Photovoltaic string fault positioning method based on dynamic threshold and topology analysis

Abstract

The invention discloses a photovoltaic string fault positioning method based on a dynamic threshold value and topology analysis. The method comprises the following steps: firstly, collecting current, voltage, temperature and illumination intensity data of a string in real time; then, generating a fault judgment threshold dynamically adjusted along with the environment based on historical data and an LSTM model; then, comparing the real-time data with a dynamic threshold value to detect an abnormal node; then, constructing a string topology network taking electrical impedance as a weight, and accurately positioning a fault point by using a Dijkstra shortest path algorithm; and finally, multi-dimensional features are extracted, specific fault types such as hot spots and open circuits are identified through a fuzzy logic classifier, and a diagnosis report is output. According to the method, the problems that the static threshold false alarm rate is high and the traditional topology analysis precision is insufficient are solved, rapid and accurate positioning and diagnosis of photovoltaic faults are realized, and the operation and maintenance efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation system operation and maintenance technology, and in particular to a method for locating photovoltaic string faults based on dynamic thresholds and topology analysis. Background Technology

[0002] String fault location in photovoltaic (PV) systems is a critical aspect of PV power generation system operation and maintenance, directly impacting power generation efficiency and equipment lifespan. Current mainstream methods rely on current-voltage (IV) curve analysis, infrared thermal imaging, and static threshold alarm technology. Among these, the static threshold method identifies faults by setting fixed thresholds (such as current deviation rate or voltage drop), while topology analysis locates faulty units through the electrical connections of the strings.

[0003] Existing technologies have the following shortcomings: First, static thresholds have poor adaptability. Photovoltaic systems are dynamically affected by environmental factors such as light intensity and temperature, and fixed thresholds are prone to false alarms or missed alarms, especially in complex environments such as rain, sandstorms, etc. Second, topology analysis accuracy is insufficient. Traditional topology models do not consider the dynamic coupling effect between strings, resulting in large errors in fault location for scenarios with multiple branches in parallel or complex connections. Third, data processing efficiency is low. Existing methods rely on manual inspection or offline analysis, which has poor real-time performance and cannot meet the rapid response requirements of large-scale power plants. Finally, the ability to distinguish fault types is weak. It is difficult to distinguish different fault types such as component aging, hot spot effect, and shading, resulting in insufficient targeting of operation and maintenance strategies. Summary of the Invention

[0004] The purpose of this invention is to provide a photovoltaic string fault location method based on dynamic threshold and topology analysis, so as to solve the technical problems of poor adaptability of static threshold, insufficient accuracy of topology analysis, low data processing efficiency and weak fault type differentiation in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A photovoltaic string fault location method based on dynamic threshold and topology analysis includes the following steps:

[0007] Real-time data acquisition of string current, voltage, temperature, and light intensity is performed, and noise filtering and data normalization are conducted using edge computing devices. Based on the mean and variance of historical data using a sliding window, an LSTM machine learning model is used to predict the normal state range of the string, and a dynamic threshold is generated by combining a two-factor adaptive correction mechanism. Real-time data is compared with the dynamic threshold to detect abnormal nodes. A string electrical topology network with electrical impedance as the weight is constructed, and the lowest cumulative impedance fault propagation path from the inverter to the abnormal node is calculated and quantified using Dijkstra's shortest path algorithm. Features of the IV curve, infrared thermal imaging data, and environmental sensor information are extracted, and a fuzzy logic classifier is used to distinguish fault types. The fault type and confidence level are output.

[0008] Furthermore, the dynamic threshold calculation formula for the two-factor adaptive correction mechanism is as follows:

[0009] Threshold(t) = α[μ h (t)±k·σ h [(t)]+β·E corr (t)

[0010] Where α is the weighting coefficient for historical data, μ h (t) represents the mean current in the sliding window, k is the confidence interval coefficient, and σ h (t) represents the standard deviation of the current, β is the environmental correction weighting coefficient, and E corr (t) represents the environmental correction term.

[0011] Furthermore, the dynamic threshold generation further includes: using an LSTM neural network model to predict the threshold range for future time periods. The LSTM neural network model includes: an input layer that receives current, voltage, temperature, and light intensity data from the past 72 hours; a hidden layer that uses a bidirectional LSTM structure containing 128 neurons, controlling the information flow through a forget gate, an input gate, and an output gate; an output layer that predicts the upper and lower bounds of the threshold range for the next 24 hours; and model training using the Adam optimizer with a learning rate of 0.001 and a loss function of mean squared error.

