A multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system and method

Abstract

The application discloses a multi-source heterogeneous data cross-scale synchronization and abnormality diagnosis system and method, and belongs to the technical field of photovoltaic system intelligent operation and maintenance. The system comprises a data acquisition module, a data synchronization module, a feature extraction and fusion module, an abnormality diagnosis module and an application display module. Millisecond-level electrical data, second-level meteorological data and minute-level environmental data are respectively acquired; a sliding time window is used to directly read the electrical data, to perform three times spline interpolation on the meteorological data, and to perform forward filling on the environmental data, so as to form a synchronous data frame on a unified time axis; electrical, meteorological and environmental features are extracted and fused, a fault type and a confidence degree are determined in combination with a device topology graph, a graph neural network and a Bayesian network, and alarm information, a to-be-reviewed prompt or a manual confirmation mark is output according to the confidence degree. The application can reduce information distortion caused by cross-scale data alignment, retain electrical transient characteristics, and improve the real-time performance and reliability of photovoltaic system abnormality diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for photovoltaic systems, specifically to a system and method for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data for photovoltaic systems, which can be used for operation monitoring and fault diagnosis of solar-fishery complementary photovoltaic systems, distributed photovoltaic power stations and centralized photovoltaic power stations. Background Technology

[0002] As the installed capacity of photovoltaic (PV) power generation continues to grow, the types of data that need to be collected and processed in PV systems are constantly increasing. PV system operation monitoring data typically includes electrical data, meteorological data, and environmental data. Electrical data includes PV string voltage, current, power, and harmonic distortion rate, reflecting the instantaneous electrical state of the PV system. Meteorological data includes irradiance, temperature, humidity, and wind speed, reflecting the impact of external meteorological conditions on the PV system's power generation performance. In applications such as fish-solar hybrid systems, environmental data also includes water quality parameters such as dissolved oxygen and pH, reflecting changes in the aquatic environment.

[0003] The different types of data mentioned above have different sampling periods and variation patterns. The sampling period for electrical data is typically in the millisecond range and may include instantaneous changes such as arc discharges and voltage spikes; the sampling period for meteorological data is typically in the second range, and its variation process has a certain degree of continuity; the sampling period for environmental data is typically in the minute range, and its variation process is relatively slow. The three types of data differ significantly in time scale.

[0004] Existing data processing methods typically employ a uniform upsampling or downsampling approach to align data across different time scales. Simply repeating low-frequency data or using a single interpolation method to increase the sampling frequency may produce intermediate values ​​that do not reflect the actual data variation patterns. Conversely, downsampling high-frequency electrical data to a lower frequency may result in the loss of transient information useful for early fault diagnosis. Furthermore, when different data sources have independent timestamps, the lack of a unified time reference can easily lead to time misalignments between different modal characteristics.

[0005] Existing anomaly diagnosis methods still have the following problems: some diagnostic methods mainly rely on single electrical data for judgment, making it difficult to distinguish between equipment failure and external environmental disturbances when irradiance or temperature changes rapidly; some diagnostic methods only output the fault type and cannot simultaneously provide the confidence level of the diagnostic result; some diagnostic models do not combine the topological correlation between photovoltaic strings, inverters and sensors, or do not make joint decisions based on topological correlation characteristics and fault probability, thus making it difficult to make full use of the correlation information between multi-source data.

[0006] Therefore, there is a need for an anomaly diagnosis system that can perform differentiated time synchronization based on the sampling period and physical change characteristics of different types of data, and can integrate multimodal features, equipment topology correlation features and fault probability. Summary of the Invention

[0007] To address the aforementioned shortcomings of existing technologies, this invention proposes a multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system and method. The aim is to solve the problems of inconsistent time scales of electrical, meteorological, and environmental data in photovoltaic systems, the tendency for unified upsampling or downsampling to cause information distortion or loss of transient information, insufficient utilization of multimodal features and equipment topology correlations, and lack of confidence quantification in diagnostic results.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A system for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data, characterized in that it includes:

[0010] The data acquisition module is used to acquire electrical data, meteorological data, and environmental data of the photovoltaic system according to a first sampling period, a second sampling period, and a third sampling period, respectively, wherein the first sampling period, the second sampling period, and the third sampling period increase sequentially;

[0011] The data synchronization module is used to directly retrieve the electrical data within a time window that slides along the time axis, with the first sampling period as the target time resolution, perform cubic spline interpolation on the meteorological data, perform forward filling on the environmental data, and align the processed electrical data, meteorological data, and environmental data into a synchronized data frame.

