Fan photovoltaic station energy efficiency diagnosis and early warning management method based on machine learning

Abstract

The invention relates to the technical field of state monitoring of new energy power generation equipment, and particularly discloses a fan photovoltaic station energy efficiency diagnosis and early warning management method based on machine learning, which comprises the following steps: firstly, collecting operation data of power generation equipment and constructing time-space correlation structure data; then, through parallel feature decoupling processing, synchronously extracting overall broadband response features of the equipment and local features of a focusing abnormal frequency band; a double-layer clustering mechanism is adopted, initial state division is carried out according to global features, and secondary abnormal sub-clusters containing significant abnormal frequency band features are identified in each type according to local features; and finally, by analyzing the frequency and intensity change trend of the abnormal sub-clusters and combining the working condition parameters, calculating the efficiency deviation degree, and realizing early efficiency degradation early warning and comprehensive evaluation. According to the method, the problem that weak early fault features are difficult to identify under strong background noise is effectively solved, and the diagnosis accuracy and the early warning timeliness are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of new energy power generation equipment state monitoring technology, in particular to a wind turbine and photovoltaic station energy efficiency diagnosis and early warning management method based on machine learning. BACKGROUND

[0002] With the global energy structure transforming to clean, the installed capacity of wind power and photovoltaic power generation continues to grow rapidly. Ensuring the efficient, stable and reliable operation of these renewable energy stations, especially the core power generation equipment (such as wind turbines and photovoltaic power generation units), has become a key to improving the economic efficiency of power stations and the safety of power grids. In this context, real-time diagnosis and early warning of the energy efficiency state of wind turbines and photovoltaic equipment is a core technical requirement for precise operation and maintenance and preventive maintenance.

[0003] The prior art has the following disadvantages: The present application solves the core technical problem of missing early and weak compound faults (such as blade hidden cracks under gust background) caused by the inability to effectively distinguish strong background noise from weak fault characteristics when analyzing wind turbine vibration data using existing clustering-based energy efficiency diagnosis methods. Specifically, when traditional methods process vibration frequency spectrum as a whole vector, the wideband energy changes caused by gusts and the like will completely drown out the weak harmonic peaks produced by defects such as cracks in a specific narrow band, causing the clustering algorithm to incorrectly classify this "sub-health coupled state" as a normal "gust response" category, thus failing to provide effective early warning before the fault occurs. SUMMARY

[0004] The purpose of the present application is to provide a wind turbine and photovoltaic station energy efficiency diagnosis and early warning management method based on machine learning to solve the problems in the above background.

[0005] The purpose of the present application can be achieved by the following technical solutions: The wind turbine and photovoltaic station energy efficiency diagnosis and early warning management method based on machine learning comprises the following steps: S1: Collecting vibration signals and operating condition parameters of the power generation equipment of the wind turbine and photovoltaic station within a continuous time period to form a state monitoring sequence; S2: Converting the vibration signals in the state monitoring sequence into frequency spectrum features of multiple time segments, taking the frequency spectrum features of each time segment as a node, and establishing connection relationships and weights between nodes according to time sequence and frequency spectrum shape similarity to create spatiotemporal correlation structure data; S3: Analyzing the spatiotemporal correlation structure data using a parallel processing method, simultaneously extracting a global feature set representing the overall wideband response of the equipment and a local feature set focusing on a preset potential abnormal frequency band, and constraining the global feature set and the local feature set to maintain low statistical correlation during extraction; S4: Based on the global feature set, the nodes are divided into multiple primary state clusters at the first level; within each primary state cluster, the nodes are divided into secondary anomalous subclusters based on the local feature set at the second level to identify the anomalous frequency band features. S5: Analyze the frequency and characteristic intensity of secondary abnormal subclusters over time, generate early performance degradation warning signals for wind turbine photovoltaic power generation equipment based on the changing trend of secondary abnormal subclusters, and output performance evaluation results in combination with operating condition parameters.

[0006] As a further aspect of the present invention: S2 specifically includes: Adaptive time segmentation is performed on the vibration signal, the length of each time segment is adjusted according to the energy change rate of the vibration signal, and spectral transformation is performed on each time segment to obtain the corresponding spectral characteristics; Each time segment's spectral features are defined as a node. Based on the temporal order of the nodes, a first-type connection is established for nodes corresponding to adjacent time segments. The spectral morphology difference between any two nodes is calculated, and a second-type connection is established for node pairs with a difference less than a preset threshold. The first type of connection is assigned a weight based on the reciprocal of the time interval, and the second type of connection is assigned a weight based on the inverse of the spectral morphology difference. All nodes and weighted connections are then fused to generate spatiotemporal relational structure data.

[0007] As a further aspect of the present invention: the step of performing spectral transformation on each time segment to obtain the corresponding spectral features specifically includes: An initial analysis window is set on the vibration signal, and the energy value of the signal within the initial analysis window is calculated. Slide the initial analysis window along the time axis to calculate the signal energy change gradient between consecutive windows in real time, and compare the energy change gradient with the preset first gradient threshold and second gradient threshold. When the energy change gradient is below the first gradient threshold, the length of the current analysis window is expanded; when the energy change gradient is above the second gradient threshold, the length of the current analysis window is shrunk; otherwise, the current analysis window length is maintained, thereby determining each final time segment. Perform spectral transformation independently on each final time segment to generate the corresponding spectral features.

