Photovoltaic operation and maintenance decision-making system and method based on data driving

Abstract

The invention relates to the field of photovoltaic operation and maintenance, and discloses a photovoltaic operation and maintenance decision-making system and method based on data driving, and the method comprises the steps: carrying out the continuous mapping of a photovoltaic power station assembly cluster through an acquisition node, and generating a data priority chain table related to the shielding sensitivity; performing cross-domain synchronous tracking of a multi-scale variation sliding window on the data priority linked list, and constructing a multi-dimensional coupling feature tensor of photovoltaic array performance degradation; based on a multi-dimensional coupling feature tensor, a self-adaptive redundancy stripping model is utilized to extract a component attenuation evolution trajectory, and power generation efficiency clustering density mutation and power entropy migration inflection points are identified; component clusters and electrical coupling chains corresponding to the potential performance degradation candidate event set are calibrated as candidate intervention areas, and a local power generation income interference topological graph is constructed; and based on the local power generation income interference topological graph, dynamically reconstructing the sensing weight, the operation and maintenance window and the trigger threshold. The method has the advantage of improving the overall power generation income of the photovoltaic power station.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of photovoltaic operation and maintenance, in particular to a photovoltaic operation and maintenance decision system and method based on data driving. BACKGROUND

[0002] At present, the operation and maintenance of photovoltaic power stations mainly relies on regular inspection and experience-driven maintenance strategies. Common operation and maintenance measures include component cleaning, inverter inspection and cable maintenance, etc. However, photovoltaic components are easily affected by environmental factors during actual operation, such as local shading caused by dust, bird droppings and fallen leaves, which can cause a decrease in component output power and thus affect the power generation efficiency of the entire photovoltaic array. The existing operation and maintenance mode mostly adopts a fixed cycle cleaning and maintenance plan, which cannot make dynamic decisions according to the actual pollution level and power generation loss. Although some photovoltaic power stations have introduced power monitoring systems, they mostly can only provide macro power generation statistics and are difficult to accurately identify small performance degradation caused by pollution, hot spots or local failure. The existing fault diagnosis methods mostly rely on artificial experience or static thresholds and fail to fully utilize the correlation analysis capability of component-level real-time data and historical operation data, resulting in delayed or redundant cleaning and maintenance decisions, increased operation and maintenance costs, and reduced overall economic benefits of photovoltaic power stations. Therefore, it is necessary to design a photovoltaic operation and maintenance decision system and method based on data driving to improve the overall power generation benefits of photovoltaic power stations. SUMMARY

[0003] In view of the deficiencies of the prior art, the present application provides a photovoltaic operation and maintenance decision system and method based on data driving, which has the advantage of improving the overall power generation benefits of photovoltaic power stations and solves the problems in the background art.

[0004] To achieve the above-mentioned purpose of improving the overall power generation benefits of photovoltaic power stations, the present application provides the following technical solution: a photovoltaic operation and maintenance decision method based on data driving, comprising the following steps:

[0005] The acquisition node continuously maps the photovoltaic power station component cluster, and based on the irradiation scene semantic slicing network, the power curve, temperature drift and dust coverage perception data stream are dislocatedly labeled to generate a data priority linked list associated with shading sensitivity;

[0006] The data priority linked list is executed with multi-scale variant sliding window cross-domain synchronous tracking, combined with historical power generation fingerprints, inverter harmonic drift trajectories and sand deposition matrices, to construct a multi-dimensional coupled feature tensor of photovoltaic array performance degradation;

[0007] Based on the multi-dimensional coupled feature tensor, the component degradation evolution trajectory is extracted using an adaptive redundant stripping model, and the potential performance degradation candidate event set is generated by identifying the power efficiency cluster density mutation and power entropy offset inflection point;

[0008] The component cluster and electrical coupling chain corresponding to the set of potential performance degradation candidate events are calibrated as candidate intervention areas, and the illumination disturbance nodes, thermal anomaly nodes and inverter control nodes in the candidate areas are fused to construct a local power generation benefit interference topology atlas;

[0009] Based on the local power generation benefit interference topology atlas, the global benefit entropy fluctuation trend in the potential failure nodes and the monitoring data are combined to dynamically reconstruct the perception weight, operation and maintenance window and trigger threshold, and generate operation and maintenance decision instructions for optimizing photovoltaic benefit.

[0010] Preferably, the process of generating a data priority linked list associated with the shading sensitivity is:

[0011] Based on the irradiation scene semantic slicing network, the spatio-temporal synchronization processing of each perception data stream is performed, and the multi-label feature labeling of the sensing data from different sources is performed by using the staggered labeling mechanism;

[0012] A shading sensitivity matrix is constructed to calculate the instantaneous influence factor and cumulative risk factor of each perception data stream on the component power generation efficiency;

[0013] The risk factor is coupled with the real-time meteorological disturbance index to obtain a shading sensitivity coefficient reflecting the local pollution characteristics of the photovoltaic array and the environment dynamics;

[0014] According to the shading sensitivity coefficient, the multi-source data stream is weighted and sorted, and the final data priority linked list used for subsequent synchronous tracking is output.

[0015] Preferably, the process of performing multi-scale variant sliding window cross-domain synchronous tracking on the data priority linked list is:

[0016] According to the shading sensitivity coefficient, the data priority linked list is hierarchically divided, the perception data higher than the set threshold is defined as a high-weight sequence, and the perception data lower than the set threshold is defined as a low-weight sequence;

[0017] The high-weight sequence is assigned a short time window, and the low-weight sequence is assigned a long time window to construct a multi-scale sliding window set;

[0018] The perception data stream is synchronously read in each sliding window, and the time stamp alignment and illumination phase correction method is used to eliminate cross-domain data delay and sampling drift;

[0019] A recursive cross-domain correlation matching algorithm is used to track the power generation power waveform form, component surface thermal image gradient and inverter harmonic amplitude variation between adjacent sliding windows to generate a cross-domain matching vector set.

