An intelligent state monitoring method for an ultrahigh-altitude light storage power station based on deep learning

Abstract

The application discloses an intelligent state monitoring method for an ultrahigh-altitude light storage power station based on deep learning, which comprises the following steps: collecting multi-source heterogeneous operation data of different types of light storage power stations, and performing data preprocessing; inputting the processed data into a state monitoring submodel based on deep learning which is built, and training and updating parameters of the model, so as to obtain the state monitoring submodel which can quickly and accurately detect the operation state of the light storage power station, and then the state monitoring submodel is used for light storage power station photovoltaic module fault identification, angle monitoring, personnel behavior and animal detection, etc., which can not only realize fault early warning of the light storage power station, prevent serious accidents of the light storage power station caused by faults, but also reduce the maintenance cost of the light storage power station, overcome the adverse environmental factors such as high altitude and low temperature, greatly reduce the huge manpower and material resources expenditure of traditional photovoltaic power plant inspection, and meet the demand of non-site supervision in actual application of the photovoltaic power plant.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent monitoring of the operating state of an ultra-high-altitude light storage power station, and more specifically relates to an intelligent state monitoring method for an ultra-high-altitude light storage power station based on deep learning. BACKGROUND

[0002] In recent years, with the rapid development of the economy, China has begun to implement the "double carbon" policy, gradually reduce the dependence on traditional fossil energy, and accelerate the development of new energy. As one of the new energies, solar energy has been rapidly developed and utilized due to its advantages of being clean and renewable. With the rapid improvement of photovoltaic power generation technology and the gradual reduction of the cost of photovoltaic cell panels, the construction scale of light storage power stations has significantly expanded. Since light storage power stations are usually built in relatively remote and harsh environments, most of the working components of light storage power stations, such as cell panels and inverters, work in outdoor environments and need to withstand the erosion of rain, snow, and low temperature, which can cause the working characteristics of light storage power stations to deteriorate, and even cause frequent failures of light storage power stations, affecting the economic benefits of photovoltaic power generation.

[0003] In order to ensure the normal operation of light storage power station equipment, light storage power stations need to arrange personnel to regularly inspect the power station equipment, and even some key equipment needs to be inspected more than once a day. With the continuous expansion of the scale of light storage power stations, the operation and maintenance cost of light storage power stations has significantly increased, and traditional regular manual maintenance cannot meet the rapid development of light storage power stations.

[0004] With the rapid development of deep learning, outstanding achievements have been made in many fields. As a key task of deep learning, target detection has made good progress in many fields such as inspection, security, medicine, and transportation. In order to reduce the operation and maintenance cost and improve the fault recognition rate, the deep learning target detection algorithm is applied to the intelligent inspection of light storage power station equipment, environment, and personnel intelligent monitoring, and a multi-target detection model that meets the demand of light storage power station intelligent inspection is established. This not only can realize the fault warning of light storage power stations to prevent serious accidents caused by faults, but also can reduce the maintenance cost of light storage power stations. In addition, the application of advanced target detection algorithms is of great significance to the intelligent development, safety, and economic improvement of light storage power stations. It not only can improve the intelligent level of light storage power station operation and maintenance, but also can further accelerate the pace of new energy development. SUMMARY

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a deep learning-based intelligent status monitoring method for ultra-high altitude photovoltaic and energy storage power stations. By establishing a multi-target detection model that meets the intelligent inspection needs of photovoltaic and energy storage power stations, advanced artificial intelligence algorithms are used to identify photovoltaic module faults and detect abnormal human behavior and animals to determine whether they are normal. This method can not only provide early warning of faults in photovoltaic and energy storage power stations and prevent serious accidents caused by faults, but also reduce the maintenance costs of photovoltaic and energy storage power stations. It overcomes harsh environmental factors such as high altitude and low temperature, greatly reduces the huge manpower and material costs of traditional photovoltaic power plant inspections, and meets the needs of off-site supervision in practical photovoltaic applications.