[0008] As a further aspect of the present invention: S3 specifically includes: The spatiotemporal correlation structure data is initialized, and an initial feature vector containing both broadband and narrowband information is generated for each node. Two independent aggregation operations are performed on the connection relationships of nodes: the first aggregation operation transmits only feature components within a wide frequency range on each connection, and the second aggregation operation transmits only feature components within a narrow frequency range that are related to a preset potential abnormal frequency band on each connection. During the repeated execution of the two aggregation operations, a decoupling constraint is introduced. The decoupling constraint is achieved by minimizing the ability to predict the result of the second aggregation operation from the result of the first aggregation operation, thereby obtaining a global feature set and a local feature set.

[0009] As a further aspect of the present invention: the generation of the initial feature vector that simultaneously contains broadband and narrowband information specifically includes: Obtain the complete spectral features corresponding to each node from the spatiotemporal correlation structure data; By integrating the spectral energy of the complete spectral characteristics within the first frequency range covering the low-frequency band, the first characteristic component characterizing the broadband information is obtained. For the complete spectral feature, within at least one pre-defined second frequency range with a width smaller than the first frequency range, the statistical distribution skewness of the spectral amplitude of the complete spectral feature is calculated to obtain at least one second feature component characterizing the narrow frequency information. The first feature component is concatenated with all the second feature components to form the initial feature vector.

[0010] As a further aspect of the present invention: S4 specifically includes: Based on the feature values ​​of each node in the global feature set, calculate the global feature differences between all nodes pairwise, and perform a first-level agglomerative partitioning of all nodes according to the global feature differences to form multiple primary state clusters; For each primary state cluster, extract the local feature set of all nodes belonging to the primary state cluster, and calculate the feature value entropy of the local feature set in the dimension corresponding to the preset potential anomaly frequency band. For a primary state cluster whose eigenvalue entropy is greater than a preset entropy threshold, the probability distribution difference of local features between any two nodes within the primary state cluster is further calculated, and the node pairs whose probability distribution difference is greater than a preset difference threshold are marked. Based on the marked node pair relationships, all nodes in the primary state cluster are divided into at least two groups, and the grouping of most anomalous frequency band feature nodes is identified as a secondary anomalous sub-cluster.

[0011] As a further aspect of the present invention: the calculation process of the eigenvalue entropy is as follows: From all nodes of the primary state cluster, extract the feature values ​​of the primary state cluster in one or more dimensions corresponding to the preset potential anomaly frequency band, and form a sequence of feature values ​​to be analyzed. The numerical range of the feature value sequence to be analyzed is divided into multiple continuous intervals. The frequency of the value in the feature value sequence falling into each interval is counted to form a probability quality distribution based on the interval. Calculate the dispersion index of the probability mass distribution and use the dispersion index as the eigenvalue entropy.

[0012] As a further aspect of the present invention: S5 specifically includes: Within a continuous monitoring period, the proportion of the number of nodes contained in the secondary abnormal sub-clusters in each period to the total number of nodes in the monitoring period is counted as the frequency of occurrence of abnormal sub-clusters within the monitoring period. The arithmetic mean of the feature values ​​of all nodes in the abnormal sub-clusters in the abnormal frequency band dimension is calculated as the intensity of abnormal features within the monitoring period. The rates of change for the frequency of occurrence of anomalous subclusters and the intensity of anomalous features are calculated on the time series, respectively. When the frequency of occurrence of abnormal subclusters and the rate of change of the intensity of abnormal features both exceed their respective thresholds for more than three consecutive monitoring periods, the early warning triggering conditions are determined to be met, and an early performance degradation early warning signal is generated. During the monitoring period in which the early warning signal is generated, the operating condition parameters are acquired synchronously, the deviation between the actual power generation efficiency and the theoretical benchmark efficiency during the monitoring period is calculated, and the early warning signal and the deviation are combined together to output the efficiency evaluation result.

[0013] As a further aspect of the present invention: the step of combining the warning signal and the deviation degree together to output a performance evaluation result specifically includes: During the monitoring period, operating condition parameters are acquired, and the corresponding historical best operating efficiency value is queried based on the operating condition parameters. The historical best operating efficiency value is used as the theoretical benchmark efficiency. The actual power generation efficiency value recorded during the monitoring period is compared with the theoretical benchmark efficiency value. The percentage of the absolute difference between the actual power generation efficiency value and the theoretical benchmark efficiency value relative to the theoretical benchmark efficiency value is calculated as the original deviation. The occurrence frequency of secondary abnormal subclusters is introduced as a confidence factor to weight and correct the original deviation, generating a comprehensive deviation. The warning signals, overall deviation, and corresponding operating condition parameters are integrated into a structured diagnostic report, which is then output as the final performance evaluation result.

[0014] The beneficial effects of this invention are: (1) This invention constructs graph structure data that integrates spatiotemporal correlations and employs parallel feature decoupling technology to effectively remove strong background interference (such as broadband vibrations caused by gusts), thereby focusing on and amplifying the weak narrowband resonance features that indicate early faults (such as microcracks in blades). The subsequent two-layer clustering mechanism first coarsely classifies based on the overall state, and then finely searches for local abnormal patterns under the same state, avoiding the omissions caused by global feature masking in traditional methods, and realizing the early detection and accurate location of progressive and complex faults.

