to suppress overfitting. , This is the set of all weight matrices in the model.

[0266] Parameter definition:

[0267] The composite loss function comprehensively measures the model's predictive performance and physical plausibility.

[0268] The loss weighting coefficients are set to 0.6, 0.3, and 0.1 respectively, which satisfies... ;

[0269] Predicted power quality index values ​​at time t (e.g.) );

[0270] : The true value of power quality indicators at time t;

[0271] Regularization loss helps suppress overfitting caused by excessively large model weights.

[0272] 3.4 Criteria for Judging Power Quality Deterioration

[0273] Based on the characteristics of high-proportion photovoltaic distribution networks (large fluctuations in photovoltaic output and unstable agricultural loads), and with reference to national standards, a degradation level assessment standard has been formulated, divided into four levels, as follows:

[0274] (1) Normal level (level 0): , No degradation;

[0275] (2) Slight deterioration (Level 1): , The degree of degradation is low, and no emergency treatment is required;

[0276] (3) Moderate degradation (Level 2): , An early warning system needs to be activated and preventative measures prepared.

[0277] (4) Severe deterioration (Level 3): , The level of degradation is high, and the prevention and control plan needs to be implemented immediately.

[0278] Step 4: Degradation Prediction

[0279] The degradation feature vector set constructed in step 2 is input into the spatiotemporal fusion prediction model in step 3, which outputs the degradation evolution trend, level, and occurrence time. The prediction results are corrected by adaptive Kalman filtering, and the contribution of degradation features is quantified using an improved Deep SHAP algorithm to identify the dominant factors, specifically including:

[0280] 4.1 Model Prediction Output

[0281] Degraded feature vector set The input spatiotemporal fusion prediction model, after feature extraction, fusion, and prediction layer output, yields the future... Time period (set as 1 hour, i.e.) , Power quality index prediction sequence (with average sampling step size) Based on the judgment criteria in step 3.4, the output is:

[0282] Deterioration and evolution trend: Positive values ​​indicate worsening degradation, while negative values ​​indicate mitigating degradation.

[0283] Deterioration level: based on The corresponding judgment criteria determine the degradation level at each moment. ;

[0284] Time of occurrence: First appearance The moment And the moment when it reaches severe degradation (Grade=3). (If it exists).

[0285] 4.2 Adaptive Kalman Filter Correction

[0286] To address model prediction errors, an adaptive Kalman filter (AKF) is used to dynamically adjust the filter parameters, correct the prediction results, and improve prediction accuracy. The core formula is as follows:

[0287] (1) Equation of state:

[0288] (2) Observation equation:

[0289] (3) Filter gain:

[0290] (4) Status update:

[0291] (5) Covariance update:

[0292] The adaptive adjustment process involves using prediction errors. Dynamic adjustment process noise covariance and observation noise covariance ,formula: , .

[0293] Parameter definition:

[0294] The predicted value at time k. Corrected predicted values;

[0295] State transition matrix, set as identity matrix (Single power quality index);

[0296] : Control matrix, set to 0;

[0297] : Control quantity, set to 0;

[0298] Process noise follows a normal distribution. ;

[0299] Observation matrix, set to 1;

[0300] Observational noise follows a normal distribution. ;

[0301] : Filter gain;

[0302] Predicting covariance, : Updated covariance;

[0303] Adaptive adjustment coefficients are all set to 0.9;

[0304] : The prediction error at time k.

[0305] 4.3 Improved Deep SHAP Algorithm (Dominant Factor Identification)

[0306] An improved Deep SHAP algorithm is adopted to quantify the contribution of each degradation feature to power quality degradation, thereby achieving quantitative identification and importance ranking of the dominant factors. The improvement involves introducing feature weight coefficients. (Same as step 2.1), optimize the accuracy of contribution calculation, core formula:

[0307]

[0308] in, The contribution of the i-th degradation feature. For the set of all degradation features, For the total number of features, For feature subset The corresponding model prediction output, The weighting coefficient for the i-th feature (same as step 2.1).

[0309] Parameter definition:

[0310] The contribution of the i-th degradation feature. This indicates that the feature promotes degradation. This indicates the suppression of degradation; the larger the absolute value, the greater the impact.

[0311] : Deterioration feature set ;

[0312] Feature subset This represents the set that does not contain the i-th feature;

[0313] Total number of features, here ;

[0314] Feature subset Predicted values ​​after inputting into the model;

[0315] : The weighting coefficient of the i-th feature, same as in step 2.1.