[0006] To achieve the above-mentioned objectives, the present invention provides a deep learning-based intelligent condition monitoring method for ultra-high altitude photovoltaic-storage power stations, characterized by comprising the following steps:

[0007] (1) Acquisition of multi-source heterogeneous data

[0008] The system collects voltage and current data, photovoltaic panel installation tilt angle data, and indoor temperature and humidity data of the photovoltaic power station in real time through sensors. It also collects image data and video signals of the photovoltaic power station in real time through drone photography, high-definition cameras, and infrared cameras. The collected multi-source heterogeneous data is used to train the intelligent status monitoring model of the photovoltaic power station.

[0009] (2) Preprocessing of multi-source heterogeneous data

[0010] (2.1) Add corresponding tags to the photovoltaic and energy storage power station operation data under different states: voltage and current data are marked as 0 if normal and 1 if abnormal; tilt angle data are marked as 0 if normal and 1 if abnormal; temperature and humidity data are marked as 0 if normal and 1 if abnormal; photovoltaic and energy storage power station images are marked as 0 if normal and 1 if abnormal; photovoltaic and energy storage power station video signals are marked as 1 if normal and 1 if abnormal; abnormal photovoltaic and energy storage power station images indicate photovoltaic module failure; abnormal photovoltaic and energy storage power station video signals indicate abnormal human behavior or detection of animals.

[0011] (2.2) The labeled sensor data is randomly divided into N equal-length subsequence datasets. The subsequence datasets are divided into training sets and test sets according to the proportion for training and testing of the corresponding state monitoring sub-model. The labeled image dataset and video signal dataset are also divided into training sets and test sets according to the proportion for training and testing of the corresponding state monitoring sub-model.

[0012] (3) Build voltage and current state monitoring sub-model, tilt angle state monitoring sub-model, temperature, humidity data state monitoring sub-model, light storage power station image state monitoring sub-model and light storage power station video signal state monitoring sub-model according to demand, each state monitoring sub-model is a convolutional neural network, and each state monitoring sub-model is trained by using a corresponding training set:

[0013] (3.1) Set the learning rate α of the convolutional neural network, and set the total number of iterations T;

[0014] (3.2) Randomly select a batch of data in the training set, input into the state monitoring sub-model for training, and the state monitoring sub-model uses a three-layer convolutional neural network (CNN) to extract hidden features of the data, and the extraction formula is:

[0015] X i+1 =ReLU(W i X i +b i )

[0016] Wherein, X i+1 is the output after the data is processed by the i-th CNN, X i is the input data, ReLU() represents an activation function, W i represents the weight matrix of the i-th CNN, and b i is the bias vector of the i-th CNN;

[0017] (3.3) The hidden features extracted by the CNN are input into the full connection layer for data dimension reduction processing, so as to facilitate the final classification of the monitoring model, and the formula is:

[0018] f dense =ReLU(dense(X i+1 ))

[0019] Wherein, dense() is a tf.layers.dense() function based on tensorflow;

[0020] (3.4) After the data features extracted by the CNN are processed by the dimension reduction of the full connection layer, the cross-entropy loss function of the optimized state monitoring sub-model is calculated according to the output f dense of the full connection layer:

[0021] f(y,p)=-(ylog(p)+(1-y)log(1-p))

[0022] Wherein, p represents the probability that the current belongs to the corresponding label y;

[0023] (3.5), the loss function value f(y, p) of the current round is compared with the loss function value of the previous round, if the loss function value of the current round is less than the previous round, the weight and bias of the CNN are directly optimized by using gradient descent method:

[0024]

[0025]

[0026] Wherein E represents error, if the loss function value of the current round is greater than the previous round, the weight and bias of the CNN are optimized by using AdamOptimizer optimizer, the specific formula of AdamOptimizer optimizer is:

[0027]

[0028]

[0029]

[0030] Wherein, The gradient of the loss function is m t The exponential moving average is h t The square gradient is β1, β2 is a constant, β1, β2∈[0, 1]; ε is a constant, default is 1e-8, θ t Indicates the updated model parameters, including the updated weight and bias of the CNN;

[0031] (3.6), judge whether the number of current training iteration rounds reaches the maximum iteration number T, if it is satisfied, the iteration stops, the updated weight and bias are taken as the parameters of the CNN, and then step (4) is entered; Otherwise, the current iteration number is added by 1, and then step (3.2) is returned to continue training the next batch of data selected from the training set;