[0012] Furthermore, the Dijkstra shortest path algorithm specifically includes: constructing a weighted string electrical topology graph, with each string node as a vertex and the electrical connections between strings as weighted edges, where the weight w(u,v) of the edge is the electrical impedance between node u and node v; initializing a distance array, setting the inverter node as the starting point with a distance of 0, and other nodes with infinite distances; maintaining unvisited nodes using a priority queue, sorted by distance from smallest to largest; iteratively updating the shortest distance between adjacent nodes, with the distance calculation formula being:

[0013] d(u,v)=d(u)+w(u,v)·(1+α·Zabn (v))

[0014] Where d(u,v) is the distance from node u to node v, w(u,v) is the edge weight, α is the impedance anomaly weight coefficient, and Z abn (v) is the impedance anomaly score of node v; after an abnormal node is detected, the backtracking path determines the shortest fault propagation path from the inverter to the fault node.

[0015] Furthermore, the detection of abnormal nodes specifically includes: establishing a multi-dimensional abnormality detection index system, including current deviation, voltage deviation, impedance deviation, and temperature deviation; and employing a sliding window abnormality detection algorithm.

[0016]

[0017] Among them, X j Let μ be the current value of the j-th indicator. j and σ j These are the historical mean and standard deviation of the indicator, ω. j For weighting coefficients; when Anomaly score When the threshold is greater than Threshold(t), it is marked as an abnormal node; an abnormal node confirmation mechanism is established, and an abnormal node is only confirmed as a faulty node when an abnormality is detected continuously for more than a set time window.

[0018] Furthermore, the calculation formula for the environmental correction term is as follows:

[0019] E corr (t)=γ·(G / G0)·e -δ·T

[0020] Where γ is the illumination correction factor, G is the current illumination intensity, G0 is the illumination intensity under standard test conditions, δ is the temperature attenuation factor, and T is the current temperature.

[0021] Furthermore, the formula for calculating the impedance anomaly score is as follows:

[0022]

[0023] Among them, Z real Z represents the actual measured impedance value. model Here, λ represents the impedance value of the theoretical model, and λ is the rate of change weighting factor. Let Z be the partial derivative of the current with respect to the voltage; the fault determination condition is: when Z... abn When the percentage is greater than 15%, it is marked as a faulty node.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. This invention, through a dynamic threshold mechanism, enables the system to automatically adjust its judgment criteria based on changes in environmental factors such as light intensity and temperature, significantly reducing the false alarm rate in complex environments. (See attached diagram.) Figure 6 As shown, compared with the static threshold method, the false alarm rate of the present invention is improved by 84.6%, 90.4%, 94.3%, 90.7%, 92.8%, 93.4% and 92.0% in environments such as sunny days, thin clouds, strong light, cloudy days, rainstorms, snowy days and sandstorms, respectively.

[0026] 2. This invention employs a topology analysis method with electrical impedance as the weight, as shown in the appendix. Figure 7 As shown, at scales of 100 nodes, 500 nodes, 1000 nodes, and 5000 nodes, the accuracy of the method of the present invention is 99.2%, 97.8%, 96.5%, and 95.1%, respectively, and the positioning error is 0.2m, 0.5m, 0.8m, and 1.2m, respectively. Compared with the accuracy of the traditional method (87.5%, 72.3%, 63.1%, and 48.7%), this represents an improvement.

[0027] 3. This invention provides the application of edge computing technology, achieving millisecond-level response and meeting the real-time monitoring needs of large-scale power plants. The response times are 12ms, 58ms, 125ms, and 680ms at scales of 100 nodes, 500 nodes, 1000 nodes, and 5000 nodes, respectively.

[0028] 4. This invention supports the accurate identification of more than 6 types of faults, such as hot spot effect, line disconnection, and shadow occlusion. It provides confidence assessment through a fuzzy logic classifier and provides targeted maintenance suggestions for operation and maintenance personnel. Attached Figure Description

[0029] To more clearly illustrate the technical solution of the present invention, the accompanying drawings are briefly described below:

[0030] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0031] Figure 2 This is a schematic diagram of the dynamic threshold generation algorithm.

[0032] Figure 3 This is a schematic diagram of topology analysis;

[0033] Figure 4 This is a schematic diagram of a fuzzy logic classifier structure;

[0034] Figure 5 System hardware deployment architecture diagram;

[0035] Figure 6The chart comparing the false alarm rates of fault detection under different environments shows the results of comparing the false alarm rates of the method of this invention with those of the static threshold method and the traditional LSTM method under seven environmental conditions: sunny day, thin clouds, strong light, cloudy day, rainstorm, snow, and dust storm.