[0316] according to The magnitude of the value is used to sort all degradation features, and the dominant factors of power quality degradation (the top 3 features with the largest contribution) are obtained.

[0317] Step 5: Model closed-loop optimization

[0318] Based on the corrected prediction results and the identification results of dominant factors, tiered early warning information and differentiated prevention and control plans are generated. Actual operation and maintenance feedback data are collected, and model parameters are dynamically optimized to achieve closed-loop iteration. Specifically, this includes:

[0319] 5.1 Generation of Tiered Early Warning and Prevention and Control Plans

[0320] Based on the corrected degradation level Generate tiered early warning information and, in conjunction with key factors, formulate differentiated prevention and control plans, specifically:

[0321] Level 1 (Slight Deterioration): Yellow alert, adjust photovoltaic output smoothing strategy (such as energy storage charge and discharge regulation);

[0322] Level 2 (Moderate Deterioration): Orange alert, activate reactive power compensation device, adjust agricultural load operation time;

[0323] Level 3 (Severe Deterioration): Red alert, cut off part of the photovoltaic output or non-critical loads, and conduct emergency line maintenance.

[0324] 5.2 Dynamic Adjustment of Model Parameters

[0325] Collect actual operation data of power distribution network In addition to operation and maintenance feedback data (effectiveness of the prevention and control plan), the error between the corrected predicted value and the actual value is calculated. ,like ( (Error threshold), the spatiotemporal fusion prediction model parameters are dynamically adjusted using the stochastic gradient descent (SGD) algorithm, formula:

[0326]

[0327] Parameter definition:

[0328] : Actual operating value of power quality indicators at time t;

[0329] Average prediction error, in percentages (%)

[0330] Error threshold, set to 3%;

[0331] Model weight matrix (including wait);

[0332] Learning rate, set to 0.001;

[0333] : The partial derivative of the composite loss function with respect to the weight matrix W.

[0334] 5.3 Knowledge distillation enables lightweight model updates

[0335] To avoid deployment difficulties caused by excessive model parameters, knowledge distillation (KD) technique is employed. A pre-trained spatiotemporal fusion prediction model is used as the teacher model to construct a lightweight student model. The knowledge (predicted probability distribution) from the teacher model is transferred to the student model, achieving lightweight model updates. The core formula is:

[0336]

[0337] Parameter definition:

[0338] Knowledge distillation loss;

[0339] Temperature coefficient, set to 10, controls the smoothness of the softening probability distribution;

[0340] Cross-entropy loss The actual label (deterioration level);

[0341] The output logits of the teacher model;

[0342] The output logits of the student model;

[0343] Mean squared error measures the difference between the probability distributions of the teacher's and student's model outputs.

[0344] 5.4 Closed-loop iterative optimization

[0345] The lightweight updated student model is used as the new spatiotemporal fusion prediction model. The process of steps 1-4 is repeated to achieve a closed-loop iteration of "data acquisition-feature extraction-prediction correction-parameter optimization", ensuring that the model prediction accuracy continues to improve and adapts to the dynamic operation characteristics of high-proportion photovoltaic distribution networks.

[0346] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the evolution of power quality degradation in high-proportion photovoltaic distribution networks, characterized in that: include: Layered dynamic data acquisition: Intelligent monitoring terminals are deployed in layers according to photovoltaic clusters, distribution network lines and agricultural load nodes to collect photovoltaic output, line electrical quantity and load data in real time; the sampling frequency is dynamically adjusted based on the photovoltaic output fluctuation rate and load change rate, and after preprocessing the collected data, a multi-dimensional time series prediction database is constructed. Degradation feature extraction: Based on the time series prediction database, an autoencoder is used to extract weighted and enhanced multidimensional degradation feature vectors, and a feature coupling coefficient is introduced to quantify the correlation between the dimensions of the degradation feature vectors, thereby constructing a degradation feature vector set; Constructing a spatiotemporal fusion prediction model: Based on a CNN-LSTM-GRU hybrid network, graph convolutional network, and physical information network, a multi-head self-attention mechanism is introduced to fuse spatiotemporal features and construct a spatiotemporal fusion prediction model; a composite loss function is designed to constrain the prediction results, and a power quality degradation judgment standard is formulated in combination with the characteristics of a high-proportion photovoltaic distribution network. Degradation prediction: Input the degradation feature vector set into the spatiotemporal fusion prediction model, output the degradation evolution trend, level and occurrence time, and combine the adaptive Kalman filter to correct the prediction results; An improved Deep SHAP algorithm is used to quantify the contribution of each degradation feature to power quality degradation, thereby achieving quantitative identification and importance ranking of the dominant factors. Model closed-loop optimization: Based on the corrected prediction results and the identification results of dominant factors, generate hierarchical early warning information and differentiated prevention and control plans and issue them for execution; collect actual operation data and operation and maintenance feedback data of the distribution network, dynamically adjust the parameters of the spatiotemporal fusion prediction model, and combine knowledge distillation to realize lightweight model updates and complete closed-loop iterative optimization.