[0032] (4), the data corresponding to the test set is used to evaluate each state monitoring sub-model trained in step (3), if the set accuracy is reached, step (5) is entered, otherwise, step (3.2) is returned to continue training;

[0033] (5) Real-time acquisition of voltage and current data of photovoltaic power station, tilt angle data of photovoltaic panel installation, indoor temperature and humidity data of photovoltaic power station, image data and video signal, and then sending them to the corresponding voltage and current status monitoring sub-model, tilt angle status monitoring sub-model, temperature and humidity data status monitoring sub-model, photovoltaic power station image status monitoring sub-model and photovoltaic power station video signal status monitoring sub-model to obtain the corresponding status. If any status monitoring sub-model outputs 1, a warning is issued, thereby realizing intelligent status monitoring of ultra-high altitude photovoltaic power station.

[0034] The objective of this invention is achieved as follows:

[0035] This invention presents a deep learning-based intelligent condition monitoring method for ultra-high altitude photovoltaic-storage power stations. It collects multi-source heterogeneous operational data from different types of photovoltaic-storage power stations, preprocesses the data, and then inputs the processed data into a pre-built deep learning-based condition monitoring sub-model. This model is then trained and its parameters are updated to obtain a condition monitoring sub-model that can quickly and accurately detect the operating status of the photovoltaic-storage power station. This sub-model can then be used for photovoltaic module fault identification, angle monitoring, and detection of human behavior and animals in photovoltaic-storage power stations. It not only enables early warning of faults in photovoltaic-storage power stations, preventing serious accidents caused by faults, but also reduces maintenance costs. It overcomes harsh environmental factors such as high altitude and low temperatures, significantly reducing the huge manpower and material costs of traditional photovoltaic power plant inspections, and meeting the needs of off-site supervision in practical applications of photovoltaic power plants.

[0036] Meanwhile, the intelligent condition monitoring method for ultra-high altitude photovoltaic-storage power stations based on deep learning in this invention also has the following beneficial effects:

[0037] (1) By using artificial intelligence technology, a deep learning-based method is adopted to extract features and identify the state of multi-source heterogeneous data of photovoltaic power station, which reduces the impact of human intervention in traditional data feature extraction methods, reduces the huge manpower and material expenditure of traditional photovoltaic power plant inspection, and improves the efficiency of power station inspection.

[0038] (2) Compared with traditional single-state monitoring, the integrated model in this invention can simultaneously complete the state monitoring of photovoltaic and energy storage power stations under multiple scenarios, making full use of various operating data of photovoltaic and energy storage power stations and improving the data utilization rate within the station.

[0039] (3) The integrated model of the present invention can be extended and migrated according to needs, ensuring the possibility of practical application of the model. Attached Figure Description

[0040] Figure 1 This is a flowchart of the intelligent condition monitoring method for ultra-high altitude photovoltaic and energy storage power stations based on deep learning, according to the present invention.

[0041] Figure 2 are schematic diagrams of various common fault types of photovoltaic modules, wherein (a) is shadow shading, (b) is hot spot failure, (c) is support failure, and (d) is module breakage;

[0042] Figure 3 is the overall technical framework of the present application. DETAILED DESCRIPTION

[0043] The specific embodiments of the present application will be described below with reference to the accompanying drawings, so that those skilled in the art can better understand the present application. It should be particularly noted that in the following description, when detailed description of known functions and designs may obscure the main content of the present application, these descriptions will be omitted here.

[0044] EMBODIMENT

[0045] Figure 1 is a flowchart of the intelligent state monitoring method for ultra-high altitude light storage power stations based on deep learning of the present application.

[0046] In this embodiment, as shown in Figure 1 the intelligent state monitoring method for ultra-high altitude light storage power stations based on deep learning of the present application, characterized in that it comprises the following steps:

[0047] Step S1: Collection of multi-source heterogeneous data

[0048] Real-time collection of voltage and current data of the light storage power station through sensors, collection of installation inclination angle data of the photovoltaic panel, collection of indoor temperature and humidity data of the light storage power station, real-time collection of image data and video signals of the light storage power station through unmanned aerial vehicle shooting, high-definition camera and infrared camera, and the collected multi-source heterogeneous data are used for training of the intelligent state monitoring model of the light storage power station.