[0015] (2) This invention combines in-depth analysis of equipment vibration characteristics with actual power generation performance. It not only identifies abnormal patterns but also quantifies the frequency and intensity trends of these patterns and sets up early warning logic based on continuous and coordinated changes, reducing false alarms caused by occasional interference. At the same time, by correlating and weighting the abnormal warnings with the actual power generation performance loss during the same period, it finally outputs a comprehensive assessment report that integrates the equipment health status and the degree of power generation performance degradation, providing an intuitive and reliable quantitative basis for operation and maintenance decisions and promoting the transformation of operation and maintenance mode from post-maintenance to predictive maintenance. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] 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.

[0019] Please see Figure 1 As shown, this invention is a machine learning-based method for energy efficiency diagnosis and early warning management of wind turbine photovoltaic power plants, comprising the following steps: S1: Collect vibration signals and operating parameters of the power generation equipment of wind turbine and photovoltaic power plants over a continuous period of time to form a condition monitoring sequence; S2: Convert the vibration signal in the state monitoring sequence into spectral features of multiple time segments, take the spectral features of each time segment as a node, and establish the connection relationship and weight between nodes according to the time sequence and spectral morphology similarity to create spatiotemporal correlation structure data; S3: Parallel processing is used to analyze spatiotemporal correlation structure data, simultaneously extracting a global feature set characterizing the overall broadband response of the equipment and a local feature set focusing on preset potential abnormal frequency bands, while constraining the global feature set and local feature set to maintain low statistical correlation during the extraction process; S4: Based on the global feature set, the nodes are divided into multiple primary state clusters at the first level; within each primary state cluster, the nodes are divided into secondary anomalous subclusters based on the local feature set at the second level to identify the anomalous frequency band features. S5: Analyze the occurrence frequency and characteristic intensity of secondary abnormal subclusters over time, generate early performance degradation warning signals for wind turbine photovoltaic power generation equipment based on the changing trend of secondary abnormal subclusters, and output performance evaluation results in combination with operating condition parameters; The power generation equipment is either a wind turbine or a photovoltaic power generation unit.

[0020] In S1, the power generation equipment, in its specific implementation, corresponds to a wind turbine or a photovoltaic power generation unit. For a wind turbine, the collected vibration signal specifically refers to a physical quantity measured by a triaxial accelerometer fixedly mounted on a key mechanical component (such as a gearbox housing or generator bearing housing) inside its nacelle. This sensor continuously records waveform data of the vibration acceleration of the mechanical structure in three orthogonal directions over time.

[0021] Meanwhile, for wind turbines, the operating parameters are acquired through their own monitoring system, specifically including: real-time wind speed and direction data measured by the anemometer on top of the nacelle; active and reactive power output data measured by power transmitters on the generator and grid sides; and generator speed, gearbox oil temperature, pitch angle, and yaw system position data provided by internal sensors. All these parameters are synchronously collected and recorded at time intervals of seconds or minutes.

[0022] When the power generation equipment is a photovoltaic power generation unit, its core components include photovoltaic strings and a grid-connected inverter. In this scenario, the acquisition of vibration signals is not a necessary step. The operating parameters mainly cover electrical and environmental quantities, specifically including: DC-side current and voltage data collected by string current and voltage sensors, AC-side power data output by the inverter, and total solar irradiance, photovoltaic module backsheet temperature, and ambient temperature data recorded by on-site environmental monitoring instruments. These parameters are also collected synchronously at intervals of seconds or minutes.

[0023] In S2, firstly, adaptive time segmentation is performed on the vibration signal. An analysis window with an initial length of 2 seconds is set at the starting point of the vibration signal. The sum of the squares of the amplitudes of all data points within this window is calculated, and then this sum is divided by the total number of data points within the window. The result is the signal energy value of this window. This window is then slid forward along the time axis by a fixed short interval of 0.1 seconds, and the signal energy value of the new window is calculated. The absolute value of the difference between the energy values ​​of two consecutive windows is divided by the time interval between the center points of these two windows, and the resulting value is defined as the signal energy change gradient. The calculated energy change gradient is compared with two pre-set thresholds: the first gradient threshold is set to a low positive number, for example, 50% of the average energy change gradient calculated from historical stable operation data; the second gradient threshold is set to a high positive number, for example, 80% of the average energy change gradient calculated from historical data during strong gust events. The judgment logic is as follows: if the current energy change gradient is lower than the first gradient threshold, it indicates that the signal is stable, so the length of the current analysis window is increased by 0.5 seconds; if it is higher than the second gradient threshold, it indicates that the signal changes drastically, so the length of the current window is decreased by 0.5 seconds; if it is between the two thresholds, the current window length remains unchanged. This process of sliding window, calculating energy change gradient, comparing with thresholds, and dynamically adjusting window length is repeated until the entire vibration signal is traversed, thereby determining a series of final time segments with potentially different lengths. For each final determined time segment, a short-time Fourier transform is applied to obtain its spectral characteristics. Specifically, the signal data points within the time segment are first multiplied by a Hanning window function, and then a discrete Fourier transform is performed on the windowed data to convert the time-domain waveform into a frequency-domain representation. The resulting amplitude spectrum sequence is the spectral characteristics corresponding to that time segment.

