Photovoltaic system state fusion decision-making method and system

Abstract

The invention discloses a photovoltaic system state fusion decision-making method and system. The method comprises the following steps: acquiring multi-mode operation data of a photovoltaic module; inputting the multi-modal operation data into the cascade sensing network, executing a self-interpretation fusion decision based on physical mechanism real-time guidance and hierarchical semantic mapping, and outputting a deep fusion feature vector and a hierarchical interpretability evidence set; wherein the cascade sensing network comprises a multi-stage feature abstraction layer, an inner core deeply embedded into a photovoltaic module physical model, and an auxiliary interpretable interaction unit bidirectionally coupled with the inner core and interacting with the inner core in real time; and based on the hierarchical interpretable evidence set, generating a decision traceability report for the decision result. Through the real-time guidance of the physical mechanism, the reasoning deviation which possibly occurs in the data driving path and deviates from the physical common sense is corrected, and the robustness and the decision reliability of the decision model under the complex abnormal working condition are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and operation and maintenance technology for photovoltaic power generation systems, specifically to a photovoltaic system state fusion decision-making method and system. Background Technology

[0002] With the rapid development and large-scale application of photovoltaic power generation technology, the intelligent operation and maintenance of photovoltaic power plants is facing increasingly severe challenges. Traditional photovoltaic system monitoring solutions mainly suffer from the following technical bottlenecks:

[0003] 1. The disconnect between data-driven models and physical mechanism models; Current mainstream deep learning-based methods for photovoltaic system condition monitoring typically involve directly inputting multimodal monitoring data (such as electrical and environmental parameters) into neural networks for end-to-end training and inference. While these data-driven models possess strong nonlinear fitting capabilities, their internal decision-making processes rely entirely on statistical regularities within the data, lacking explicit modeling and utilization of the underlying physical operating mechanisms of the photovoltaic system. This leads to the model's tendency to make erroneous decisions that violate physical principles when faced with abnormal operating conditions outside the training data distribution, and the basis for these decisions is difficult to trace.

[0004] 2. Insufficient model interpretability and a crisis of trust in operation and maintenance; Existing Explainable Artificial Intelligence (XAI) techniques, such as SHAP and LIME, are typically used as post-analysis tools to perform attribution analysis on the final input-output relationships of a trained "black box" model. These methods have significant limitations: Passivity: The interpretation process is independent of the model's real-time reasoning process, and it is impossible to actively intervene or correct errors in real time during the reasoning process.

[0005] Fragmentation: It can only provide the static contribution of input features to the final output result, and cannot fully reveal the propagation path and dependency of abnormal signals from the underlying physical features to the high-level semantic patterns.

[0006] Lack of physical semantics: Attribution results mostly remain at the feature weight level, making it difficult to transform them into fault location criteria with clear physical meaning that can be directly understood and operated by maintenance personnel.

[0007] This "black box" characteristic severely hinders the practical application and trust establishment of artificial intelligence models in industrial operation and maintenance scenarios that require high reliability and traceability.

[0008] 3. The superficiality and homogenization of existing integration solutions; Some existing technologies attempt to introduce "attention mechanisms" or "multi-stage fusion" to improve model performance, but most still remain at the level of weighting the importance of features from different modalities or simply splicing them together. These methods fail to achieve deep integration and dynamic interaction between data-driven learning and domain physics knowledge in the model reasoning process, and also fail to endogenously generate decision-making basis with clear physical interpretability while outputting decisions.

[0009] The invention patent with patent application publication number CN120994567A discloses a fault diagnosis and recovery verification method, device, equipment and medium. The paper sets up an attribution analysis module to analyze the decision logic of the intelligent diagnosis model and generate an interpretable evidence chain, but it belongs to ex post-event attribution analysis. Summary of the Invention

[0010] The technical problem to be solved by this invention is to provide a novel fusion decision-making method that can introduce physical mechanisms in real time during the model reasoning process and generate process-oriented explanations with physical semantics, so as to fundamentally improve the reliability, interpretability and operation and maintenance practicality of photovoltaic system condition monitoring models.

[0011] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation of physical mechanisms includes: Acquire multimodal operation data of photovoltaic modules; Multimodal operating data is input into a cascaded sensing network to perform self-explanatory fusion decision-making based on real-time guidance of physical mechanisms and hierarchical semantic mapping, and outputs a deep fusion feature vector and a hierarchical set of interpretable evidence. The cascaded sensing network includes a multi-level feature abstraction layer and a kernel deeply embedded in the physical model of the photovoltaic module, and an auxiliary interpretable interactive unit that is bidirectionally coupled to the kernel and interacts in real time. The deep fusion feature vector is input into the task module, and the decision result of the photovoltaic module status is output. Based on a hierarchical set of interpretable evidence, a decision source tracing report is generated for the decision results.

[0012] In this embodiment, the cascaded sensing network includes at least two levels. The first-level sensing network is used to execute two calculation paths based on multimodal operation data: data-driven and physical-driven based on the physical model of photovoltaic modules. It also performs real-time physical-data interaction and credibility assessment, obtains the fused feature vector after real-time guidance and correction by physical mechanism, and outputs interpretable evidence. In this embodiment, the second-level perception network performs hierarchical state mapping based on the fusion feature vector output by the previous level, under the constraints of physical mechanisms, to obtain a deep fusion feature vector; and accumulates evidence chains to obtain a hierarchical set of interpretable evidence.