[0012] The feature extraction and fusion module is used to extract electrical features, meteorological features and environmental features from the synchronous data frame, and fuse the electrical features, meteorological features and environmental features into a fused feature vector;

[0013] The anomaly diagnosis module includes a topology graph construction part, a graph neural network unit, a Bayesian network unit, and a joint decision unit. The topology graph construction part is used to construct a device topology graph based on the connection relationships between devices in the photovoltaic system. The graph neural network unit is used to process the device topology graph and output a topology feature vector. The Bayesian network unit is used to determine the posterior probability of each fault type based on a preset fault mode library and the fused feature vector. The joint decision unit is used to perform weighted fusion of the topology feature vector and the posterior probability to determine the final fault type. The anomaly diagnosis module is also used to determine a confidence score based on the posterior probability of each fault type.

[0014] Furthermore, the data acquisition module includes an electrical data acquisition unit, a meteorological data acquisition unit, and an environmental data acquisition unit; the electrical data acquisition unit acquires the voltage, current, power, and harmonic distortion rate of the photovoltaic string through an IV acquisition device or an inverter communication interface; the meteorological data acquisition unit acquires irradiance, temperature, humidity, and wind speed through a illuminance sensor, a temperature and humidity sensor, and a wind speed sensor; the environmental data acquisition unit acquires dissolved oxygen and pH value through a water quality sensor.

[0015] Furthermore, the first sampling period is 1 millisecond, the second sampling period is 1 second, and the third sampling period is 1 minute; the length of the time window is 100 milliseconds, the step size of the time window sliding along the time axis is 1 millisecond, and the time resolution of the synchronization data frame is 1 millisecond.

[0016] Furthermore, the data synchronization module selects the measured meteorological data values ​​adjacent to the current target time before and after the current target time, and determines the meteorological data interpolation value at the current target time based on the measured meteorological data values ​​and the first derivative continuity condition and second derivative continuity condition of the adjacent interpolation interval.

[0017] Furthermore, the electrical features include harmonic distortion rate, power change rate, and voltage drop depth; the meteorological features include irradiance change rate and temperature change rate; and the environmental features include pH change rate. The feature extraction and fusion module concatenates the electrical features, meteorological features, and environmental features in a preset order to form the fused feature vector. Alternatively, an attention fusion mechanism can be used, which adjusts the importance of different categories of features in the fusion process through learnable attention weights, and forms the fused feature vector according to the attention weights.

[0018] Furthermore, the nodes of the device topology graph include photovoltaic string nodes, inverter nodes, meteorological sensor nodes, and environmental sensor nodes. The edges of the device topology graph include electrical connection edges between the inverter and the photovoltaic string, and data association edges between the meteorological sensor and the environmental sensor and the inverter. The graph neural network unit adopts a two-layer graph convolutional network, adds self-connection relationships of each node to the adjacency matrix of the device topology graph, and propagates the node feature matrix layer by layer according to the adjacency matrix after adding the self-connection relationships to obtain the topology feature vector containing the spatial topology association information between devices.

[0019] Furthermore, the preset fault mode library includes feature modes corresponding to inverter faults, partial shading, sudden changes in water quality, compound faults, and normal operation; the Bayesian network unit determines the prior probability of each fault type based on historical fault data, determines the likelihood information corresponding to each fault type through maximum likelihood estimation, and determines the posterior probability of each fault type based on the prior probability, the likelihood information, and the fused feature vector; the anomaly diagnosis module determines the highest posterior probability among each fault type as the confidence score.

[0020] Furthermore, it includes an application display module connected to the anomaly diagnosis module; the anomaly diagnosis module generates an alarm message when the confidence score is not lower than 90%, generates a verification prompt when the confidence score is not lower than 70% but lower than 90%, and generates a manual confirmation mark when the confidence score is lower than 70%; the application display module is used to display the final fault type and the confidence score, and output the alarm message, the verification prompt, or the manual confirmation mark.