[0036] Figure 7 The comparison chart shows the positioning accuracy of photovoltaic arrays of different sizes, illustrating the positioning accuracy and positioning error of the method of this invention and the traditional method in arrays with scales of 100 nodes, 500 nodes, 1000 nodes, and 5000 nodes. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the present invention clearer, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of protection of the invention.

[0038] Example 1

[0039] This embodiment aims to specifically disclose the complete technical solution of a photovoltaic string fault location method based on dynamic threshold and topology analysis.

[0040] Reference Figure 1 This embodiment provides a photovoltaic string fault location method based on dynamic threshold and topology analysis. The method specifically includes the following steps:

[0041] Step S101: Data Acquisition and Preprocessing

[0042] Basic data acquisition and preprocessing

[0043] Real-time data acquisition of string current, voltage, temperature, and light intensity is performed, and noise filtering and data normalization are conducted using edge computing devices. High-precision sensors are employed: HIOKIPW3390 current sensor (accuracy ±0.2%), Fluke289 voltage sensor (accuracy ±0.1%), SensirionSHT45 temperature sensor (accuracy ±0.1℃), and a light intensity sensor with a measurement range of 0-2000W / m. 2 .

[0044] • Specific parameters and conditions: Sampling frequency 100Hz, data transmission adopts Modbus RTU protocol, edge computing gateway is based on NVIDIA Jetson AGXXavier platform, with 64GB RAM and 32 TOPSAI computing power.

[0045] Algorithm and Model: Kalman filter is used for noise filtering.

[0046] Step S102: Dynamic threshold generation

[0047] Based on the mean and variance of historical data obtained through a sliding window, an LSTM machine learning model is used to predict the normal range of string states, and a dynamic threshold is generated by combining a two-factor adaptive correction mechanism. (Refer to...) Figure 2 The sliding window size is set to 72 hours, containing 259,200 data points.

[0048] • LSTM prediction model: An LSTM neural network model is used to predict threshold intervals for future time periods. The specific structure is as follows:

[0049] Input layer: Receives historical 72-hour data on current, voltage, temperature, and light intensity; data dimension is 4×72.

[0050] o Hidden layer: Employs a bidirectional LSTM structure, containing 128 neurons, and uses a forgetting gate f t =σ(W f ·[h t-1 ,x t ]+b f ), Input gate i t =σ(W i ·[h t-1 ,x t ]+b i ) and output gate o t =σ(W o ·[h t-1 ,x t ]+b o Control information flow

[0051] Output layer: Predicts the upper and lower bounds of the threshold range for the next 24 hours.

[0052] Model training: The Adam optimizer was used with a learning rate of 0.001 and the loss function was the mean squared error.

[0053] • Specific parameters and conditions: historical data weighting coefficient α = 0.7, environmental correction weighting coefficient β = 0.3, confidence interval coefficient k = 2.58 (corresponding to 99% confidence level), illumination correction coefficient γ = 0.15, temperature attenuation coefficient δ = 0.023.

[0054] Algorithm and Model: The formula for calculating the dynamic threshold is:

[0055] Threshold(t) = α[μ h (t)±k·σ h [(t)]+β·E corr (t)

[0056] Among them, the average current of the sliding window Current standard deviation Environmental remediation items Where G0 = 1000 W / m 2 The light intensity is for standard test conditions.

[0057] Step S103: Anomaly Detection

[0058] Real-time data is compared with dynamic thresholds to detect abnormal nodes. Specifically, the abnormal node detection described in this invention involves establishing a multi-dimensional abnormal detection index system, including current deviation, voltage deviation, impedance deviation, and temperature deviation; and employing a sliding window abnormal detection algorithm.

[0059]

[0060] Among them, X j Let μ be the current value of the j-th indicator. j and σ j These are the historical mean and standard deviation of the indicator, ω. j These are weighting coefficients. Set the anomaly detection threshold; when anomalies occur... score When the threshold is greater than Threshold(t), the node is marked as an anomalous node. Furthermore, an anomalous node confirmation mechanism is established, confirming a node as faulty only when anomalies are continuously detected for more than a set time window, in order to avoid misjudgments caused by instantaneous fluctuations.

[0061] Step S104: Topology Analysis and Fault Location

[0062] A string electrical topology network with electrical impedance as the weight is constructed, and Dijkstra's shortest path algorithm is used to identify abnormal nodes and calculate the lowest cumulative impedance fault propagation path from the inverter to the node.