2. The method for predicting power quality degradation evolution applicable to high-proportion photovoltaic distribution networks according to claim 1, characterized in that: Layered dynamic data acquisition specifically includes: Intelligent monitoring terminals are deployed across three levels: photovoltaic clusters, distribution network lines, and agricultural load nodes. First, raw power quality data from multiple dimensions, including photovoltaic output, line electrical quantities, and agricultural load power, are collected with an initial step size. Then, by calculating the photovoltaic output fluctuation rate and agricultural load change rate, and combining this with preset high and low frequency thresholds, the sampling step size is dynamically adjusted to balance the capture of key fluctuation information with the reduction of data redundancy. Subsequently, outliers are removed and corrected using the 3σ criterion, noise is removed by moving average filtering, and data dimension differences are eliminated through normalization, thus completing the cleaning and standardization of the raw data. Finally, the preprocessed multi-feature data is integrated according to timestamps to construct a multi-dimensional time-series prediction database.

3. The method for predicting power quality degradation evolution applicable to high-proportion photovoltaic distribution networks according to claim 1, characterized in that: Degradation feature extraction specifically includes: Using a time-series database as input, a weighted augmented autoencoder is constructed. By assigning differentiated weights to different input features, the expression of features related to power quality degradation is strengthened and interference from irrelevant features is suppressed. The original electrical data is compressed into a degradation feature vector by the encoder, and the autoencoder parameters are optimized using reconstruction loss. Subsequently, the linear correlation between each degradation feature dimension is quantified by the feature coupling coefficient to identify redundant features. Finally, strongly correlated redundant features are eliminated by using the coupling coefficient threshold to simplify the feature dimensions and form a low-redundancy, highly representative degradation feature vector set.

4. The power quality degradation evolution prediction method applicable to high-proportion photovoltaic distribution networks according to claim 1, characterized in that: The construction of a spatiotemporal fusion prediction model specifically includes: Three basic networks—a CNN-LSTM-GRU hybrid network, a graph convolutional network (GCN), and a physical information network (PINN)—were constructed to extract temporal and local spatial correlation features of degradation characteristics and global spatial correlation features driven by the distribution network topology, respectively. Physical constraint losses were formed by embedding power balance and voltage constraints into the power grid's physical equations to ensure that the prediction results conform to the power grid's operating rules. A multi-head self-attention mechanism was then used to fuse local and global spatiotemporal features to complete the spatiotemporal fusion prediction model. Subsequently, a composite loss function was constructed that integrates prediction error, physical constraints, and regularization, and the model was optimized according to preset weight coefficients to improve prediction accuracy while suppressing overfitting. Finally, combining the characteristics of a high-proportion photovoltaic distribution network and agricultural loads, a four-level power quality degradation judgment standard was formulated based on voltage deviation and total harmonic distortion rate, achieving accurate prediction and level discrimination of power quality degradation evolution.

5. The power quality degradation evolution prediction method applicable to high-proportion photovoltaic distribution networks according to claim 1, characterized in that: Degradation prediction specifically includes: The degradation feature vector set is input into the spatiotemporal fusion prediction model, which outputs a predicted sequence of power quality indicators for a future period. Combined with the preset four-level degradation judgment criteria, the degradation evolution trend, degradation level at each moment, and the time of first degradation and severe degradation are clarified. To address the model prediction error, an adaptive Kalman filter is used to correct the prediction results and improve prediction accuracy by dynamically adjusting the covariance of process noise and observation noise. Finally, an improved Deep SHAP algorithm with feature weight coefficients is used to quantify the contribution of each degradation feature to power quality degradation. By ranking the absolute values ​​of contribution, the dominant factors affecting degradation are identified, thus completing the accurate prediction, error correction, and location of dominant factors of power quality degradation.