[0021] Furthermore, the data acquisition module and the data synchronization module are deployed at the edge of the photovoltaic power station. The edge is also used to perform feature extraction, Bayesian network inference, and alarm push. The cloud is used to train and update the graph neural network model, store historical data and fault cases, and process samples that require manual confirmation uploaded by the edge. The edge communicates with the cloud through message queue telemetry transmission protocol or hypertext transfer protocol.

[0022] This invention also provides a method for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data, characterized by comprising:

[0023] Electrical data, meteorological data, and environmental data of the photovoltaic system are collected according to the first sampling period, the second sampling period, and the third sampling period, respectively, with the first sampling period, the second sampling period, and the third sampling period increasing sequentially.

[0024] Within a time window that slides along the time axis, with the first sampling period as the target time resolution, the electrical data is directly used, cubic spline interpolation is performed on the meteorological data, forward filling is performed on the environmental data, and the processed electrical data, meteorological data and environmental data are aligned into a synchronous data frame.

[0025] Electrical features, meteorological features, and environmental features are extracted from the synchronous data frames, and the electrical features, meteorological features, and environmental features are fused into a fused feature vector;

[0026] A device topology graph is constructed based on the connection relationships between devices in the photovoltaic system. The device topology graph is then processed using a graph neural network to obtain topology feature vectors.

[0027] Based on the preset fault mode library and the fused feature vector, the posterior probability of each fault type is determined using a Bayesian network.

[0028] The feature vector and the posterior probability are weighted and fused, and the final fault type is determined based on the weighted fusion result. The confidence score is determined based on the posterior probability of each fault type.

[0029] Furthermore, the first sampling period is 1 millisecond, the second sampling period is 1 second, and the third sampling period is 1 minute; the length of the time window is 100 milliseconds, the step size of the time window sliding along the time axis is 1 millisecond, and the time resolution of the synchronization data frame is 1 millisecond.

[0030] Further, cubic spline interpolation is performed on the meteorological data, including: selecting measured meteorological data values ​​that are located before and after the current target time and adjacent to the current target time; and determining the interpolated meteorological data value for the current target time based on the measured meteorological data values ​​and the first derivative continuity condition and second derivative continuity condition of the adjacent interpolation interval.

[0031] Furthermore, the electrical features include harmonic distortion rate, power change rate, and voltage drop depth; the meteorological features include irradiance change rate and temperature change rate; and the environmental features include pH change rate. The electrical features, meteorological features, and environmental features are concatenated in a preset order to form the fused feature vector; or, the importance of different categories of features in the fusion process is adjusted through an attention fusion mechanism, and the fused feature vector is formed according to the corresponding attention weights.

[0032] Further, constructing the device topology graph includes: using photovoltaic strings, inverters, meteorological sensors, and environmental sensors as nodes; establishing electrical connection edges based on the electrical connection relationship between the inverter and the photovoltaic strings, and establishing data association edges based on the data association relationship between the meteorological sensors, environmental sensors, and the inverter; processing the device topology graph using a two-layer graph convolutional network, adding self-connection relationships of each node to the adjacency matrix of the device topology graph, and propagating the node feature matrix layer by layer based on the adjacency matrix after adding the self-connection relationships to obtain the topology feature vector containing spatial topology association information between devices.

[0033] Furthermore, the preset fault mode library includes feature modes corresponding to inverter faults, partial shading, sudden changes in water quality, compound faults, and normal operation; the prior probability of each fault type is determined based on historical fault data, the likelihood information corresponding to each fault type is determined through maximum likelihood estimation, and the posterior probability of each fault type is determined based on the prior probability, the likelihood information, and the fused feature vector; the highest posterior probability among each fault type is determined as the confidence score; an alarm message is generated when the confidence score is not lower than 90%, a verification prompt is generated when the confidence score is not lower than 70% but lower than 90%, and a manual confirmation mark is generated when the confidence score is lower than 70%.

[0034] Furthermore, the method performs data acquisition, cross-scale synchronization, feature extraction, Bayesian network inference, and alarm push at the edge of the photovoltaic power station, trains and updates the graph neural network model, stores historical data and fault cases in the cloud, and processes samples requiring manual confirmation uploaded by the edge. The edge communicates with the cloud via message queue telemetry transmission protocol or hypertext transfer protocol, and the edge compresses the data to be uploaded before transmitting it to the cloud.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] (1) According to the sampling period and physical change characteristics of electrical data, meteorological data and environmental data, the present invention adopts direct use, cubic spline interpolation and forward filling respectively, so that data of different time scales are aligned on a unified time axis, while retaining transient information in electrical data and reducing human distortion of meteorological data and environmental data.

