Remote fault diagnosis method and system for solar device

Abstract

The invention discloses a remote fault diagnosis method and system for a solar device, and relates to the field of remote fault diagnose.The remote fault diagnosis system comprises an acquisition module, a preprocessing module, a fusion module, a fault module and a prediction module. Constructing a dynamic space-time hypergraph to form a space-time hypergraph tensor, outputting a complementation adversarial tensor, constructing an equipment dynamic topological network, generating a topological feature tensor, performing decomposition analysis by utilizing a space-time graph convolutional network and a wavelet packet, performing fusion analysis by utilizing a gating interaction mechanism, and establishing a solar device fault hybrid kernel model through a hybrid kernel function by utilizing a support vector machine. The model outputs a corresponding fault level state, a multivariable LSTM network is utilized to construct a solar device fault time sequence prediction model, a solar device fault regression analysis model is constructed, the model inputs real-time parameters to output a fault probability, and early warning is set according to different threshold values, so that accurate fault level judgment can be ensured.

Description

Technical Field

[0001] The present invention relates to the field of remote fault diagnosis, and in particular to a remote fault diagnosis method and system for a solar device. Background Art

[0002] With the increasing scale of solar power installations, failures in core components (such as energy storage systems, inverters, and solar panels) can significantly impact system stability and power generation efficiency. Traditional fault diagnosis methods rely on on-site personnel and single-sensor data, making them difficult to handle the dynamic interactions between devices and the complexity of multi-source, heterogeneous monitoring data. Consequently, remote fault diagnosis methods and systems for solar power installations have emerged.

[0003] However, existing remote fault diagnosis methods and systems for solar devices are unable to accurately perform deep modeling of variable correlations in multi-source data, resulting in accumulated errors in subsequent analysis. Existing methods also ignore temporal causality (such as the conduction effect of parameter fluctuations) and physical coupling between devices (such as circuit connections), making it difficult to capture complex fault causes. A single feature extraction method cannot fully characterize the fault mode, resulting in inaccurate fault level determination and a lack of differentiated analysis of short-term sudden failures, mid-term performance degradation, and long-term trends in solar devices.

[0004] In order to solve the above-mentioned defects, a technical solution is now provided. Summary of the Invention

[0005] In order to solve the technical problems raised by the above background technology, the present invention is proposed. The embodiments of the present invention provide a remote fault diagnosis method and system for a solar device.

[0006] The purpose of the present invention can be achieved by the following technical solution: A remote fault diagnosis method for a solar device comprises the following steps: Step S100: Obtaining solar panel monitoring parameters, inverter monitoring parameters, energy storage system monitoring parameters, and charge controller monitoring parameters through the provided sensors to form a multi-source heterogeneous monitoring parameter set for the solar device; Step S200: A multivariate outlier detection model is constructed to remove abnormal data values, and a dynamic spatiotemporal hypergraph of the solar device is constructed to form a spatiotemporal hypergraph tensor. The generator and discriminator are used to perform tensor convolution, LSTM, and loss function optimization to finally output a completed adversarial tensor, completing data preprocessing. Step S300: Input the completed adversarial tensor to construct a dynamic topology network of the device, generate a topology feature tensor, use the spatiotemporal graph convolutional network and wavelet packet decomposition analysis, and use the gated interaction mechanism for fusion analysis to obtain the final fusion optimized coding feature vector; Step S400: Using a support vector machine to establish a hybrid kernel model of solar device faults through a hybrid kernel function. The model determines a specific fault level state based on a fault probability threshold and outputs the corresponding fault level state. Step S500: A multivariate LSTM network is used to construct a solar device failure time series prediction model to obtain future parameter prediction values, and a solar device failure regression analysis model is constructed. The model inputs real-time parameters and outputs failure probability, and sets short-term, medium-term, and long-term warnings based on different thresholds.

[0007] Furthermore, the step of completing the adversarial tensor analysis is as follows: Step S201: Constructing a multivariate outlier detection model to remove abnormal data values, specifically including: , where x represents a p-dimensional data point vector, including the features in the solar device multi-source heterogeneous monitoring parameter set, and the vector composed of the solar device multi-source heterogeneous monitoring parameter set, µ is the p-dimensional mean vector, and S is the p×p covariance matrix. Represents the Mahalanobis distance, T represents the transpose of the matrix, if Then the data point vector is determined to be abnormal and the data point vector is removed, where represents the 99% quantile of the p-dimensional chi-square distribution; Step S202: Construct a spatiotemporal hypergraph of solar device states, using solar panels, inverters, energy storage systems, and charge controllers as device nodes, and construct hyperedges v l Connect the device node to the device g i and g j Perform physical coupling function fphys processing to obtain the physical coupling relationship value, and perform physical coupling function fphys processing on the device g i and g j Perform Granger causality test to obtain the temporal causality value, multiply the physical coupling relationship value and the temporal causality value by the learnable weight respectively, and perform Sigmoid normalization function on the result of the sum of the two to obtain the hyperedge v l The weight at time t generates the time slice adjacency tensor B(t)∈R 4×4×6 , where dimensions 1-2 represent node pairs consisting of 4 devices, the third dimension corresponds to 6 hyperedge types, R represents a set of real numbers, and the element value Bijk(t) is the hyperedge v l The weight at time t, Bijk(t)=ω(v l , t), stack the T time slice tensors according to the time dimension to form a spatiotemporal hypergraph tensor X, X=Stack(B(t1),B(t2),...,B(tT)), and X∈R T×4×4×6 , Stack represents stacking tensor operations; The weights are dynamically calculated from the physical relationship between the parameters: ω(v l, t) = σ(β×fphys(g i , g j )+ɑ×Granger(g i →g j ), where fphys(*) is the physical coupling function, Granger(*) is the Granger causality test value, which quantifies the temporal causality between device parameters, β and ɑ are learnable weights, and σ(*) is the Sigmoid normalization function; Step S203: The generator G inputs the spatiotemporal hypergraph tensor X and the latent variable K. With the help of the spatiotemporal hypergraph tensor and the corresponding LSTM hidden state at the previous moment, the current hidden state is obtained through tensor convolution and LSTM processing, and the completed tensor Xgen is generated with the latent variable K. The discriminator D inputs the completed tensor Xgen, performs feature extraction through multi-scale three-dimensional convolution, and then uses the Sigmoid function to output the probability value of the real tensor. The Tucker norm difference between the completed tensor Xgen and the real tensor Xreal is calculated through the loss function. The generator and the discriminator optimize their respective parameters through the back-propagation algorithm, and continuously iterate and adjust until the value of the loss function is less than the set threshold. The operation is terminated and the completed adversarial tensor Xfinal is output.