[0020] Preferably, the process of constructing a multi-dimensional coupling feature tensor of photovoltaic array performance degradation is:

[0021] Retrieving power fingerprint template set matching current component cluster model, inverter topology and operating environment from historical database, dynamically comparing template set with real-time power waveform collected in sliding window, and outputting waveform similarity vector;

[0022] Performing exponential weighted recursive smoothing on real-time collected inverter harmonic and power factor data, extracting high-frequency oscillation characteristics reflecting short-time disturbance and long-term drift characteristics describing trend change, forming dual-channel coupled feature vector;

[0023] Building wind and sand deposition matrix collected by environmental sensor array, calculating interference degree of dust of different particle sizes on light transmittance and component surface heat dissipation through particle size band analysis method, and obtaining environmental interference weight vector;

[0024] Mapping waveform similarity vector, dual-channel coupled feature vector and environmental interference weight vector to multi-dimensional feature embedding space through tensorization embedding mechanism to form initial embedding tensor;

[0025] Using multi-core kernel function fusion algorithm to perform nonlinear feature superposition on initial embedding tensor to generate multi-dimensional coupled feature tensor of photovoltaic array performance degradation.

[0026] Preferably, the component degradation evolution trajectory process is extracted by using an adaptive redundancy stripping model as follows:

[0027] Adaptive network combining gated recurrent unit and time sequence attention mechanism is used to sequence encode multi-dimensional coupled feature tensor;

[0028] Embedding data priority list in the encoding process, and using priority gate unit to dynamically adjust the hidden state update rate of different perception data streams, so as to strip redundant information;

[0029] Aggregating and analyzing continuous hidden states to generate a number of candidate degradation evolution trajectory segments;

[0030] Using the joint discriminant criterion of density peak clustering identification and power entropy offset detection to evaluate the similarity and degenerative mutation between candidate degradation trajectory segments, and outputting the component degradation evolution trajectory.

[0031] Preferably, the process of generating potential performance degradation candidate event set is as follows:

[0032] Dynamically time warping comparing component degradation evolution trajectory with standard healthy operation trajectory library, calculating trajectory shape deviation and time synchronization error;

[0033] Building a double-threshold abnormal score function, wherein the first threshold is based on clustering density mutation detection, and the second threshold is based on power entropy offset inflection point amplitude detection;

[0034] If the trajectory sub-sequence deviation degree and the anomaly score exceed the corresponding threshold value at the same time, the trajectory sub-sequence is marked as a potential degradation hidden danger segment;

[0035] The adjacent potential hidden danger segments are spatiotemporally merged to form a degradation continuous section, and a potential performance degradation candidate event set is output.

[0036] Preferably, the process of marking the component cluster and the electrical coupling chain corresponding to the potential performance degradation candidate event set as a candidate intervention area is:

[0037] According to the component cluster index involved in the candidate event set, an electrical coupling chain is generated;

[0038] The illumination disturbance nodes, thermal anomaly nodes and inverter control nodes involved in the electrical coupling chain are mapped into a heterogeneous node set;

[0039] A multi-modal edge structure is established between the heterogeneous node set based on an event-triggered edge weight update strategy, and a dynamic disturbance intensity score of the difference in perception features between nodes is added;

[0040] Core disturbance source nodes and key coupling paths in the heterogeneous graph are identified through joint analysis of spectral clustering and node centrality;

[0041] A subgraph containing core nodes and key paths is marked as a candidate intervention area.

[0042] Preferably, the process of constructing a local power generation revenue disturbance topology graph is:

[0043] Based on the topology structure of the candidate intervention area, a historical revenue loss matrix and a current real-time power generation deviation degree distribution are superimposed;

[0044] The edge weight between nodes is dynamically adjusted, the revenue sensitivity weight of the key path is calculated, and a local revenue risk influence matrix is constructed;

[0045] A node revenue priority label is introduced on the heterogeneous node set, and a risk-revenue joint analysis algorithm is used to generate a node revenue disturbance influence value;

[0046] A multi-scale graph clustering method is used to extract a key disturbance subgraph in the local topology, identify the main coupling path of the change in power generation revenue, and output a local power generation revenue disturbance topology graph.

[0047] Preferably, the process of generating a photovoltaic revenue optimization operation and maintenance decision instruction is:

[0048] Based on the historical power generation records and current state data of all nodes, the revenue state entropy of the main variable in the multi-dimensional coupling feature tensor is calculated, and a global revenue entropy fluctuation trend sequence is constructed;

[0049] The global income entropy fluctuation trend sequence is fused and analyzed with the local power generation income interference topological atlas to identify an income evolution risk section where the critical path is located, and a path-level risk cumulative value is calculated;

[0050] According to the risk cumulative value of the critical path and the node income priority label, the sampling frequency, the perception weight and the operation and maintenance window of the corresponding node are dynamically adjusted, and a photovoltaic income optimization operation and maintenance decision instruction is output.

[0051] A data-driven photovoltaic operation and maintenance decision system comprises:

[0052] The shading labeling module collects the power curve, temperature drift and dust coverage of the photovoltaic component cluster, and generates a data priority linked list associated with the shading sensitivity based on the irradiation scene semantic slicing network;

[0053] The sliding window tracking module performs multi-scale variant sliding window cross-domain synchronous tracking on the data priority linked list, and fuses the historical power generation fingerprint, harmonic drift trajectory and wind sand deposition matrix to construct a performance degradation feature tensor;

[0054] The trajectory extraction module identifies the power efficiency cluster density mutation and power entropy offset inflection point based on the multi-dimensional coupled feature tensor, extracts the component degradation evolution trajectory, and generates a performance degradation candidate event set;

[0055] The region identification module labels the component cluster and electrical coupling chain corresponding to the performance degradation candidate event set as a candidate intervention region, and fuses the light disturbance node, thermal anomaly node and inverter control node to construct an interference topological atlas:

[0056] The optimization decision module dynamically reconstructs the perception weight, operation and maintenance window and trigger threshold in combination with the local power generation income interference topological atlas and the global income entropy fluctuation trend, and generates a photovoltaic income optimization operation and maintenance decision instruction.