[0024] Next, nodes and connections are constructed based on the aforementioned spectral features. Each transformed spectral feature sequence of time segments is defined as an independent node. Based on the chronological order of the time segments corresponding to these nodes on the time axis, connections are established between all nodes that are directly adjacent in time. This type of connection is called the first type of connection, used to characterize the natural and continuous evolution of device states in the time dimension. Furthermore, connections based on state similarity need to be established. The degree of difference between the spectral feature sequences of any two nodes (regardless of whether their times are adjacent) is calculated. Specifically, the amplitude values ​​corresponding to each corresponding frequency point in the two sequences are subtracted, and the absolute value of the difference is taken. Then, these absolute values ​​at all frequency points are summed, and finally, this sum is divided by the total number of frequency points to obtain a single scalar value characterizing the overall morphological difference. This value is called the spectral morphological difference degree. A connection is established for all node pairs whose calculated spectral morphology difference is less than a preset threshold. Such connections are called second-type connections. The preset threshold can be determined by statistical methods. For example, the spectral morphology difference values ​​of all node pairs are arranged in ascending order, and the values ​​in the top 15% are taken as the threshold.

[0025] Finally, weights are assigned to all connections to quantify their relationship strength, and these weights are then merged to generate the final data structure. For each pair of first-type connections determined by temporal adjacency, the weight is set to 1 divided by the quotient of the time difference between the center points of the corresponding time segments of the two nodes. For each pair of second-type connections determined by spectral similarity, the weight is set to 1 divided by the quotient of the calculated spectral morphology difference between the two nodes. By summarizing and combining all nodes obtained through the above process, all first-type connections with the aforementioned calculated weights, and all second-type connections with the aforementioned calculated weights, a complete spatiotemporal relational structure data is created. In this data structure, nodes carry spectral feature information, connections reflect temporal proximity or state similarity relationships, and weights quantitatively describe the relative strength of different relationships.

[0026] In step S3, the spatiotemporal correlation structure data is first initialized, generating an initial feature vector for each node that integrates broadband and narrowband information. The specific implementation is as follows: The complete spectral features corresponding to each node are obtained from the spatiotemporal correlation structure data. This feature is a sequence containing signal amplitude values ​​corresponding to multiple discrete frequency points within the range from 0 Hz to the Nyquist frequency range. A first frequency range covering the main low-frequency band is defined, for example, from 0 Hz to 5 Hz. Within this first frequency range, the complete spectral feature sequence of the node is processed: the amplitude values ​​corresponding to all frequency points within this range are squared, and then all squared results are summed. This sum represents the total energy of the node's signal within this broadband range, serving as the first feature component, reflecting the overall vibration response intensity of the equipment. Next, at least one second frequency range is defined, with each second frequency range having a significantly smaller width than the first frequency range, for example, a width of 0.5 Hz, and its center frequency is pre-set as a potential abnormal frequency band based on the equipment structure (such as the blade's natural frequency). For each second frequency range, the amplitude values ​​falling within this range in the complete spectral feature sequence are extracted to form a subsequence. To calculate the statistical distribution skewness of the subsequence: First, calculate the mean of all amplitude values ​​in the subsequence; then, calculate the difference between each amplitude value and the mean, take the cube of this difference, and then calculate the average of all these cubed values; finally, divide this average by the cube of the standard deviation of the subsequence. The result is the statistical distribution skewness of the spectral amplitude within the second frequency range, which serves as a second characteristic component. This skewness value reflects the symmetry of the energy distribution within a narrow frequency band; an asymmetric distribution may suggest the presence of resonance. The calculated first characteristic component and all (e.g., three) second characteristic components are sequentially concatenated to form a multidimensional vector, which is the initial characteristic vector of the node.

[0027] After initialization, two independent and parallel aggregation operations are performed on the connections between nodes in the spatiotemporal relational structure data to update the node feature representations. The first aggregation operation is specifically used to transmit and aggregate broadband information: for each connection in the spatiotemporal relational structure data (including first and second type connections), only the first feature component (i.e., broadband energy component) in the initial feature vectors of the nodes at both ends of the connection is multiplied by the weight of its corresponding connection, which serves as the message transmitted along that connection. For each node, the broadband message values ​​transmitted from all its neighboring connections (including inflow and outflow) are weighted and summed, where the weighting coefficients are the weights of the corresponding connections. Then, this weighted sum is linearly combined (e.g., added) with the first feature component in the node's own initial feature vector to form a new feature representation of the node that contains neighborhood broadband context information. This process can be iterated multiple times (e.g., three times), with the broadband part of the node features output from the previous round used as new input in each iteration, ultimately obtaining the stable broadband features of each node. These features of all nodes constitute the global feature set. The second aggregation operation is executed in parallel with the first, but focuses only on narrow-frequency information: for each connection, only the second feature components corresponding to the preset potential abnormal frequency bands in the initial feature vectors of the nodes at both ends of the connection are passed. The logic of passing and aggregation is the same as the first operation, but the operation objects are the individual narrow-frequency feature components. After multiple iterations, a stable feature representation of each node containing neighborhood narrow-frequency context information is obtained, and the features of all nodes constitute a local feature set.