[0008] Furthermore, the physical coupling relationship value analysis step: Equipment i and g j The Pearson correlation coefficient is calculated using historical monitoring data from the parameter set. When it is greater than the set threshold, the parameter is retained, and then the key parameters with physical connections are screened out. The function parameters are determined with the help of the least squares method to establish physical relationships. These parameters are constructed into a matrix and standardized. The covariance matrix is calculated, and its eigenvalues are decomposed. The principal component with a cumulative contribution rate greater than the threshold is selected. If the coefficient of the original parameter in the principal component is greater than the set threshold, the parameter is attributed to the key influencing factor to form a simplified key parameter subset. The data set of the simplified key parameter subset is divided into a training set and a test set in a ratio of 7:3. The random forest regression model is initialized and the initial parameters are set. The simplified key parameter subset of the training set is used as input and the target physical coupling value is used as output for training. The model learns the mapping relationship and predicts the physical coupling value of the test set. Through the mean square error, if it is greater than the set threshold, the model parameters are adjusted. Different parameter combinations are preset by grid search. The optimal parameters are found through cross-validation to obtain the optimized model. The input device g i and g j The actual parameter set of the physical coupling relationship value fphys (g i , g j ).

[0009] Furthermore, the temporal causality value analysis step:

[0010] Collection equipment i and g j The time series data of the parameters are aligned and the optimal lag order is determined by using the information criterion to construct the device-free g i The autoregressive model of the influence includes x j,t =η0+∑η k ×x j,t-k +ϵ 1，t , x j,t Indicates device g j At the time t, a parameter is observed, k represents the index variable of the lag order, η0 represents the constant term, η k represents the autoregressive coefficient, ϵ 1，t Represents the error term, constructs a device containing g i The autoregressive model of the influence includes x j,t =δ0+∑δ k ×x j,t-k +∑Φ k ×x i,t-k +ϵ 2，t , δ0 represents a constant term, η k and Φ k represents the autoregressive coefficient, ϵ 2，t Represents the error term, calculates the residual sum of squares of the two models, and constructs the F statistic, where , where p is the lag order, N is the total number of observations of time series data, SR1 and SR2 are respectively i and contains equipment g i The residual sum of squares of the affected model follows an F distribution with degrees of freedom (p,N−2p−1). If F>F Ω (p, N−2p−1), then reject the null hypothesis and mark the F value at this time as the temporal causality value. Otherwise, the null hypothesis cannot be rejected, set the F value at this time to zero, and mark it as the temporal causality value to obtain the corresponding temporal causality value.

[0011] Furthermore, the final fusion optimization encoding feature vector analysis step:

[0012] The time domain dynamic coding vector Z1 and the frequency domain topology coding vector Z2 are nonlinearly fused using the gated interaction mechanism, and the fusion strength of the two is adaptively controlled using dynamic gating weights to obtain a fused feature coding vector. The fused feature coding vector is used to extract temporal dynamic information through a multi-head self-attention mechanism, and a dynamic adjustment vector is generated by lightweight convolution. The dynamic adjustment vector is used to perform cross-attention optimization on the time domain dynamic coding vector Z1 and the frequency domain topology coding vector Z2 to obtain the optimized time domain dynamic coding vector and frequency domain topology coding vector. The optimized time domain dynamic coding vector and frequency domain topology coding vector are adaptively weighted fused by dynamically calculating weights to obtain the final fused optimized coding feature vector Zc.

[0013] Furthermore, the time domain dynamic coding vector Z1 and the frequency domain topology coding vector Z2 analysis steps are:

[0014] The topological feature tensor Xtopo is processed by the spatiotemporal graph convolutional network, and the time domain dynamic coding vector Z1 is generated through the feature extraction formula. The topological feature tensor Xtopo is decomposed into wavelet packets, and the wavelet packet entropy of the topological feature tensor is obtained through the wavelet packet entropy formula. The wavelet packet entropy of the topological feature tensor is subjected to the attention aggregation formula to generate the frequency domain topological coding vector Z2.

[0015] Furthermore, the topological feature tensor Xtopo analysis steps are:

[0016] Input completion adversarial tensor Xfinal constructs a dynamic topology network of devices. For each time slice t, the connection relationship between devices is extracted from the hyperedge weight to form a weighted graph Q(t)=(G,E,M), where G is a node and E is a hyperedge E={v l}, M is the comprehensive interaction strength, and the weighted graph of each time slice is constructed by increasing the weight threshold to construct a filter complex, track the birth and death thresholds of topological features in the filtering process, and generate a persistent homology barcode Barcode (Q (t)). i g j Calculate the persistence strength Pij(t) of its connection relationship, map the persistence strength matrix to a tensor channel, stack the channels of all time slices along the time axis, and obtain the topological feature tensor Xtopo, Xtopo∈R T×4×4×1 .

[0017] Furthermore, the corresponding fault level status analysis step:

[0018] A hybrid kernel model of solar device faults is established using a support vector machine through a hybrid kernel function, and a hierarchical classification strategy is adopted: the first-level classifier distinguishes the major fault categories based on the weight vector and bias, and the second-level classifier further refines the subclasses. The model outputs the fault status value through the decision function and maps the value to the fault probability. The model is optimized through the loss function, and the optimization process is accelerated by sequence minimization. The model determines the specific fault level status according to the fault probability thresholds JZ1 and JZ2, and the corresponding levels are normal, warning and severe. The final fusion optimization coding feature Zc calculated by the model input outputs the corresponding fault level status.

[0019] Furthermore, the solar device fault regression analysis model analysis steps are: The processed multi-source heterogeneous monitoring parameter set of solar energy devices is used to construct the time window sequence parameter Xt∈R Tw ×d1 , where Tw is the historical time step, d1 is the feature dimension, and a multivariate LSTM network is used to build a solar device failure time series prediction model, including: ht=LSTM(Xt,ht−1), S t+Δt =Ws×ht+bt, where ht represents the hidden state at the current time t, Xt represents the input feature of the multi-source heterogeneous monitoring parameter set at time t, S t+Δt Represents the parameter prediction value at the future Δt moment, Ws represents the output weight matrix, bt represents the output bias term, and the mean square error loss function Lts=1 / N×∑ Training model, where N represents the number of samples, represents the true value of the future parameter, and S represents the model predicted value; The predicted future parameters S output by the time series forecasting model t+Δt , finally fuse the optimized encoded feature vector Zc and the current time window sequence parameter Xt in real time, merge them into the input comprehensive feature Xcomb, build a solar device fault regression analysis model, set the target variable: the fault probability Pfault in the future Δt, and Pfault∈[0,1], including the gradient boosting tree to establish a nonlinear mapping between comprehensive features and fault probability: Pfault=∑ γm×hm(Xcomb), where M is the number of trees in the regression analysis model, γm is the weight of the mth tree, and hm is the mapping function of the mth tree to Xcomb. The model is trained by minimizing the cross-entropy loss function. The loss function is specifically LReg=-∑[b×logP+(1-b)×log(1−P)], where b is the true fault label, b=1 for fault and b=0 for no fault, and P is the fault probability predicted by the model. The trained model inputs real-time parameters and outputs the fault probability Pfault in the future Δt time. For short-term warnings, if the output fault probability Pfault is greater than the set threshold YH1, a red emergency response is initiated. For medium-term warnings, if the cumulative probability is greater than the set threshold YH2, an orange alarm is triggered. For long-term warnings, if the trend slope dPfault / dt is greater than the set threshold YH3, a yellow alarm is pushed.