[0049] Step S2: Preprocessing of multi-source heterogeneous data

[0050] Step S2.1: Adding corresponding labels to the light storage power station operation data

[0051] Adding corresponding labels to the light storage power station operation data under different states: for voltage and current data, normal is marked as 0 and abnormal state is marked as 1; for inclination angle data, normal is marked as 0 and abnormal is marked as 1; for temperature and humidity data, normal is marked as 0 and abnormal state is marked as 1; for light storage power station image, normal is marked as 0 and abnormal state is marked as 1; for light storage power station video signal, normal is marked as 1 and abnormal state is marked as 1; light storage power station image abnormality is photovoltaic module failure, and light storage power station video signal abnormality is personnel behavior abnormality or detection of animals.

[0052] Step S2.2: Dividing training set and test set

[0053] The labeled sensor data is randomly divided into N equal-length subsequence data sets, and the subsequence data sets are proportionally divided into training sets and test sets for training and testing of the corresponding state monitoring sub-models. The labeled image data set and the video signal data set are also proportionally divided into training sets and test sets for training and testing of the corresponding state monitoring sub-models.

[0054] Step S3: building and training multiple corresponding state monitoring sub-models

[0055] The voltage and current state monitoring sub-model, the inclination angle state monitoring sub-model, the temperature and humidity data state monitoring sub-model, the image state monitoring sub-model of the optical storage power station, and the video signal state monitoring sub-model of the optical storage power station are built according to requirements, and each state monitoring sub-model is a convolutional neural network. Each state monitoring sub-model is trained using the corresponding training set.

[0056] Step S3.1: setting the learning rate of the convolutional neural network, the total number of iterations, and initializing

[0057] The learning rate of the convolutional neural network is set to a = 0.0001, the total number of iterations T is set to 10000, and the weights and biases of the convolutional neural network CNN are initialized.

[0058] Step S3.2: extracting hidden features

[0059] A batch of data is randomly selected from the training set and input into the state monitoring sub-model for training. The state monitoring sub-model extracts hidden features of the data using a three-layer convolutional neural network (CNN). The extraction formula is:

[0060] X i+1 =ReLU(W i X i +b i )

[0061] Where X i+1 is the output of the data after being processed by the i-th CNN, X i is the input data, ReLU() represents the activation function, W i represents the weight matrix of the i-th CNN, b i is the bias vector of the i-th CNN, and the output dimension of the last layer of CNN is 512.

[0062] Step S3.3: inputting hidden features to a fully connected layer for data dimension reduction

[0063] The hidden features extracted by the CNN are input to a fully connected layer for data dimension reduction, which facilitates the final classification of the monitoring model. The formula is:

[0064] f dense = ReLU(dense(X i+1 ))

[0065] where dense() is the tf.layers.dense() function based on tensorflow.

[0066] Step S3.4: Calculate the cross-entropy loss function of the optimization state monitoring sub-model

[0067] After the data features extracted by the CNN are processed by the dimension reduction of the fully connected layer, the cross-entropy loss function of the optimization state monitoring sub-model is calculated according to the output f dense of the fully connected layer.

[0068] f(y, p) = -(ylog(p) + (1-y)log(1-p))

[0069] where p represents the probability that the current belongs to the corresponding label y.

[0070] Step S3.5: If the cross-entropy loss function is smaller than the previous round, optimize the weights and biases of the CNN

[0071] The cross-entropy loss function value f(y, p) of the current round is compared with the loss function value of the previous round. If the loss function value of the current round is smaller than the previous round, the weights and biases of the CNN are directly optimized by using the gradient descent method:

[0072]

[0073]

[0074] where E represents the error. If the loss function value of the current round is greater than the previous round, the weights and biases of the CNN are optimized by using the AdamOptimizer optimizer. The specific formula of the AdamOptimizer optimizer is:

[0075]

[0076]

[0077]

[0078] where is the gradient of the loss function, m t is the exponential moving average, h t is the square gradient, β1 and β2 are constants, β1 and β2 ∈ [0, 1]; ε is a constant, default is 1e-8, θ tdenote the updated model parameters, including the updated weights and biases of the CNN.