[0028] To achieve low statistical correlation between the global and local feature sets, a decoupling constraint needs to be introduced and optimized during the iterative execution of the two aggregation operations. This constraint is implemented by attempting to predict the corresponding local features (narrow-band features) using the global features (i.e., wideband features) of all nodes in the current batch after each iteration. This prediction can be achieved by establishing a simple linear mapping relationship, i.e., finding a set of coefficients such that the difference between the predicted value obtained by linearly combining the global features and the true local feature value is as small as possible. The squared average of the difference between this predicted value and the true value is defined as the prediction error. The goal of the decoupling constraint is to maximize this prediction error, or in other words, minimize the ability to predict local features from global features, by adjusting the internal parameters of the two aggregation operations. By iteratively solving for maximizing this prediction error along with the information transfer optimization objective of the aggregation operation itself, the learned global feature representation can be forced to contain as little information as possible related to local features, thus achieving statistical decoupling between the two types of features. After sufficient iterative optimization, the final output is a global feature set and a local feature set with low statistical correlation.

[0029] In step S4, the aim is to perform hierarchical state classification and anomaly identification on the node feature set output in step S3. Its core objective is to further refine the analysis within each category, building upon the initial classification based on the overall operating state (global features), to uncover local anomaly patterns that may indicate early, minor faults. This step first performs coarse-grained classification of nodes based on global features, forming several primary state clusters. Subsequently, within each primary state cluster, fine-grained analysis is performed based on local features to identify secondary anomaly subclusters containing significant anomaly frequency band features, thus completing the diagnostic process from macro-state classification to micro-anomaly detection.

[0030] First, a first-level node partitioning is performed to form multiple primary state clusters. Specifically, this is implemented based on a global feature set, where each node corresponds to a feature vector representing its overall broadband response. The global feature differences between all pairs of nodes are calculated, using Euclidean distance: for any two nodes... and Let their global eigenvectors be: and ;in, Represents a node global feature vectors Represents a node global feature vectors express The eigenvalues ​​of dimension express The The eigenvalues ​​of the dimension; then the node and Global feature differences between Calculated using the following formula: ; in, express The first eigenvalue, express The second eigenvalue, express The first eigenvalue, express The eigenvalues ​​of the second dimension; This formula calculates the square root of the sum of the squares of the differences between the global feature vectors of two nodes across all dimensions. A larger value indicates a greater difference in the overall operational states of the two nodes. This is done after obtaining all node pairs... Then, an agglomerative hierarchical clustering method is used to partition all nodes: initially, each node is treated as an independent cluster; then, the nodes that are closest to each other (i.e., the clusters are merged iteratively) are merged. The two clusters with the smallest values ​​are selected for further processing until a preset stopping condition is met (e.g., the minimum distance between clusters exceeds a preset threshold, or the number of clusters reaches a preset value). Here, the distance between two clusters is defined as the distance between all pairs of nodes in the two clusters. The average value. Through this process, all nodes are classified into several (e.g., 3 to 5) primary state clusters, and the nodes in each cluster have similar global operating characteristics.

[0031] Subsequently, for each primary state cluster obtained from the partitioning, the discrete stability of its internal states is evaluated. The core of this evaluation is calculating the eigenvalue entropy of the cluster in the dimension corresponding to the preset potential anomaly frequency band. The specific process consists of three steps: First, from all nodes included in the current primary state cluster, feature values ​​of these nodes in one or more specific dimensions corresponding to the preset potential anomaly frequency band are extracted from their local feature sets. For example, if three potential anomaly frequency bands are preset, three feature values ​​corresponding to these three frequency bands are extracted from the local feature vector of each node, and these feature values ​​of all nodes are aggregated to form a feature value sequence to be analyzed. Second, the feature value sequence is discretized. The maximum and minimum values ​​of all values ​​in the sequence are found to determine its numerical range. This numerical range is divided into multiple (e.g., 10) continuous and non-overlapping intervals. Each value in the feature value sequence is counted to fall into one of these 10 intervals, and the proportion of values ​​falling into each interval to the total number of values ​​in the sequence is calculated, thus forming a probability quality distribution based on intervals. ,in This indicates that the value falls into the first... The third step is to calculate the probability of each interval. The Shannon entropy formula from information theory is used for calculation here: ; in, The total number of intervals (10 in this example). For the value to fall into the first The probability of each interval. This represents the base-2 logarithm. The entropy value is calculated when the eigenvalues ​​are highly concentrated (e.g., most values ​​fall into a few intervals). Lower; when the eigenvalues ​​are very uniformly distributed (e.g., uniformly distributed across all intervals), the entropy value is lower. The entropy threshold is relatively high. The preset entropy threshold can be determined based on the entropy distribution of multiple primary state clusters in historical normal operation data; for example, it can be the 80th percentile value after historical entropy values ​​are sorted in ascending order. If the characteristic entropy of the current primary state cluster... If the entropy exceeds the preset threshold, it is determined that the local feature distribution of the nodes in the cluster is discrete and has low stability, and there may be heterogeneous sub-patterns, requiring further analysis.