[0020] As a preferred embodiment of the present invention, a remote fault diagnosis system for a solar device includes: an acquisition module, a preprocessing module, a fusion module, a fault module, and a prediction module. The acquisition module is used to obtain solar panel monitoring parameters, inverter monitoring parameters, energy storage system monitoring parameters, and charge controller monitoring parameters through provided sensors to form a multi-source heterogeneous monitoring parameter set for the solar device; The preprocessing module is used to remove abnormal data values by building a multivariate outlier detection model, constructing a dynamic spatiotemporal hypergraph of the solar installation to form a spatiotemporal hypergraph tensor X. Using the generator and discriminator, after tensor convolution, LSTM, and loss function optimization, the final output is a completed adversarial tensor. The fusion module is used to input the completed adversarial tensor to construct the device dynamic topology network, generate the topological feature tensor, use the spatiotemporal graph convolutional network and wavelet packet decomposition analysis, and use the gated interaction mechanism to fuse the analysis to obtain the final fused optimized coding feature vector; The fault module uses a support vector machine to establish a hybrid kernel model of solar device faults through a hybrid kernel function. The model determines the specific fault level state based on the fault probability threshold and outputs the corresponding fault level state; The prediction module uses a multivariate LSTM network to build a solar device failure time series prediction model to obtain future parameter prediction values, and build a solar device failure regression analysis model. The model inputs real-time parameters and outputs failure probability, and sets short-term, medium-term and long-term warnings based on different thresholds.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention obtains solar panel monitoring parameters, inverter monitoring parameters, energy storage system monitoring parameters, and charge controller monitoring parameters through the provided sensors to form a multi-source heterogeneous monitoring parameter set for solar devices. Abnormal data values are removed by constructing a multivariate outlier detection model, and a dynamic spatiotemporal hypergraph of the solar device is constructed to form a spatiotemporal hypergraph tensor. The generator and discriminator are used to perform tensor convolution, LSTM, and loss function optimization to finally output a completed adversarial tensor to complete data preprocessing. The completed adversarial tensor is input to construct a dynamic topological network of the device and generate a topological feature tensor. The spatiotemporal graph convolution network and wavelet packet decomposition analysis are used, and the gated interaction mechanism is used for fusion analysis to obtain the final fused optimized coding feature vector. The method can accurately perform deep modeling of variable correlations on multi-source data, reduce the accumulation of subsequent analysis errors, and simultaneously integrate temporal causality and physical coupling between devices to capture complex fault causes and comprehensively characterize fault modes. 2. The present invention establishes a hybrid kernel model of solar device failure by using a support vector machine through a hybrid kernel function. The model determines the specific fault level state according to the fault probability threshold and outputs the corresponding fault level state. A multivariate LSTM network is used to construct a solar device failure time series prediction model to obtain future parameter prediction values, and a solar device failure regression analysis model is constructed. The model inputs real-time parameters and outputs failure probability. Short-term, medium-term and long-term warnings are set according to different thresholds, which can accurately determine the fault level and accurately perform differentiated analysis of short-term sudden failures, medium-term performance degradation and long-term trends of solar devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. The following drawings are not intentionally scaled to the actual size, and the focus is on illustrating the main purpose of the present invention.

[0023] Figure 1 is a flow chart of the method of the present invention; Figure 2 is a system block diagram of the present invention; Figure 3 It is a process framework diagram of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts also fall within the scope of protection of the present invention.

[0025] like Figure 1 、 Figure 3As shown, a remote fault diagnosis method for a solar device includes the following steps: Step S100: Obtaining solar panel monitoring parameters, inverter monitoring parameters, energy storage system monitoring parameters, and charge controller monitoring parameters through the provided sensors to form a multi-source heterogeneous monitoring parameter set for the solar device; Specifically, the monitoring parameters of the solar panel are open-circuit voltage, operating current, output power, branch voltage extreme difference rate, and light utilization rate. The branch voltage extreme difference rate is the branch maximum voltage minus the branch minimum voltage divided by the branch average voltage. The monitoring parameters of the inverter are input voltage, output frequency, conversion efficiency, internal temperature, and harmonic distortion. Harmonic distortion refers to the ratio of harmonic components to fundamental components in an electrical signal (current or voltage). The monitoring parameters of the energy storage system are discharge current, remaining capacity, battery temperature, and internal resistance growth rate. The monitoring parameters of the charge controller are photovoltaic input voltage, battery terminal voltage, and load terminal current. Step S200: A multivariate outlier detection model is constructed to remove abnormal data values, and a dynamic spatiotemporal hypergraph of the solar device is constructed to form a spatiotemporal hypergraph tensor X. Using the generator and discriminator, tensor convolution, LSTM, and loss function optimization are performed to finally output the completed adversarial tensor Xfinal, completing data preprocessing. Step S300: Input the completed adversarial tensor to construct a dynamic topology network of the device, generate a topology feature tensor, use the spatiotemporal graph convolutional network and wavelet packet decomposition analysis, and use the gated interaction mechanism for fusion analysis to obtain the final fusion optimized coding feature vector Zc; Step S400: Using a support vector machine to establish a hybrid kernel model of solar device faults through a hybrid kernel function. The model determines a specific fault level state based on a fault probability threshold and outputs the corresponding fault level state. Step S500: A multivariate LSTM network is used to construct a solar device failure time series prediction model to obtain future parameter prediction values, and a solar device failure regression analysis model is constructed. The model inputs real-time parameters and outputs failure probability, and sets short-term, medium-term, and long-term warnings based on different thresholds.