[0013] In this embodiment, the fused feature vector, after real-time guidance and correction based on physical mechanisms, is obtained, and interpretable evidence is output, including: Using light intensity feature vector as query vector Using the current characteristic vector of photovoltaic modules as a key vector Value vector Under the data-driven computation path, the first attention weight matrix and data interaction feature vectors ; Based on the physical model of photovoltaic modules, according to the current time Light intensity Photovoltaic module output voltage and photovoltaic module temperature Real-time calculation of theoretically expected current value And construct the corresponding theoretical current eigenvector. ; with the same light intensity feature vector as query vector Using the theoretical current eigenvector as a key vector Value vector Calculate the second attention weight matrix under the physical driving path. and theoretical interaction feature vectors ; Calculate the difference between the attention weight matrices under the two paths to obtain the physical guidance deviation matrix. ; and based on the physical guidance deviation matrix Generate a real-time physical reliability coefficient matrix And used to adjust the first attention weight matrix Make corrections; Using the modified first attention weight matrix For the actual value vector The weighted values ​​are then used to output a fused feature vector that has been corrected in real time by physical mechanisms. ; At the same time, the physical guidance deviation matrix Actual current Compared with the theoretical expected current value scalar deviation As explanatory evidence Output.

[0014] In this embodiment, a real-time physical reliability coefficient matrix is ​​generated, which is expressed by the following formula: ; The first attention weight matrix is ​​corrected using the following formula: ; The output fused feature vector, after real-time correction guided by physical mechanisms, is expressed by the following formula: ; In the formula, This is the real-time physical reliability coefficient matrix. For the Sigmoid function, A scale factor greater than zero. This is the physical guidance deviation matrix. This is the corrected first attention weight matrix. For element-wise multiplication, This is the second attention weight matrix. This is the first attention weight matrix. To fuse feature vectors, This is the actual value vector.

[0015] In this embodiment, obtaining deep fusion feature vectors and hierarchical interpretability evidence includes: The fused feature vector output from the previous stage Enter the first The layer feature abstraction layer yields the abstract feature output of that layer. ; Output abstract features The input is sent to the k-th layer's supplementary interpretable interaction unit; this supplementary interpretable interaction unit, based on sharing the same photovoltaic module physical model kernel as the first-level perception network, translates the abstract features of the current layer... Mapping or inverse reasoning into an intermediate theoretical state feature that should appear under ideal physical conditions and corresponds to the semantics of the current layer. ; Calculate the first Layer abstract features Characteristics of intermediate theoretical states Abstract feature differences and this abstract feature difference As the first Adding interpretable evidence at each layer Accumulate; Based on abstract feature differences Abstract features Perform adaptive modulation based on physical mechanisms to output the first... Layer fusion feature vector And pass it to the next layer.

[0016] In this embodiment, obtaining a hierarchical set of interpretable evidence includes: Repeatedly acquire the fused feature vector in the second-level perceptual network The process ultimately outputs a deep fusion feature vector. It also obtains a hierarchical set of interpretable evidence organized in processing order, ranging from underlying physical relationships to high-level semantic patterns. ,in, This is the last level.

[0017] In this embodiment, intermediate theoretical state features are obtained. This can be achieved in the following ways: An auxiliary decoder is trained synchronously at each feature abstraction layer. The auxiliary decoder abstracts features from the current layer. As input, the output of the auxiliary decoder is the intermediate theoretical state feature. .

[0018] In this embodiment, based on abstract feature differences Abstract features Perform adaptive modulation based on physical mechanisms to output the first... Layer fusion feature vector This can be achieved in the following ways: ; ; In the formula, For modulation gate vector, For element-wise multiplication, For the Sigmoid function, , These are two learnable parameters. For abstract features, This represents the characteristics of an intermediate theoretical state.

[0019] In this embodiment, a decision attribution report is generated based on a hierarchical set of interpretable evidence, including: Initial anomaly localization: based on first-level interpretability evidence By judging scalar deviation Whether the current exceeds a preset threshold is used to directly and quantitatively determine whether the current is abnormal; at the same time, the physical guidance deviation matrix is ​​analyzed. For elements with significant median values, qualitatively identify which patterns of correlation between illumination and current characteristics violate physical expectations; Tracing the anomalous propagation: Continuing to analyze subsequent high-level evidence. By observing the differences in abstract features By examining the order and intensity of events in the chain of evidence, we can trace how the initial current anomaly was associated with, amplified, or transformed into more complex failure modes in the higher semantic space of the model. Report generation: Based on the initial anomaly location and anomaly propagation tracing analysis, a decision-making source tracing report is automatically generated.

[0020] This invention also provides a system applying the above-described photovoltaic system state fusion decision-making method based on physical mechanisms and real-time guidance and self-explanation, comprising: The multimodal data module is used to acquire multimodal operating data of photovoltaic modules; The self-explaining fusion decision module is used to input multimodal operating data into the cascaded sensing network, execute self-explaining fusion decisions based on real-time guidance of physical mechanisms and hierarchical semantic mapping, and output a deep fusion feature vector and a hierarchical set of interpretable evidence. The cascaded sensing network includes a multi-level feature abstraction layer and a kernel deeply embedded in the physical model of the photovoltaic module, and an auxiliary interpretable interaction unit that is bidirectionally coupled to the kernel and interacts in real time. The state decision module is used to input the deep fusion feature vector into the task module and output the decision result of the photovoltaic module state; The decision tracing and report generation module is used to generate decision tracing reports based on a hierarchical set of interpretable evidence.

