Photovoltaic system multi-channel anomaly detection method and system based on sliding window mechanism

Abstract

The invention discloses a photovoltaic system multichannel anomaly detection method and system based on a sliding window mechanism, and the method comprises the steps: generating a training set through a sliding window based on preprocessed historical data, inputting the training set into a multilayer feedforward neural network model and a deep learning model based on a self-attention mechanism for training, the latest data of the photovoltaic system is input into the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism for anomaly detection, so that rolling update and state synchronization of training samples can be realized, and the adaptability of the model to slow-varying characteristics such as seasonal change and power degradation is improved; moreover, the fitting advantage of the multi-layer feedforward neural network model in short-time modeling and the performance of the deep learning model based on the self-attention mechanism in long-term dependence modeling are fused, and the local fluctuation and long-term trend change in the operation of the photovoltaic system can be captured at the same time, thereby improving the real-time performance and precision of the anomaly detection of the photovoltaic system.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic system anomaly detection technology, and in particular to a method and system for multi-channel anomaly detection of photovoltaic systems based on a sliding window mechanism. Background Technology

[0002] With the widespread application of photovoltaic power generation in distributed energy systems, the operating status of large-scale photovoltaic power plants is becoming increasingly complex. These plants involve a large number of components, monitoring channels, and high-dimensional, non-stationary, and strongly time-series-dependent operational data. During long-term operation, factors such as environmental changes, component aging, and external shading can trigger abnormal power output, thereby affecting power generation efficiency and system safety. Therefore, efficient and accurate anomaly detection technologies are urgently needed.

[0003] Traditional fault detection methods mostly rely on fixed thresholds, static models, or manual inspections, making it difficult to adapt to dynamic changes in the system and often resulting in detection lag and high false alarm rates. In recent years, machine learning has been increasingly used in photovoltaic state modeling and anomaly identification. However, existing research is mostly focused on the analysis of single model structures or specific channels, lacking the ability to model the overall operating status of multi-channel photovoltaic systems and perform parallel diagnostics. Furthermore, its timeliness and stability in non-stationary scenarios remain insufficient.

[0004] It is particularly noteworthy that traditional methods generally use static training samples, which are difficult to reflect the actual operating state of the system in real time. Photovoltaic systems have strong temporal characteristics and dynamic evolution processes. The lack of an effective sample update mechanism will lead to insufficient model perception capabilities, and thus fail to accurately and in real time detect anomalies in photovoltaic systems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for multi-channel anomaly detection of photovoltaic systems based on a sliding window mechanism, which can improve the real-time performance and accuracy of anomaly detection in photovoltaic systems.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for multi-channel anomaly detection in a photovoltaic system based on a sliding window mechanism includes the following steps: Historical data of the photovoltaic system is acquired, and the historical data is preprocessed to obtain preprocessed historical data; A training set is generated using a sliding window based on the preprocessed historical data. The training set is input into the multilayer feedforward neural network model and the deep learning model based on the self-attention mechanism for training, to obtain the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism. Obtain the latest data of the photovoltaic system; The latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection, and anomaly detection results are obtained.

[0007] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A multi-channel anomaly detection system for a photovoltaic system based on a sliding window mechanism includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it performs the following steps: Historical data of the photovoltaic system is acquired, and the historical data is preprocessed to obtain preprocessed historical data; A training set is generated using a sliding window based on the preprocessed historical data. The training set is input into the multilayer feedforward neural network model and the deep learning model based on the self-attention mechanism for training, to obtain the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism. Obtain the latest data of the photovoltaic system; The latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection, and anomaly detection results are obtained.