[0037] (2) The present invention extracts electrical features, meteorological features and environmental features from the synchronized data, and uses serial fusion or attention fusion to form a fusion feature vector. It can utilize the complementary relationship between different modal data to reduce the possibility of being affected by external environmental disturbances when diagnosing based solely on a single electrical data.

[0038] (3) The present invention extracts the spatial topological correlation features between photovoltaic strings, inverters and sensors through graph neural networks, determines the posterior probability of different fault types through Bayesian networks, and makes joint decisions on topological feature vectors and posterior probabilities, so that the diagnostic process considers both equipment topological relationships and fault uncertainties.

[0039] (4) The present invention adopts a collaborative deployment method between the edge and the cloud, performing data collection, cross-scale synchronization and real-time inference at the edge, and performing model training, updating and historical data management at the cloud, which is conducive to balancing real-time response and model computing needs.

[0040] The present invention will be further described below with reference to the accompanying drawings and specific examples. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the overall architecture of the multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system and method of the present invention;

[0042] Figure 2 This is a schematic diagram of the sliding time window dynamic interpolation process of the present invention;

[0043] Figure 3 This is a schematic diagram of the structure of the joint diagnostic model of graph neural network and Bayesian network of the present invention;

[0044] Figure 4 This is a schematic diagram of the edge-cloud collaborative deployment architecture of the present invention. Detailed Implementation

[0045] 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.

[0046] Example 1

[0047] like Figure 1 As shown, the multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system of this embodiment includes a data acquisition module, a data synchronization module, a feature extraction and fusion module, an anomaly diagnosis module, and an application display module for displaying diagnostic results and pushing alarm information.

[0048] The data acquisition module is installed at the photovoltaic power station site and includes an electrical data acquisition unit, a meteorological data acquisition unit, and an environmental data acquisition unit.

[0049] The electrical data acquisition unit collects the voltage (V), current (I), power (P), and harmonic distortion (THD) of each photovoltaic string via an IV acquisition device or inverter communication interface, with a sampling period of 1 millisecond. This electrical data reflects the instantaneous operating status of the photovoltaic strings and the inverter.

[0050] The meteorological data acquisition unit collects irradiance (G), temperature (T), humidity (RH), and wind speed (WS) with a sampling period of 1 second using illuminance sensors, temperature and humidity sensors, and wind speed sensors.

[0051] The environmental data acquisition unit uses water quality sensors to collect dissolved oxygen (DO) and pH values ​​at 1-minute sampling intervals. The environmental data is primarily used in applications requiring aquatic environment monitoring, such as fisheries-solar hybrid systems.

[0052] The three types of data, each carrying a collection timestamp, are written to the collection buffer. The data synchronization module reads the three types of data from the collection buffer and aligns them to the same timeline using dynamic interpolation via a sliding time window. The feature extraction and fusion module extracts multimodal features from the synchronized data and forms a fused feature vector. The anomaly diagnosis module performs anomaly diagnosis based on the fused feature vector, device topology association information, and a preset fault mode library. The application display module displays the fault type and confidence score, and pushes corresponding information based on the confidence score.

[0053] Example 2

[0054] like Figure 2 As shown, the method for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data in this implementation will perform cross-scale synchronization of multi-source heterogeneous data according to the following steps.

[0055] Step S201: Set the time window length W to 100 milliseconds, the sliding step size S to 1 millisecond, and set the target time resolution to 1 millisecond.

[0056] Step S202: Move the time window along the time axis with a sliding step of 1 millisecond. For each time window with a time range of [t, t+W], set the target time with an interval of 1 millisecond.

[0057] Step S203: For electrical data with a sampling period of 1 millisecond, the data value with the corresponding timestamp of the target time is directly read from the acquisition buffer without interpolation. This process avoids the loss of transient information such as arc discharge and voltage spikes due to downsampling or smoothing.