[0021] Compared with existing technologies, the beneficial effects of this invention are fundamentally different from the existing technical solutions that simply connect the feature extraction module and the interpretability module or perform post-hoc attribution analysis. The core of this invention lies in constructing a "forward propagation network with deep embedding of physical mechanisms." This network is not composed of independent "feature extractors" and "interpreters" pieced together, but rather, within each processing layer, the feature learning process and the physical calculation process are atomically integrated and interact in real time. Specifically, each layer's "auxiliary interpretability interaction unit" is deeply embedded in the same photovoltaic physical model kernel. Its core function is to simultaneously calculate the theoretical state that the feature at that layer should have under ideal physical conditions at the same moment of feature abstraction, and immediately use the difference between the two to guide and correct the propagation direction of the feature. This "feature-physics" dual-thread real-time collaborative mechanism is the key to achieving highly reliable decision-making and intrinsic interpretability.

[0022] This invention breaks down the barriers between data-driven models and physical mechanism models, dynamically introducing and utilizing the physical operation knowledge of photovoltaic systems during the real-time inference process of deep learning models.

[0023] This invention realizes the "white-boxing" of the decision-making process of the decision-making model, so that the model not only outputs the decision result, but also simultaneously outputs a complete, coherent chain of decision-making basis with physical semantics.

[0024] This invention improves the robustness and reliability of decision-making models under complex and abnormal conditions. Through real-time guidance of physical mechanisms, it corrects reasoning biases that may deviate from physical common sense in data-driven paths.

[0025] This invention embeds the physical model as a computable kernel deeply into the forward propagation of each network layer, enabling real-time guidance. Unlike existing technologies that simply connect feature extraction modules and interpretability modules in a phased concept generation process, or rely on post-hoc attribution analysis, the interpretability of this invention is endogenous and proactive in the reasoning process.

[0026] Existing technologies typically rely on data statistics to perform "data concept induction" and generate artificially defined semantic tags. However, the interpretable interactive unit of this invention is based on first-principles physics, and its core is to perform "physical state mapping" and calculate theoretically expected features. .

[0027] The decision optimization of this invention is based on physical deviation ( The former dynamically modulates features, while existing technologies typically rely on path probability selection based on concept vectors. The former achieves real-time calibration of the internal representation of the neural network based on physical principles, while the latter involves probabilistic inference under a fixed structure.

[0028] This invention deeply integrates all innovative points under the unified mechanism of "real-time guidance and self-explanation of physical mechanisms". Through rigorous mathematical construction, it forms a solution with high originality and technical barriers. Correspondingly, this avoids the defects of traditional solutions such as functional stacking and scattered highlights. Attached Figure Description

[0029] Figure 1 This is a flowchart of a photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation of physical mechanisms, according to an embodiment of the present invention.

[0030] Figure 2 This is an architecture diagram of the first-level perception network according to an embodiment of the present invention.

[0031] Figure 3 This is a comparison chart of the physical guidance attention correction effects in embodiments of the present invention.

[0032] Figure 4 This is a schematic diagram illustrating the hierarchical evidence chain accumulation and propagation in an embodiment of the present invention.

[0033] Figure 5 This is an example diagram of a hot spot fault decision tracing report according to an embodiment of the present invention.

[0034] Figure 6(a) is a radar chart showing the capabilities of the self-explanatory fusion decision model of the present invention in terms of computational efficiency, model interpretability, deployment complexity, prediction accuracy, noise resistance, and generalization ability.

[0035] Figure 6(b) is a comparison of the performance indicators of the self-explanatory fusion decision model of the present invention and the existing decision model.

[0036] Figure 6(c) is a comparison chart of the comprehensive scores of the self-explanatory fusion decision model of the present invention and the existing decision model.

[0037] Figure 6(d) is a comparison of the explanatory trade-off distribution between the self-explanatory fusion decision model of the present invention and the existing decision model.

[0038] Figure 6(e) is a comparison of the training time of the self-explanatory fusion decision model of the present invention and the existing decision model. Detailed Implementation

[0039] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0040] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0041] Please see Figure 1 As shown, this invention provides a photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation of physical mechanisms, comprising: S10: Acquire multi-mode operation data of photovoltaic modules.

[0042] In one embodiment of the present invention, the multimodal operation data of the photovoltaic module at the current moment is acquired in real time, including at least the light intensity. Photovoltaic module output current Photovoltaic module output voltage and photovoltaic module temperature These data constitute the raw input for subsequent fusion and decision-making.

[0043] S20 inputs multimodal operation data into the cascaded sensing network, executes self-explanatory fusion decision based on real-time guidance of physical mechanisms and hierarchical semantic mapping, and outputs a deep fusion feature vector and a hierarchical interpretable evidence set; wherein, the cascaded sensing network includes a multi-level feature abstraction layer and a kernel deeply embedded in the physical model of the photovoltaic module, and an auxiliary interpretable interaction unit bidirectionally coupled with the kernel and interacting in real time.

[0044] In one embodiment of the present invention, this step is implemented through a cascaded model. The cascaded sensing network consists of multiple layers, each of which is a fusion of "physical sensing and feature modulation", including a feature abstraction layer and an auxiliary interpretable interactive unit that is "bidirectionally coupled and interacts with it in real time".

