An intelligent monitoring system and method for implementing dynamic performance adjustment for a distributed photovoltaic network

Abstract

The application discloses an intelligent monitoring system and method for realizing dynamic performance adjustment of a distributed photovoltaic network, and relates to the technical field of smart grids.The system comprises a data acquisition module, which is used for collecting image data of photovoltaic panels and operation data of photovoltaic power stations in real time; a task scheduling module, which is used for coordinating dynamic transmission of the image data and the operation data between edge computing nodes and the cloud; a data processing module, which is used for pre-processing and feature extraction of the image data and the operation data; a dynamic estimation module, which is used for inputting the extracted features into a sparse regression model and predicting the dynamic state of the photovoltaic network by using an unscented Kalman filter; and a monitoring feedback module, which is used for remotely and dynamically adjusting the angle or position of the photovoltaic panels, starting and stopping equipment and / or adjusting parameters according to the prediction result of the dynamic state.The application can not only significantly improve the overall operation efficiency of the photovoltaic network, but also reduce operation and maintenance costs and prolong the service life of the equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart grid, in particular to an intelligent monitoring system and method for realizing dynamic performance adjustment of distributed photovoltaic network. BACKGROUND

[0002] Distributed photovoltaic network is an important part of modern renewable energy solutions, mainly refers to the photovoltaic power generation system constructed in the user's near end, such as roof, community, factory, etc., which uses solar energy as energy source, reduces the loss in the process of power transmission, and improves the directness and efficiency of energy use.

[0003] At present, photovoltaic power generation network has been widely applied due to its characteristics of clean and renewable. However, with the expansion of the scale of distributed photovoltaic power station, the monitoring and maintenance of photovoltaic network become more and more complex and arduous. The traditional photovoltaic monitoring system mainly relies on a single central node for data collection and processing, which not only increases the burden of the central server, but also easily leads to delay problem, especially in the scenario of large-scale network, the efficiency of data processing is difficult to guarantee. At the same time, with the dynamic state change of photovoltaic panel, the traditional monitoring system cannot quickly respond and adjust the running state of the equipment, which leads to the dynamic performance of photovoltaic panel cannot be optimally adjusted, affecting the power generation efficiency and equipment life. In addition, the lack of real-time adjustment of intelligent system often needs more human input, and does not have the ability of automatic optimization. SUMMARY

[0004] In order to solve the above problems, the present application proposes an intelligent monitoring system and method for realizing dynamic performance adjustment of distributed photovoltaic network, which optimizes the dynamic transmission of image data and running data between edge computing nodes and cloud by real-time adjusting the hierarchical processing rules of data, significantly improves the data processing efficiency and network response speed.

[0005] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0006] An intelligent monitoring system for realizing dynamic performance adjustment of distributed photovoltaic network, the system comprises:

[0007] A data acquisition module for acquiring image data of photovoltaic panel in target area in real time through high-resolution remote sensing technology, acquiring running data of photovoltaic power station in real time through distributed sensor, and transmitting to edge computing node or cloud through wired or wireless mode; the running data includes voltage, current, power and temperature information;

[0008] A task scheduling module for coordinating the dynamic transmission of image data and running data between edge computing nodes and cloud;

[0009] a data processing module, configured to pre-process and extract features from the image data and operation data;

[0010] a dynamic estimation module, configured to input the extracted features into a sparse regression model and predict a dynamic state of the photovoltaic network by using an unscented Kalman filter;

[0011] a monitoring feedback module, configured to remotely dynamically adjust an angle or position of the photovoltaic panel, start or stop the equipment and / or adjust parameters according to a prediction result of the dynamic state, and feed back the adjustment result to the system, so as to continuously optimize the state estimation model and the adjustment strategy;

[0012] a report statistical module, configured to generate a daily report and statistical data, and help a user understand a running state and economic benefits of the photovoltaic network.

[0013] As a preferred scheme of the present application, the task scheduling module comprises:

[0014] an intermediate data buffer, configured to temporarily store the collected image data and operation data, and wait for a hierarchical processing decision of the edge-cloud collaborative optimization model;

[0015] a hierarchical processing unit, configured to perform hierarchical processing on the image data and operation data based on the edge-cloud collaborative optimization model, so as to reduce a computing load of the center node;

[0016] a dynamic adjustment unit, configured to adjust a hierarchical processing rule of the data in real time according to a load condition of the edge computing node and a difference in network delay of the cloud, so as to prevent server overload and make full use of cloud computing resources for task processing.

[0017] As a preferred scheme of the present application, the edge-cloud collaborative optimization model specifically comprises:

[0018] a coordination factor D of the edge computing node and the cloud processing capability is set, and is used for globally balancing distribution of data processing tasks between the edge computing node and the cloud;

[0019] wherein:

[0020]

[0021] In the formula, G represents a data processing efficiency of a single edge computing node, M represents a mean delay of data exchange between multiple edge computing nodes, C represents a response time of a cloud unit time processing data volume, and N represents a number of edge computing nodes in the photovoltaic network.