[0058] Step S204: For meteorological data with a sampling period of 1 second, determine the measured meteorological data points located before and after the current target time and adjacent to the current target time, and construct a cubic spline interpolation function by combining the first and second derivative continuity conditions of the data within adjacent sampling intervals. For the i-th interpolation interval containing the current target time, the interpolated meteorological data value is expressed as:

[0059]

[0060] in, For the current target time, For the first The initial measured time of each interpolation interval , , and The coefficients are determined based on measured meteorological data and the continuity conditions of the first and second derivatives of adjacent interpolation intervals.

[0061] Step S205: For environmental data with a sampling period of 1 minute, a forward padding method is used, where the most recent measured environmental data value, no later than the current target time, is used as the environmental data padding value for the current target time, i.e.:

[0062]

[0063] in, It is the most recent measured value of environmental data no later than the current target time.

[0064] Step S206: Combine the electrical data, meteorological data interpolation values, and environmental data filler values ​​at the same target time into a synchronization data frame. The synchronization data frame can be represented as:

[0065]

[0066] Step S207: Repeat steps S202 to S206 according to the sliding step size until the synchronization of the data to be processed is completed.

[0067] When verifying the synchronization error, a synchronization pulse signal with a known timestamp can be input to different acquisition channels. The data of each channel is then processed using the synchronization method described above. The output timestamp is compared with the known timestamp, and the average absolute error and maximum absolute error are calculated.

[0068] Example 3

[0069] The feature extraction and fusion module extracts electrical features, meteorological features, and environmental features from the synchronous data frames.

[0070] Electrical characteristics include harmonic distortion rate Power change rate and voltage sag depth. Among them, harmonic distortion rate... Take the data from the current synchronized data frame The power change rate is determined based on the power difference and time interval between adjacent synchronous data frames; voltage sag depth. Represented as:

[0071]

[0072] in, This is the reference voltage under normal operating conditions. This is the current voltage.

[0073] Meteorological characteristics include the rate of change of irradiance and the rate of change of temperature. The rate of change of irradiance is determined based on the difference in irradiance between adjacent synchronized data frames and the time interval, while the rate of change of temperature is determined based on the difference in temperature between adjacent synchronized data frames and the time interval.

[0074] Environmental characteristics include The rate of change of value, the The rate of change in pH is determined based on the pH difference between adjacent environmental data and the corresponding time interval. In an alternative embodiment, the rate of change in dissolved oxygen can also be determined in the same manner and used as an environmental characteristic.

[0075] In one implementation, the feature extraction and fusion module employs a series fusion method, splicing harmonic distortion rate, power change rate, voltage drop depth, irradiance change rate, temperature change rate, and pH change rate in a preset order to form a fused feature vector. :

[0076]

[0077] In another implementation, the feature extraction and fusion module employs an attention fusion mechanism. For the first... Class features Its attention weight Represented as:

[0078]

[0079] The fused feature vector is represented as:

[0080]

[0081] in, These are learnable weight parameters. For bias parameters, This represents the normalized exponential function. Through training... and This allows the attention weights for different categories of features to adapt to the input data.

[0082] In one diagnostic criterion setting method, a harmonic distortion rate greater than 7% is marked as a harmonic anomaly; an irradiance change rate greater than 200 W / m² / s is marked as a rapid irradiance change; and an absolute value of the pH change rate greater than 0.1 / min is marked as a sudden change in water quality parameters. These markings, along with other features, are input into the anomaly diagnosis module, rather than determining the final fault type based solely on a single threshold.

[0083] Example 4

[0084] like Figure 3 As shown, the anomaly diagnosis module includes a topology graph construction part, a graph neural network unit, a Bayesian network unit, and a joint decision-making unit.

[0085] The topology graph construction section represents the photovoltaic field area as a topology graph:

[0086]

[0087] Node set Includes photovoltaic string nodes, inverter nodes, meteorological sensor nodes, and environmental sensor nodes; edge set This includes the electrical connection edges between the inverter and the photovoltaic string, as well as the data association edges between the weather and environmental sensors and the inverter. An adjacency matrix is ​​established based on the connection relationships between the nodes. .