[0045] In this embodiment, the cascaded sensing network specifically includes at least two levels. The first-level sensing network executes two computational paths based on multimodal operational data: data-driven and physics-driven based on the photovoltaic module physical model. It performs real-time physical-data interaction and credibility assessment, obtains a fused feature vector after real-time guidance and correction by the physical mechanism, and outputs interpretable evidence. The second-level sensing network, based on the fused feature vector output by the previous level, performs hierarchical state mapping under physical mechanism constraints to obtain a deep fused feature vector; and accumulates evidence chains to obtain a hierarchical set of interpretable evidence. When the cascaded sensing network has more than two levels, its architecture is the same as the design of the second-level sensing network.

[0046] In this embodiment, the first level, physical-guided feature correction and preliminary interpretation, is centered on designing a parallel dual-path attention computing mechanism and performing real-time interaction and correction based on it. Specifically, it acquires the fused feature vector after real-time guidance and correction by the physical mechanism, and outputs interpretability evidence, including: S211, using the light intensity feature vector as query vector Using the current characteristic vector of photovoltaic modules as a key vector Value vector Under the data-driven computation path, the first attention weight matrix and data interaction feature vectors .

[0047] In this embodiment, the data-driven path learns the correlations between features from the observed data (the multimodal runtime data obtained in step S10) and performs calculations using a standard scaled dot product attention mechanism: ; ; In the formula, The dimension of the key vector, used for scaling. This is the matrix transpose. This can be understood as learning the attention weight matrix of "how light focuses on electric current". The features are obtained based on this weighted fusion. The activation function, along with the formula for obtaining the data interaction feature vector, can be understood as a representation of its associated interpretable interaction unit.

[0048] S212, based on the physical model of the photovoltaic module, according to the current time Light intensity Photovoltaic module output voltage and photovoltaic module temperature Real-time calculation of theoretically expected current value And construct the corresponding theoretical current eigenvector. ; with the same light intensity feature vector as query vector Using the theoretical current eigenvector as a key vector Value vector Calculate the second attention weight matrix under the physical driving path. and theoretical interaction feature vectors .

[0049] In this embodiment, the physical driving path is calculated based on the physical principles of the photovoltaic system, and also through the standard scaled dot product attention mechanism: ; ; in, This represents the attentional pattern of "how light should focus on electric current" under ideal physical conditions.

[0050] In this embodiment, the theoretically expected current value The calculation can be performed using any of the following methods: 1) Based on the equivalent circuit model of a single diode in a photovoltaic module, combined with the current light intensity Photovoltaic module output voltage and photovoltaic module temperature The solution is obtained through numerical iteration.

[0051] 2) It is calculated in real time by a neural network pre-trained on historical normal operating data (based on historical multimodal operating data) and constrained by physical equations.

[0052] S213, calculate the difference between the attention weight matrices under the two paths, and obtain the physical guidance deviation matrix. ; and based on the physical guidance deviation matrix Generate a real-time physical reliability coefficient matrix And used to adjust the first attention weight matrix Make corrections.

[0053] In this embodiment, interpretable interaction and correction are key to achieving real-time guidance. Specifically, the system calculates the absolute difference in attention weights between the two paths: ; This physical guidance deviation matrix It directly quantifies the real-time deviation between the observed data patterns and the patterns expected based on physical principles, thus constituting a native explanatory signal with strong explanatory power that directly corresponds to the physical mechanism.

[0054] In this embodiment, based on Generate a real-time physical credibility coefficient matrix : ; in, For the Sigmoid function, is a scale factor greater than zero. Correspondingly, the physical meaning of this formula is: when When it is very large, The value will approach 0, indicating a decrease in the system's trust in the data-driven path; conversely, when... When I was very young, A value close to 1 indicates that the data-driven model is consistent with physical expectations, and the data-driven path has high credibility.

[0055] More specifically, scale factor It is a trainable parameter or a parameter that is used during system initialization based on historical data. and The typical distribution is set to a constant greater than 1 to amplify the moderating effect of the deviation.

[0056] In this embodiment, using The attention weights are then adjusted to obtain the final, physically guided attention weights. : ; In the formula, This is the corrected first attention weight matrix. This is an element-wise multiplication. The formula implements dynamic weight fusion, and the modified formula implements a dynamic attention fusion mechanism based on physical credibility: a physical guidance bias matrix. The larger the value, the higher the real-time physical reliability coefficient matrix. The smaller the value (approaching 0), the more the corrected second attention weight matrix... Physical guidance deviation matrix The smaller the value, the more the first attention weight matrix of the data-driven path is preserved. This is equivalent to embedding an online feedback loop based on first physics principles into the feedforward channel of model inference. It can sense the deviation between the data-driven mode and the physical expectations in real time, and dynamically adjust the dependence on the data-driven signals accordingly, thereby suppressing possible non-physical inference biases at the source.

[0057] like Figure 3 As shown, this invention proposes a real-time guided and self-explained fusion decision-making method with deep embedding of physical mechanisms. By constructing a cascaded sensing network that is "data-driven" and "physical-driven" in parallel, the theoretical state is calculated synchronously in each layer of feature abstraction, and the feature propagation path is dynamically corrected according to physical deviations. Finally, reliable decisions and hierarchical evidence chains with physical semantics are output, realizing high reliability, high interpretability and strong generalization ability of photovoltaic system state monitoring.

[0058] S214, using the modified first attention weight matrix For the actual value vector The weighted values ​​are then used to output a fused feature vector that has been corrected in real time by physical mechanisms. .