[0032] For primary state clusters where the eigenvalue entropy exceeds a preset entropy threshold, a more refined second-level partitioning will be performed. First, it is necessary to quantify the pattern differences in local features between any two nodes within the cluster. To do this, the probability distribution differences of local features between each pair of nodes are calculated. For two nodes within the cluster... and We extract the feature values ​​of each node across all dimensions of its local feature set, and treat the local feature vector of each node as an empirical distribution. To quantify the difference between these two distributions, we use Jensen-Shannon divergence. The specific calculation process is as follows: Let the node... The local feature distribution is ,node The for First, calculate their average distribution: Then, calculate separately. and Between divergence ,as well as and Between divergence The KL divergence is calculated as follows: for a discrete probability distribution... Relative to another distribution ,That Then the node and Jensen-Shannon divergence between Given by the following formula: ; This formula calculates The value is a non-negative symmetric metric; a larger value indicates a greater difference in the local operating modes of the two nodes. The calculation is performed for all node pairs. After setting the value, a preset difference threshold is set, which can be based on all calculated values. The statistical distribution of the values ​​is determined, for example, taking all The median of the values. Node pairs with values ​​greater than this preset difference threshold are marked, meaning that these node pairs are significantly different in local feature patterns.

[0033] Finally, based on the node pair relationships marked above, the current primary state cluster is subdivided. All nodes within the cluster are considered as a graph, where each node is the original node. If two node pairs are marked (i.e., ... If the value is greater than a threshold, an edge is connected between the corresponding two nodes in the graph. Using graph connectivity analysis (e.g., finding connected components), all nodes can be divided into at least two unconnected groups. After division, attribute evaluation is performed on each group: the number of nodes within each group whose feature values ​​in the dimension corresponding to a preset potential anomaly frequency band exceed the historical normal range of that dimension (e.g., more than twice the standard deviation of the historical mean) is counted. The group containing the largest number of such "anomaly frequency band feature nodes" is identified as a "secondary anomaly subcluster." This subcluster represents a set of nodes that share a specific local anomaly feature in a context of similar overall state (belonging to the same primary state cluster), thereby accurately locating potential early, weak fault modes.

[0034] In step S5, firstly, the secondary anomalous subclusters are quantitatively tracked over time. A fixed monitoring period is set, for example, 24 hours. At the end of each monitoring period, all secondary anomalous subclusters identified in step S4 within that period are counted. The frequency of occurrence of anomalous subclusters in each monitoring period is calculated as follows: the total number of nodes contained in all secondary anomalous subclusters within that period is accumulated, and then this total number is divided by the total number of nodes generated in step S2 within that monitoring period; the resulting ratio is the frequency of occurrence. Simultaneously, the anomalous feature intensity within each monitoring period is calculated: for each secondary anomalous subcluster within that period, the specific values ​​of all nodes within that subcluster in the local feature dimension corresponding to the preset potential anomalous frequency band in which they were identified are extracted, and the arithmetic mean of these values ​​is calculated; if multiple subclusters exist within a period, the arithmetic mean of these subclusters is averaged again to obtain a single value representing the overall anomalous feature intensity of that period.

[0035] Secondly, trend analysis is performed based on the aforementioned frequency and intensity sequences to trigger early warnings. The frequency of occurrence of anomalous subclusters and the intensity of anomalous features are considered as two independent time series. For each series, its rate of change is calculated: based on the value of the current monitoring period, subtract the value of the previous monitoring period, and then divide by the value of the previous monitoring period; the resulting percentage is the instantaneous rate of change of the series in the current period. Two thresholds are set: a frequency rate of change threshold and an intensity rate of change threshold. These two thresholds can be determined based on the statistical characteristics of these two rate of change sequences during historical stable operation; for example, both can be set to +10%. The early warning triggering logic is as follows: when the instantaneous rate of change of the frequency of occurrence of anomalous subclusters exceeds the frequency rate of change threshold for more than three consecutive monitoring periods, and the instantaneous rate of change of the intensity of anomalous features also exceeds the intensity rate of change threshold within the same three consecutive monitoring periods, the early warning triggering condition is met. At this time, a clear early performance degradation early warning signal is generated, which at least includes the trigger time and trigger indicator information.

[0036] Next, within the monitoring period triggered by the early warning, a precise quantitative assessment of performance loss is conducted. Simultaneously, actual operating parameters within this monitoring period are acquired, such as average wind speed, solar irradiance, and ambient temperature. Based on these specific operating parameters, the highest power generation efficiency or maximum output power value achieved by the power generation equipment under the same or similar operating conditions is retrieved from the historical operating database. This historical best value is used as the theoretical benchmark performance under the current operating conditions. The average power generation efficiency or total power generation actually recorded within the current monitoring period is calculated, and this actual power generation performance value is compared with the theoretical benchmark performance value. The absolute difference between the two is calculated, then this absolute difference is divided by the theoretical benchmark performance value, and multiplied by 100% to obtain the original deviation, which directly reflects the percentage performance loss of the power generation equipment relative to its historical best level under the current operating conditions.