[0008] The beneficial effects of this invention are as follows: Historical data of the photovoltaic system is preprocessed to obtain preprocessed historical data. A training set is generated using a sliding window based on the preprocessed historical data. The training set is input into a multilayer feedforward neural network model and a deep learning model based on a self-attention mechanism for training. The latest data of the photovoltaic system is obtained. The latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on a self-attention mechanism for anomaly detection, obtaining anomaly detection results. By constructing the input samples of the model through a sliding window, the rolling update and state synchronization of training samples can be achieved, improving the model's adaptability to slowly changing characteristics such as seasonal variations and power degradation. Furthermore, by combining the fitting advantages of the multilayer feedforward neural network model in short-term modeling with the performance of the deep learning model based on a self-attention mechanism in long-term dependency modeling, it can simultaneously capture local fluctuations and long-term trend changes in the operation of the photovoltaic system. It has significant potential in anomaly identification of multi-channel photovoltaic systems, thereby improving the real-time performance and accuracy of photovoltaic system anomaly detection. Attached Figure Description

[0009] Figure 1 This is a flowchart of a multi-channel anomaly detection method for a photovoltaic system based on a sliding window mechanism, according to an embodiment of the present invention. Figure 2This is a schematic diagram of a multi-channel anomaly detection system for a photovoltaic system based on a sliding window mechanism, according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the Transformer input sample construction under sliding window driving in a multi-channel anomaly detection method for photovoltaic systems based on a sliding window mechanism according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the actual power output timing of each photovoltaic string in a multi-channel anomaly detection method for a photovoltaic system based on a sliding window mechanism according to an embodiment of the present invention. Figure 5 This is a schematic diagram of photovoltaic string power prediction based on a BP neural network model in a multi-channel anomaly detection method for photovoltaic systems based on a sliding window mechanism according to an embodiment of the present invention. Figure 6 This is a schematic diagram of photovoltaic string power prediction based on the Transformer model in a multi-channel anomaly detection method for photovoltaic systems based on a sliding window mechanism according to an embodiment of the present invention. Figure 7 This is a Pearson correlation curve of PV output at adjacent time points in a multi-channel anomaly detection method for a photovoltaic system based on a sliding window mechanism according to an embodiment of the present invention. Figure 8 This is a distribution diagram of the daily average power and output fluctuation of each photovoltaic branch in a multi-channel anomaly detection method for a photovoltaic system based on a sliding window mechanism according to an embodiment of the present invention. Figure 9 This is a schematic diagram comparing the actual power of photovoltaic strings PV5 and PV6 with the prediction results of the BP neural network model in a multi-channel anomaly detection method for photovoltaic systems based on a sliding window mechanism according to an embodiment of the present invention. Figure 10 This is a schematic diagram comparing the actual power of photovoltaic strings PV5 and PV6 with the prediction results of the Transformer model in a multi-channel anomaly detection method for photovoltaic systems based on a sliding window mechanism according to an embodiment of the present invention. Figure 11 This is a schematic diagram of the rolling prediction and anomaly identification results of a multi-branch photovoltaic system in a multi-channel anomaly detection method for photovoltaic systems based on a sliding window mechanism according to an embodiment of the present invention. Figure 12 This is a schematic diagram of photovoltaic system operation data and anomaly point labeling results in a photovoltaic system multi-channel anomaly detection method based on a sliding window mechanism according to an embodiment of the present invention. Detailed Implementation

[0010] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0011] Before detailing the embodiments of this application, some related concepts will first be explained: Sliding window: A fixed-size window that slides across a data stream to process the data within it; Multilayer feedforward (BP) neural network model: A multilayer feedforward network that achieves self-optimization through the error backpropagation algorithm, consisting of an input layer, hidden layers (which may contain multiple layers) and an output layer; Deep learning (Transformer) models based on self-attention mechanisms: They abandon the sequential computation method of traditional recurrent neural networks (RNNs) and directly capture the global dependencies between elements in a sequence through self-attention mechanisms, achieving parallel processing and significantly improving training efficiency and long-distance dependency modeling capabilities.

[0012] In existing technologies, most methods for detecting anomalies in photovoltaic systems rely on fixed thresholds, static models, or manual inspections, which are difficult to adapt to dynamic changes in the system and often lead to detection delays and high false alarm rates.