[0057] Compared with the prior art, the present application provides a data-driven photovoltaic operation and maintenance decision system and method, which has the following beneficial effects:

[0058] This invention achieves precise tracking of key sensing flows by semantically slicing and misaligning multi-source heterogeneous sensing data from photovoltaic power plants, enabling refined identification of the shading sensitivity of component clusters and dynamic generation of data priority lists. Combined with cross-domain synchronous tracking using multi-scale variable sliding windows and multi-dimensional coupling analysis of historical power generation fingerprints, harmonic drift trajectories, and aeolian deposition matrices, a high-dimensional feature tensor of photovoltaic array performance degradation is constructed, helping to reveal potential complex degradation modes. Furthermore, an adaptive redundancy stripping model is used to extract component degradation evolution trajectories, and potential performance degradation candidate events are identified through cluster density mutations and power entropy inflection points, enabling early detection of potential problems. Based on this, electrical coupling chain mapping is performed on candidate intervention areas to construct a local power generation revenue interference topology map integrating illumination disturbances, thermal anomalies, and inverter control, thereby accurately locating core interference source nodes and key coupling paths. Finally, by dynamically reconstructing sensing weights, operation and maintenance windows, and trigger thresholds based on global revenue entropy fluctuation trends, optimal revenue-oriented operation and maintenance decision instructions can be generated. This method can realize a closed-loop process from bottom-level perception and performance modeling to revenue interference topology and decision optimization, which can significantly improve the precision, intelligence and revenue maximization of photovoltaic power plant operation and maintenance, reduce operation and maintenance costs, reduce performance loss and improve long-term power generation efficiency and equipment life. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the method of the present invention;

[0060] Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example 1: Please refer to Figure 1 As shown in the figure, a data-driven photovoltaic operation and maintenance decision-making method according to an embodiment of the present invention includes the following steps:

[0063] S1: The acquisition node continuously maps the photovoltaic power station module clusters and performs misaligned annotation on the power curve, temperature drift and dust coverage perception data stream based on the irradiation scene semantic slicing network, generating a data priority linked list associated with shading sensitivity.

[0064] The process of generating a data priority linked list associated with occlusion sensitivity in S1 is as follows:

[0065] The photovoltaic module cluster is collected by a multi-modal perception component to form perception data streams such as power, voltage, current, temperature and dust coverage;

[0066] The spatial and temporal synchronization processing is performed on each perception data stream based on the irradiation scene semantic slicing network, and a staggered labeling mechanism is used to label the multi-label features of the sensing data from different sources. The timestamp alignment is performed on the perception data streams such as power, voltage, current, temperature and dust coverage, so as to eliminate the synchronization error caused by the inconsistent collection cycle of different sensors. The staggered labeling mechanism is applied to compare the dynamic response of the irradiation scene slice and the real-time power curve in time sequence, label the corresponding relationship between the power sudden drop point and the temperature drift peak, and add the pollution factor feature label. A multi-label data set containing scene semantic categories, module power generation dynamic response and pollution mode labels is formed;

[0067] A shading sensitivity matrix is constructed to calculate the instantaneous influence factor and cumulative risk factor of each perception data stream on the module power generation efficiency. On the basis of the multi-label data set, the short-term fluctuation features (such as peak rate and standard deviation) and long-term trend features (such as mean shift and cumulative deviation) of each type of sensing data (power, voltage, current, temperature and dust coverage) are extracted. The deviation rate of the real-time curve is calculated by matching the standard power curve of the same type of module in a non-polluted environment through the historical power generation fingerprint database. The time correlation between the dust coverage and the power curve deviation rate is compared to calculate the power generation efficiency loss proportion corresponding to the unit dust deposition rate to form the instantaneous influence factor. The cumulative risk factor of dust and local shading on the long-term performance degradation of the module is derived by combining the input and output harmonic drift of the inverter and the surface temperature rise trend of the module. The instantaneous influence factor and the cumulative risk factor of all sensing dimensions are arranged in a matrix according to the spatial position of the module to form a shading sensitivity matrix of module number x sensing data type;

[0068] The risk factor and real-time weather disturbance index are coupled to score the shading sensitivity coefficient reflecting the local pollution characteristics of the photovoltaic array and the environment dynamics. Real-time weather data of the power station is obtained, including external environmental parameters such as irradiance, wind speed, wind direction, air humidity and suspended particulate matter concentration. The weather disturbance index and the instantaneous influence factor in the shading sensitivity matrix are weighted and coupled, for example, when the wind speed is too low, the weight of dust deposition on power generation is increased, and the additional risk of dust agglomeration effect is considered in high humidity environment. The comprehensive environmental disturbance weight of each module is calculated by a weight redistribution algorithm based on principal component analysis. The instantaneous influence factor and the cumulative risk factor of each module are weighted and summed based on the comprehensive environmental disturbance weight to obtain the shading sensitivity coefficient of each module. The higher the shading sensitivity coefficient value, the more significant the impact of local pollution or environmental disturbance on the module cluster.

[0069] According to the shading sensitivity coefficient, the multi-source data stream is weighted and sorted, and a data priority linked list is output for subsequent synchronous tracking.

[0070] S2: Cross-domain synchronous tracking of the data priority linked list is performed by a multi-scale variant sliding window, historical power generation fingerprints, inverter harmonic drift trajectories, and wind and sand deposition matrices are combined to construct a multi-dimensional coupling feature tensor of photovoltaic array performance degradation.