[0063] Implementation of Dijkstra's shortest path algorithm:

[0064] First, a weighted string electrical topology graph is constructed, with each string node as a vertex and the electrical connections between strings as weighted edges, where the edge weight represents the electrical impedance between nodes. A distance array is initialized, setting the inverter node as the starting point with a distance of 0, and other nodes with infinite distances. A priority queue is used to maintain unvisited nodes, sorted in ascending order of distance.

[0065] Iteratively update the shortest distance between adjacent nodes. The distance calculation formula is:

[0066] d(u,v)=d(u)+w(u,v)·(1+α·Z abn (v))

[0067] Where d(u,v) is the distance from node u to node v, w(u,v) is the edge weight, α is the impedance anomaly weight coefficient, and Z abn(v) represents the impedance anomaly score of node v.

[0068] The impedance anomaly score Z abn The calculation formula is:

[0069]

[0070] Among them, Z real Z represents the actual measured impedance value. model Here, λ represents the impedance value of the theoretical model, and λ is the rate of change weighting factor. Let Z be the partial derivative of the current with respect to the voltage. The fault determination condition is: when Z... abn When the percentage is greater than 15%, it is marked as a faulty node.

[0071] After detecting an abnormal node, the backtracking path determines the shortest fault propagation path from the inverter to the faulty node.

[0072] Step S105: Fault Feature Extraction and Classification

[0073] Extract fault features and output the fault type and confidence level using a fuzzy logic classifier. (Refer to...) Figure 4 The eigenvector includes the current deviation rate I. dev Voltage deviation rate V dev Temperature difference ΔT and impedance anomaly score Z abn .

[0074] • Specific parameters and conditions: Supports identification of 6 fault types: normal, hot spot effect, line breakage, shading, component aging, and poor contact. The classification confidence threshold is set to 0.85. Feature input vector x = [I dev V dev ,ΔI,Z abn ].

[0075] • Fuzzy rule base: Designed based on the rules in the technical briefing, it includes the following core rules:

[0076] Rule 1: IFI dev =High ANDV dev =Low ANDΔT=High THEN Hot Spot Effect

[0077] Rule 2: IFI dev =Zero ANDV dev =High ANDΔT=Low THEN Line disconnected

[0078] Rule 3: IFI dev =Middle ANDV dev =Middle AND ΔT=Low THEN occlusion

[0079] Rule 4: IFI dev=High ANDV dev =High ANDΔT=High THEN Component Aging

[0080] Algorithm and Model: The fuzzy logic classifier uses a triangular membership function.

[0081]

[0082] Where a, b, and c are the left endpoint, vertex, and right endpoint parameters of the trigonometric function, respectively, and the feature input vector is x = [I dev V dev ,ΔT,Z abn ].

[0083] Step S107: Output the result

[0084] Outputs a fault location map, fault type report, and maintenance recommendations. The system generates a comprehensive report including fault node coordinates, fault type, confidence level, impact range, and recommended maintenance measures.

[0085] Working principle and function

[0086] This method, implemented in this embodiment, adaptively adjusts the judgment criteria based on historical data and environmental factors through a dynamic threshold mechanism, effectively addressing the dynamic characteristics of photovoltaic systems. Topology analysis combined with Dijkstra's shortest path algorithm identifies abnormal nodes and quantifies fault propagation paths. A fuzzy logic classifier achieves accurate fault type identification through multi-feature fusion.

[0087] Technical effect

[0088] By adopting the above method, the following beneficial effects can be achieved:

[0089] 1. Significantly improved environmental adaptability: (Refer to...) Figure 6 The dynamic threshold mechanism enables the system to automatically adapt to different environmental conditions. For example... Figure 6 As shown, under sunny conditions, the false alarm rate of the method of this invention is only 0.8%, while the false alarm rate of the traditional static threshold method is 5.2%; under dusty weather conditions, the false alarm rate of the method of this invention is 4.2%, while the traditional method is as high as 52.3%; under cloudy weather conditions, the false alarm rate of the method of this invention is 2.3%, while the traditional method reaches 24.8%. Compared with the traditional method, the false alarm rate of this invention is reduced under different environmental conditions, demonstrating excellent environmental adaptability.

[0090] 2. Significantly improved fault location accuracy: (Refer to...) Figure 7The comparison chart shows the positioning accuracy of photovoltaic arrays of different sizes. At scales of 100 nodes, 500 nodes, 1000 nodes, and 5000 nodes, the accuracy of the method of this invention is 99.2%, 97.8%, 96.5%, and 95.1%, respectively, with positioning errors of 0.2m, 0.5m, 0.8m, and 1.2m, respectively. This represents an improvement over the accuracy of traditional methods (87.5%, 72.3%, 63.1%, and 48.7%).