[0013] To at least solve the above problems, please refer to Figure 1 This invention provides a method for multi-channel anomaly detection in photovoltaic systems based on a sliding window mechanism, comprising the following steps: Historical data of the photovoltaic system is acquired, and the historical data is preprocessed to obtain preprocessed historical data; A training set is generated using a sliding window based on the preprocessed historical data. The training set is input into the multilayer feedforward neural network model and the deep learning model based on the self-attention mechanism for training, to obtain the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism. Obtain the latest data of the photovoltaic system; The latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection, and anomaly detection results are obtained.

[0014] As can be seen from the above description, the beneficial effects of the present invention are as follows: Historical data of the photovoltaic system is preprocessed to obtain preprocessed historical data. A training set is generated using a sliding window based on the preprocessed historical data. The training set is input into a multilayer feedforward neural network model and a deep learning model based on a self-attention mechanism for training. The latest data of the photovoltaic system is obtained. The latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on a self-attention mechanism for anomaly detection, and anomaly detection results are obtained. By constructing the input samples of the model through a sliding window, the rolling update and state synchronization of the training samples can be realized, improving the model's adaptability to slow-changing characteristics such as seasonal changes and power degradation. Furthermore, by combining the fitting advantages of the multilayer feedforward neural network model in short-term modeling with the performance of the deep learning model based on a self-attention mechanism in long-term dependency modeling, it can simultaneously capture local fluctuations and long-term trend changes in the operation of the photovoltaic system. It has significant potential in anomaly identification of multi-channel photovoltaic systems, thereby improving the real-time performance and accuracy of photovoltaic system anomaly detection.

[0015] Furthermore, generating a training set using a sliding window based on the preprocessed historical data includes: The Pearson correlation coefficient was used to analyze the degree of linear correlation between the preprocessed historical data and the performance of the photovoltaic system; Filter the highly correlated historical data whose linear correlation is higher than a preset value from the preprocessed historical data; Set a fixed-length time window; The time window is used to continuously slide on the highly relevant historical data in chronological order. Each time the data is slideped, the continuous data within the time window is extracted as a training sample. A training set is generated based on all the training samples.

[0016] As described above, highly correlated historical data with a linear correlation higher than the preset value are selected from the preprocessed historical data to reduce the input dimension and improve training efficiency. Then, a sliding time window mechanism is introduced to solve the problem that existing anomaly detection methods rely on static samples and cannot perceive changes in system state in real time. This enables the model to dynamically construct inputs based on the data of the current time period, thereby improving the response capability to power anomalies under non-stationary operating conditions.

[0017] Furthermore, the latest data is input into the trained multi-layer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection, and the anomaly detection results include: Calculate the variance of the latest data; Based on the variance of the historical data and the variance of the latest data, it is determined whether there is an anomaly in the volatility of the latest data. If not, the latest data is compared with the historical data of the same time period to determine whether the data change of the latest data and the historical data of the same time period is within a preset range. If so, the latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism for prediction to obtain the first predicted value and the second predicted value. Obtain the actual observation value corresponding to the latest data; Calculate the first error between the first predicted value and the actual observed value, and calculate the second error between the second predicted value and the actual observed value; If both the first error and the second error exceed a preset percentage, and the first predicted value and the second predicted value show a preset number of abnormal samples consecutively, and the abnormal duration of the abnormal samples exceeds a preset time, then an anomaly is determined to exist, and the anomaly level is a severe anomaly. If the first error or the second error exceeds a preset percentage, and the duration of the abnormal sample in the first predicted value or the second predicted value does not exceed a preset time, then an anomaly is determined to exist, and the anomaly level is moderate. If neither the first error nor the second error exceeds a preset percentage, then various other detection methods are used to detect anomalies in the latest data. If at least two other detection methods mark the latest data as abnormal, then an anomaly is determined to exist, and the anomaly level is minor; otherwise, no anomaly is determined to exist.