[0022] When the coordination factor D is greater than a preset first threshold value, the image data or operation data is preferentially transmitted from the intermediate data buffer to the cloud for processing, otherwise the image data or operation data is transmitted to the edge computing node for subsequent processing.

[0023] As a preferred scheme of the present application, the hierarchical processing rule of the data is adjusted in real time according to the load condition of the edge computing node and the difference of the network delay of the cloud, specifically:

[0024] whether the difference ΔLoad between the current data processing load of the edge computing node and the preset safety threshold is greater than a preset difference δ;

[0025] a standard value η is set, and the emergency coefficient E of the data uploading to the cloud is evaluated based on the performance parameter of the edge computing node and the current network delay condition;

[0026] when ΔLoad> δ and E≥ η, the storage time of the local data is increased, and the frequency of data transmission to the edge computing node is reduced;

[0027] when ΔLoad≤ δ or E< η, the frequency of data uploading to the cloud is increased.

[0028] As a preferred scheme of the present application, the data processing module comprises:

[0029] an image preprocessing unit, configured to preprocess the image data, including cropping irrelevant areas, adjusting resolution to unify input image size, color normalization, removing image noise and image enhancement;

[0030] a running data preprocessing unit, configured to perform preprocessing operation on the running data, including data cleaning, missing value processing and data standardization;

[0031] a feature extraction unit, configured to extract high-dimensional features from the preprocessed image data through a deep learning model, and extract time series features from the preprocessed running data through an LSTM algorithm.

[0032] As a preferred scheme of the present application, the image enhancement method comprises:

[0033] the following steps are independently performed on each color channel of the image:

[0034] Gaussian blur processing is performed on the input image to simulate different blur conditions in the imaging process, and the standard deviation σ of the Gaussian blur is set to 2 to remove high-frequency details of a specific scale;

[0035] discrete wavelet transform is performed on the blurred image to decompose the image into coefficients of multiple scales and directions, and the wavelet transform formula is:

[0036]

[0037] In the formula, W(f)(x,s) is the wavelet transform result of the input image f at the position x under the scale s; f(u) represents the pixel value of the input image f at the position u; ψ is a wavelet base function; s is a scale parameter; and u is a spatial position;

[0038] Randomly perturbing the wavelet coefficients of the image, randomly setting 20% of the coefficients as 0, and the original, unperturbed wavelet coefficient vector z s Multiplying the mask matrix M for controlling the coefficient perturbation, and denoted as:

[0039] z′ s =z s ·M;

[0040] In the formula, z′ s is the wavelet coefficient vector after the perturbation; the elements of M are 0 or 1, wherein 0 represents that the coefficient is perturbed, and 1 represents that the coefficient remains unchanged;

[0041] Performing inverse discrete wavelet transform on the perturbed wavelet coefficients to reconstruct the image, and the reconstruction formula is:

[0042]

[0043] In the formula, f1(x,y) represents the value of the reconstructed image at the pixel point (x,y); W -1 is a wavelet inverse transform operator.

[0044] As a preferred scheme of the present application, the deep learning model comprises:

[0045] Inputting the preprocessed image data into the deep learning model, and training using a convolutional neural network, and the convolution operation is defined as:

[0046]

[0047] In the formula, f2(x,y) is the output value of the image at the pixel point (x,y) after the convolution operation; f(x-i,y-j) is the value of the input image at the pixel point (x-i,y-j); k is a convolution kernel; and k(i,j) represents the weight value of the convolution kernel at the position (i,j);

[0048] For each scale s of the image, the response of the model to the details of the image of different scales is analyzed through a wavelet scale attribution method, and the Sobol index Ss is used to calculate the degree of dependence of the model on the scale details, and the formula is:

[0049]

[0050] In the formula, Var(f(x, y)) is the variance of the original image f(x, y), and Var(f'(x, y, s)) is the variance of the perturbed image f'(x, y, s) after removing the scale s details.

[0051] The dependence of each scale s is summarized as a scale embedding vector z=(z1, z2, …, z L )∈R L , wherein z L represents the dependence of the model on the scale s L .

[0052] The weight of the model on different scales is adjusted according to the value in the scale embedding vector z, specifically: the loss function of the model is modified in the training stage, so that the dependence of the model on different scale features matches its importance, and the modified loss function L is:

[0053]

[0054] In the formula, w s is the weight of the scale s, and L s is the loss value at the scale s.

[0055] As a preferred scheme of the present application, the dynamic estimation module comprises:

[0056] A sparse matrix construction unit is configured to construct a sparse regression model and calculate a sparse coefficient matrix Π, which is represented as:

[0057]

[0058] , wherein, represents the system state; Θ(X, U) is a feature function library based on the feature matrix X and the input matrix U, and the feature matrix X includes high-dimensional features extracted by a deep learning model and time sequence features extracted by an LSTM algorithm;

[0059] A sparsification processing unit is configured to perform sparsification processing on the sparse coefficient matrix Π using a sequential threshold least squares method according to a preset sparsity threshold, and set non-key coefficients to zero.