[0037] Finally, a comprehensive performance evaluation result is generated and output. To improve the reliability and stability of the evaluation result, the frequency of occurrence of abnormal subclusters calculated in the current monitoring period is introduced as a reliability factor to correct the original deviation. The specific weighted correction method is as follows: a baseline reliability coefficient is preset, for example, 1.0; the frequency of occurrence of abnormal subclusters (a value between 0 and 1) is multiplied by an amplification factor greater than 1 (for example, 5), and then added to the baseline reliability coefficient to obtain the final reliability weight; this weight is multiplied by the original deviation to obtain the comprehensive deviation. This processing gives a higher confidence weight to the performance deviation indicated by the frequent occurrence of abnormal patterns. Finally, the generated early performance degradation warning signal, the calculated comprehensive deviation, and the operating condition parameters of the current monitoring period used as the analysis background are integrated into a structured data record or report document as the complete performance evaluation result output.

[0038] The working principle of this invention is as follows: First, vibration signals and operating parameters of power generation equipment (wind turbines or photovoltaic power generation units) are collected over a continuous time period to form a condition monitoring sequence. Then, adaptive time segmentation and spectral transformation are performed on the vibration signals, defining the spectral characteristics of each time segment as a node, and weighted connections are established based on time proximity and spectral similarity to create spatiotemporally correlated structural data. Next, this structural data is analyzed using parallel processing, simultaneously extracting a global feature set characterizing the overall broadband response of the equipment and a local feature set focusing on preset potential abnormal frequency bands. During the extraction process, optimization and decoupling constraints are implemented. The conditions ensure that the two types of features maintain low statistical correlation. Subsequently, the nodes are divided into multiple primary state clusters based on the global feature set at the first level. Within each cluster, a second level division is performed based on the local feature set. By calculating the entropy of the feature value and the difference in probability distribution between node pairs, secondary abnormal subclusters containing significant abnormal frequency band features are identified. Finally, the changing trends of the occurrence frequency and feature intensity of the secondary abnormal subclusters over time are analyzed. When the rate of change of both exceeds the threshold within a continuous monitoring period, an early warning is triggered. The deviation between the actual performance and the theoretical benchmark is calculated in combination with the operating condition parameters, generating a comprehensive early performance degradation warning signal and performance evaluation results.

[0039] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A machine learning-based method for energy efficiency diagnosis and early warning management of wind turbine photovoltaic power plants, characterized in that, Includes the following steps: S1: Collect vibration signals and operating parameters of the power generation equipment of wind turbine and photovoltaic power plants over a continuous period of time to form a condition monitoring sequence; S2: Convert the vibration signal in the state monitoring sequence into spectral features of multiple time segments, take the spectral features of each time segment as a node, and establish the connection relationship and weight between nodes according to the time sequence and spectral morphology similarity to create spatiotemporal correlation structure data; S3: Parallel processing is used to analyze spatiotemporal correlation structure data, simultaneously extracting a global feature set characterizing the overall broadband response of the equipment and a local feature set focusing on preset potential abnormal frequency bands, while constraining the global feature set and local feature set to maintain low statistical correlation during the extraction process; S4: Based on the global feature set, the nodes are divided into multiple primary state clusters at the first level; within each primary state cluster, the nodes are divided into secondary anomalous subclusters based on the local feature set at the second level to identify the anomalous frequency band features. S5: Analyze the frequency and characteristic intensity of secondary abnormal subclusters over time, generate early performance degradation warning signals for wind turbine photovoltaic power generation equipment based on the changing trend of secondary abnormal subclusters, and output performance evaluation results in combination with operating condition parameters.

2. The method for energy efficiency diagnosis and early warning management of wind turbine photovoltaic power plants based on machine learning according to claim 1, characterized in that, S2 specifically includes: Adaptive time segmentation is performed on the vibration signal, the length of each time segment is adjusted according to the energy change rate of the vibration signal, and spectral transformation is performed on each time segment to obtain the corresponding spectral characteristics; Each time segment's spectral features are defined as a node. Based on the temporal order of the nodes, a first-type connection is established for nodes corresponding to adjacent time segments. The spectral morphology difference between any two nodes is calculated, and a second-type connection is established for node pairs with a difference less than a preset threshold. The first type of connection is assigned a weight based on the reciprocal of the time interval, and the second type of connection is assigned a weight based on the inverse of the spectral morphology difference. All nodes and weighted connections are then fused to generate spatiotemporal relational structure data.

3. The method for energy efficiency diagnosis and early warning management of wind turbine photovoltaic power plants based on machine learning according to claim 2, characterized in that, The step of performing a spectral transformation on each time segment to obtain the corresponding spectral features specifically includes: An initial analysis window is set on the vibration signal, and the energy value of the signal within the initial analysis window is calculated. Slide the initial analysis window along the time axis to calculate the signal energy change gradient between consecutive windows in real time, and compare the energy change gradient with the preset first gradient threshold and second gradient threshold. When the energy change gradient is below the first gradient threshold, the length of the current analysis window is expanded; when the energy change gradient is above the second gradient threshold, the length of the current analysis window is shrunk; otherwise, the current analysis window length is maintained, thereby determining each final time segment. Perform spectral transformation independently on each final time segment to generate the corresponding spectral features.