[0059] In this embodiment, the corrected weights are used to adjust the actual observation vector. Weighting is performed to output the first-level fused feature vector. : ; S215, simultaneously, the physical guidance deviation matrix Actual current Compared with the theoretical expected current value scalar deviation As explanatory evidence Output.

[0060] In this embodiment, the system simultaneously outputs the first layer of interpretability evidence. ,in This is the scalar deviation between the actual current and the theoretical current. (Explanational evidence) It captured the most basic physical relationship anomalies.

[0061] In one embodiment of the present invention, the first Layers: Hierarchical abstraction under physical constraints and endogenous evidence. Starting from the second level, self-explanatory fusion decision models begin to enter the hierarchical abstraction stage, such as... Figure 4 This section showcases an example of hierarchical evidence chain accumulation and propagation, presented in the form of five types of analytical charts: basic electricity consumption trend chart, anomaly detection chart, holiday impact chart, weekly electricity consumption pattern chart, and monthly electricity consumption trend chart. This visually presents the complete tracing chain from basic data trends to high-level semantic patterns. Figure 4The weekly electricity consumption pattern diagram shows the blue dashed line representing the weekly average and the orange dashed line representing the leading edge reference line for electricity efficiency. It should be noted that the analytical dimensions and network levels do not strictly correspond one-to-one; rather, it reorganizes and visualizes the evidence from an operations and maintenance perspective. (The last sentence appears to be incomplete and possibly refers to obtaining data from a different source.) Figure 4 The hierarchical evidence chain shown accumulates and propagates data; specifically, each level performs the following operations: S221, the fused feature vector output from the previous stage. Enter the first The layer feature abstraction layer yields the abstract feature output of that layer.

[0062] In this embodiment, higher-order feature abstraction (multimodal coding) involves taking the fused feature vector output from the previous stage. This, along with other modal features currently available that have not participated in previous deep fusion, is input into a feature abstraction layer. Here, "other modal features" refers to features other than the core electrical and environmental modalities (…). Auxiliary diagnostic information that may exist in the monitoring of external photovoltaic systems, such as: infrared thermal imaging characteristics reflecting the temperature distribution of photovoltaic module panels, electroluminescent image characteristics revealing internal defects, or environmental wind speed / humidity characteristics that affect heat dissipation and efficiency.

[0063] In this embodiment, the feature abstraction layer is essentially a multimodal feature encoder, such as a Transformer encoder layer or a graph convolutional layer. Its core function is to perform unified representation learning, nonlinear transformation, and high-order semantic compression on the input multi-source heterogeneous features, and output an encoded feature vector containing richer abstract semantics. .this Will be used simultaneously for: a) Input to the interpretable interactive unit at the same layer for subsequent physical theory state mapping and comparison.

[0064] b) Pass down the results of this level of abstraction. Note: Within this layer, if an attention mechanism is used, the query, key, and value vectors can be designed to be different linear projections of the fusion features to be encoded. The purpose is to automatically learn the internal dependencies between high-level semantic feature elements, which is different from the "light-current" attention design with clear physical orientation in the first level.

[0065] S222, Output abstract features The input is sent to the k-th layer's supplementary interpretable interaction unit; this supplementary interpretable interaction unit, based on sharing the same photovoltaic module physical model kernel as the first-level perception network, translates the abstract features of the current layer... Mapping or inverse reasoning into an intermediate theoretical state feature that should appear under ideal physical conditions and corresponds to the semantics of the current layer. .

[0066] In this embodiment, intermediate theoretical state characteristics This is achieved by simultaneously training a lightweight auxiliary decoder at each feature abstraction layer. This decoder abstracts features from the current layer. The decoder takes as input an attempt to reconstruct or predict the ideal intermediate features that should be generated based on the physical model of the photovoltaic module; the output of the decoder is the intermediate theoretical state feature. This operation ensures that even in high-dimensional abstract spaces, the evolution of features remains constrained by fundamental physical equations.

[0067] S223, calculate the... Layer abstract features Characteristics of intermediate theoretical states Abstract feature differences and this abstract feature difference As the first Adding interpretable evidence at each layer Accumulate.

[0068] In this embodiment, difference calculation and evidence generation calculate the difference between abstract features and theoretical state features: ; Correspondingly, this difference Recorded as the first Newly added interpretable evidence This evidence quantifies the degree of deviation between the high-order semantic features extracted by the current feature abstraction layer and the ideal state expected by physical theory, and the degree to which the current data-driven model deviates from physical principles, and is immediately recorded as the [number missing]. Layer of evidence .

[0069] At the same time, this difference The feedback is used to modulate the features themselves, and the features are corrected to the theoretical state through a gating mechanism, resulting in the output. This process demonstrates that interpretation (evidence generation) and decision-making (feature optimization) are completed simultaneously in the same forward propagation step and are mutually causal.

[0070] S224, Based on abstract feature differences Abstract features Perform adaptive modulation based on physical mechanisms to output the first... Layer fusion feature vector And pass it to the next layer.

[0071] In this embodiment, a modulation gate vector is calculated: ; The modulated characteristics are: ; In the formula, For modulation gate vector, For element-wise multiplication, For the Sigmoid function, , These are two learnable parameters.

[0072] Repeat steps S221~S224, and after multiple processing layers, finally output the deep fusion feature vector. It also obtains a hierarchical set of interpretable evidence organized in processing order, ranging from underlying physical relationships to high-level semantic patterns. ,in, This is the last level.

