A fault alarm device for solar power generation

Abstract

A fault alarm device for solar power generation belongs to the field of solar fault monitoring. The system includes an edge monitoring module, a multimodal data fusion module, a fault classification module, a trend prediction module, a dynamic adjustment module and an alarm response module. The present invention improves the accuracy of fault identification by integrating multidimensional features such as photovoltaic panel current, voltage, power, ambient temperature and light intensity by using a deep learning fusion model. By introducing an attention mechanism and dynamically adjusting the alarm threshold, it effectively distinguishes environmental fluctuations from actual faults, reduces the false alarm rate and missed alarm rate, and can adapt to the complex and changeable photovoltaic system operating environment without manual intervention.

Description

Technical Field

[0001] The present invention belongs to the technical field of solar energy fault monitoring, and more specifically, particularly relates to a fault alarm device for solar power generation. Background Art

[0002] With the widespread application of solar power generation, the stability and efficiency of its system operation have become the core issues of operation and maintenance. However, traditional fault alarm methods have the following problems in terms of real-time, accuracy and adaptability. Traditional methods rely on regular manual inspections or simple threshold alarm strategies, which are difficult to capture sudden faults or dynamic changes in a timely manner. The system has a slow response speed, which may cause small faults to evolve into large-scale losses, such as component burnout or grid feedback failure. Simple fixed threshold settings ignore the dynamic effects of weather conditions, operating environment and equipment aging on parameters. Some normal fluctuations are misjudged as faults, while minor anomalies may not be detected in time. It is difficult to distinguish different fault types and provide accurate fault location and diagnosis.

[0003] The system usually uses static rules to set thresholds, which makes it difficult to cope with complex and changing environmental conditions. For example, in cloudy weather or when light intensity changes rapidly, system performance may fluctuate, and traditional methods are prone to false alarms. It is impossible to grade alarms according to the severity of the fault, resulting in all anomalies being handled with the same priority. Traditional alarm methods are mostly based on static data or historical data, which makes it difficult to predict future equipment status and detect potential faults in advance. It is difficult to take preventive measures early to avoid further losses. A new fault alarm method based on intelligent algorithms is urgently needed to solve the above problems and improve the operating efficiency and maintenance level of solar power generation systems. Summary of the invention

[0004] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0005] The present invention is proposed in view of the above-mentioned or existing problems of solar power generation fault alarm systems.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] The embodiment of the present invention provides a fault alarm device for solar power generation, including: an edge monitoring module, which is used to deploy edge computing equipment, and the edge computing equipment runs a neural network model to analyze the collected data and detect preliminary anomalies; a multimodal data fusion module, which is used to collect multimodal data, including electrical parameters, environmental data and visual data, and integrate different types of data using a deep learning fusion model; a fault classification module, which is used to deploy a pre-trained deep learning classification model on a central server to classify feature data uploaded by edge nodes; a trend prediction module, which is used to analyze historical data based on an LSTM time series prediction model, predict future parameter change trends, and generate a health score according to the operating status of the equipment; a dynamic adjustment module, which is used to group and analyze the operating data in combination with a K-means clustering algorithm, and dynamically adjust the alarm threshold according to different weather conditions and equipment status; an alarm response module, which is used to implement hierarchical alarms according to the analysis results of the intelligent algorithm.

[0008] As a preferred solution of the fault alarm system for solar power generation described in the present invention, the deployment of edge computing equipment includes deploying photovoltaic panel output ends, junction boxes and inverter peripheries at the nodes of the photovoltaic array, and establishing a Mesh network between the devices.

[0009] As a preferred solution of the solar power generation fault alarm system of the present invention, the edge computing device runs a neural network model to analyze the collected data and detect preliminary anomalies, including:

[0010] The neural network model inputs the feature vector F, the network weight matrix is ​​W, the activation function is σ, and the output formula of the single-layer neural network is:

[0011] y=σ(W·F+b)

[0012] Among them, W is the weight matrix, b is the bias vector, and σ is the activation function;

[0013] The output vector y of the model is the probability distribution of various types of anomalies:

[0014] y=[P 正常 ,P 过流 ,P 过压 ,P 温度异常 ,P 遮挡 ]

[0015] By setting the threshold θ, we can determine whether there is an abnormality:

[0016]

[0017] As a preferred solution of the solar power generation fault alarm system of the present invention, the fusion model using deep learning to integrate different types of data includes:

[0018] The deep learning fusion model introduces the attention mechanism to dynamically adjust the weights of each data source. Assume that the feature vector is:

[0019] F=[f 电流 ,f 电压 ,f 温度 ,f 光照 ,f 图像 ]

[0020] Use the attention mechanism to calculate the importance weight α of each type of feature:

[0021]

[0022] Among them, α i is the feature f i The weight of i is the feature f i The score, W a is the weight matrix, f i is the representation vector of the i-th feature, b a is the bias vector, exp(e i ) is for e i Index operations;

[0023] The final fusion features are:

[0024]

[0025] Based on this method, key data under different fault types are dynamically captured.