[0071] The cross-domain synchronous tracking process of the data priority linked list in S2 is as follows:

[0072] According to the shading sensitivity coefficient, the data priority linked list is hierarchically divided, and the perception data higher than the set threshold is defined as a high-weight sequence, and the perception data lower than the set threshold is defined as a low-weight sequence;

[0073] The high-weight sequence is assigned a short time window to capture rapid degradation trends, and the low-weight sequence is assigned a long time window to track chronic degradation characteristics, and a multi-scale sliding window set is constructed;

[0074] In each sliding window, the perception data stream is synchronously read, and the time stamp alignment and illumination phase correction method is used to eliminate cross-domain data delay and sampling drift; the perception data involved in each sliding window is time-sequentially sorted, including power, voltage, current, temperature, and dust coverage degree, etc. multi-modal perception information; the time stamp alignment algorithm is used to correct the sensor data of different sampling periods to a unified time reference, and if there is data missing, interpolation is used to complete it, so as to eliminate the time offset problem caused by inconsistent sampling frequencies; according to the real-time solar incident angle and irradiance variation curve of the location of the photovoltaic module, the illumination phase correction is performed, and the power waveform phase drift problem caused by illumination change lag is corrected; the sampling drift of the sensing data located in different component clusters is corrected, which is mapped to the same time reference frame, ensuring that the cross-region and cross-sensor data have consistent time characteristics; finally, the multi-source data segment in the sliding window is obtained after time alignment and illumination phase correction;

[0075] The recursive cross-domain correlation matching algorithm is adopted to track the power waveform morphology, component surface thermal image gradient and inverter harmonic amplitude variation between adjacent sliding windows, and a cross-domain matching vector set is generated; the key features of the multi-source perception data in each sliding window are extracted, including the peak value, valley value, fluctuation amplitude, slope change rate of the power waveform, the spatial gradient distribution of the component surface thermal image temperature, and the main frequency band energy value of the inverter harmonic signal; the current sliding window and the next sliding window are selected as a group for sliding window comparison, the power waveform of the two is dynamically time-aligned, the similarity of the comparison curve form is compared, and the waveform shift rate is calculated; for the thermal image gradient and the harmonic frequency band energy features, the change rate and the distribution difference between adjacent sliding windows are calculated, and the multi-dimensional similarity is judged by the joint measurement method of cosine similarity and Euclidean distance; the cross-domain correlation analysis is performed on all sliding window pairs in a recursive manner, the feature matching results of the continuous sliding windows are continuously accumulated and updated, and an evolution feature sequence with time extension is formed; the matching results of each pair of sliding windows are encoded into cross-domain matching vectors, including the sliding window number, the power waveform similarity, the thermal image gradient shift rate and the harmonic amplitude difference degree, and are summarized into a complete cross-domain matching vector set.

[0076] The process of constructing the multi-dimensional coupling feature tensor of the performance degradation of the photovoltaic array in S2 is:

[0077] The power fingerprint template group matched with the current component cluster model, inverter topology and operating environment is retrieved from the historical database, the template group is dynamically compared with the real-time power waveform collected in the sliding window, and a waveform similarity vector is output;

[0078] The exponential weighted recursive smoothing processing is performed on the real-time collected inverter harmonic and power factor data, the high-frequency oscillation features reflecting short-term disturbance and the long-term drift features describing trend change are extracted, and a double-channel coupling feature vector is formed; the harmonic amplitude sequence and the power factor time series data of the input and output ends of the inverter are read from the sliding window, and the extreme discrete points existing in the data are first removed by median filtering; the exponential weighted recursive smoothing algorithm is used to process the harmonic sequence and the power factor sequence respectively, and higher weight is given to the recent data, so as to enhance the sensitivity of short-term dynamic fluctuation; the harmonic amplitude sequence after smoothing is subjected to spectral decomposition, and the energy concentration degree of the high-frequency band and the frequency band energy distribution uniformity are extracted as short-term disturbance features; the drift rate and stability index of the long-term evolution trend of the power factor are extracted, which are used to describe the performance degradation trend of the inverter and the component cluster in a long period; the high-frequency oscillation features and the long-term drift features are combined into a double-channel coupling feature vector, which reflects both instantaneous electrical disturbance and chronic degradation mode;

[0079] The wind-sand deposition matrix collected based on the environmental sensor array is constructed, the interference degree of dust of different particle sizes on the light transmittance and the heat dissipation of the component surface is calculated by the particle size band analysis method, and an environmental interference weight vector is obtained; the data of the environmental sensor array arranged at different positions of the power station are read, including wind speed, wind direction, air particulate matter concentration, particle size distribution, and component surface deposition thickness and other information; the dust particles are classified and counted according to the particle size distribution, for example, the dust is divided into three categories of fine particle size, medium particle size and coarse particle; the deposition amount per unit area of dust of different particle sizes on the component surface is calculated, and combined with the experimentally calibrated light transmittance attenuation coefficient, the weakening proportion of each particle size dust on the light incidence is calculated; at the same time, according to the thermal resistance effect formed by dust of different particle sizes on the component surface, the inhibition degree of the heat dissipation capacity of the component surface is calculated, and the thermal anomaly amplification interference coefficient is obtained; the light transmittance weakening proportion and the heat dissipation interference coefficient are weighted and fused to form an environmental interference weight vector reflecting the real-time performance influence of environmental pollution on the component;

[0080] The waveform similarity vector, the double-channel coupling feature vector and the environmental interference weight vector are uniformly mapped to a multi-dimensional feature embedding space through a tensorization embedding mechanism to form an initial embedding tensor; each dimension of the waveform similarity vector is unfolded in time sequence to construct a feature subspace reflecting the deviation degree of the power curve form; the short-term disturbance feature and the long-term drift feature of the double-channel coupling feature vector are split into two parallel channels and are respectively mapped to an electrical disturbance feature subspace and a trend degradation feature subspace; the light weakening proportion and the heat dissipation interference coefficient of the environmental interference weight vector are mapped to an environmental disturbance feature subspace; the three feature subspaces are uniformly aligned according to the time axis, the component cluster number axis and the feature category axis to construct a three-dimensional tensor coordinate system; through the tensorization embedding mechanism, the three types of feature vectors are embedded in the unified multi-dimensional feature embedding space according to their respective weights to form an initial embedding tensor, which completely records the comprehensive state of the multi-source information of the power, the inverter, the electrical and the environment;

[0081] The initial embedding tensor is subjected to a multi-kernel function fusion algorithm for nonlinear feature superposition to generate a photovoltaic array performance degradation multi-dimensional coupling feature tensor; a plurality of kernel functions suitable for different feature subspaces are selected, such as a Gaussian kernel suitable for curve form features, a polynomial kernel suitable for trend features, and a radial basis kernel suitable for environmental interference features; the kernel mapping results are calculated in each subspace of the initial embedding tensor respectively, the nonlinear correlation is extracted, and the originally linearly independent multi-source features are converted to high dimensions; the kernel mapping results of all subspaces are fused and superposed according to the weight coefficients, the main nonlinear interaction information is retained, the redundant dimensions are eliminated, and the separability of the features and the sensitivity to degradation modes are enhanced.