[0026] The specific analysis of the data preprocessing in step S200 is as follows: Step S201: Constructing a multivariate outlier detection model to remove abnormal data values, specifically including: , where x represents a p-dimensional data point vector, including each feature in the solar device multi-source heterogeneous monitoring parameter set, a vector composed of the solar device multi-source heterogeneous monitoring parameter set, µ is a p-dimensional mean vector, composed of the average values of each feature, and S is a p×p covariance matrix. Represents the Mahalanobis distance, which measures the distance between the data point x and the mean vector μ, and T represents the transpose of the matrix. If Then the data point vector is determined to be abnormal and the data point vector is removed, where represents the 99% quantile of the p-dimensional chi-square distribution; Step S202: Construct a spatiotemporal hypergraph of the solar device state, with the solar panels, inverters, energy storage systems, and charge controllers as device nodes, where the device nodes G = {g1, g2, ..., g4}, g1 is the node solar panel, and after removing the abnormal parameter data values, specifically the open circuit voltage, operating current, output power, branch voltage extreme difference rate, and light utilization rate, determine the g2, g3, and g4 device nodes in turn, and construct a hyperedge v l Connect the device node to the device g i and g j Perform physical coupling function fphys processing to obtain the physical coupling relationship value, and perform physical coupling function fphys processing on the device g i and g j Perform Granger causality test to obtain the temporal causality value, multiply the physical coupling relationship value and the temporal causality value by the learnable weight respectively, and perform Sigmoid normalization function on the result of the sum of the two to obtain the hyperedge v l The weight at time t generates the time slice adjacency tensor B(t)∈R 4×4×6 , where dimensions 1-2 represent node pairs consisting of 4 devices, the third dimension corresponds to 6 hyperedge types, R represents a set of real numbers, and the element value Bijk(t) is the hyperedge v l The weight at time t, specifically Bijk(t)=ω(v l , t), stack the T time slice tensors according to the time dimension to form a spatiotemporal hypergraph tensor X, X=Stack(B(t1),B(t2),...,B(tT)), and X∈R T×4×4×6 , Stack represents stacking tensor operations; Specifically, the weight is dynamically calculated based on the physical relationship between parameters: ω(v l , t) = σ(β×fphys(g i , g j )+ɑ×Granger(g i →g j ), where fphys(*) is the physical coupling function, Granger(*) is the Granger causality test value, which quantifies the temporal causality between device parameters, β and ɑ are learnable weights, and σ(*) is the Sigmoid normalization function; More specifically, the physical coupling relationship value fphys(g i , g j ) solution steps: Equipment i and g jThe Pearson correlation coefficient is calculated using historical monitoring data from the parameter set. When it is greater than the set threshold, the parameter is retained, and then the key parameters with physical connections are screened out. The function parameters are determined with the help of the least squares method to establish physical relationships. These parameters are constructed into a matrix and standardized. The covariance matrix is calculated, and its eigenvalues are decomposed. The principal component with a cumulative contribution rate greater than the threshold is selected. If the coefficient of the original parameter in the principal component is greater than the set threshold, the parameter is attributed to the key influencing factor to form a simplified key parameter subset. The data set of the simplified key parameter subset is divided into a training set and a test set in a ratio of 7:3. The random forest regression model is initialized and the initial parameters are set. The simplified key parameter subset of the training set is used as input and the target physical coupling value is used as output for training. The model learns the mapping relationship and predicts the physical coupling value of the test set. Through the mean square error, if it is greater than the set threshold, the model parameters are adjusted. Different parameter combinations are preset by grid search. The optimal parameters are found through cross-validation to obtain the optimized model. The input device g i and g j The actual parameter set of the physical coupling relationship value fphys (g i , g j ); More specifically, the steps for solving the temporal causality value are: Collection equipment i and g j The time series data of the parameters are aligned and the optimal lag order is determined by using the information criterion to construct the device-free g i The autoregressive model of the influence includes x j,t =η0+∑η k ×x j,t-k +ϵ 1，t , x j,t Indicates device g j At the time t, a parameter is observed, k represents the index variable of the lag order, η0 represents the constant term, η k represents the autoregressive coefficient, ϵ 1，t Represents the error term, constructs a device containing g i The autoregressive model of the influence includes x j,t =δ0+∑δ k ×x j,t-k +∑Φ k ×x i,t-k +ϵ 2，t , δ0 represents a constant term, η k and Φ k represents the autoregressive coefficient, ϵ 2，t Represents the error term, calculates the residual sum of squares of the two models, and constructs the F statistic, where , where p is the lag order, N is the total number of observations of time series data, SR1 and SR2 are respectivelyi and contains equipment g i The residual sum of squares of the affected model follows an F distribution with degrees of freedom (p,N−2p−1). If F>F Ω (p,N−2p−1), then reject the null hypothesis and mark the F value at this time as the temporal causality value. Otherwise, the null hypothesis cannot be rejected, set the F value at this time to zero, and mark it as the temporal causality value.

[0027] Step S203: The generator G inputs the spatiotemporal hypergraph tensor X and the latent variable K. With the help of the spatiotemporal hypergraph tensor and the corresponding LSTM hidden state at the previous moment, the current hidden state is obtained through tensor convolution and LSTM processing, and the completed tensor Xgen is generated with the latent variable K. The discriminator D inputs the completed tensor Xgen, performs feature extraction through multi-scale three-dimensional convolution, and then uses the Sigmoid function to output the probability value of the real tensor. The Tucker norm difference between the completed tensor Xgen and the real tensor Xreal is calculated through the loss function. The generator and the discriminator optimize their respective parameters through the back-propagation algorithm, and continuously iterate and adjust until the value of the loss function is less than the set threshold. The operation is terminated and the completed adversarial tensor Xfinal is output.

[0028] Specifically, the tensor convolution-LSTM generates the completion tensor Xgen, specifically: Wt=LSTM(TensorConv(Xt-1, Wt-1)), Xgen=G(Wt, K), where Xt-1 is the spatiotemporal hypergraph tensor at the previous moment, Wt-1 is the hidden state of the LSTM at the previous moment, LSTM(*) is a long short-term memory network, TensorConv(*) is a tensor convolution operation, G(Wt, K) is the completed tensor output by the generator, and the hidden variable K is the implicit encoding of the device state or environmental factors; Specifically, the probability value P of the true tensor is: P=D(Xgen)=Sigmoid(∑Conv3D s (Xgen)), where Conv3D s (*) is the s-th three-dimensional convolution operation, and Sigmoid (*) is the activation function; Specifically, the loss function L: Generator G: LG = -E[logD(Xgen)] + γ1 Discriminator D: LD = -E[logD(Xreal)] -E[log(1-D(Xgen)] + γ2 , where E represents the mathematical expectation, γ1 and γ2 represent hyperparameters, represents the Tucker norm, and Xreal is the real tensor set in the database.