[0079] Step S3.6: determining whether the maximum number of iterations is reached

[0080] It is determined whether the current number of training iteration rounds reaches the maximum number of iterations T, if yes, the iteration is stopped, the updated weights and biases are taken as the parameters of the CNN, and then step S4 is entered; otherwise, the current number of iterations is increased by 1, and step S3.2 is returned to continue training on the next batch of data selected from the training set.

[0081] Step S4: determining whether the set accuracy is reached

[0082] The state monitoring sub-models trained in step S3 are evaluated using the data of the corresponding test set, if the set accuracy is reached, step S5 is entered, otherwise, step S3.2 is returned to continue training.

[0083] Step S4: real-time collection of the data of the light storage power station and inputting the data into the corresponding state monitoring sub-model for state monitoring

[0084] The voltage and current data of the light storage power station, the installation inclination angle data of the photovoltaic panel, the indoor temperature and humidity data of the light storage power station, the image data and the video signal are collected, and then are respectively sent into the corresponding voltage and current state monitoring sub-model, the inclination angle state monitoring sub-model, the temperature and humidity data state monitoring sub-model, the light storage power station image state monitoring sub-model and the light storage power station video signal state monitoring sub-model, to obtain the corresponding state, if any one of the state monitoring sub-models outputs 1, a warning is issued, so that the intelligent state monitoring of the light storage power station at an ultra-high altitude is realized.

[0085] In this embodiment, the data collected by the light storage power station is tested, Figure 2 is a schematic diagram of various common photovoltaic module fault types, including four states of shadow shielding, hot spot failure, support failure and module damage, and the module damage includes photovoltaic panel cracks and damage.

[0086] The application is built under the windows 10 operating system, a state monitoring integrated model based on CNN is built in the Pycharm3.6 (64-bit) software by using the programming framework based on Tensorflow, and the training process of the model is accelerated by GPU; the results show that the method can well realize the identification and monitoring of various abnormal conditions of the photovoltaic module.

[0087] Compared with other methods, the application can solve the problem of difficult inspection and fault discovery of new energy light storage power station, can timely investigate problem points, effectively improve utilization hours by about 5%, can effectively improve the inspection efficiency of new energy outdoor equipment, improve the efficiency by more than 20%, can reduce the loss of power caused by personnel monitoring missing faults, and reduce the loss of power by about 5%.

[0088] Figure 3 The whole technical framework of the application includes a data acquisition layer, a data layer and an intelligent state perception layer, from data acquisition, through the acquired multi-source heterogeneous data, light storage power station situation awareness is carried out, finally, multi-target remote state monitoring and early warning of the light storage power station is realized, the personnel in the station can formulate corresponding inspection strategy according to the monitoring result, unnecessary inspection cost is saved, and the scientificity, logicality and feasibility of the method of the application are fully proved.

[0089] Although the above describes the specific embodiments of the application in order to facilitate the understanding of the application by the person skilled in the art, it should be clear that the application is not limited to the scope of the specific embodiments, and for the person skilled in the art, all kinds of changes are obvious within the spirit and scope of the application defined and determined by the appended claims, and all kinds of application and creation using the concept of the application are included in the protection.