[0091] 3. Significantly optimized computational efficiency: The response time of the 5000-node array is controlled within 680ms, meeting the real-time monitoring requirements of large-scale power plants.

[0092] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A photovoltaic string fault location method based on dynamic threshold and topology analysis, characterized in that, Includes the following steps: Real-time acquisition of string current, voltage, temperature, and light intensity data, followed by noise filtering and data normalization processing via edge computing devices; Based on the mean and variance of historical data obtained through a sliding window, an LSTM machine learning model is used to predict the normal state range of the inverter string. A dynamic threshold is generated by combining a two-factor adaptive correction mechanism. Real-time data is compared with the dynamic threshold to detect abnormal nodes. An electrical topology network of the inverter string is constructed with electrical impedance as the weight, and the lowest cumulative impedance fault propagation path from the inverter to the abnormal node is calculated and quantified by combining Dijkstra's shortest path algorithm. Features of the IV curve, infrared thermal imaging data, and environmental sensor information are extracted, and the fault type is distinguished by a fuzzy logic classifier. The fault type and confidence level are output.

2. The method according to claim 1, characterized in that, The dynamic threshold calculation formula for the two-factor adaptive correction mechanism is as follows: Threshold(t)=α[μ h (t)±k·σ h (t)]+β·E corr (t) Where α is the weighting coefficient for historical data, μ h (t) represents the mean current in the sliding window, k is the confidence interval coefficient, and σ h (t) represents the standard deviation of the current, β is the environmental correction weighting coefficient, and E corr (t) represents the environmental correction term.

3. The method according to claim 1, characterized in that, The dynamic threshold generation further includes: using an LSTM neural network model to predict the threshold range for future time periods. The LSTM neural network model includes: an input layer that receives current, voltage, temperature, and light intensity data from the past 72 hours; a hidden layer that uses a bidirectional LSTM structure containing 128 neurons, controlling the information flow through a forget gate, an input gate, and an output gate; and an output layer that predicts the upper and lower bounds of the threshold range for the next 24 hours. The model is trained using the Adam optimizer with a learning rate of 0.001 and a loss function of mean squared error.

4. The method according to claim 1, characterized in that, The Dijkstra shortest path algorithm specifically includes: constructing a weighted string electrical topology graph, where each string node is a vertex and the electrical connections between strings are weighted edges, where the weight w(u,v) of an edge is the electrical impedance between node u and node v; initializing a distance array, setting the inverter node as the starting point with a distance of 0, and other nodes with infinite distances; maintaining unvisited nodes using a priority queue, sorted by distance from smallest to largest; and iteratively updating the shortest distance between adjacent nodes, using the following distance calculation formula: d(u,v)=d(u)+w(u,v)·(1+α·Z abn (v)) Where d(u,v) is the distance from node u to node v, w(u,v) is the edge weight, α is the impedance anomaly weight coefficient, and Z abn (v) is the impedance anomaly score of node v; after an abnormal node is detected, the backtracking path determines the shortest fault propagation path from the inverter to the fault node.

5. The method according to claim 2, characterized in that, The specific steps for detecting abnormal nodes include: establishing a multi-dimensional abnormal detection index system, including current deviation, voltage deviation, impedance deviation, and temperature deviation; and employing a sliding window abnormal detection algorithm. Among them, X j Let μ be the current value of the j-th indicator. j and σ j These are the historical mean and standard deviation of the indicator, ω. j The weighting coefficient is set; the anomaly detection threshold is set when Anomaly score When the threshold is greater than Threshold(t), it is marked as an abnormal node; an abnormal node confirmation mechanism is established, and an abnormal node is only confirmed as a faulty node when an abnormality is detected continuously for more than a set time window.

6. The method according to claim 2, characterized in that, The formula for calculating the environmental correction item is as follows: E corr (t)=γ·(G / G0)·e -δ·T Where γ is the illumination correction factor, G is the current illumination intensity, G0 is the illumination intensity under standard test conditions, δ is the temperature attenuation factor, and T is the current temperature.

7. The method according to claim 4, characterized in that, The formula for calculating the impedance anomaly score is as follows: Among them, Z real Z represents the actual measured impedance value. model Here, λ represents the impedance value of the theoretical model, and λ is the rate of change weighting factor. Let Z be the partial derivative of the current with respect to the voltage; the fault determination condition is: when Z... abn When the percentage is greater than 15%, it is marked as a faulty node.

Images