[0026] As a preferred solution of the solar power generation fault alarm system of the present invention, the pre-trained deep learning classification model is deployed on the central server to classify the feature data uploaded by the edge node, including:

[0027] The classifier uses the Softmax function to map the fused feature vector to the probability of each fault category:

[0028]

[0029] Among them, z i is the score of the i-th fault type, n is the total number of all fault types;

[0030] Based on feature recognition of deep learning models, fault types include photovoltaic panel failure, electrical system failure, and environmental failure.

[0031] As a preferred solution of the solar power generation fault alarm system of the present invention, wherein: the pre-trained deep learning classification model is deployed on the central server to classify the feature data uploaded by the edge node, and further includes:

[0032] Input the thermal image, use CNN to extract the texture features of the high-temperature area, compare the extracted texture features with the features of normal samples, calculate the difference score, and when the proportion of high-temperature areas or the temperature difference is abnormal, the classifier outputs a photovoltaic panel hot spot fault;

[0033] Edge detection and texture extraction are performed on the visible light image of the photovoltaic panel. The slender morphology and discontinuous characteristics of the crack are identified through the convolution layer. If the crack feature score is higher than the threshold, it is classified as a photovoltaic panel crack fault.

[0034] As a preferred solution of the solar power generation fault alarm system of the present invention, the LSTM-based time series prediction model analyzes historical data, predicts future parameter change trends, and generates a health score according to the equipment operation status, including:

[0035] The LSTM unit captures the long-term dependency and short-term variation characteristics of the data. The formula is as follows:

[0036] Forget Gate:

[0037] f t =σ(W f ·[h t-1 , x t ]+b f )

[0038] Input Gate:

[0039] i t =σ(W i ·[h t-1 ,x t +b i ),

[0040] Through the trained LSTM model, input the sequence data of a period of time, predict the parameter change trend in the future period, perform differential processing on the predicted value, calculate the parameter change rate, and detect abnormal change trend in combination with historical patterns;

[0041] Based on the equipment's operating status and prediction parameters, a scoring mechanism is used to quantify the equipment's health status. The formula is as follows:

[0042] H=w1·S perf +w2·S stab +w3·S hist

[0043] Among them, Sperf is the performance score, S stab is the stability score, S hist is the historical fault score, w1, w2, w3 are weights.

[0044] As a preferred solution of the solar power generation fault alarm system of the present invention, the scoring mechanism is used to quantify the health status of the equipment, including:

[0045] The performance score is based on the deviation of the predicted current from the rated current and is defined as:

[0046]

[0047] The stability score is calculated based on the standard deviation of the predicted current fluctuations:

[0048]

[0049] The historical failure scores are:

[0050] S hist =100-(historical failure times × penalty factor)

[0051] A comprehensive health score H is generated through the scoring and graded according to thresholds: if the score is between 80-100, it is normal and no maintenance is required; if the score is between 50-80, it is slightly abnormal and requires inspection; if the score is between 0-50, it is seriously abnormal and requires immediate treatment.

[0052] As a preferred solution of the solar power generation fault alarm system of the present invention, the K-means clustering algorithm is used to group and analyze the operation data, and the alarm threshold is dynamically adjusted according to different weather conditions and equipment status, including:

[0053] By minimizing the distance from each data point to its cluster center, the data is divided into k groups. The optimal k is determined according to the demand settings and combined with the elbow rule. According to the clustering results, the data is divided into different groups, including sunny high light group, sunny low light group, cloudy group, equipment high load group and equipment low load group.