[0060] A state prediction unit is configured to predict the system state using an unscented Kalman filter, and predict the state of the system at k+1 according to a state transition equation The formula is:

[0061]

[0062] In the formula, represents the estimated state of the system at k; u kis the input control signal at time k; g represents the nonlinear state transition function of the system; w k is the process noise;

[0063] The covariance updating unit is configured to update the covariance matrix according to the state prediction result, and is denoted as:

[0064] P k+1 = FP k F T +Q;

[0065] In the formula, P k is the state covariance matrix at the current time k; P k+1 is the state covariance matrix at time k+1; F is the state transition matrix; T represents the matrix transpose; and Q is the process noise covariance matrix.

[0066] The measurement updating unit is configured to correct the predicted state according to the Kalman gain K, and the calculation formula of the Kalman gain is:

[0067] K = P k+1 H T (HP k+1 H T +R) -1 ;

[0068] In the formula, H is the measurement matrix; and R is the measurement noise covariance matrix.

[0069] The state estimation is then updated as:

[0070]

[0071] In the formula, y k+1 is the actual measurement value at time k+1.

[0072] An intelligent monitoring method for implementing dynamic performance adjustment on a distributed photovoltaic network, the method comprising:

[0073] Image data of photovoltaic panels in a target area is collected in real time by high-resolution remote sensing technology, and operation data of the photovoltaic power station is collected in real time by distributed sensors, and the image data and operation data are temporarily stored in an intermediate data buffer;

[0074] The image data and operation data are processed in stages based on an edge-cloud collaborative optimization model, the staging processing rules of the data are adjusted in real time according to the load conditions of the edge computing nodes and the differences in network delays of the cloud, and the dynamic transmission of the image data and operation data between the edge computing nodes and the cloud is coordinated;

[0075] The image data is preprocessed, including cropping irrelevant areas, adjusting the resolution to unify the input image size, color normalization, removing image noise, and image enhancement;

[0076] The operation data is preprocessed, including data cleaning, missing value processing and data standardization;

[0077] High-dimensional features are extracted from the preprocessed image data through a deep learning model, and time sequence features are extracted from the preprocessed operation data through an LSTM algorithm;

[0078] The extracted high-dimensional features and time sequence features are input into a sparse regression model, and a unscented Kalman filter is used to predict the dynamic state of the photovoltaic network;

[0079] According to the prediction result of the dynamic state, the angle or position of the photovoltaic panel is dynamically adjusted remotely, the equipment is started or stopped and / or the parameters are adjusted, and the adjustment result is fed back to the system, so as to continuously optimize the state estimation model and the adjustment strategy.

[0080] The beneficial effects of the present application are:

[0081] The data is processed by the edge cloud collaborative optimization model, which effectively disperses the computing load of the center node, avoids the risk of server overload, fully utilizes the cloud computing resources, and ensures the efficient operation of the system; according to the load of the edge computing node and the difference of the cloud network delay, the data transmission frequency is dynamically adjusted, so that the data transmission is more flexible, and the possibility of network congestion is reduced; the image preprocessing module is adopted, the system can crop, enhance and denoise the image data of the photovoltaic panel, improve the accuracy of image recognition, and provide high-quality input for subsequent feature extraction; the dynamic state of the photovoltaic network is predicted in real time through the deep learning model and the unscented Kalman filter, the system can remotely dynamically adjust the angle of the photovoltaic panel, the equipment start-stop and other parameters, so as to optimize the efficiency of photovoltaic power generation; the monitoring feedback module automatically adjusts according to the prediction result, reduces the need for manual intervention, improves the intelligent level of the system, and can continuously optimize the state estimation model and the adjustment strategy. BRIEF DESCRIPTION OF DRAWINGS

[0082] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0083] Among them:

[0084] Figure 1 The system modular structure of the present application;

[0085] Figure 2 The method flowchart in the embodiment of the present application. DETAILED DESCRIPTION

[0086] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art belong to the scope of protection of the present application.

[0087] As shown in Figure 1 , one embodiment of the present application provides an intelligent monitoring system for realizing dynamic performance adjustment of a distributed photovoltaic network, comprising:

[0088] (1) Data acquisition module

[0089] used for collecting image data of photovoltaic panels in a target area in real time through high-resolution remote sensing technology, collecting operation data of a photovoltaic power station in real time through distributed sensors, and transmitting the operation data to an edge computing node or a cloud end through wired or wireless mode; the operation data includes voltage, current, power, temperature and other information;

[0090] By reasonably distributing data to a central server and an edge computing node, the overall efficiency of the system can be improved, the real-time requirement can be ensured, and the accuracy of complex computing tasks is not sacrificed.