[0088] The graph neural network unit uses a two-layer graph convolutional network. The layer-by-layer propagation relationship of the graph convolutional network is represented as:

[0089]

[0090] The adjacency matrix after adding node self-connection relationships is represented as follows:

[0091]

[0092] In the formula, The identity matrix; and the adjacency matrix The corresponding degree matrix satisfy:

[0093]

[0094] The feature matrix of the nodes in the l-th layer; Let be the learnable weight matrix of the l-th layer; For the activation function, ReLU activation is used in this embodiment. After two layers of graph convolution processing, the output is a topological feature vector containing spatial topological association information between devices. .

[0095] Bayesian network units are equipped with a preset fault mode library. The fault mode library includes the following fault types and their corresponding characteristic patterns:

[0096]

[0097] Bayesian network units determine the prior probability of each fault type based on historical fault data, and determine the likelihood function corresponding to different fault types through maximum likelihood estimation. Fault Types The posterior probability is expressed as:

[0098]

[0099] Where X is the fused feature vector, P( ) represents the fault type The prior probability, In the fault type The likelihood of the fused feature vector X under the given conditions. Fault type The posterior probability.

[0100] The joint decision-making unit performs a weighted fusion of the topological feature vector output by the graph neural network unit and the posterior probability output by the Bayesian network unit. The output of the graph neural network unit is then mapped by a mapping function. Converted into topology judgment values ​​corresponding to each fault type, the final fault type Represented as:

[0101]

[0102] in, This represents the fusion weight between the Bayesian posterior probability and the topological decision value. The topological feature vector corresponds to the fault type. The topology judgment value.

[0103] Confidence score Determined based on the posterior probability of each fault type:

[0104]

[0105] When the confidence score is not lower than 90%, the anomaly diagnosis module generates an alarm message and pushes it to the application display module; when the confidence score is not lower than 70% but lower than 90%, a prompt for review is generated; when the confidence score is lower than 70%, a mark requiring manual confirmation is generated. Samples with low confidence scores can be uploaded to the cloud as input for subsequent model updates.

[0106] Example 5

[0107] like Figure 4 As shown, the system adopts a collaborative deployment approach between the edge and the cloud.

[0108] The edge computing unit is deployed at the photovoltaic power station site to run the data acquisition and synchronization modules, and performs feature extraction, real-time Bayesian network inference, and alarm push notifications. The edge computing unit can be an edge computing device that meets the corresponding data acquisition and computing requirements.

[0109] Deployed in the cloud on a remote server, it is used to train and update graph neural network models, store historical data and failure cases, and process samples requiring manual verification uploaded from edge devices. After the samples requiring manual verification are reviewed, they are added to the training dataset for incremental learning of the model.

[0110] The edge device communicates with the cloud via message queue telemetry transmission protocol or hypertext transfer protocol. The edge device compresses the data to be uploaded before transmitting it to the cloud to reduce bandwidth consumption; updated model parameters obtained from cloud training can be distributed to the edge device. The topological feature vector output by the graph neural network unit and the posterior probability output by the Bayesian network unit are passed to the joint decision unit, which determines the final fault type and confidence score.

[0111] During system operation, the data acquisition module continuously collects electrical, meteorological, and environmental data; the data synchronization module performs cross-scale synchronization of the three types of data according to a sliding time window; the feature extraction and fusion module generates fused feature vectors; the anomaly diagnosis module completes topological association analysis, Bayesian inference, and joint decision-making; and the application display module displays the diagnostic results and outputs alarm information, prompts for review, or marks requiring manual confirmation based on the confidence score, thus forming a complete processing flow from data acquisition, time synchronization, feature fusion to anomaly diagnosis and alarm output.

[0112] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims of the present invention.

Claims

1. A system for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data, characterized in that, include: The data acquisition module is used to acquire electrical data, meteorological data, and environmental data of the photovoltaic system according to a first sampling period, a second sampling period, and a third sampling period, respectively, wherein the first sampling period, the second sampling period, and the third sampling period increase sequentially; The data synchronization module is used to directly retrieve the electrical data within a time window that slides along the time axis, with the first sampling period as the target time resolution, perform cubic spline interpolation on the meteorological data, perform forward filling on the environmental data, and align the processed electrical data, meteorological data, and environmental data into a synchronized data frame. The feature extraction and fusion module is used to extract electrical features, meteorological features and environmental features from the synchronous data frame, and fuse the electrical features, meteorological features and environmental features into a fused feature vector; The anomaly diagnosis module includes a topology graph construction part, a graph neural network unit, a Bayesian network unit, and a joint decision unit. The topology graph construction part is used to construct a device topology graph based on the connection relationships between devices in the photovoltaic system. The graph neural network unit is used to process the device topology graph and output a topology feature vector. The Bayesian network unit is used to determine the posterior probability of each fault type based on a preset fault mode library and the fused feature vector. The joint decision unit is used to perform weighted fusion of the topology feature vector and the posterior probability to determine the final fault type. The anomaly diagnosis module is also used to determine a confidence score based on the posterior probability of each fault type.