[0082] S3: Based on the multi-dimensional coupled feature tensor, the adaptive redundant stripping model is used to extract the component attenuation evolution trajectory, and the potential performance degradation candidate event set is generated by identifying the power entropy inflection point and the power efficiency cluster density mutation.

[0083] The process of extracting component attenuation evolution trajectory in S3 using an adaptive redundant stripping model is as follows:

[0084] An adaptive network combining a gated recurrent unit and a timing attention mechanism is used to sequence encode the multi-dimensional coupled feature tensor; the multi-dimensional coupled feature tensor of the photovoltaic array performance attenuation output by the previous step is read, which includes time dimension, component cluster number dimension, and power, electrical, environmental, and other feature dimensions; the tensor is time dimensionally expanded, and each time slice is input as a sequence unit, while retaining the component cluster number as an additional index label to distinguish the features of components at different physical locations; the expanded multi-source feature sequence is input into a timing encoding network composed of multiple layers of gated recurrent units, which uses the gating mechanism to capture long-term dependencies and short-term fluctuations while avoiding the problem of gradient vanishing or explosion in long sequence input in traditional recurrent structures; a timing attention mechanism is added to the output layer of the recurrent unit, which adaptively weights the feature contribution of different time slices, making the encoded hidden state pay more attention to key time periods with significant degradation patterns; a timing hidden state sequence processed by the gated recurrent unit and attention weighting is output;

[0085] In the encoding process, a data priority list is embedded, and a priority gating unit is used to dynamically adjust the hidden state update rate of different perception data streams, thereby stripping redundant information; a data priority list with shading sensitivity coefficients generated in the previous step is read, and each perception data in the list contains a weight label and a priority tracking identifier; when each time slice is input into the recurrent unit, the priority label of the corresponding perception data is input into the priority gating unit at the same time, and the gating unit controls the hidden state update rate of the recurrent unit through the weight coefficient; for high-priority data, the gating unit increases the update frequency and feature retention rate of the hidden state to ensure that critical degradation signals are not diluted; for low-priority data, the hidden state update rate is reduced, and it is treated as redundant information and gradually attenuates its influence; in continuous sequence calculation, the priority gating unit adaptively adjusts the weight contribution of each data stream, thereby stripping noise and redundant signals from the timing encoding and retaining only features that have a key effect on component performance degradation;

[0086] The continuous hidden states are aggregated for analysis to generate several candidate attenuation evolution track segments; the hidden state sequence after priority gating stripping is segmented and aggregated in time sequence, and the length of each aggregation window can be adaptively adjusted according to the typical time scale of component degradation; the hidden states in each window are statistically analyzed to extract key indicators describing the attenuation trend, such as power mean drop rate, feature change slope, thermal image gradient cumulative value, harmonic disturbance amplification factor, etc.; the aggregated feature indicators are combined into track segments, and each track segment represents the local degradation evolution mode of the component cluster in a certain time period; to avoid short-term fluctuations being misjudged as degradation segments, a smoothing constraint condition is introduced to discriminate the continuity of adjacent time period track segments, and if the fluctuation amplitude does not exceed the preset threshold, the adjacent time period track segments are merged into the same track segment; finally, several candidate attenuation evolution track segments are obtained;

[0087] The joint discrimination criterion of density peak clustering identification and power entropy offset detection is used to evaluate the similarity and degradation mutation between the candidate attenuation track segments, and the component attenuation evolution track with time labeling, sensitivity label and degradation amplitude label is output.

[0088] The process of generating a set of potential performance degradation candidate events in S3 is:

[0089] The component attenuation evolution track is compared with the standard healthy operation track library by dynamic time warping, and the track shape deviation and time synchronization error are calculated; the output component attenuation evolution track is read, each track contains time sequence, power output feature, temperature drift, dust coverage and harmonic disturbance, etc., and is labeled with time label and degradation amplitude label; the standard healthy operation track library corresponding to the component model, installation area, inverter topology and operating climate condition is retrieved from the historical database, and the track library stores the power waveform, thermal distribution feature and harmonic frequency distribution of the component in the ideal state; to solve the inconsistency problem of the actual degradation track and the standard healthy track in time scale and sampling frequency, the dynamic time warping method is adopted, and the optimal alignment path of the two tracks at different time points is calculated to obtain the matching path with minimum shape difference; based on the dynamic matching path, the track shape deviation of each time slice is calculated, that is, the power waveform difference, temperature thermal image distribution difference and harmonic disturbance amplitude difference of the degradation track and the healthy track at the same time point; at the same time, the synchronization error of the track in time dimension is calculated, that is, the time delay between the degradation track reaching the same feature state and the standard track, which is used to measure the advance or lag of the degradation speed; a track comparison result matrix is output;

[0090] A double-threshold anomaly scoring function is constructed, in which the first threshold is based on cluster density mutation detection, and the second threshold is based on power entropy offset inflection point amplitude detection; a sliding window analysis is performed on the trajectory shape deviation degree sequence, the deviation degree distribution in each window is mapped to a high-dimensional similarity space, and the density estimation value of the local point is calculated; by using the cluster density mutation detection method, the significant change point of the deviation degree density between the continuous windows is identified, and the density increase amplitude of the change point is defined as the density mutation score; at the same time, the information entropy sequence of the trajectory power output feature is calculated, the trend curve of the entropy change with time is extracted, and the inflection point position and the inflection point amplitude of the entropy change are found out through first-order derivative and second-order derivative analysis; after the density mutation score and the entropy inflection point amplitude are normalized, a double-threshold anomaly scoring function is constructed: the first threshold is used to judge whether the abnormal cluster behavior of the trajectory shape in the spatial distribution is significant; the second threshold is used to judge whether the mutation degree of the trajectory power entropy in the time evolution reaches the degradation feature level; the baseline of the two thresholds is determined by the statistical characteristics of the standard healthy trajectory, and can be adaptively adjusted according to different power station environments;

[0091] If the trajectory subsequence deviation degree and the anomaly score exceed the corresponding threshold value at the same time, the trajectory subsequence is marked as a potential degradation hidden danger segment;

[0092] Adjacent potential hidden danger segments are spatiotemporally merged to form a degradation continuous section, and a potential performance degradation candidate event set containing component cluster identification, degradation level and positioning information is output.