[0091] (2) Task scheduling module

[0092] used for coordinating dynamic transmission of image data and operation data between an edge computing node and a cloud end server, including an intermediate data buffer, a hierarchical processing unit and a dynamic adjustment unit;

[0093] The intermediate data buffer is used for temporarily storing collected image data and operation data, and waiting for hierarchical processing decisions of an edge-cloud collaborative optimization model;

[0094] Before the edge-cloud collaborative optimization model decides whether the data is processed by the cloud end or the edge node, the collected data needs an intermediate storage or transmission waiting mechanism to ensure that the data will not be lost before the edge-cloud processing path is determined. This mechanism is completed by the intermediate data buffer or the task scheduling system.

[0095] The hierarchical processing unit is used for hierarchical processing of image data and operation data based on the edge-cloud collaborative optimization model, so as to reduce the computing load of the central node;

[0096] The edge-cloud collaborative optimization model specifically comprises:

[0097] A coordination factor D of the edge computing node and the cloud processing capability is set to globally balance the allocation of data processing tasks between the edge computing node and the cloud;

[0098] wherein:

[0099]

[0100] In the formula, G represents the data processing efficiency of a single edge computing node, representing the effectiveness of each edge device independently processing local or near data, with a value ranging from 0 to 10, the optimal value being 8, and the ideal value being the maximum, to reflect the highest processing level; M represents the average delay of data exchange between multiple edge computing nodes, with a value ranging from 0 to 150 ms, the optimal value being 2 ms, representing the communication quality of the network, the lower the delay, the more efficient the data transmission; C is the response time of the cloud processing data volume per unit time, i.e. the time of the central server in receiving and analyzing the information cluster, ranging from 50 to 1000 ms, the optimal value being 65 ms, the smaller the C value, the faster the cloud processing speed, which can quickly respond to a large amount of data; N is the number of edge computing nodes in the photovoltaic network, the larger the value of N, the stronger the edge computing capability of the system, which can share more computing tasks;

[0101] When the score of the coordination factor D rises, it means that the cloud can also provide strong support while bearing high data load, and reduces the possibility of end-to-end delay, enhances the self-regulation function in dynamic change situation and the tolerance and compatibility of different hardware interfaces or services. Therefore, when the coordination factor D is greater than a preset first threshold value (determined according to different application scenarios or network states), the image data or running data is transmitted from the intermediate data buffer to the cloud for processing, otherwise it is transmitted to the edge computing node for subsequent processing.

[0102] For example, assume a photovoltaic power station containing 300 edge devices, each with good performance, such as G being 7 points; in addition, it is assumed that the communication speed between all machines in the area is very fast, only about 2 ms (M); and the cloud computing service is also very agile, responding within 65 ms (C). Now substitute the actual numbers into the formula calculation to get: D = (7 + 2 + 65) / 300 x 100 = 24.67%. Since this number is relatively low, it shows that even with a large number of sensor nodes and a large increase in data volume, good interactive response performance and direct visual display can still be maintained, which verifies that the edge-cloud collaborative optimization model of the application can successfully operate to reach the preset standard in real application fields.

[0103] The dynamic adjustment unit is used to adjust the hierarchical processing rules of data in real time according to the load condition of the edge computing node and the difference of network delay in the cloud, so as to prevent server overload and make full use of cloud computing resources for task processing, specifically:

[0104] whether the difference ΔLoad between the current data processing load (such as CPU occupancy rate, bandwidth utilization or data processing task quantity) of the edge computing node and the preset safety threshold is greater than the preset difference δ;

[0105] The key of this step is to monitor and quantify the computing capacity burden of the edge computing node and the actual distance of this burden from the safety operation boundary in real time; for example, if a safety data processing speed of 10 GB / s is defined in advance for an industrial automation scene and the current measurement value is 9 GB / s, δ can be preset as 0.7 GB / s according to system requirements; if the actual load reaches 9.4 GB / s, ΔLoad is 0.4 GB / s, and ΔLoad < δ when δ is set as 0.7 GB / s, which indicates that the device is running well and has not exceeded the expected acceptable load range;

[0106] Based on the performance parameters of the edge computing node and the urgency coefficient E of the current network delay condition evaluation data uploaded to the cloud;

[0107] According to the specific resource condition of the edge computing node, such as the speed of the processor or the storage capacity, the number and activity state of the connected Internet of Things devices, and combined with the communication delay to the cloud at this moment, the demand level and the important and urgent degree of uploading data to the cloud are comprehensively evaluated, and the evaluation standard can consider that the remaining processing capacity of the edge device and the real-time communication time between the device and the central management cloud platform are comprehensively determined by the emergency coefficient E;