[0073] In this embodiment, a three-level sensing network is applied. The self-explanatory fusion decision model ultimately outputs a deep fusion feature vector. and a complete hierarchical set of interpretable evidence. Specifically, this hierarchical set of interpretable evidence records the entire chain of anomalous signals, from the underlying physical relationships to the high-level semantic patterns, in the order of processing.

[0074] S30 inputs the deep fusion feature vector into the task module and outputs the decision result of the photovoltaic module status.

[0075] In one embodiment of the present invention, the state decision and result output are: the final deep fusion feature vector containing physical guidance information is generated. The input is a lightweight, task-specific module, such as a fully connected classification or regression layer, which performs photovoltaic system operating state classification or future output prediction, and outputs the final decision result. .

[0076] S40 generates a decision attribution report based on a hierarchical set of interpretable evidence.

[0077] In one embodiment of the present invention, decision tracing and report generation: such as... Figure 5 As shown, this step utilizes a hierarchical set of interpretable evidence. On the decision results To make attributions and explanations.

[0078] Initial anomaly localization: System analysis of the first layer of interpretable evidence By judging the deviation between actual and theoretical current Whether the current output exceeds a preset threshold is directly and quantitatively determined to be abnormal. Simultaneously, the physical guidance deviation matrix is ​​analyzed. For elements with significant median values, qualitative identification can be used to determine which patterns of correlation between illumination and current characteristics deviate from physical expectations. Combining these two approaches allows the root cause of the problem to be pinpointed to the most fundamental and specific physical anomalies, such as "under 850W / m² illumination, the current response is only 53% of the expected value."

[0079] Tracing the anomalous propagation: Continuing to analyze subsequent high-level evidence. The differences in abstract features of each layer as evidence The high-order semantic features extracted by this layer, such as "thermoelectric coupling mode" and "power curve morphology," were quantified to determine their deviation from their theoretical ideal state. This was achieved by observing the differences in abstract features. By examining the order and intensity changes in the chain of evidence, we can trace how the initial current anomaly is associated, amplified, or transformed into more complex fault modes in the higher-level semantic space of the model. Fault modes include "abnormal thermoelectric coupling" and "hot spot fault characteristics."

[0080] Report Generation: Based on the initial anomaly localization and anomaly propagation tracing analysis, a decision-making source tracing report is automatically generated. This report not only provides conclusions but also clearly demonstrates the progression from the "localized" physical layer anomaly to the "traced" high-level semantic patterns, culminating in the final decision outcome. A complete and understandable chain of reasoning.

[0081] Please refer to Figures 6(a) to 6(e), which present a comparison of the performance and interpretability advantages of the self-explaining fusion decision model of this invention (represented by the physics-guided model in the figures) and existing decision models. This comparison demonstrates the performance and interpretability advantages of the proposed model compared to traditional methods. Specifically, Figure 6(a) is a radar chart showing the capabilities of the self-explaining fusion decision model of this invention in terms of computational efficiency, model interpretability, deployment complexity, prediction accuracy, noise resistance, and generalization ability; Figure 6(b) is a performance index comparison chart between the self-explaining fusion decision model of this invention and existing decision models; Figure 6(c) is a comprehensive score comparison chart between the self-explaining fusion decision model of this invention and existing decision models; Figure 6(d) is a comparison chart of the interpretability trade-off distribution between the self-explaining fusion decision model of this invention and existing decision models; and Figure 6(e) is a training time comparison chart between the self-explaining fusion decision model of this invention and existing decision models. Existing decision models include existing simple statistical models, physics-guided models, machine learning models, and deep learning models. As shown in the figure, the model of this invention significantly outperforms traditional data-driven models in key indicators such as fault detection accuracy and anomaly robustness. It can also simultaneously output a hierarchical chain of evidence with physical meaning, achieving a leap from "black box decision-making" to "white box tracing," effectively improving the credibility and practicality of photovoltaic intelligent operation and maintenance. In Figure 6(d), the Pareto front depicts the optimal trade-off boundary between performance and interpretability. Models located on the Pareto front (simple statistical model, physics-guided model, machine learning model, and deep learning model) represent the optimal choices under different needs. The Pareto front verifies the core innovative value of the proposed "real-time guidance and self-explanation of physical mechanisms" method: it successfully breaks the traditional trade-off dilemma of "high accuracy inevitably leading to low interpretability." The physics-guided model, located in the middle of the Pareto front, is the only solution that simultaneously possesses high interpretability and good predictive performance—compared to the deep learning model, interpretability is improved by 183%; compared to the simple statistical model, performance is improved by 8.3%. This result strongly demonstrates that the technical approach of introducing photovoltaic physical mechanisms in real time for feature guidance and endogenously generating hierarchical interpretable evidence can achieve competitive prediction accuracy while maintaining the "white-box" nature of the decision-making process. It fundamentally solves the core bottlenecks of existing technologies, such as the separation of data-driven approaches from physical mechanisms, insufficient interpretability, and shallow integration, and provides a highly reliable and traceable decision-making basis for the intelligent operation and maintenance of photovoltaic systems.

[0082] In another embodiment, a module-level intelligent monitoring application scenario is given: real-time diagnosis of photovoltaic module hot spot faults based on physical mechanisms.

[0083] Suppose a photovoltaic module develops a hot spot due to partial shading. This manifests as follows: under sufficient sunlight, the output current of the photovoltaic module is significantly lower than expected, and the local temperature rises.

[0084] S10, Data Acquisition: The system collects data from this component at the current moment: light intensity. Output current Output voltage Photovoltaic module backsheet temperature .