[0093] S4: The component cluster and the electrical coupling chain corresponding to the potential performance degradation candidate event set are marked as a candidate intervention area, and the illumination disturbance node, the thermal anomaly node and the inverter control node in the candidate area are fused to construct a local power generation benefit interference topology map.

[0094] The process of marking the component cluster and the electrical coupling chain corresponding to the potential performance degradation candidate event set as the candidate intervention area in S4 is as follows:

[0095] According to the component cluster index involved in the candidate event set, an electrical coupling chain is generated;

[0096] The illumination disturbance node, the thermal anomaly node and the inverter control node involved in the electrical coupling chain are mapped into a heterogeneous node set; each electrical coupling chain is traversed, and all physical nodes contained in the link are subjected to attribute screening:

[0097] Illumination disturbance node: a node marked by an irradiance sensor or an optical monitoring unit, having a light shielding rate, a local reflectivity or a dust deposition thickness feature;

[0098] Thermal anomaly node: a component cluster node whose temperature rise exceeds a threshold value monitored by an infrared thermal imaging or a temperature sensor, having a thermal distribution uneven feature;

[0099] Inverter control node: the inverter and its upstream controller node, containing dynamic control parameters such as power factor setting, harmonic suppression and voltage regulation;

[0100] The screened nodes are uniformly numbered and type-labeled to form a heterogeneous node set, in which each node is attached with a node type label, a node attribute vector (illumination disturbance characteristics, thermal anomaly indicators, control parameter states, etc.), and its connection position in the electrical coupling chain. This heterogeneous node set breaks the traditional single structure that relies only on component clusters, forming a multi-modal information set containing illumination disturbances, thermal anomalies, and control nodes.

[0101] A multi-modal edge structure is established between the heterogeneous node sets based on an event-triggered edge weight update strategy, and the dynamic interference intensity score of the perceived feature difference between nodes is added. Each pair of nodes in the heterogeneous node set is traversed to determine whether there is an electrical coupling, spatial adjacency, or control signal transmission relationship between them. If any of the relationships are met, a multi-modal edge is established between the two nodes. For each edge, the perceived feature vectors of the two end nodes are collected, and the feature difference between the nodes is calculated, including illumination disturbance difference (such as local irradiance gradient), thermal anomaly difference (such as temperature gradient and thermal diffusion rate), inverter control parameter difference (such as harmonic content change rate, power factor offset). The node feature difference is coupled with the degradation level, degradation amplitude, and duration in the candidate event set to generate an event-triggered intensity coefficient, which is used to measure the influence degree of the candidate event on the node relationship. Through the edge weight update strategy, the event-triggered intensity coefficient is superimposed with the node basic connection weight to obtain the dynamic interference intensity score of the edge, and the score is added to the edge attribute. Finally, a heterogeneous interference graph containing a multi-modal edge structure is obtained, and each edge of the graph contains the node relationship type, the basic topological weight, and the interference intensity score after event triggering.

[0102] Core interference source nodes and key coupling paths in the heterogeneous graph are identified through spectral clustering and node centrality joint analysis. The subgraph containing the core nodes and key paths is designated as the candidate intervention area.

[0103] The process of constructing a local power generation revenue interference topology graph in S4 is as follows:

[0104] Based on the topological structure of the candidate intervention area, the historical revenue loss matrix and the current real-time power generation deviation degree distribution are superimposed.

[0105] The edge weight between nodes is dynamically adjusted, the benefit sensitivity weight of the critical path is calculated, and a local benefit risk influence matrix is constructed. All edges in the topology structure of the candidate intervention area are traversed, the historical benefit loss value and the real-time deviation value of the two end nodes are taken, and the benefit difference is calculated. If the historical benefit loss of the two nodes is highly correlated, the basic weight of the edge is increased. If the time synchronization exists in the real-time deviation of the two nodes, the immediate weight of the edge is also increased. Finally, the preliminary updated edge weight matrix is obtained. The benefit sensitivity modeling is performed on each edge:

[0106] Direct sensitivity: the direct influence degree of the benefit change of the two end nodes of the edge on the local benefit;

[0107] Propagation sensitivity: the probability of the benefit change of the two end nodes of the edge spreading to other nodes along the electrically coupled chain;

[0108] The direct sensitivity and the propagation sensitivity are combined in a weighted manner to obtain the comprehensive benefit sensitivity weight of the edge. Through the shortest path and the maximum flow analysis method, all critical paths from the core interference source node to the inverter or bus node are found. The benefit sensitivity weight of all edges of each critical path is accumulated to obtain the path benefit sensitivity value. The sensitivity values of all critical paths are mapped into a benefit risk influence matrix, and the rows and columns of the matrix represent different paths, and the matrix elements represent the benefit risk coupling degree between the paths;

[0109] The node benefit priority label is introduced on the heterogeneous node set, and the risk-benefit joint analysis algorithm is used to generate the node benefit interference influence value. The node benefit priority label is assigned to each node in combination with the historical benefit loss, the real-time deviation and the appearance frequency in the critical path. The high-priority node is usually the core node with the most concentrated benefit loss and the greatest impact on the local power generation benefit. For each node, the benefit sensitivity in the local benefit risk influence matrix, the centrality index (betweenness centrality, eigenvector centrality, etc.) in the topology graph, and the interference intensity with adjacent nodes are comprehensively calculated. A node risk-benefit joint scoring function is constructed: the risk part is determined by the influence and the potential failure probability of the node in the interference propagation chain; the benefit part is determined by the historical loss proportion and the real-time deviation proportion of the node; the risk score and the benefit score are fused through a nonlinear fusion to generate the node benefit interference influence value; the benefit interference influence values of all nodes are clustered or segmented to divide them into multiple levels, and are stored jointly with the node benefit priority label;

[0110] The key interference subgraph in the local topology is extracted through the multi-scale graph clustering method, the main coupling path of the power generation benefit change is identified, and the local power generation benefit interference topology atlas with benefit sensitivity annotation and risk weight stratification is output.