[0108] Exemplarily, In the formula, P represents the remaining processing capacity of the edge device, the unit is GHz; S represents the remaining storage capacity, the unit is GB; O represents the communication delay between the device and the cloud, the unit is ms; A represents the number and activity proportion of the currently connected Internet of Things devices; α, β, γ, ξ are weight coefficients of each factor respectively, and these weights can be dynamically adjusted according to the application scene; the higher the value of E is, the stronger the urgency of uploading data to the cloud is;

[0109] When the gap ΔLoad exceeds the preset difference value (ΔLoad > δ) and the urgency coefficient E reaches the emergency threshold (E ≥ η, η is a standard value determined in advance for dividing high and low priorities), the data transmission frequency of the edge device is reduced, and the storage time of local data is increased to reduce the amount of data transmitted to the cloud; for example, assuming that η is located at 0.6, data below this value can be processed later, when the system calculates that the urgency coefficient is 0.7 and the edge computing node is already very busy (assuming that the ΔLoad = 1 GB / s > δ = 0.7 GB / s at this time), the data reporting and temporary storage measures will be slowed down;

[0110] But on the contrary, once the calculated data transmission tension is insufficient, that is, when the gap ΔLoad is within the preset difference value range (ΔLoad ≤ δ) or the urgency coefficient E is lower than the emergency threshold (E < η), it means that there is space to upload more data as soon as possible for higher-level algorithms to use, and the frequency of data upload to the cloud can be increased to utilize the powerful computing power of the cloud for higher-level data analysis and strategy generation. As mentioned earlier, when the E calculation result shows that only 0.4 is lower than the limit 0.6, although there may be a little light burden, it will still tend to send to the remote platform immediately. This method not only takes advantage of the advantages of on-site immediate response, but also balances the value of centralized deep data analysis. By adjusting the load sharing scheme of edge and cloud computing, the purpose of maximizing the overall efficiency of the system is achieved.

[0111] Using the coordination factor D, the allocation of data processing tasks can be intelligently adjusted according to the processing capacity of the edge device and the cloud, network delay and other factors, improving the response speed of the system; by detecting the load of the edge device and the network delay, the upload frequency of data and the allocation of processing tasks can be adjusted in real time, ensuring the stability and adaptability of the system under different operating conditions.

[0112] (3) Data processing module

[0113] For pre-processing and feature extraction of image data and operation data, including image preprocessing unit, operation data preprocessing unit and feature extraction unit;

[0114] The image preprocessing unit is used for pre-processing the image data, including cropping irrelevant areas, adjusting the resolution to unify the input image size, color normalization, removing image noise and image enhancement;

[0115] Specifically, the method of image enhancement includes:

[0116] The following steps are independently performed on each color channel of the image:

[0117] The input image is subjected to Gaussian blur processing to simulate different blur conditions in the imaging process, and the standard deviation σ of the Gaussian blur is set to 2 to remove high-frequency details of a certain scale;

[0118] The image after the blur processing is subjected to a discrete wavelet transform, and the image is decomposed into coefficients in multiple scales and directions, and the wavelet transform formula is:

[0119]

[0120] In the formula, W(f)(x,s) is the wavelet transform result of the input image f at position x in scale s; f(u) represents the pixel value of the input image f at position u; ψ is a wavelet base function; s is a scale parameter; and u is a spatial position;

[0121] The wavelet coefficients of the image are subjected to random disturbance, and 20% of the coefficients are randomly set to 0, and the original un-disturbed wavelet coefficient vector z s is multiplied by the mask matrix M for controlling coefficient disturbance, and is represented as:

[0122] z′ s = z s ·M;

[0123] In the formula, z′ s is the wavelet coefficient vector after disturbance; the elements of M are 0 or 1, wherein 0 represents that the coefficient is disturbed, and 1 represents that the coefficient remains unchanged;

[0124] An inverse discrete wavelet transform is performed on the disturbed wavelet coefficients to reconstruct the image, and the reconstruction formula is:

[0125]

[0126] In the formula, f1(x,y) represents the value of the reconstructed image at pixel point (x,y); W -1 is a wavelet inverse transform operator.

[0127] The running data preprocessing unit is used for performing preprocessing operations on the running data, including data cleaning, missing value processing and data standardization;

[0128] The feature extraction unit is used for extracting high-dimensional features from the preprocessed image data through a deep learning model, and extracting time sequence features from the preprocessed running data through an LSTM algorithm;

[0129] The deep learning model includes:

[0130] The preprocessed image data is input into the deep learning model, and a convolutional neural network is used for training, and the convolution operation is defined as:

[0131]

[0132] In the formula, f2(x, y) is the output value of the image at pixel point (x, y) after the convolution operation; f(x-i, y-j) is the value of the input image at pixel point (x-i, y-j); k is the convolution kernel; k(i, j) represents the weight value of the convolution kernel at position (i, j);

[0133] For each scale s of the image, the response of the model to the details of the image at different scales is analyzed by the wavelet scale attribution method, and the degree of dependence of the model on the scale details is calculated using the Sobol index Ss, and the formula is:

[0134]

[0135] In the formula, Var(f(x, y)) is the variance of the original image f(x, y), and Var(f'(x, y, s)) is the variance of the perturbed image f'(x, y, s) after removing the scale s details;

[0136] The degree of dependence of each scale s is summarized into a scale embedding vector z=(z1, z2, …, z L )∈R L , wherein z L represents the degree of dependence of the model on the scale s L ;

[0137] According to the values in the scale embedding vector z, the weights of the model at different scales are adjusted, specifically: in the training phase, the loss function of the model is modified, so that the dependence of the model on different scale features matches the importance thereof, and then the modified loss function L is:

[0138]

[0139] In the formula, w s is the weight of the scale s, and L s is the loss value at the scale s.