[0085] S20, Integrated Decision-Making Process.

[0086] First-level processing: Physical drive path calculation: based on single diode model and current Calculate the theoretical expected current Accordingly, features are constructed and attention calculations are performed to obtain... It represents the attention distribution pattern that the current characteristics should have under 850W / m² illumination.

[0087] Data-driven path computation: based on actual observations Calculations yielded It represents the "light-current" correlation pattern learned from real data.

[0088] Interaction and correction: Calculated A large matrix value indicates that the pattern of "illuminance-related current" in the actual data deviates significantly from physical expectations. Specifically, the real-time physical reliability coefficient matrix is ​​calculated using the Sigmoid function. .

[0089]

[0090] The above formula should be corrected accordingly. Because... Very small Large, corrected attention weights It mainly inherits the weights of the physical path. (In this embodiment) The contribution coefficient is 0.8, which reflects the physical guidance: when the data is abnormal, the decision model trusts physical principles more.

[0091] Subsequent layer processing, combined with Figure 4 It carried the information that "the current is seriously low" and The fusion characteristics of the evidence are then incorporated into subsequent levels.

[0092] At a certain feature abstraction layer, the model may extract a pattern of "high current deviation and high temperature occurring together," and output abstract features. Furthermore, the interpretable interaction unit attached to this layer maps intermediate theoretical state features based on the physical kernel. .

[0093] Calculate abstract feature differences Discover abstract feature output It deviates significantly from theoretical values ​​in the "thermal-electric coupling" dimension, generating evidence. This indicates the presence of an abnormal thermoelectric coupling effect.

[0094] After multiple layers of abstraction, the feature vectors are finally deeply fused. It incorporates high-level semantics such as "strong light illumination, low current, high temperature, and abnormal thermoelectric coupling".

[0095] S30, State Decision. The classifier is based on deeply fused feature vectors. The photovoltaic module is classified as "severely abnormal (hot spot)" with a confidence level of 98%.

[0096] S40, Decision-making origination (combined with...) Figure 5 The system generates a report: Chain of evidence: Level 1 sensing network (physical relationship layer): detected a severe attenuation of the "light-current response relationship" (only 53% of the expected value), and the actual current was 3.7A lower than the theoretical value (initial anomaly location).

[0097] The second-level sensing network (thermoelectric coupling layer) detected an abnormal correlation pattern between "current deviation" and "temperature rise", which does not conform to the normal thermo-induced current decay curve (abnormal pattern evolution).

[0098] The third-level perception network (fault semantic layer) abstracts the fault mode characteristics of "local hot spots".

[0099] Ultimately, the decision to identify "hot spot failure" was primarily based on anomalies in the physical relationships of the first-level sensing network, supported by evidence of anomalous patterns in subsequent layers. Accordingly, the report recommends prioritizing checks for localized occlusion or damage to components.

[0100] Please see Figures 1 to 5 As shown, the present invention also provides a system for applying the above-described photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation of physical mechanisms, comprising: The multimodal data module is used to acquire multimodal operating data of photovoltaic modules.

[0101] The self-explanatory fusion decision module is used to input multimodal operating data into the cascaded sensing network, perform self-explanatory fusion decision based on real-time guidance of physical mechanisms and hierarchical semantic mapping, and output a deep fusion feature vector and a hierarchical interpretable evidence set. The cascaded sensing network includes a multi-level feature abstraction layer and a kernel deeply embedded in the physical model of the photovoltaic module, and an auxiliary interpretable interaction unit that is bidirectionally coupled to the kernel and interacts in real time.

[0102] The state decision module is used to input the deep fusion feature vector into the task module and output the decision result of the photovoltaic module state.

[0103] The decision tracing and report generation module is used to generate decision tracing reports based on a hierarchical set of interpretable evidence.

[0104] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0105] The above-described embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above-described embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation of physical mechanisms, characterized in that, include: Acquire multimodal operation data of photovoltaic modules; Multimodal operating data is input into a cascaded sensing network to perform self-explanatory fusion decision-making based on real-time guidance of physical mechanisms and hierarchical semantic mapping, and outputs a deep fusion feature vector and a hierarchical set of interpretable evidence. The cascaded sensing network includes a multi-level feature abstraction layer and a kernel deeply embedded in the physical model of the photovoltaic module, and an auxiliary interpretable interactive unit that is bidirectionally coupled to the kernel and interacts in real time. The deep fusion feature vector is input into the task module, and the decision result of the photovoltaic module status is output. Based on a hierarchical set of interpretable evidence, a decision source tracing report is generated for the decision results.

2. The photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation according to claim 1, characterized in that, The cascaded sensing network consists of at least two levels. The first-level sensing network is used to execute two computational paths based on multimodal operating data: data-driven and physical-driven based on the physical model of photovoltaic modules. It also performs real-time physical-data interaction and credibility assessment, obtains the fused feature vector after real-time guidance and correction by physical mechanisms, and outputs interpretable evidence. The second-level perception network, based on the fusion feature vector output from the previous level, performs hierarchical state mapping under physical mechanism constraints to obtain a deep fusion feature vector; and accumulates evidence chains to obtain a hierarchical set of interpretable evidence.