[0111] S5: Based on the local power generation revenue interference topology map, combined with the potential failure nodes and the global revenue entropy fluctuation trend in the monitoring data, the perception weight, operation and maintenance window and trigger threshold are dynamically reconstructed, and the operation and maintenance decision instruction of photovoltaic revenue optimization is generated.

[0112] The operation and maintenance decision instruction of photovoltaic revenue optimization generated in S5 is:

[0113] Based on the historical power generation records and current state data of all nodes, the revenue state entropy of the main variable in the multi-dimensional coupling characteristic tensor is calculated, and the global revenue entropy fluctuation trend sequence is constructed;

[0114] The global revenue entropy fluctuation trend sequence and the local power generation revenue interference topology map are analyzed and fused, the revenue evolution risk section where the key path is located is identified, and the path level risk cumulative value is calculated;

[0115] According to the risk cumulative value of the key path and the node revenue priority label, the sampling frequency, perception weight and operation and maintenance window of the corresponding node are dynamically adjusted, and the photovoltaic revenue optimization operation and maintenance decision instruction is output.

[0116] Embodiment 2: as shown in Figure 2 A data-driven photovoltaic operation and maintenance decision system includes:

[0117] The occlusion labeling module: collects the power curve, temperature drift and dust coverage of the photovoltaic component cluster, and generates a data priority linked list associated with the occlusion sensitivity based on the irradiation scene semantic slicing network;

[0118] The sliding window tracking module: performs multi-scale variant sliding window cross-domain synchronous tracking on the data priority linked list, and fuses the historical power generation fingerprint, harmonic drift trajectory and wind and sand deposition matrix to construct the performance degradation feature tensor;

[0119] The trajectory extraction module: identifies the power efficiency cluster density mutation and power entropy offset inflection point based on the multi-dimensional coupling characteristic tensor, extracts the component degradation evolution trajectory and generates the performance degradation candidate event set;

[0120] The region identification module: labels the component cluster and electrical coupling chain corresponding to the performance degradation candidate event set as the candidate intervention region, and fuses the light disturbance node, thermal abnormal node and inverter control node to construct the interference topology map:

[0121] Optimization decision module: combined with the local power generation revenue interference topology map and the global revenue entropy fluctuation trend, the perception weight, operation and maintenance window and trigger threshold are dynamically reconstructed, and the photovoltaic revenue optimization operation and maintenance decision instruction is generated.

[0122] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0123] While the embodiments of the application have been shown and described herein, it is to be understood that the scope of the application, jointly pointed out in the appended claims, is not to be limited to the above-described embodiments but can be otherwise variously changed, modified, replaced, and altered within the principles and spirit of the present application.

Claims

1. A data-driven photovoltaic operation and maintenance decision method, characterized in that, The method comprises the following steps: The collection node continuously maps the photovoltaic power station component cluster, and based on the irradiation scene semantic slicing network, misplaces labels on the power curve, temperature drift and dust coverage perception data stream, generates a data priority linked list associated with the shading sensitivity; The data priority linked list is executed by a multi-scale variant sliding window cross-domain synchronous tracking, combined with the historical power generation fingerprint, the inverter harmonic drift trajectory and the sand deposition matrix, a multi-dimensional coupled feature tensor of the photovoltaic array performance degradation is constructed; Based on the multi-dimensional coupled feature tensor, the adaptive redundant stripping model is used to extract the component degradation evolution trajectory, the power efficiency cluster density mutation and power entropy offset inflection point are identified to generate a potential performance degradation candidate event set; The component cluster and electrical coupling chain corresponding to the potential performance degradation candidate event set are marked as a candidate intervention area, and the light disturbance node, thermal anomaly node and inverter control node in the candidate area are fused to construct a local power generation benefit interference topology map; Based on the local power generation benefit interference topology map, combined with the potential failure node and the global benefit entropy fluctuation trend in the monitoring data, the perception weight, operation and maintenance window and trigger threshold are dynamically reconstructed, and the operation and maintenance decision instruction of the photovoltaic benefit optimization is generated.

2. The data-driven photovoltaic operation and maintenance decision system and method according to claim 1, characterized in that, The process of generating a data priority linked list associated with the shading sensitivity is: Based on the irradiation scene semantic slicing network, the spatio-temporal synchronous processing of each perception data stream is performed, and the multi-label feature labeling of the sensing data from different sources is performed by using the misplacement labeling mechanism; An shading sensitivity matrix is constructed, and the instantaneous influence factor and cumulative risk factor of each perception data stream on the component power generation efficiency are calculated; The risk factor is coupled with the real-time meteorological disturbance index to obtain the shading sensitivity coefficient reflecting the local pollution characteristics of the photovoltaic array and the environment dynamics; According to the shading sensitivity coefficient, the multi-source data stream is weighted and sorted, and the final data priority linked list used for subsequent synchronous tracking is output.

3. The data-driven photovoltaic operation and maintenance decision method according to claim 2, characterized in that, The process of executing multi-scale variant sliding window cross-domain synchronous tracking on the data priority linked list is: According to the shading sensitivity coefficient, the data priority linked list is hierarchically divided, the perception data higher than the set threshold is defined as a high weight sequence, and the perception data lower than the set threshold is defined as a low weight sequence; The high weight sequence is assigned a short time window, and the low weight sequence is assigned a long time window to construct a multi-scale sliding window set; In each sliding window, the perception data stream is synchronously read, and the time stamp alignment and light phase correction method is used to eliminate the cross-domain data delay and sampling drift; A recursive cross-domain correlation matching algorithm is used to track the power generation power waveform form, component surface thermal image gradient and inverter harmonic amplitude variation between adjacent sliding windows to generate a cross-domain matching vector set.