[0029] Specifically, by constructing a multivariate outlier detection model, outliers in the multi-source heterogeneous monitoring data of solar energy devices are removed. Taking devices as nodes, the hyperedge weights are determined through physical coupling functions and Granger causality tests, and a dynamic spatiotemporal hypergraph of solar energy devices is constructed. The dynamic interaction relationship between devices is modeled and the correlation characteristics of multiple dimensions (timing dependence, physical connection) are captured to provide a basis for structuring spatiotemporal data into tensors. A generative adversarial network is used to allow the generator to generate a completion tensor through tensor convolution and LSTM processing. The generation ability is used to complete the missing data in the spatiotemporal hypergraph tensor, retain the spatiotemporal structure and dynamic characteristics of the data, and convert the multi-source heterogeneous monitoring data into a structured, noise-free and complete spatiotemporal tensor, laying the foundation for subsequent dynamic topology analysis, feature extraction and fault prediction of equipment.

[0030] The specific analysis of the final fusion optimization coding feature vector Zc in step S300 is as follows: Input completion adversarial tensor Xfinal constructs a dynamic topology network of devices. For each time slice t, the connection relationship between devices is extracted from the hyperedge weight to form a weighted graph Q(t)=(G,E,M), where G is a node and E is a hyperedge E={v l}, M is the comprehensive interaction strength, and the weighted graph of each time slice is constructed by increasing the weight threshold to construct a filter complex, track the birth and death thresholds of topological features in the filtering process, and generate a persistent homology barcode Barcode (Q (t)). i g j Calculate the persistence strength Pij(t) of its connection relationship, map the persistence strength matrix to a tensor channel, stack the channels of all time slices along the time axis, and obtain the topological feature tensor Xtopo, Xtopo∈R T×4×4×1 ; Specifically, the analysis of the comprehensive interaction intensity M is as follows: Mij(t)=∑ Bijk(t).

[0031] The persistence strength Pij(t) is analyzed as follows: Pij(t)=∑ (bk-dk)×ζij(k), where dk is the weight threshold when topological feature k is born, bk is the weight threshold when topological feature k is dead, and ζij(k) is the indicator function. If topological feature k involves devices i and j, then ζij(k)=1, otherwise it is 0.

[0032] The topological feature tensor Xtopo is processed by the spatiotemporal graph convolutional network, and the time domain dynamic coding vector Z1 is generated through the feature extraction formula. The topological feature tensor Xtopo is decomposed by wavelet packets, and the wavelet packet entropy of the topological feature tensor is obtained through the wavelet packet entropy formula. The wavelet packet entropy of the topological feature tensor is subjected to the attention aggregation formula to generate the frequency domain topological coding vector Z2. Specifically, the feature extraction formula is Z1=STGCN(Xtopo)=∑ (H -1 / 2 A k H -1 / 2 ×Xtopo×Wk), where STGCN is a spatiotemporal graph convolutional network, A k is the dynamic adjacency matrix derived from the hyperedge weights, H is the degree matrix, Wk is the learnable parameter matrix, k is the sum index, and the maximum value is K; Specifically, the wavelet packet entropy formula is: E i，q (t)=- ,in are the coefficients after wavelet packet decomposition, the characteristics of different frequency sub-bands i, q and time t, is the square of the norm; The attention aggregation formula is Z2=Softmax(Q×K T / d 1 / 2 )V, where Q,K,V=MLP(E i，q ), where Q, K, V are the query, key, and value in the attention mechanism, d is the dimension parameter, MLP (*) is a multi-layer perceptron, and Softmax is the activation function.

[0033] The time domain dynamic coding vector Z1 and the frequency domain topology coding vector Z2 are nonlinearly fused using a gated interaction mechanism. The fusion strength of the two is adaptively controlled using dynamic gating weights to obtain a fused feature coding vector. The fused feature coding vector is used to extract temporal dynamic information through a multi-head self-attention mechanism, and a dynamic adjustment vector is generated by lightweight convolution. The dynamic adjustment vector is used to perform cross-attention optimization on the time domain dynamic coding vector Z1 and the frequency domain topology coding vector Z2 to obtain the optimized time domain dynamic coding vector and frequency domain topology coding vector. The optimized time domain dynamic coding vector and the frequency domain topology coding vector are adaptively weighted fused by dynamically calculating weights to obtain the final fused optimized coding feature vector Zc. Specifically, the above process of obtaining the final fusion optimization coding feature can be expressed as follows: g = σ (Wg × [Z1; Z2] + bg); Zj=g⊙tanh(Wf×Z1)+(1-g)⊙tanh(Wf×Z2); Q3=Wq×Zj, K3=Wk×Zj, V3=Wv×Zj; Attention(Q3, K3, V3)=softmax(Q3×K3 T / dk 1 / 2 +mv)V3; mv=Conv1D 1×1 (Zj); Z1′=LayerNorm(Attention(Q1,K1+mv,V1); Q1=Wq1×Z1, K1=Wk1×V1; Z2′=LayerNorm(Attention(Q2,K2+mv,V2); Q2=Wq2×Z2, K2=Wk2×V2; Ω=σ(WΩ[Z1′, Z2′]+bΩ; Zc=Ω⊙Z1′+(1-Ω)⊙Z2′; Among them, g is the dynamic gating weight, σ is the Sigmoid function, Wg, Wf, Wq, Wk, Wv, Wq1, Wk1, Wq2, Wk2 and WΩ are the corresponding learnable weight matrices, bg and bΩ are bias terms, ⊙ is element-by-element multiplication, Zj is the fused feature encoding vector, tanh is the hyperbolic tangent activation function, Q3, K3, V3 are query, key, and value vectors respectively, mv is the dynamic adjustment vector, dk is the dimension of the key vector, Conv1D 1×1 is the one-dimensional convolution in deep learning, LayerNorm is the layer normalization function, Q2, K2, and V2 are the query, key, and value vectors after Z2 mapping, respectively. Q1, K1, and V1 are the query, key, and value vectors after Z1 mapping, respectively. Ω is the dynamic fusion weight, and Zc is the final fused optimized encoding feature vector.

[0034] Specifically, the input completion adversarial tensor is used to construct a dynamic topological network of the device, the connection relationship is extracted to generate a persistent homology barcode, the persistence strength is calculated to obtain the topological feature tensor, and the time domain and frequency domain coding vectors are obtained respectively through spatiotemporal graph convolution, wavelet packet decomposition and attention aggregation. Then, nonlinear fusion and optimization are achieved through gated interaction, multi-head self-attention, etc., and finally the final fused optimized coding feature vector is obtained. The final fused optimized coding feature vector is a bridge from data preprocessing to the establishment of a solar device fault model and a solar device fault regression analysis model. It extracts and fuses the spatiotemporal and frequency domain features, enhances the feature expression capability, and makes the model output results more accurate.