2. The multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system according to claim 1, characterized in that, The data acquisition module includes an electrical data acquisition unit, a meteorological data acquisition unit, and an environmental data acquisition unit. The electrical data acquisition unit acquires the voltage, current, power, and harmonic distortion rate of the photovoltaic string through an IV acquisition device or an inverter communication interface. The meteorological data acquisition unit acquires irradiance, temperature, humidity, and wind speed through a illuminance sensor, a temperature and humidity sensor, and a wind speed sensor. The environmental data acquisition unit acquires dissolved oxygen and pH value through a water quality sensor.

3. The multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system according to claim 1, characterized in that, The first sampling period is 1 millisecond, the second sampling period is 1 second, and the third sampling period is 1 minute; the length of the time window is 100 milliseconds, the step size of the time window sliding along the time axis is 1 millisecond, and the time resolution of the synchronization data frame is 1 millisecond.

4. The multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system according to claim 1, characterized in that, The data synchronization module selects the measured meteorological data values ​​adjacent to the current target time before and after the current target time, and determines the meteorological data interpolation for the current target time based on the measured meteorological data values ​​and the continuity conditions of the first and second derivatives of the adjacent interpolation intervals.

5. The multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system according to claim 1, characterized in that, The electrical features include harmonic distortion rate, power change rate, and voltage drop depth; the meteorological features include irradiance change rate and temperature change rate; and the environmental features include pH change rate. The feature extraction and fusion module concatenates the electrical features, meteorological features, and environmental features in a preset order to form the fused feature vector. Alternatively, an attention fusion mechanism can be used, which adjusts the importance of different categories of features in the fusion process through learnable attention weights, and forms the fused feature vector according to the attention weights.

6. The multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system according to claim 1, characterized in that, The nodes of the device topology graph include photovoltaic string nodes, inverter nodes, meteorological sensor nodes, and environmental sensor nodes. The edges of the device topology graph include electrical connection edges between the inverter and the photovoltaic string, and data association edges between the meteorological sensor and the environmental sensor and the inverter. The graph neural network unit adopts a two-layer graph convolutional network, adds self-connection relationships of each node to the adjacency matrix of the device topology graph, and propagates the node feature matrix layer by layer according to the adjacency matrix after adding the self-connection relationships to obtain the topology feature vector containing the spatial topology association information between devices.

7. The multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system according to claim 1, characterized in that, The preset fault mode library includes feature modes corresponding to inverter faults, partial shading, sudden changes in water quality, compound faults, and normal operation; the Bayesian network unit determines the prior probability of each fault type based on historical fault data, determines the likelihood information corresponding to each fault type through maximum likelihood estimation, and determines the posterior probability of each fault type based on the prior probability, the likelihood information, and the fused feature vector; the anomaly diagnosis module determines the highest posterior probability among each fault type as the confidence score.

8. The multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system according to claim 1, characterized in that, It also includes an application display module connected to the anomaly diagnosis module; the anomaly diagnosis module generates an alarm message when the confidence score is not lower than 90%, generates a verification prompt when the confidence score is not lower than 70% but lower than 90%, and generates a manual confirmation mark when the confidence score is lower than 70%; the application display module is used to display the final fault type and the confidence score, and output the alarm message, the verification prompt or the manual confirmation mark.

9. The multi-source heterogeneous data cross-scale synchronization and anomaly diagnosis system according to claim 1, characterized in that, The data acquisition module and the data synchronization module are deployed at the edge of the photovoltaic power station. The edge is also used to perform feature extraction, Bayesian network inference, and alarm push. The cloud is used to train and update the graph neural network model, store historical data and fault cases, and process samples that require manual confirmation uploaded by the edge. The edge communicates with the cloud through message queue telemetry transmission protocol or hypertext transmission protocol.