[0054] Adjustments are made based on clustering groups. Each cluster group represents a specific operating state. For example, the high-efficiency group on sunny days requires a higher current threshold, while the cloudy group requires a lower power limit. The feature mean and standard deviation within each group are calculated:

[0055]

[0056] Among them, μ i is the mean of the ith feature, and σ is the standard deviation of the ith feature;

[0057] Set the dynamic threshold range based on group characteristics:

[0058] Threshold range = [μ i -k1σ i , μ i +k2σ i ]

[0059] Among them, k1 and k2 are empirical setting factors;

[0060] When new data arrives, determine which cluster group it belongs to, call the dynamic threshold range of the group to update the alarm rules, and detect the device status in real time based on the dynamically adjusted threshold range. If the characteristic value exceeds the dynamic range, an alarm is triggered. If it is close to the boundary value by 10%, it is recorded as a potential risk and an early warning is issued.

[0061] As a preferred solution of the solar power generation fault alarm system of the present invention, wherein:

[0062] The implementation of graded alarms based on the analysis results of the intelligent algorithm includes:

[0063] Use the attention mechanism to assign weights to features and combine multiple abnormal parameters to generate a health score:

[0064]

[0065] Among them, w i is the weight of the i-th feature, V i is the actual value, V normal,i is a normal value;

[0066] Based on the analysis results, the anomalies are divided into the following three categories:

[0067]

[0068] If the fault severity is >30%, a level 1 alarm is triggered; if the fault severity is between 10%-30%, a level 2 alarm is triggered; if the fault severity is <10%, a level 3 alarm is triggered.

[0069] The beneficial effects of the present invention are as follows: the present invention improves the accuracy of fault identification by integrating multi-dimensional features such as photovoltaic panel current, voltage, power, ambient temperature and light intensity through the use of deep learning fusion models. By introducing an attention mechanism and dynamically adjusting the alarm threshold, it effectively distinguishes environmental fluctuations from actual faults, reduces the false alarm rate and missed alarm rate, and can adapt to the complex and changeable photovoltaic system operating environment without manual intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0071] Figure 1 A schematic structural diagram of a fault alarm device for solar power generation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0073] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0074] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0075] Example

[0076] Reference below Figure 1 , which is an embodiment of the present invention.

[0077] S1: Edge monitoring module, used to deploy edge computing devices, which run neural network models to analyze collected data and detect preliminary anomalies.

[0078] Preferably, deploying edge computing equipment includes deploying photovoltaic panel output terminals, junction boxes, and inverter peripheries at nodes of the photovoltaic array, and establishing a Mesh network between the devices.

[0079] Preferably, the neural network model inputs a feature vector F, the network weight matrix is ​​W, the activation function is σ, and the output formula of the single-layer neural network is:

[0080] y=σ(W·F+b)

[0081] Among them, W is the weight matrix, b is the bias vector, and σ is the activation function;

[0082] The output vector y of the model is the probability distribution of various types of anomalies:

[0083] y=[P normal, P overcurrent, P overvoltage, P temperature abnormal, P shielding]

[0084] By setting the threshold θ, we can determine whether there is an abnormality:

[0085]

[0086] Furthermore, F = [Icurrent, Vvoltage, Ttemperature, Girradiance, ...] is a multidimensional vector that contains multiple sensor data of the photovoltaic system. For example, Icurrent is the output current of the photovoltaic module, Vvoltage is the output voltage of the photovoltaic module, Ttemperature is the temperature of the photovoltaic panel or device, and Girradiance is the light intensity. The activation function uses the Sigmoid function or the ReLU function to convert the linear combination into a nonlinear output. The form of the Sigmoid function is:

[0087]

[0088] In this embodiment, the activation function σ is used to perform nonlinear transformation on the linear output of the neural network to ensure the fitting ability of the network.

[0089] Furthermore, a threshold value θ is set to determine whether an abnormality exists. The current state can be determined based on the probability values ​​of each category in the output vector y. The specific implementation steps are as follows:

[0090] For each type of fault (overcurrent, overvoltage, temperature anomaly, shielding, etc.), a threshold θ is set. If the probability of a certain type of fault P i Exceeding the threshold value of this type of fault θ i , then the fault is considered to exist, for example, if P 过流 >θ 过流 , it is determined that the system has an overcurrent fault; if P 过压 >θ 过压 , it is determined that an overvoltage fault has occurred in the system.

[0091] S2: Multimodal data fusion module, used to collect multimodal data, including electrical parameters, environmental data and visual data, and use deep learning fusion models to integrate different types of data.