[0035] The specific steps S400 and the detailed analysis of the solar device failure model are as follows: The support vector machine is used to process the nonlinearity of high-dimensional features through a hybrid kernel function to establish a hybrid kernel model for solar device faults, balancing nonlinear classification capabilities and computational efficiency. A hierarchical classification strategy is adopted: the first-level classifier distinguishes major fault categories based on weight vectors and biases, such as electrical or mechanical faults, and the second-level classifier further refines subcategories, such as voltage sag or component aging. The model outputs fault status values through a decision function and maps the values to fault probabilities, which are optimized through a loss function. The optimization process accelerates training through sequence minimization. The model determines the specific fault level status based on the fault probability thresholds JZ1 and JZ2, and the corresponding levels are normal, warning, and severe. Specifically, values greater than JZ1 correspond to severe, values between JZ1 and JZ2 correspond to warnings, and values less than JZ2 correspond to normal. The model inputs the final fusion optimization encoding feature Zc calculated to output the corresponding fault level status; Specifically, the above process of determining a specific fault level state can be expressed as follows: F(Zc （i） , Zc （j） ) = exp (-ψ1 )+ψ2×Zc （i） ×Zc （j） ; f(Zc （i） )=w1 T ×ϕ(Zc （i） )+b1; Pi=1 / (1+exp(-f(Zc （i） ))); L=∑ max(0,1−yi(w1 T ×ϕ(Zc （i） )+b1)+λ× ;

[0036] Where ψ1 is the control kernel width coefficient, Zc （i） 、Zc （j） The final fusion optimization encoding feature vector of the sample, w1 is the weight vector of the first-level classifier, b1 is the bias of the first-level classifier, F(*) is the kernel function, exp is the exponential function with the natural parameter e as the base, ψ2 is the linear term weight, Pi is the probability that the sample belongs to a certain fault category, f(*) is the fault state value of the sample belonging to a certain fault category, L is the loss function, yi is the sample label, ϕ(*) represents the kernel mapping, and λ is the regularization coefficient.

[0037] Specifically, the hybrid kernel function balances the nonlinear classification capability and computational efficiency, and the hierarchical classification strategy refines the fault classification from large categories to subcategories, which can improve the accuracy of fault diagnosis and quickly input features to output fault levels.

[0038] The processed multi-source heterogeneous monitoring parameter set of solar energy devices is used to construct the time window sequence parameter Xt∈R Tw ×d1 , where Tw is the historical time step, d1 is the feature dimension, and a multivariate LSTM network is used to build a solar device failure time series prediction model, including: ht=LSTM(Xt,ht−1), S t+Δt =Ws×ht+bt, where ht represents the hidden state at the current time t, Xt represents the input feature of the multi-source heterogeneous monitoring parameter set at time t, S t+Δt Represents the parameter prediction value at the future Δt moment, Ws represents the output weight matrix, bt represents the output bias term, and the mean square error loss function Lts=1 / N×∑ Training model, where N represents the number of samples, represents the true value of the future parameter, and S represents the model predicted value; The predicted future parameters S output by the time series forecasting model t+Δt , finally fuse the optimized encoded feature vector Zc and the current time window sequence parameter Xt in real time, merge them into the input comprehensive feature Xcomb, build a solar device fault regression analysis model, set the target variable: the fault probability Pfault in the future Δt, and Pfault∈[0,1], including the gradient boosting tree to establish a nonlinear mapping between comprehensive features and fault probability: Pfault=∑ γm×hm(Xcomb), where M is the number of trees in the regression analysis model, γm is the weight of the mth tree, and hm is the mapping function of the mth tree to Xcomb. The model is trained by minimizing the cross-entropy loss function. The loss function is specifically LReg=-∑[b×logP+(1-b)×log(1−P)], where b is the true fault label, b=1 for fault and b=0 for no fault, and P is the fault probability predicted by the model. The trained model inputs real-time parameters and outputs the fault probability Pfault within the future Δt time. For short-term warnings, if the output fault probability Pfault is greater than the set threshold YH1, a red emergency response is initiated. For medium-term warnings, if the cumulative probability is greater than the set threshold YH2, an orange alarm is triggered. For long-term warnings, if the trend slope dPfault / dt is greater than the set threshold YH3, a yellow warning is pushed. Specifically, short-term warnings, medium-term warnings, and long-term warnings are set according to different solar energy devices, for example, within 1 hour, 6 hours, and 24 hours, respectively.

[0039] like Figure 3 As shown, a remote fault diagnosis system for solar energy devices includes an acquisition module, a preprocessing module, a fusion module, a fault module and a prediction module; The acquisition module is used to obtain the monitoring parameters of the solar panel, inverter, energy storage system and charge controller through the set sensors, forming a multi-source heterogeneous monitoring parameter set of the solar device; The preprocessing module is used to remove abnormal data values by building a multivariate outlier detection model, constructing a dynamic spatiotemporal hypergraph of the solar installation to form a spatiotemporal hypergraph tensor X. Using the generator and discriminator, after tensor convolution, LSTM, and loss function optimization, the final output is the completed adversarial tensor Xfinal. The fusion module is used to input the completed adversarial tensor to construct the device dynamic topology network, generate the topological feature tensor, use the spatiotemporal graph convolutional network and wavelet packet decomposition analysis, and use the gated interaction mechanism to fuse the analysis to obtain the final fused optimized coding feature vector Zc; The fault module uses a support vector machine to establish a hybrid kernel model of solar device faults through a hybrid kernel function. The model determines the specific fault level state based on the fault probability threshold and outputs the corresponding fault level state; The prediction module uses a multivariate LSTM network to build a solar device failure time series prediction model to obtain future parameter prediction values, and build a solar device failure regression analysis model. The model inputs real-time parameters and outputs failure probability, and sets short-term, medium-term and long-term warnings based on different thresholds.

[0040] The above is an illustration of the present invention and should not be considered as limiting thereof. Although several exemplary embodiments of the present invention have been described, it will be readily understood by those skilled in the art that many modifications may be made to the exemplary embodiments without departing from the novel teachings and advantages of the present invention. Therefore, all such modifications are intended to be included within the scope of the present invention as defined by the claims. It should be understood that the above is an illustration of the present invention and should not be considered as being limited to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present invention is defined by the claims and their equivalents.