4. The data-driven photovoltaic operation and maintenance decision method according to claim 3, characterized in that, The process of constructing a multi-dimensional coupled feature tensor of the photovoltaic array performance degradation is: From the historical database, the power generation fingerprint template group matching the current component cluster model, inverter topology and running environment is retrieved, the template group is dynamically compared with the real-time power waveform collected in the sliding window, and a waveform similarity vector is output. The exponentially weighted recursive smoothing processing is performed on the real-time collected inverter harmonic and power factor data, high-frequency oscillation characteristics reflecting short-time disturbance and long-term drift characteristics describing trend change are extracted, and a double-channel coupled feature vector is formed; A wind-sand deposition matrix based on an environmental sensor array is constructed, the interference degree of dust of different particle sizes on light transmittance and component surface heat dissipation is calculated through a particle size band analysis method, and an environmental interference weight vector is obtained; The waveform similarity vector, double-channel coupled feature vector and environmental interference weight vector are uniformly mapped to a multi-dimensional feature embedding space through a tensorization embedding mechanism to form an initial embedding tensor; A multi-core kernel function fusion algorithm is used to perform nonlinear feature superposition on the initial embedding tensor to generate a multi-dimensional coupled feature tensor of photovoltaic array performance degradation.

5. The data-driven photovoltaic operation and maintenance decision method according to claim 4, characterized in that, The component degradation evolution trajectory process is extracted using an adaptive redundancy stripping model as follows: An adaptive network combining a gated recurrent unit and a timing attention mechanism is used to sequence encode the multi-dimensional coupled feature tensor; In the encoding process, a data priority linked list is embedded, and a priority gating unit is used to dynamically adjust the hidden state update rate of different perception data streams to strip redundant information; The continuous hidden state is aggregated and analyzed to generate several candidate degradation evolution trajectory segments; A joint discrimination criterion of density peak value clustering identification and power entropy offset detection is used to evaluate the similarity and degenerative mutation between the candidate degradation trajectory segments, and the component degradation evolution trajectory is output.

6. The data-driven photovoltaic operation and maintenance decision method according to claim 5, characterized in that, The process of generating a potential performance degradation candidate event set is as follows: The component degradation evolution trajectory is compared with a standard healthy operation trajectory library through dynamic time warping, the trajectory shape deviation and time synchronization error are calculated; A double-threshold abnormal score function is constructed, wherein the first threshold is based on clustering density mutation detection, and the second threshold is based on power entropy offset inflection point amplitude detection; If the trajectory sub-sequence deviation and abnormal score exceed the corresponding thresholds at the same time, the trajectory sub-sequence is marked as a potential degradation hazard segment; Adjacent potential hazard segments are spatiotemporally merged to form a degradation continuous section, and a potential performance degradation candidate event set is output.

7. The data-driven photovoltaic operation and maintenance decision method according to claim 6, characterized in that, The process of labeling the component cluster and electrical coupling chain corresponding to the potential performance degradation candidate event set as a candidate intervention area is as follows: An electrical coupling chain is generated according to the component cluster index involved in the candidate event set; The light disturbance nodes, thermal anomaly nodes and inverter control nodes involved in the electrical coupling chain are mapped to a heterogeneous node set; A multi-modal edge structure is established between the heterogeneous node set based on an event-triggered edge weight update strategy, and a dynamic interference strength score of the difference in perception features between nodes is added; The core interference source nodes and key coupling paths in the heterogeneous graph are identified through spectral clustering and node centrality joint analysis; The subgraph containing the core nodes and key paths is labeled as the candidate intervention area.

8. The data-driven photovoltaic operation and maintenance decision method according to claim 7, characterized in that, The process of constructing a local power generation revenue interference topology graph is as follows: Based on the topology structure of the candidate intervention area, the historical revenue loss matrix and the current real-time generation deviation distribution are superimposed; The edge weight between nodes is dynamically adjusted, the revenue sensitivity weight of the key path is calculated, and a local revenue risk influence matrix is constructed; The node benefit priority label is introduced on the heterogeneous node set, and a risk-benefit joint analysis algorithm is used to generate a node benefit interference impact value; The key interference subgraph in the local topology is extracted by the multi-scale graph clustering method, the main coupling path of the change of power generation benefit is identified, and the local power generation benefit interference topology atlas is output.

9. The data-driven photovoltaic operation and maintenance decision method according to claim 8, characterized in that, The process of generating photovoltaic benefit optimization operation and maintenance decision instructions is as follows: Based on the historical power generation records and current state data of all nodes, the benefit state entropy of the main variable in the multi-dimensional coupling feature tensor is calculated, and a global benefit entropy fluctuation trend sequence is constructed; The global benefit entropy fluctuation trend sequence and the local power generation benefit interference topology atlas are fused and analyzed to identify the benefit evolution risk section where the key path is located, and the path-level risk cumulative value is calculated; According to the risk cumulative value of the key path and the node benefit priority label, the sampling frequency, the perception weight and the operation and maintenance window of the corresponding node are dynamically adjusted, and the photovoltaic benefit optimization operation and maintenance decision instructions are output.

10. A data-driven photovoltaic operation and maintenance decision system, applied to the method of any one of claims 1-9, characterized in that, It includes: Shading annotation module: collect the power curve, temperature drift and dust coverage of the photovoltaic module cluster, and generate a data priority linked list associated with shading sensitivity based on the irradiation scene semantic slicing network; Sliding window tracking module: perform multi-scale variant sliding window cross-domain synchronous tracking on the data priority linked list, and fuse the historical power generation fingerprint, harmonic drift trajectory and wind and sand deposition matrix to construct a performance degradation feature tensor; Trajectory extraction module: identify the power efficiency cluster density mutation and power entropy offset inflection point based on the multi-dimensional coupling feature tensor, extract the component degradation evolution trajectory, and generate a performance degradation candidate event set; Region identification module: label the component cluster and electrical coupling chain corresponding to the performance degradation candidate event set as the candidate intervention region, and fuse the light disturbance node, thermal anomaly node and inverter control node to construct an interference topology atlas: Optimization decision module: combine the local power generation benefit interference topology atlas and the global benefit entropy fluctuation trend to dynamically reconstruct the perception weight, operation and maintenance window and trigger threshold, and generate photovoltaic benefit optimization operation and maintenance decision instructions.

Images