[0092] Preferably, the deep learning fusion model introduces an attention mechanism to dynamically adjust the weights of each data source. Assume that the feature vector is:

[0093] F=[f 电流 ,f 电压 ,f 温度 ,f光照 ,f 图像 ]

[0094] Use the attention mechanism to calculate the importance weight α of each type of feature:

[0095]

[0096] Among them, α i is the feature f i The weight of i is the feature f i The score, W a is the weight matrix, f i is the representation vector of the i-th feature, b a is the bias vector, exp(e i ) is for e i Index operations;

[0097] The final fusion features are:

[0098]

[0099] Based on this method, key data under different fault types are dynamically captured.

[0100] Furthermore, suppose that a photovoltaic system sensor collects the following data: 电流 =5.1A, f 电压 =310V, f 温度 =45℃, f 光照 =800W / m 2 , fimage = [occluded image features], these feature vectors are input into the neural network, and after the attention mechanism is calculated, the weight of each feature is obtained. For example, after calculation, the following attention weights are obtained:

[0101] α=[0.1,0.15,0.2,0.05,0.5]

[0102] This means that the image feature f 图像 The importance of this fault type is the highest, with a weight of 0.5, followed by temperature and voltage features. Next, by weighted summing up each feature, we get the fusion feature F. 融合 :

[0103] F 融合 =0.1·f 电流 +0.15 f 电压 +0.2 f 温度 +0.05 f 光照 +0.5 f 图像

[0104] The fusion feature is input into the deep neural network for classification, and finally the fault classification result is obtained, for example:

[0105] y=[P 正常 =0.05,P 过流 =0.1,P 过压 =0.02,P 温度异常 =0.02,P 遮挡 =0.8]

[0106] Based on this output, it can be determined that an occlusion fault has occurred in the system.

[0107] S3: Fault classification module, which is used to deploy the pre-trained deep learning classification model on the central server to classify the feature data uploaded by the edge nodes.

[0108] Preferably, a pre-trained deep learning classification model is deployed on the central server to classify the feature data uploaded by the edge nodes, including:

[0109] The classifier uses the Softmax function to map the fused feature vector to the probability of each fault category:

[0110]

[0111] Among them, z i is the score of the i-th fault type, n is the total number of all fault types;

[0112] Based on feature recognition of deep learning models, fault types include photovoltaic panel failure, electrical system failure, and environmental failure.

[0113] Preferably, a thermal image is input, and texture features of the high-temperature area are extracted using CNN. The extracted texture features are compared with features of normal samples, and a difference score is calculated. When the proportion of the high-temperature area or the temperature difference is abnormal, the classifier outputs a photovoltaic panel hot spot fault;

[0114] Edge detection and texture extraction are performed on the visible light image of the photovoltaic panel. The slender morphology and discontinuous characteristics of the crack are identified through the convolution layer. If the crack feature score is higher than the threshold, it is classified as a photovoltaic panel crack fault.

[0115] Furthermore, the input is a preprocessed image of a photovoltaic panel, an image matrix of size H×W, and multiple convolutional layers are used to automatically extract the features of the image. In these convolutional layers, the filter learns how to recognize the elongated morphology, edge features, and texture features of the cracks. A 3×3 or 5×5 convolution kernel can be used for feature extraction. Through pooling operations, the size of the feature map is reduced, the computational efficiency is improved, and the robustness of the features is enhanced.

[0116] By processing the output of the convolutional neural network, the score of the crack feature can be obtained. The higher the score, the greater the possibility of cracks in the image. The score of the crack is obtained by calculating the similarity between the feature vector and the predefined "crack feature vector". For example, using the cosine similarity formula:

[0117]

[0118] Among them, F 图像 is the feature vector extracted by CNN, F 裂纹 is the predefined crack eigenvector;

[0119] If the crack feature score exceeds the preset threshold θ, it is considered that there is a crack in the image. The selection of the threshold can be trained and optimized based on a large number of normal and crack samples to ensure high-precision detection.

[0120] S4: Trend prediction module, which is used to analyze historical data based on the LSTM time series prediction model, predict future parameter change trends, and generate health scores based on the equipment operation status.

[0121] Preferably, the LSTM unit captures the long-term dependency and short-term variation characteristics of the data, and the formula is as follows:

[0122] Forget Gate:

[0123] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0124] Input Gate:

[0125] i t =σ(W i ·[h t-1 , x t ]+b i ),

[0126] Through the trained LSTM model, input the sequence data of a period of time, predict the parameter change trend in the future period, perform differential processing on the predicted value, calculate the parameter change rate, and detect abnormal change trend in combination with historical patterns;

[0127] Based on the equipment's operating status and prediction parameters, a scoring mechanism is used to quantify the equipment's health status. The formula is as follows:

[0128] H=w1·S perf +w2·S stab +w3·S hist

[0129] Among them, S perf is the performance score, S stab is the stability score, S hist is the historical fault score, w1, w2, w3 are weights.