6. The method for predicting power quality degradation evolution applicable to high-proportion photovoltaic distribution networks according to claim 1, characterized in that: Model closed-loop optimization specifically includes: Based on the corrected degradation prediction results and the identification results of the dominant factors, graded early warning and differentiated prevention and control plans are generated according to different degradation levels, and targeted prevention and control measures such as photovoltaic output adjustment, reactive power compensation, and load shedding are taken. Subsequently, actual operation data and operation and maintenance feedback data of the distribution network are collected to calculate the prediction error. If the error exceeds the preset threshold, the parameters of the spatiotemporal fusion prediction model are dynamically adjusted using the stochastic gradient descent algorithm. In order to solve the deployment problem caused by too many model parameters, knowledge distillation technology is used to transfer prediction knowledge to achieve lightweight model update by using the original model as the teacher model and the lightweight model as the student model. Finally, the updated lightweight model is used as the new prediction model, and the whole process of power quality degradation evolution prediction is repeated to form a closed loop iteration.

7. The method for predicting power quality degradation evolution applicable to high-proportion photovoltaic distribution networks according to claim 2, characterized in that: The raw power quality data includes: Photovoltaic power output data ; Line electrical quantity data: Voltage Current Power factor Total harmonic distortion ; Agricultural load data: Active load reactive load .

8. The power quality degradation evolution prediction method applicable to high-proportion photovoltaic distribution networks according to claim 3, characterized in that: The characteristic coupling coefficient is calculated as follows: Introducing characteristic coupling coefficients Quantization of degraded feature vectors The correlation between the i-th and j-th dimensions, to avoid feature redundancy, is expressed by the formula: in, Let be the covariance between the i-th and j-th features. , Let be the variances of the i-th and j-th features, respectively. The formula is: in: : The coupling coefficient between the i-th and j-th dimension degradation features, with a value range of [-1, 1]. The larger the value, the higher the correlation. Degraded eigenvectors at time t The i-th dimension component; : The mean of the i-th dimension of degradation feature, ; Covariance function: measures the degree of linear correlation between two features; Variance function: measures the degree of dispersion of a single feature.

9. The method for predicting the evolution of power quality degradation in a high-proportion photovoltaic distribution network according to claim 5, characterized in that: The improved Deep SHAP algorithm is specifically as follows: Introducing feature weight coefficients optimizes the accuracy of contribution calculation. The core formula is: in: The contribution of the i-th degradation feature. This indicates that the feature promotes degradation. This indicates the suppression of degradation; the larger the absolute value, the greater the impact. : Set of deterioration features; Feature subset This represents the set that does not contain the i-th feature; Total number of features; Feature subset Predicted values ​​after inputting into the model; : The weighting coefficient of the i-th feature.

10. A system for implementing the power quality degradation evolution prediction method for high-proportion photovoltaic distribution networks as described in any one of claims 1 to 9, characterized in that: include: Layered data acquisition module: used to deploy intelligent monitoring terminals in layers according to photovoltaic clusters, distribution network lines and agricultural load nodes, collect multi-dimensional operation data in real time, dynamically adjust the sampling frequency, preprocess the data and build a time series prediction database; Degradation Feature Extraction Module: Based on the time-series prediction database, it extracts multi-dimensional degradation feature vectors through a weighted augmented autoencoder, calculates feature coupling coefficients, removes redundant features, and constructs a set of degradation feature vectors. Spatiotemporal fusion prediction module: used to construct a spatiotemporal fusion prediction model based on CNN-LSTM-GRU hybrid network, graph convolutional network and physical information network, introduce multi-head self-attention mechanism to fuse spatiotemporal features, design composite loss function to constrain model training, and formulate power quality degradation judgment criteria. Degradation prediction and correction module: It is used to input the degradation feature vector set into the spatiotemporal fusion prediction model, output the degradation evolution trend, level and occurrence time, correct the prediction results through adaptive Kalman filtering, and use the improved Deep SHAP algorithm to quantify the contribution of degradation features, and complete the identification and ranking of dominant factors. Closed-loop optimization module: Based on the corrected prediction results and the identification results of the dominant factors, it generates graded early warning information and differentiated prevention and control plans and issues them for execution, collects actual operation and maintenance feedback data, dynamically adjusts model parameters, realizes lightweight model updates through knowledge distillation, and completes closed-loop iterative optimization; Main control module: Electrically connected to each of the above modules, used to control the coordinated work of each module, schedule data transmission and processing flow, ensure the stable operation of the entire system and realize the function of predicting the evolution of power quality degradation.

Images