Claims

1. A remote fault diagnosis method for a solar device, characterized in that: The following steps are involved: Step S100: Acquire solar panel monitoring parameters, inverter monitoring parameters, energy storage system monitoring parameters, and charge controller monitoring parameters to form a multi-source heterogeneous monitoring parameter set for solar devices; Step S200: A multivariate outlier detection model is constructed to remove abnormal data values, and a dynamic spatiotemporal hypergraph of the solar device is constructed to form a spatiotemporal hypergraph tensor. The generator and discriminator are used to perform tensor convolution, LSTM, and loss function optimization to finally output a completed adversarial tensor, completing data preprocessing. Step S300: Input the completed adversarial tensor to construct a dynamic topology network of the device, generate a topology feature tensor, use the spatiotemporal graph convolutional network and wavelet packet decomposition analysis, and use the gated interaction mechanism for fusion analysis to obtain the final fusion optimized coding feature vector; Step S400: Using a support vector machine to establish a hybrid kernel model of solar device faults through a hybrid kernel function. The model determines a specific fault level state based on a fault probability threshold and outputs the corresponding fault level state. Step S500: A multivariate LSTM network is used to construct a solar device failure time series prediction model to obtain future parameter prediction values, and a solar device failure regression analysis model is constructed. The model inputs real-time parameters and outputs failure probability, and sets short-term, medium-term, and long-term warnings based on different thresholds.

2. A remote fault diagnosis method for a solar device according to claim 1, characterized in that: The steps of completing the adversarial tensor analysis are as follows: Step S201: Constructing a multivariate outlier detection model to remove abnormal data values, including: , where x represents a p-dimensional data point vector, including the features in the solar device multi-source heterogeneous monitoring parameter set, and the vector composed of the solar device multi-source heterogeneous monitoring parameter set, µ is the p-dimensional mean vector, and S is the p×p covariance matrix. Represents the Mahalanobis distance, T represents the transpose of the matrix, if Then the data point vector is determined to be abnormal and the data point vector is removed, where represents the 99% quantile of the p-dimensional chi-square distribution; Step S202: Construct a spatiotemporal hypergraph of solar device states, using solar panels, inverters, energy storage systems, and charge controllers as device nodes, and construct hyperedges v l Connect the device node to the device g i and g j Perform physical coupling function fphys processing to obtain the physical coupling relationship value, and perform physical coupling function fphys processing on the device g i and g j Perform Granger causality test to obtain the temporal causality value, multiply the physical coupling relationship value and the temporal causality value by the learnable weight respectively, and perform Sigmoid normalization function on the result of the sum of the two to obtain the hyperedge v l The weight at time t generates the time slice adjacency tensor B(t)∈R 4×4×6 , where dimensions 1-2 represent node pairs consisting of 4 devices, the third dimension corresponds to 6 hyperedge types, R represents a set of real numbers, and the element value Bijk(t) is the hyperedge v l The weight at time t, Bijk(t)=ω(v l , t), stack the T time slice tensors according to the time dimension to form a spatiotemporal hypergraph tensor X, X=Stack(B(t1),B(t2),...,B(tT)), and X∈R T×4×4×6 , Stack represents stacking tensor operations; Step S203: The generator G inputs the spatiotemporal hypergraph tensor X and the latent variable K. With the help of the spatiotemporal hypergraph tensor and the corresponding LSTM hidden state at the previous moment, the current hidden state is obtained through tensor convolution and LSTM processing, and the completed tensor Xgen is generated with the latent variable K. The discriminator D inputs the completed tensor Xgen, performs feature extraction through multi-scale three-dimensional convolution, and then uses the Sigmoid function to output the probability value of the real tensor. The Tucker norm difference between the completed tensor Xgen and the real tensor Xreal is calculated through the loss function. The generator and the discriminator optimize their respective parameters through the back-propagation algorithm, and continuously iterate and adjust until the value of the loss function is less than the set threshold. The operation is terminated and the completed adversarial tensor Xfinal is output.

3. A remote fault diagnosis method for a solar device according to claim 2, characterized in that: The physical coupling relationship value analysis steps: Equipment i and g j The Pearson correlation coefficient is calculated using historical monitoring data from the parameter set. When it is greater than the set threshold, the parameter is retained, and then the key parameters with physical connections are screened out. The function parameters are determined with the help of the least squares method to establish physical relationships. These parameters are constructed into a matrix and standardized. The covariance matrix is calculated, and its eigenvalues are decomposed. The principal component with a cumulative contribution rate greater than the threshold is selected. If the coefficient of the original parameter in the principal component is greater than the set threshold, the parameter is attributed to the key influencing factor to form a simplified key parameter subset. The data set of the simplified key parameter subset is divided into a training set and a test set in a ratio of 7:

3. The random forest regression model is initialized and the initial parameters are set. The simplified key parameter subset of the training set is used as input and the target physical coupling value is used as output for training. The model learns the mapping relationship and predicts the physical coupling value of the test set. Through the mean square error, if it is greater than the set threshold, the model parameters are adjusted. Different parameter combinations are preset by grid search. The optimal parameters are found through cross-validation to obtain the optimized model. The input device g i and g j The actual parameter set of the physical coupling relationship value fphys (g i , g j ).

4. A remote fault diagnosis method for a solar device according to claim 2, characterized in that: The temporal causality value analysis steps are: Collection equipment i and g j The time series data of the parameters are aligned and the optimal lag order is determined by using the information criterion to construct the device-free g i The autoregressive model of the influence includes x j,t =η0+∑η k ×x j,t-k +ϵ 1，t , x j,t Indicates device g j At the time t, a parameter is observed, k represents the index variable of the lag order, η0 represents the constant term, η k represents the autoregressive coefficient, ϵ 1，t Represents the error term, constructs a device containing g i The autoregressive model of the influence includes x j,t =δ0+∑δ k ×x j,t-k +∑Φ k ×x i,t-k +ϵ 2，t , δ0 represents a constant term, η k and Φ k represents the autoregressive coefficient, ϵ 2，t Represents the error term, calculates the residual sum of squares of the two models, and constructs the F statistic, where , where p is the lag order, N is the total number of observations of time series data, SR1 and SR2 are respectively i and contains equipment g i The residual sum of squares of the affected model follows an F distribution with degrees of freedom (p,N−2p−1). If F>F Ω (p, N−2p−1), then reject the null hypothesis and mark the F value at this time as the temporal causality value. Otherwise, the null hypothesis cannot be rejected, set the F value at this time to zero, and mark it as the temporal causality value to obtain the corresponding temporal causality value.