[0130] Preferably, the performance score is based on the deviation of the predicted current from the rated current, defined as:

[0131]

[0132] The stability score is calculated based on the standard deviation of the predicted current fluctuations:

[0133]

[0134] The historical failure scores are:

[0135] S hist =100-(historical failure times × penalty factor)

[0136] A comprehensive health score H is generated through the scoring and graded according to thresholds: if the score is between 80-100, it is normal and no maintenance is required; if the score is between 50-80, it is slightly abnormal and requires inspection; if the score is between 0-50, it is seriously abnormal and requires immediate treatment.

[0137] Furthermore, assuming that the predicted current I pred 9.6A, rated current I rated The current fluctuation standard deviation (σ(Ipred) is 10.0A, the historical fault record is 3 times, and the temperature image analysis identifies the presence of hot spots (temperature anomalies) in the local area of ​​the photovoltaic panel.

[0138] Performance Rating S perf =96, stability score S stab =97, historical fault score S hist =70. The deep learning model dynamically adjusts the weights based on the current operating environment and historical data of the device. For example, the weight of the stability score may be increased to 0.4, and the performance score and historical fault score may be 0.3 respectively.

[0139] H=0.4·97+0.3·96+0.3·70=38.8+28.8+21=88.6

[0140] According to the comprehensive health score H=88.6, the health status of the equipment is healthy and no maintenance is required.

[0141] S5: Dynamic adjustment module, used to combine K-means clustering algorithm to group and analyze the operation data, and dynamically adjust the alarm threshold according to different weather conditions and equipment status.

[0142] Preferably, the data is divided into k groups by minimizing the distance from each data point to the center of its cluster. The optimal k is determined according to the setting requirements and the elbow rule. According to the clustering results, the data is divided into different groups, including a sunny high light group, a sunny low light group, a cloudy group, a device high load group, and a device low load group.

[0143] Adjustments are made based on clustering groups. Each cluster group represents a specific operating state. For example, the high-efficiency group on sunny days requires a higher current threshold, while the cloudy group requires a lower power limit. The feature mean and standard deviation within each group are calculated.

[0144]

[0145] Among them, μ i is the mean of the ith feature, and σ is the standard deviation of the ith feature;

[0146] Set the dynamic threshold range based on group characteristics:

[0147] Threshold range = [μ i -k1σ i , μ i +k2σ i ]

[0148] Among them, k1 and k2 are empirical setting factors;

[0149] When new data arrives, determine which cluster group it belongs to, call the dynamic threshold range of the group to update the alarm rules, and detect the device status in real time based on the dynamically adjusted threshold range. If the characteristic value exceeds the dynamic range, an alarm is triggered. If it is close to the boundary value by 10%, it is recorded as a potential risk and an early warning is issued.

[0150] Furthermore, the following operating data were collected: current 8.5A, voltage 220V, light intensity 900W / m 2 , the device load is 70%, and after K-means clustering, the device is assigned to the sunny high light group. In this group, the current and voltage thresholds are set as follows: the upper current limit is 9.2A (mean + 1.5 times standard deviation), the lower current limit is 7.8A (mean - 1.5 times standard deviation), the upper voltage limit is 230V (mean + 1.5 times standard deviation), and the lower voltage limit is 210V (mean - 1.5 times standard deviation).

[0151] The current is 8.5A and the voltage is 220V. According to the dynamic range, both the current and voltage are within the normal range, so the device is in normal condition. Assume that the next data point is current 9.0A, voltage 230V, and light intensity 950W / m 2At this time, both the current and voltage are close to the upper limit (within 10%), so the system will issue a potential risk warning to remind the operator to check.

[0152] If the new data point is a current of 9.5 A (exceeding the upper limit), the system will trigger a fault alarm to notify the maintenance personnel to handle the possible overcurrent problem.

[0153] S6: Alarm response module, used to implement hierarchical alarms according to the analysis results of the intelligent algorithm.