[0140] (4) Dynamic estimation module

[0141] used for inputting the extracted features into a sparse regression model and predicting the dynamic state of the photovoltaic network using an unscented Kalman filter;

[0142] In one embodiment, the dynamic estimation module comprises:

[0143] A sparse matrix construction unit is configured to construct a sparse regression model and calculate a sparse coefficient matrix Π, which is represented as:

[0144]

[0145] , wherein, represents the system state; Θ(X, U) is a feature function library based on a feature matrix X and an input matrix U, the feature matrix X including high-dimensional features extracted by a deep learning model and time-series features extracted by an LSTM algorithm;

[0146] The sparse processing unit is configured to perform sparse processing on the sparse coefficient matrix Π by using a sequential threshold least square method according to a preset sparse threshold, and set non-critical coefficients to zero.

[0147] The state prediction unit is configured to predict the system state by using an unscented Kalman filter, and predict the state of the system at the k+1 time according to a state transition equation The formula is:

[0148]

[0149] In the formula, represents the estimated state of the system at the k time; u k is an input control signal at the k time; g represents a nonlinear state transition function of the system; w k is process noise;

[0150] The covariance updating unit is configured to update a covariance matrix according to the state prediction result, and the formula is:

[0151] P k+1 = FP k F T +Q;

[0152] In the formula, P k is a state covariance matrix at the current time k; P k+1 is a state covariance matrix at the k+1 time; F is a state transition matrix; T represents matrix transposition; and Q is a process noise covariance matrix.

[0153] The measurement updating unit is configured to correct the predicted state according to a Kalman gain K, and the formula for calculating the Kalman gain is:

[0154] K=P k+1 H T (HP k+1 H T +R) -1 ;

[0155] In the formula, H is a measurement matrix; and R is a measurement noise covariance matrix.

[0156] The state estimation is then updated:

[0157]

[0158] In the formula, y k+1 is an actual measurement value at the k+1 time.

[0159] (5) Monitoring feedback module

[0160] Remote dynamic adjustment of the angle or position of the photovoltaic panel according to the prediction result of the dynamic state, starting and stopping the equipment and / or adjusting the parameters, and feeding back the adjustment result to the system, continuously optimizing the state estimation model and adjustment strategy.

[0161] (6) Report statistics module

[0162] For generating daily reports and statistical data, helping users understand the operation and economic benefits of the photovoltaic network.

[0163] As Figure 2 shown, another embodiment of the present application provides an intelligent monitoring method for dynamic performance adjustment of a distributed photovoltaic network, comprising the following steps:

[0164] Real-time collection of image data of photovoltaic panels in the target area through high-resolution remote sensing technology, real-time collection of operation data of photovoltaic power stations through distributed sensors, and temporary storage of image data and operation data in an intermediate data buffer;

[0165] Based on the edge cloud collaborative optimization model, the image data and operation data are processed in stages, the classification processing rules of the data are adjusted in real time according to the load of the edge computing nodes and the difference in network delay of the cloud, and the dynamic transmission of the image data and operation data between the edge computing nodes and the cloud is coordinated;

[0166] The image data is preprocessed, including cropping irrelevant areas, adjusting the resolution to unify the input image size, color normalization, removing image noise and image enhancement;

[0167] The operation data is preprocessed, including data cleaning, missing value processing and data standardization;

[0168] High-dimensional features are extracted from the preprocessed image data through a deep learning model, and time series features are extracted from the preprocessed operation data through an LSTM algorithm;

[0169] The extracted high-dimensional features and time series features are input into a sparse regression model, and an unscented Kalman filter is used to predict the dynamic state of the photovoltaic network;

[0170] Remote dynamic adjustment of the angle or position of the photovoltaic panel according to the prediction result of the dynamic state, starting and stopping the equipment and / or adjusting the parameters, and feeding back the adjustment result to the system, continuously optimizing the state estimation model and adjustment strategy

[0171] In summary, the application optimizes the dynamic transmission of image data and operation data between the edge computing node and the cloud by adjusting the hierarchical processing rules of data in real time, significantly improves the data processing efficiency and network response speed, not only can significantly improve the overall operation efficiency of the photovoltaic network, but also can reduce the operation and maintenance cost, prolong the service life of the equipment, and has wide application prospect.