4. The method for energy efficiency diagnosis and early warning management of wind turbine photovoltaic power plants based on machine learning according to claim 1, characterized in that, S3 specifically includes: The spatiotemporal correlation structure data is initialized, and an initial feature vector containing both broadband and narrowband information is generated for each node. Two independent aggregation operations are performed on the connection relationships of nodes: the first aggregation operation transmits only feature components within a wide frequency range on each connection, and the second aggregation operation transmits only feature components within a narrow frequency range that are related to a preset potential abnormal frequency band on each connection. During the repeated execution of the two aggregation operations, a decoupling constraint is introduced. The decoupling constraint is achieved by minimizing the ability to predict the result of the second aggregation operation from the result of the first aggregation operation, thereby obtaining a global feature set and a local feature set.

5. The machine learning-based energy efficiency diagnosis and early warning management method for wind turbine photovoltaic power plants according to claim 4, characterized in that, The generation of the initial feature vector, which simultaneously contains both broadband and narrowband information, specifically includes: Obtain the complete spectral features corresponding to each node from the spatiotemporal correlation structure data; By integrating the spectral energy of the complete spectral characteristics within the first frequency range covering the low-frequency band, the first characteristic component characterizing the broadband information is obtained. For the complete spectral feature, within at least one pre-defined second frequency range with a width smaller than the first frequency range, the statistical distribution skewness of the spectral amplitude of the complete spectral feature is calculated to obtain at least one second feature component characterizing the narrow frequency information. The first feature component is concatenated with all the second feature components to form the initial feature vector.

6. The method for energy efficiency diagnosis and early warning management of wind turbine photovoltaic power plants based on machine learning according to claim 1, characterized in that, S4 specifically includes: Based on the feature values ​​of each node in the global feature set, calculate the global feature differences between all nodes pairwise, and perform a first-level agglomerative partitioning of all nodes according to the global feature differences to form multiple primary state clusters; For each primary state cluster, extract the local feature set of all nodes belonging to the primary state cluster, and calculate the feature value entropy of the local feature set in the dimension corresponding to the preset potential anomaly frequency band. For a primary state cluster whose eigenvalue entropy is greater than a preset entropy threshold, the probability distribution difference of local features between any two nodes within the primary state cluster is further calculated, and the node pairs whose probability distribution difference is greater than a preset difference threshold are marked. Based on the marked node pair relationships, all nodes in the primary state cluster are divided into at least two groups, and the grouping of most anomalous frequency band feature nodes is identified as a secondary anomalous sub-cluster.

7. The method for energy efficiency diagnosis and early warning management of wind turbine photovoltaic power plants based on machine learning according to claim 6, characterized in that, The calculation process of the eigenvalue entropy is as follows: From all nodes of the primary state cluster, extract the feature values ​​of the primary state cluster in one or more dimensions corresponding to the preset potential anomaly frequency band, and form a sequence of feature values ​​to be analyzed. The numerical range of the feature value sequence to be analyzed is divided into multiple continuous intervals. The frequency of the value in the feature value sequence falling into each interval is counted to form a probability quality distribution based on the interval. Calculate the dispersion index of the probability mass distribution and use the dispersion index as the eigenvalue entropy.

8. The method for energy efficiency diagnosis and early warning management of wind turbine photovoltaic power plants based on machine learning according to claim 1, characterized in that, S5 specifically includes: Within a continuous monitoring period, the proportion of the number of nodes contained in the secondary abnormal sub-clusters in each period to the total number of nodes in the monitoring period is counted as the frequency of occurrence of abnormal sub-clusters within the monitoring period. The arithmetic mean of the feature values ​​of all nodes in the abnormal sub-clusters in the abnormal frequency band dimension is calculated as the intensity of abnormal features within the monitoring period. The rates of change for the frequency of occurrence of anomalous subclusters and the intensity of anomalous features are calculated on the time series, respectively. When the frequency of occurrence of abnormal subclusters and the rate of change of the intensity of abnormal features both exceed their respective thresholds for more than three consecutive monitoring periods, the early warning triggering conditions are determined to be met, and an early performance degradation early warning signal is generated. During the monitoring period in which the early warning signal is generated, the operating condition parameters are acquired synchronously, the deviation between the actual power generation efficiency and the theoretical benchmark efficiency during the monitoring period is calculated, and the early warning signal and the deviation are combined together to output the efficiency evaluation result.

9. The method for energy efficiency diagnosis and early warning management of wind turbine photovoltaic power plants based on machine learning according to claim 8, characterized in that, The process of combining the warning signal and the deviation rate into a performance evaluation result for output specifically includes: During the monitoring period, operating condition parameters are acquired, and the corresponding historical best operating efficiency value is queried based on the operating condition parameters. The historical best operating efficiency value is used as the theoretical benchmark efficiency. The actual power generation efficiency value recorded during the monitoring period is compared with the theoretical benchmark efficiency value. The percentage of the absolute difference between the actual power generation efficiency value and the theoretical benchmark efficiency value relative to the theoretical benchmark efficiency value is calculated as the original deviation. The occurrence frequency of secondary abnormal subclusters is introduced as a confidence factor to weight and correct the original deviation, generating a comprehensive deviation. The warning signals, overall deviation, and corresponding operating condition parameters are integrated into a structured diagnostic report, which is then output as the final performance evaluation result.

Images