5. A remote fault diagnosis method for a solar device according to claim 1, characterized in that: The final fusion optimization coding feature vector analysis steps: The time domain dynamic coding vector Z1 and the frequency domain topology coding vector Z2 are nonlinearly fused using the gated interaction mechanism, and the fusion strength of the two is adaptively controlled using dynamic gating weights to obtain a fused feature coding vector. The fused feature coding vector is used to extract temporal dynamic information through a multi-head self-attention mechanism, and a dynamic adjustment vector is generated by lightweight convolution. The dynamic adjustment vector is used to perform cross-attention optimization on the time domain dynamic coding vector Z1 and the frequency domain topology coding vector Z2 to obtain the optimized time domain dynamic coding vector and frequency domain topology coding vector. The optimized time domain dynamic coding vector and frequency domain topology coding vector are adaptively weighted fused by dynamically calculating weights to obtain the final fused optimized coding feature vector Zc.

6. A remote fault diagnosis method for a solar device according to claim 5, characterized in that: The analysis steps of the time domain dynamic coding vector Z1 and the frequency domain topology coding vector Z2 are as follows: The topological feature tensor Xtopo is processed by the spatiotemporal graph convolutional network, and the time domain dynamic coding vector Z1 is generated through the feature extraction formula. The topological feature tensor Xtopo is decomposed into wavelet packets, and the wavelet packet entropy of the topological feature tensor is obtained through the wavelet packet entropy formula. The wavelet packet entropy of the topological feature tensor is subjected to the attention aggregation formula to generate the frequency domain topological coding vector Z2.

7. A remote fault diagnosis method for a solar device according to claim 6, characterized in that: The topological characteristic tensor Xtopo analysis steps are: Input completion adversarial tensor Xfinal constructs a dynamic topology network of devices. For each time slice t, the connection relationship between devices is extracted from the hyperedge weight to form a weighted graph Q(t)=(G,E,M), where G is a node and E is a hyperedge E={v l }, M is the comprehensive interaction strength, and the weighted graph of each time slice is constructed by increasing the weight threshold to construct a filter complex, track the birth and death thresholds of topological features in the filtering process, and generate a persistent homology barcode Barcode (Q (t)). i g j Calculate the persistence strength Pij(t) of its connection relationship, map the persistence strength matrix to a tensor channel, stack the channels of all time slices along the time axis, and obtain the topological feature tensor Xtopo, Xtopo∈R T×4×4×1 .

8. A remote fault diagnosis method for a solar device according to claim 1, characterized in that: The corresponding fault level status analysis steps: A hybrid kernel model of solar device faults is established using a support vector machine through a hybrid kernel function, and a hierarchical classification strategy is adopted: the first-level classifier distinguishes the major fault categories based on the weight vector and bias, and the second-level classifier further refines the subclasses. The model outputs the fault status value through the decision function and maps the value to the fault probability. The model is optimized through the loss function, and the optimization process is accelerated by sequence minimization. The model determines the specific fault level status according to the fault probability thresholds JZ1 and JZ2, and the corresponding levels are normal, warning and severe. The final fusion optimization coding feature Zc calculated by the model input outputs the corresponding fault level status.

9. A remote fault diagnosis method for a solar device according to claim 1, characterized in that: The analysis steps of the solar device fault regression analysis model are as follows: The processed multi-source heterogeneous monitoring parameter set of solar energy devices is used to construct the time window sequence parameter Xt∈R Tw ×d1 , where Tw is the historical time step, d1 is the feature dimension, and a multivariate LSTM network is used to build a solar device failure time series prediction model, including: ht=LSTM(Xt,ht−1), S t+Δt =Ws×ht+bt, where ht represents the hidden state at the current time t, Xt represents the input feature of the multi-source heterogeneous monitoring parameter set at time t, S t+Δt Represents the parameter prediction value at the future Δt moment, Ws represents the output weight matrix, bt represents the output bias term, and the mean square error loss function Lts=1 / N×∑ Training model, where N represents the number of samples, represents the true value of the future parameter, and S represents the model predicted value; The predicted future parameters S output by the time series forecasting model t+Δt , finally fuse the optimized encoded feature vector Zc and the current time window sequence parameter Xt in real time, merge them into the input comprehensive feature Xcomb, build a solar device fault regression analysis model, set the target variable: the fault probability Pfault in the future Δt, and Pfault∈[0,1], including the gradient boosting tree to establish a nonlinear mapping between comprehensive features and fault probability: Pfault=∑ γm×hm(Xcomb), where M is the number of trees in the regression analysis model, γm is the weight of the mth tree, and hm is the mapping function of the mth tree to Xcomb. The model is trained by minimizing the cross-entropy loss function. The loss function is specifically LReg=-∑[b×logP+(1-b)×log(1−P)], where b is the true fault label, b=1 for fault and b=0 for no fault, and P is the fault probability predicted by the model. The trained model inputs real-time parameters and outputs the fault probability Pfault in the future Δt time. For short-term warnings, if the output fault probability Pfault is greater than the set threshold YH1, a red emergency response is initiated. For medium-term warnings, if the cumulative probability is greater than the set threshold YH2, an orange alarm is triggered. For long-term warnings, if the trend slope dPfault / dt is greater than the set threshold YH3, a yellow alarm is pushed.

10. A remote fault diagnosis system for solar energy devices, characterized in that A remote fault diagnosis method for a solar device according to claim 1, comprising: an acquisition module, a preprocessing module, a fusion module, a fault module, and a prediction module, wherein the acquisition module is configured to acquire solar panel monitoring parameters, inverter monitoring parameters, energy storage system monitoring parameters, and charge controller monitoring parameters through provided sensors to form a multi-source heterogeneous monitoring parameter set for the solar device; The preprocessing module is used to remove abnormal data values by building a multivariate outlier detection model, constructing a dynamic spatiotemporal hypergraph of the solar installation to form a spatiotemporal hypergraph tensor X. Using the generator and discriminator, after tensor convolution, LSTM, and loss function optimization, the final output is a completed adversarial tensor. The fusion module is used to input the completed adversarial tensor to construct the device dynamic topology network, generate the topological feature tensor, use the spatiotemporal graph convolutional network and wavelet packet decomposition analysis, and use the gated interaction mechanism to fuse the analysis to obtain the final fused optimized coding feature vector; The fault module uses a support vector machine to establish a hybrid kernel model of solar device faults through a hybrid kernel function. The model determines the specific fault level state based on the fault probability threshold and outputs the corresponding fault level state; The prediction module uses a multivariate LSTM network to build a solar device failure time series prediction model to obtain future parameter prediction values, and build a solar device failure regression analysis model. The model inputs real-time parameters and outputs failure probability, and sets short-term, medium-term and long-term warnings based on different thresholds.

Images