Claims

1. A method for intelligent condition monitoring of ultra-high altitude photovoltaic-storage power stations based on deep learning, characterized in that, Includes the following steps: (1) Acquisition of multi-source heterogeneous data The system collects voltage and current data, photovoltaic panel installation tilt angle data, and indoor temperature and humidity data of the photovoltaic power station in real time through sensors. It also collects image data and video signals of the photovoltaic power station in real time through drone photography, high-definition cameras, and infrared cameras. The collected multi-source heterogeneous data is used to train the intelligent status monitoring model of the photovoltaic power station. (2) Preprocessing of multi-source heterogeneous data (2.1) Add corresponding tags to the photovoltaic and energy storage power station operation data under different states: voltage and current data are marked as 0 if normal and 1 if abnormal; tilt angle data are marked as 0 if normal and 1 if abnormal; temperature and humidity data are marked as 0 if normal and 1 if abnormal; photovoltaic and energy storage power station images are marked as 0 if normal and 1 if abnormal; photovoltaic and energy storage power station video signals are marked as 1 if normal and 1 if abnormal; abnormal photovoltaic and energy storage power station images indicate photovoltaic module failure; abnormal photovoltaic and energy storage power station video signals indicate abnormal human behavior or detection of animals. (2.2) The labeled sensor data is randomly divided into N equal-length subsequence datasets. The subsequence datasets are divided into training sets and test sets according to the proportion for training and testing of the corresponding state monitoring sub-model. The labeled image dataset and video signal dataset are also divided into training sets and test sets according to the proportion for training and testing of the corresponding state monitoring sub-model. (3) Build voltage and current status monitoring sub-models, tilt angle status monitoring sub-models, temperature and humidity data status monitoring sub-models, photovoltaic-storage power station image status monitoring sub-models, and photovoltaic-storage power station video signal status monitoring sub-models according to requirements. Each status monitoring sub-model is a convolutional neural network, and each status monitoring sub-model is trained using the corresponding training set. (3.1) Set the learning rate α of the convolutional neural network and the total number of iterations T; (3.2) Randomly select a batch of data from the training set and input it into the state monitoring sub-model for training. The state monitoring sub-model uses a three-layer convolutional neural network (CNN) to extract the hidden features of the data. The extraction formula is as follows: X i+1 =ReLU(W i X i +b i ) Among them, X i+1 X is the output after the data has been processed by the i-th CNN layer. i For the input data, ReLU() represents the activation function, W i Let b represent the weight matrix of the i-th CNN layer. i Let be the bias vector of the i-th layer of the CNN; (3.3) The hidden features extracted by the CNN are input into the fully connected layer for data dimensionality reduction, which facilitates monitoring the final classification of the model. The formula is as follows: f dense <ReLU(dense(X i+1 )) Among them, `dense()` is based on the `tf.layers.dense()` function in TensorFlow; (3.4) After the data features extracted by CNN are processed by the dimensionality reduction of the fully connected layer, the output f of the fully connected layer is used to further refine the data. dense Calculate the cross-entropy loss function for the optimized state monitoring sub-model: f(y,p)=-(ylog(p)+(1-y)log(1-p)) Where p represents the probability that the current label is y; (3.5) Compare the loss function value f(y,p) of the current round with the loss function value of the previous round. If the loss function value of the current round is less than that of the previous round, then use gradient descent to directly optimize the weights and biases of the CNN: Where E represents the error, if the loss function value of the current round is greater than that of the previous round, the AdamOptimizer optimizes the weights and biases of the CNN. The specific formula for the AdamOptimizer optimization is as follows: in, Let m be the gradient of the loss function. t h is the exponential moving average. t The gradient is squared, β1 and β2 are constants, β1 and β2 ∈ [0,1]; ε is a constant, defaulting to 1e-8, θ t This represents the updated model parameters, including the updated weights and biases of the CNN. (3.6) Determine whether the current training iteration rounds have reached the maximum iteration count T. If so, stop the iteration, use the updated weights and biases as the parameters of the CNN, and then proceed to step (4); otherwise, increment the current iteration count by 1, return to step (3.2), and continue to select the next batch of data in the training set for training. (4) Use the data of the corresponding test set to evaluate each state monitoring sub-model trained in step (3). If the set accuracy is achieved, proceed to step (5); otherwise, return to step (3.2) to continue training. (5) Real-time acquisition of voltage and current data of photovoltaic power station, tilt angle data of photovoltaic panel installation, indoor temperature and humidity data of photovoltaic power station, image data and video signal, and then sending them to the corresponding voltage and current status monitoring sub-model, tilt angle status monitoring sub-model, temperature and humidity data status monitoring sub-model, photovoltaic power station image status monitoring sub-model and photovoltaic power station video signal status monitoring sub-model to obtain the corresponding status. If any status monitoring sub-model outputs 1, a warning is issued, thereby realizing intelligent status monitoring of ultra-high altitude photovoltaic power station.

2. The intelligent condition monitoring method for ultra-high altitude photovoltaic-storage power stations based on deep learning according to claim 1, characterized in that, The photovoltaic module failures mentioned above include four states: shading, hot spot failure, support failure, and module damage.

Images