[0154] Preferably, the attention mechanism is used to assign weights to features, and multiple abnormal parameters are integrated to generate a health score:

[0155]

[0156] where w i is the weight of the i-th feature, V i is the actual value, and V normal,i is the normal value;

[0157] According to the analysis results, the anomalies are classified into the following three categories:

[0158]

[0159] If the fault severity > 30%, trigger a first-level alarm; if the fault severity is between 10% - 30%, trigger a second-level alarm; if the fault severity < 10%, trigger a third-level alarm.

[0160] Furthermore, multiply the anomaly degree of each feature by the corresponding weight w i to calculate the weighted health score S:

[0161]

[0162] where w i is the weight of each feature, ∣V i -V normal ,i is the anomaly degree of the feature V i ;

[0163] According to the value of the health score S, the equipment faults are divided into three levels:

[0164] First-level alarm (severe fault): When the health score S > 30%, the equipment is in a severe fault state, trigger a first-level alarm, and immediate handling is required;

[0165] Second-level alarm (minor fault): When the health score 10% < S ≤ 30%, the equipment has a minor fault, inspection and handling are required, and trigger a second-level alarm;

[0166] Level 3 alarm (potential risk): When the health score S ≤ 10%, the device is in normal condition, but there is a potential risk, triggering a level 3 alarm for early warning.

[0167] The following data was collected:

[0168] Current V1 = 8.5A, normal value V normal,1 = = 9.0A, voltage V2 = 220V, normal value V normal,2 =225V, temperature V3 = 70℃, normal value V normal,3 =65℃, light intensity V4=800W / m 2 , normal value V normal,4 =1000W / m 2 , assuming that the weight calculated by the attention mechanism is:

[0169] w1=0.2, w2=0.3, w3=0.25, w4=0.25, then the health score S=8.7%. The system determines that the status of the equipment is a level 3 alarm (potential risk) and issues an early warning to remind the operator to check.

[0170] In the description of the present invention, it should be noted that the terms “first”, “second” and “third” are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0171] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0172] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0173] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0174] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0175] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

[0176] In addition, although the operations of the method of the present invention are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

Claims

1. A fault alarm device for solar power generation, characterized in that: include: The edge monitoring module is used to deploy edge computing devices, which run neural network models to analyze collected data and detect preliminary anomalies; Multimodal data fusion module, which is used to collect multimodal data, including electrical parameters, environmental data, and visual data, and integrate different types of data using a deep learning fusion model; The fault classification module is used to deploy a pre-trained deep learning classification model on the central server to classify the feature data uploaded by the edge nodes; The trend prediction module is used to analyze historical data based on the LSTM time series prediction model, predict future parameter change trends, and generate health scores based on the equipment operation status; Dynamic adjustment module, which is used to group and analyze the operation data in combination with the K-means clustering algorithm, and dynamically adjust the alarm threshold according to different weather conditions and equipment status; The alarm response module is used to implement graded alarms based on the analysis results of the intelligent algorithm.

2. The solar power generation fault alarm system according to claim 1, characterized in that: The deployment of edge computing equipment includes deploying photovoltaic panel output terminals, junction boxes, and inverter peripheries at the nodes of the photovoltaic array, and establishing a Mesh network between the devices.

3. The solar power generation fault alarm system according to claim 1, characterized in that: The edge computing device runs a neural network model to analyze the collected data and detect preliminary anomalies, including: The neural network model inputs the feature vector F, the network weight matrix is ​​W, the activation function is σ, and the output formula of the single-layer neural network is: y=σ(W·F+b) Among them, W is the weight matrix, b is the bias vector, and σ is the activation function; The output vector y of the model is the probability distribution of various types of anomalies: y=[P 正常 ,P 过流 ,P 过压 ,P 温度异常 ,P 遮挡 ] By setting the threshold θ, we can determine whether there is an abnormality:

4. The solar power generation fault alarm system according to claim 1, characterized in that: The fusion model using deep learning integrates different types of data, including: The deep learning fusion model introduces the attention mechanism to dynamically adjust the weights of each data source. Assume that the feature vector is: F=[f 电流 ,f 电压 ,f 温度 ,f 光照 ,f 图像 ] Use the attention mechanism to calculate the importance weight α of each type of feature: Among them, α i is the feature f i The weight of i is the feature f i The score, W a is the weight matrix, f i is the representation vector of the i-th feature, b a is the bias vector, exp(e i ) is for e i Index operations; The final fusion features are: Based on this method, key data under different fault types are dynamically captured.