[0172] The above merely describes a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An intelligent monitoring system for implementing dynamic performance adjustment for a distributed photovoltaic network, characterized in that, The system comprises: a data acquisition module for acquiring image data of photovoltaic panels in a target area in real time through high-resolution remote sensing technology, acquiring operation data of the photovoltaic power station in real time through distributed sensors, and transmitting the data to an edge computing node or a cloud through wired or wireless means; the operation data comprises voltage, current, power and temperature information; a task scheduling module for coordinating dynamic transmission of the image data and the operation data between the edge computing node and the cloud, adjusting hierarchical processing rules of the data in real time according to differences in load conditions of the edge computing node and network delay of the cloud, and specifically: detecting whether a gap ΔLoad between a current data processing load of the edge computing node and a preset safety threshold is greater than a preset difference δ; setting a standard value η, and evaluating an urgency coefficient E of data uploading to the cloud based on performance parameters of the edge computing node and current network delay conditions; when ΔLoad>δ and E≥η, increasing a storage time of local data and reducing a frequency of data transmission to the edge computing node; and when ΔLoad≤δ or E<η, increasing a frequency of data uploading to the cloud; a data processing module for pre-processing and feature extraction of the image data and the operation data, extracting high-dimensional features from the pre-processed image data through a deep learning model, and the deep learning model comprising: inputting the pre-processed image data into the deep learning model, training the model using a convolutional neural network, and defining a convolution operation as: wherein f2(x,y) is an output value of the image at a pixel point (x,y) after the convolution operation; f(x-i,y-j) is a value of an input image at a pixel point (x-i,y-j); k is a convolution kernel; and k(i,j) represents a weight value of the convolution kernel at a position (i,j); for each scale s of the image, analyzing a response of the model to details of different scales of the image through a wavelet scale attribution method, and calculating a degree of dependence of the model on the scale details using a Sobol index Ss, and the formula is: wherein Var(f(x,y)) is a variance of an original image f(x,y), and Var(f'(x,y,s)) is a variance of a perturbed image f'(x,y,s) after removing details of a scale s; The dependency of each scale s is summarized into a scale embedding vector z = (z1, z2, ..., z L )∈R L , where z L Indicates that the model is at scale s L the degree of dependence on adjusting weights of the model at different scales according to values in a scale embedding vector z, and specifically: correcting a loss function of the model in a training stage, so that a degree of dependence of the model on different scale features matches an importance of the features, and then a corrected loss function L is: where w s is the weight of scale s, L s is the loss value at scale s; a dynamic estimation module for inputting the extracted features into a sparse regression model and predicting a dynamic state of the photovoltaic network using an unscented Kalman filter; a monitoring feedback module for remotely and dynamically adjusting an angle or a position of the photovoltaic panel according to a prediction result of the dynamic state, starting and stopping equipment and / or adjusting parameters, and feeding back the adjustment result to the system, so as to continuously optimize a state estimation model and an adjustment strategy; a report statistical module for generating daily reports and statistical data to help users understand operation conditions and economic benefits of the photovoltaic network.

2. The intelligent monitoring system for achieving dynamic performance adjustment of distributed photovoltaic network according to claim 1, characterized in that, The task scheduling module comprises: an intermediate data buffer for temporarily storing the acquired image data and the operation data, and waiting for hierarchical processing decisions of an edge-cloud collaborative optimization model. The hierarchical processing unit is configured to perform hierarchical processing on the image data and the operation data based on an edge-cloud collaborative optimization model, so as to reduce the computing load of the central node. The dynamic adjustment unit is configured to adjust the hierarchical processing rules of the data in real time according to the load conditions of the edge computing nodes and the differences in network delays of the cloud, so as to prevent server overload and make full use of cloud computing resources for task processing.

3. The intelligent monitoring system for achieving dynamic performance adjustment of distributed photovoltaic network according to claim 2, characterized in that, The edge-cloud collaborative optimization model specifically comprises: A coordination factor D of the edge computing nodes and the cloud processing capability is set, which is used for globally balancing the allocation of data processing tasks between the edge computing nodes and the cloud; In the formula, G represents the data processing efficiency of a single edge computing node; M represents the average delay of data exchange between multiple edge computing nodes; C is the response time of the cloud for processing data per unit time; and N is the number of edge computing nodes in the photovoltaic network. When the coordination factor D is greater than a preset first threshold value, the image data or the operation data is preferentially transmitted from the intermediate data buffer to the cloud for processing, otherwise the image data or the operation data is transmitted to the edge computing node for subsequent processing. The data processing module comprises:

4. The intelligent monitoring system for achieving dynamic performance adjustment of distributed photovoltaic network according to claim 1, characterized in that, An image preprocessing unit is configured to preprocess the image data, including cropping irrelevant areas, adjusting the resolution to unify the input image size, color normalization, removing image noise and image enhancement. An operation data preprocessing unit is configured to perform preprocessing operations on the operation data, including data cleaning, missing value processing and data standardization. A feature extraction unit is configured to extract high-dimensional features from the preprocessed image data through a deep learning model and extract time series features from the preprocessed operation data through an LSTM algorithm. The image enhancement method comprises:

5. The intelligent monitoring system for achieving dynamic performance adjustment of distributed photovoltaic network according to claim 4, characterized in that, The following steps are independently performed on each color channel of the image: Gaussian blur processing is performed on the input image to simulate different blur conditions in the imaging process, and the standard deviation σ of the Gaussian blur is set to 2 to remove high-frequency details of a specific scale; Discrete wavelet transform is performed on the blurred image to decompose the image into coefficients of multiple scales and directions, and the wavelet transform formula is: In the formula, W(f)(x,s) is the wavelet transform result of the input image f at scale s and position x; f(u) represents the pixel value of the input image f at position u; ψ is a wavelet basis function; s is a scale parameter; and u is a spatial position; Inverse discrete wavelet transform is performed on the perturbed wavelet coefficients to reconstruct the image, and the reconstruction formula is: Randomly perturb the wavelet coefficients of the image, randomly set 20% of the coefficients to 0, the original unperturbed wavelet coefficient vector z s Multiply by the mask matrix M that controls the perturbation of the coefficients, denoted as: z' s = z s • M; where z' = z + M s is the vector of perturbed wavelet coefficients; the elements of M are 0 or 1, where 0 indicates that the coefficient is perturbed and 1 indicates that the coefficient is left unchanged; The dynamic estimation module comprises: In the formula, f1(x, y) is a value of the reconstructed image at a pixel point (x, y); W -1 is a wavelet inverse transform operator.

6. The intelligent monitoring system for achieving dynamic performance adjustment for distributed photovoltaic network according to claim 1, characterized in that, A sparse matrix construction unit is configured to construct a sparse regression model and calculate a sparse coefficient matrix Π, which is represented as: A sparsification processing unit is configured to perform sparsification processing on the sparse coefficient matrix Π using a sequential threshold least squares method according to a preset sparsification threshold, and set non-critical coefficients to zero; wherein, represents a system state; Θ(X, U) is a feature function library based on a feature matrix X and an input matrix U, the feature matrix X including high-dimensional features extracted by a deep learning model and time-series features extracted by an LSTM algorithm; A covariance update unit is configured to update the covariance matrix according to the state prediction result, which is represented as: A state prediction unit is configured to predict the state of the system using an unscented Kalman filter, and predict the state of the system at k+1 according to a state transition equation The formula is: wherein x(k) represents the estimated state of the system at time k; u k is the input control signal at time k; g represents the nonlinear state transition function of the system; w k is the process noise; A measurement update unit is configured to correct the predicted state according to the Kalman gain K, and the calculation formula of the Kalman gain is: P k+1 = FP k F T + Q; where P k is the state covariance matrix at the current time k; P k+1 is the state covariance matrix at time k+1 ; F is the state transition matrix; T denotes matrix transpose; Q is the process noise covariance matrix; In the formula, H is a measurement matrix; and R is a measurement noise covariance matrix; K = P k+1 H T (HP k+1 H T +R) -1 ; The state estimation is then updated: The method comprises: where y is the actual measurement at time k + 1. k+1 is the actual measurement at time k + 1.

7. A monitoring method for an intelligent monitoring system for achieving dynamic performance adjustment for a distributed photovoltaic network according to any one of claims 1-6, characterized in that, ​ Image data of the photovoltaic panels in the target area is collected in real time by high-resolution remote sensing technology, and operation data of the photovoltaic power station is collected in real time by distributed sensors, and the image data and operation data are temporarily stored in an intermediate data buffer; The image data and operation data are processed hierarchically based on an edge-cloud collaborative optimization model, the hierarchical processing rules of the data are adjusted in real time according to the load conditions of the edge computing nodes and the differences in network delay of the cloud, and the dynamic transmission of the image data and operation data between the edge computing nodes and the cloud is coordinated; The image data is preprocessed, including cropping irrelevant areas, adjusting the resolution to unify the input image size, color normalization, removing image noise and image enhancement; The operation data is preprocessed, including data cleaning, missing value processing and data standardization; High-dimensional features are extracted from the preprocessed image data by a deep learning model, and time series features are extracted from the preprocessed operation data by an LSTM algorithm; The extracted high-dimensional features and time series features are input into a sparse regression model, and a unscented Kalman filter is used to predict the dynamic state of the photovoltaic network; The angle or position of the photovoltaic panels is dynamically adjusted remotely according to the prediction result of the dynamic state, the equipment is started or stopped and / or the parameters are adjusted, and the adjustment result is fed back to the system, and the state estimation model and the adjustment strategy are continuously optimized.

Images