10. A method for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data, characterized in that, include: Electrical data, meteorological data, and environmental data of the photovoltaic system are collected according to the first sampling period, the second sampling period, and the third sampling period, respectively, with the first sampling period, the second sampling period, and the third sampling period increasing sequentially. Within a time window that slides along the time axis, with the first sampling period as the target time resolution, the electrical data is directly used, cubic spline interpolation is performed on the meteorological data, forward filling is performed on the environmental data, and the processed electrical data, meteorological data and environmental data are aligned into a synchronous data frame. Electrical features, meteorological features, and environmental features are extracted from the synchronous data frames, and the electrical features, meteorological features, and environmental features are fused into a fused feature vector; A device topology graph is constructed based on the connection relationships between devices in the photovoltaic system. The device topology graph is then processed using a graph neural network to obtain topology feature vectors. Based on the preset fault mode library and the fused feature vector, the posterior probability of each fault type is determined using a Bayesian network. The feature vector and the posterior probability are weighted and fused, and the final fault type is determined based on the weighted fusion result. The confidence score is determined based on the posterior probability of each fault type.

11. The method for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data according to claim 10, characterized in that, The first sampling period is 1 millisecond, the second sampling period is 1 second, and the third sampling period is 1 minute; the length of the time window is 100 milliseconds, the step size of the time window sliding along the time axis is 1 millisecond, and the time resolution of the synchronization data frame is 1 millisecond.

12. The method for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data according to claim 10, characterized in that, The meteorological data is subjected to cubic spline interpolation, which includes: selecting measured meteorological data values ​​that are located before and after the current target time and adjacent to the current target time; and determining the meteorological data interpolation value for the current target time based on the measured meteorological data values ​​and the first derivative continuity condition and second derivative continuity condition of the adjacent interpolation interval.

13. The method for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data according to claim 10, characterized in that, The electrical features include harmonic distortion rate, power change rate, and voltage drop depth; the meteorological features include irradiance change rate and temperature change rate; and the environmental features include pH change rate. The electrical features, meteorological features, and environmental features are concatenated in a preset order to form the fused feature vector; or, the importance of different categories of features in the fusion process is adjusted through an attention fusion mechanism, and the fused feature vector is formed according to the corresponding attention weights.

14. The method for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data according to claim 10, characterized in that, Constructing the device topology graph includes: using photovoltaic strings, inverters, meteorological sensors, and environmental sensors as nodes; establishing electrical connection edges based on the electrical connection relationship between the inverter and the photovoltaic strings, and establishing data association edges based on the data association relationship between the meteorological sensors, environmental sensors, and the inverter; processing the device topology graph using a two-layer graph convolutional network, adding self-connection relationships of each node to the adjacency matrix of the device topology graph, and propagating the node feature matrix layer by layer based on the adjacency matrix after adding the self-connection relationships to obtain the topology feature vector containing spatial topology association information between devices.

15. The method for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data according to claim 10, characterized in that, The preset fault mode library includes feature modes corresponding to inverter faults, partial shading, sudden changes in water quality, compound faults, and normal operation. The prior probability of each fault type is determined based on historical fault data, and the likelihood information corresponding to each fault type is determined through maximum likelihood estimation. The posterior probability of each fault type is determined based on the prior probability, the likelihood information, and the fused feature vector. The highest posterior probability among all fault types is determined as the confidence score. An alarm message is generated when the confidence score is not lower than 90%, a verification prompt is generated when the confidence score is not lower than 70% but lower than 90%, and a manual confirmation mark is generated when the confidence score is lower than 70%.

16. The method for cross-scale synchronization and anomaly diagnosis of multi-source heterogeneous data according to claim 10, characterized in that, At the edge of the photovoltaic power station, data acquisition, cross-scale synchronization, feature extraction, Bayesian network inference, and alarm push are performed. In the cloud, graph neural network models are trained and updated, historical data and fault cases are stored, and samples requiring manual confirmation uploaded by the edge are processed. The edge communicates with the cloud via message queue telemetry transmission protocol or hypertext transfer protocol. The edge compresses the data to be uploaded and transmits it to the cloud.

Images