3. The photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation according to claim 2, characterized in that, Obtain the fused feature vector after real-time correction guided by physical mechanisms, and output interpretable evidence, including: Using light intensity feature vector as query vector Using the current characteristic vector of photovoltaic modules as a key vector Value vector Under the data-driven computation path, the first attention weight matrix and data interaction feature vectors ; Based on the physical model of photovoltaic modules, according to the current time Light intensity Photovoltaic module output voltage and photovoltaic module temperature Real-time calculation of theoretically expected current value And construct the corresponding theoretical current eigenvector. ; with the same light intensity feature vector as query vector Using the theoretical current eigenvector as a key vector Value vector Calculate the second attention weight matrix under the physical driving path. and theoretical interaction feature vectors ; Calculate the difference between the attention weight matrices under the two paths to obtain the physical guidance deviation matrix. ; and based on the physical guidance deviation matrix Generate a real-time physical reliability coefficient matrix And used to adjust the first attention weight matrix Make corrections; Using the modified first attention weight matrix For the actual value vector The weighted values ​​are then used to output a fused feature vector that has been corrected in real time by physical mechanisms. ; At the same time, the physical guidance deviation matrix Actual current Compared with the theoretical expected current value scalar deviation As explanatory evidence Output.

4. The photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation according to claim 3, characterized in that, The real-time physical reliability coefficient matrix is ​​generated and expressed using the following formula: ； The first attention weight matrix is ​​corrected using the following formula: ； The output fused feature vector, after real-time correction guided by physical mechanisms, is expressed by the following formula: ； In the formula, This is the real-time physical reliability coefficient matrix. For the Sigmoid function, A scale factor greater than zero. This is the physical guidance deviation matrix. This is the corrected first attention weight matrix. For element-wise multiplication, This is the second attention weight matrix. This is the first attention weight matrix. To fuse feature vectors, This is the actual value vector.

5. The photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation according to claim 2, characterized in that, Obtaining deep fusion feature vectors and hierarchical interpretability evidence, including: The fused feature vector output from the previous stage Enter the first The layer feature abstraction layer yields the abstract feature output of that layer. ; Output abstract features The input is sent to the k-th layer's supplementary interpretable interaction unit; this supplementary interpretable interaction unit, based on sharing the same photovoltaic module physical model kernel as the first-level perception network, interprets the abstract features of the current layer. Mapping or inverse reasoning into an intermediate theoretical state feature that should appear under ideal physical conditions and corresponds to the semantics of the current layer. ; Calculate the first Layer abstract features Characteristics of intermediate theoretical states Abstract feature differences and this abstract feature difference As the first Adding interpretable evidence at each layer Accumulate; Based on abstract feature differences Abstract features Perform adaptive modulation based on physical mechanisms to output the first... Layer fusion feature vector And pass it to the next layer.

6. The photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation according to claim 5, characterized in that, Obtain a hierarchical set of interpretable evidence, including: Repeatedly acquire the fused feature vector in the second-level perceptual network The process ultimately outputs a deep fusion feature vector. It also obtains a hierarchical set of interpretable evidence organized in processing order, ranging from underlying physical relationships to high-level semantic patterns. ,in, This is the last level.

7. The photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation according to claim 5, characterized in that, Obtaining intermediate theoretical state features This can be achieved in the following ways: An auxiliary decoder is trained synchronously at each feature abstraction layer. The auxiliary decoder abstracts features from the current layer. As input, the output of the auxiliary decoder is the intermediate theoretical state feature. .

8. The photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation according to claim 5, characterized in that, Based on abstract feature differences Abstract features Perform adaptive modulation based on physical mechanisms to output the first... Layer fusion feature vector This can be achieved in the following ways: ； ； In the formula, The modulation gate vector, For element-wise multiplication, For the Sigmoid function, , These are two learnable parameters. For abstract features, This represents the characteristics of an intermediate theoretical state.

9. The photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation according to claim 1, characterized in that, Based on a hierarchical set of interpretable evidence, a decision attribution report is generated, including: Initial anomaly localization: based on first-level interpretability evidence By judging scalar deviation Whether the current exceeds a preset threshold is used to directly and quantitatively determine whether the current is abnormal; at the same time, the physical guidance deviation matrix is ​​analyzed. For elements with significant median values, qualitatively identify which patterns of correlation between illumination and current characteristics violate physical expectations; Tracing the anomalous propagation: Continuing to analyze subsequent high-level evidence. By observing the differences in abstract features By examining the order and intensity of events in the chain of evidence, we can trace how the initial current anomaly was associated with, amplified, or transformed into more complex failure modes in the higher semantic space of the model. Report generation: Based on the initial anomaly location and anomaly propagation tracing analysis, a decision-making source tracing report is automatically generated.

10. A system applying the photovoltaic system state fusion decision-making method based on real-time guidance and self-explanation according to any one of claims 1-9, characterized in that, include: The multimodal data module is used to acquire multimodal operating data of photovoltaic modules; The self-explaining fusion decision module is used to input multimodal operating data into the cascaded sensing network, execute self-explaining fusion decisions based on real-time guidance of physical mechanisms and hierarchical semantic mapping, and output a deep fusion feature vector and a hierarchical set of interpretable evidence. The cascaded sensing network includes a multi-level feature abstraction layer and a kernel deeply embedded in the physical model of the photovoltaic module, and an auxiliary interpretable interaction unit that is bidirectionally coupled to the kernel and interacts in real time. The state decision module is used to input the deep fusion feature vector into the task module and output the decision result of the photovoltaic module state; The decision tracing and report generation module is used to generate decision tracing reports based on a hierarchical set of interpretable evidence.

Images