5. The solar power generation fault alarm system according to claim 1, characterized in that: The pre-trained deep learning classification model is deployed on the central server to classify the feature data uploaded by the edge node, including: The classifier uses the Softmax function to map the fused feature vector to the probability of each fault category: Among them, z i is the score of the i-th fault type, n is the total number of all fault types; Based on feature recognition of deep learning models, fault types include photovoltaic panel failure, electrical system failure, and environmental failure.

6. The solar power generation fault alarm system according to claim 1, characterized in that: The method of deploying a pre-trained deep learning classification model on the central server to classify the feature data uploaded by the edge node also includes: Input the thermal image, use CNN to extract the texture features of the high-temperature area, compare the extracted texture features with the features of normal samples, calculate the difference score, and when the proportion of high-temperature areas or the temperature difference is abnormal, the classifier outputs a photovoltaic panel hot spot fault; Edge detection and texture extraction are performed on the visible light image of the photovoltaic panel. The slender morphology and discontinuous characteristics of the crack are identified through the convolution layer. If the crack feature score is higher than the threshold, it is classified as a photovoltaic panel crack fault.

7. The solar power generation fault alarm system according to claim 1, characterized in that: The LSTM-based time series prediction model analyzes historical data, predicts future parameter change trends, and generates a health score based on the equipment operating status, including: The LSTM unit captures the long-term dependency and short-term variation characteristics of the data. The formula is as follows: Forget 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 ), Through the trained LSTM model, input the sequence data of a period of time, predict the parameter change trend in the future period, perform differential processing on the predicted value, calculate the parameter change rate, and detect abnormal change trend in combination with historical patterns; Based on the equipment's operating status and prediction parameters, a scoring mechanism is used to quantify the equipment's health status. The formula is as follows: H=w1·S perf +w2·S stab + w 3·S hist Among them, S pref is the performance score, S stab is the stability score, S hist is the historical fault score, w1, w2, w3 are weights.

8. The solar power generation fault alarm system according to claim 7, characterized in that: The scoring mechanism used to quantify the health status of the equipment includes: The performance score is based on the deviation of the predicted current from the rated current and is defined as: The stability score is calculated based on the standard deviation of the predicted current fluctuations: The historical failure scores are: S hist =100-(historical failure times × penalty factor) A comprehensive health score H is generated through the scoring and graded according to thresholds: if the score is between 80-100, it is normal and no maintenance is required; if the score is between 50-80, it is slightly abnormal and requires inspection; if the score is between 0-50, it is seriously abnormal and requires immediate treatment.

9. The solar power generation fault alarm system according to claim 1, characterized in that: The K-means clustering algorithm is used to group and analyze the operation data, and the alarm threshold is dynamically adjusted according to different weather conditions and equipment status, including: By minimizing the distance from each data point to its cluster center, the data is divided into k groups. The optimal k is determined according to the demand settings and combined with the elbow rule. According to the clustering results, the data is divided into different groups, including sunny high light group, sunny low light group, cloudy group, equipment high load group and equipment low load group. Adjustments are made based on clustering groups. Each cluster group represents a specific operating state. For example, the high-efficiency group on sunny days requires a higher current threshold, while the cloudy group requires a lower power limit. The feature mean and standard deviation within each group are calculated: Among them, μi is the mean of the i-th feature, σ is the standard deviation of the i-th feature; Set the dynamic threshold range based on group characteristics: threshold range=[m i -k1σ i ,m i +k2σ i ] Among them, k1 and k2 are empirical setting factors; When new data arrives, determine which cluster group it belongs to, call the dynamic threshold range of the group to update the alarm rules, and detect the device status in real time based on the dynamically adjusted threshold range. If the characteristic value exceeds the dynamic range, an alarm is triggered. If it is close to the boundary value by 10%, it is recorded as a potential risk and an early warning is issued.

10. The solar power generation fault alarm system according to claim 1, characterized in that: The implementation of graded alarms based on the analysis results of the intelligent algorithm includes: Use the attention mechanism to assign weights to features and combine multiple abnormal parameters to generate a health score: Among them, w i is the weight of the i-th feature, V i is the actual value, V normal，i is a normal value; Based on the analysis results, the anomalies are divided into the following three categories: If the fault severity is >30%, a level 1 alarm is triggered; if the fault severity is between 10% and 30%, a level 2 alarm is triggered; if the fault severity is <10%, a level 3 alarm is triggered.

Images