[0018] As described above, combining the BP neural network model and the Transformer model for anomaly detection is effective. The BP neural network model is suitable for short-term power prediction (10 minutes to 24 hours), optimizing network weights through backpropagation. The Transformer model is more suitable for long-term time-series modeling, capable of identifying periodic trends and long-term changes, and relying on self-attention mechanisms to improve its ability to learn complex relationships between multiple variables. By combining the BP neural network model's ability to quickly model short-term disturbances with the Transformer model's advantages in long-sequence dependency modeling, a dual-channel parallel structure is constructed to achieve collaborative modeling of local changes and global trends in the photovoltaic system. This effectively improves the accuracy and coverage of anomaly detection in the photovoltaic system. Furthermore, during anomaly detection, not only is the existence of anomalies determined based on the prediction error and predicted value, but the anomaly level can also be determined, improving the reliability of anomaly detection.

[0019] Further, the latest data is input into the trained multi-layer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection. After obtaining the anomaly detection result, the method further includes: If the anomaly detection result indicates the presence of an anomaly and the anomaly level is minor, a first response strategy is generated. The first response strategy is to automatically remove the anomaly label if the data returns to normal after observation. If the anomaly detection result indicates the presence of an anomaly and the anomaly level is moderate, a second response strategy is generated. The second response strategy is to retain the anomaly record and increase the response strategy level if the anomaly recurs. If the anomaly detection result indicates the presence of an anomaly, and the anomaly level is severe, a third response strategy is generated. The third response strategy is to immediately trigger an alarm and initiate a maintenance and troubleshooting process.

[0020] As can be seen from the above description, there are corresponding response strategies for different abnormal detection results of photovoltaic systems, which can address different abnormal situations in a targeted manner and have stronger adaptability and deployment flexibility.

[0021] Further, the latest data is input into the trained multi-layer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection. After obtaining the anomaly detection result, the method further includes: Store the abnormal data based on the abnormality detection results; Determine whether the anomaly detection result triggers a preset model update rule. If so, retrain the trained multilayer feedforward neural network model and / or the trained deep learning model based on the anomaly data.

[0022] As described above, after anomaly detection, when the anomaly detection result triggers the preset model update rule, the stored anomaly data is used to retrain the trained BP neural network model and / or the trained Transformer model. Through online updates, the continuous anomaly detection effect of the model is optimized, ensuring the stability and reliability of anomaly detection.

[0023] Please refer to Figure 2 Another embodiment of the present invention provides a multi-channel anomaly detection system for photovoltaic systems based on a sliding window mechanism, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it performs the following steps: Historical data of the photovoltaic system is acquired, and the historical data is preprocessed to obtain preprocessed historical data; A training set is generated using a sliding window based on the preprocessed historical data. The training set is input into the multilayer feedforward neural network model and the deep learning model based on the self-attention mechanism for training, to obtain the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism. Obtain the latest data of the photovoltaic system; The latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection, and anomaly detection results are obtained.

[0024] As can be seen from the above description, the beneficial effects of the present invention are as follows: Historical data of the photovoltaic system is preprocessed to obtain preprocessed historical data. A training set is generated using a sliding window based on the preprocessed historical data. The training set is input into a multilayer feedforward neural network model and a deep learning model based on a self-attention mechanism for training. The latest data of the photovoltaic system is obtained. The latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on a self-attention mechanism for anomaly detection, and anomaly detection results are obtained. By constructing the input samples of the model through a sliding window, the rolling update and state synchronization of the training samples can be realized, improving the model's adaptability to slow-changing characteristics such as seasonal changes and power degradation. Furthermore, by combining the fitting advantages of the multilayer feedforward neural network model in short-term modeling with the performance of the deep learning model based on a self-attention mechanism in long-term dependency modeling, it can simultaneously capture local fluctuations and long-term trend changes in the operation of the photovoltaic system. It has significant potential in anomaly identification of multi-channel photovoltaic systems, thereby improving the real-time performance and accuracy of photovoltaic system anomaly detection.

[0025] Furthermore, generating a training set using a sliding window based on the preprocessed historical data includes: The Pearson correlation coefficient was used to analyze the degree of linear correlation between the preprocessed historical data and the performance of the photovoltaic system; Filter the highly correlated historical data whose linear correlation is higher than a preset value from the preprocessed historical data; Set a fixed-length time window; The time window is used to continuously slide on the highly relevant historical data in chronological order. Each time the data is slideped, the continuous data within the time window is extracted as a training sample. A training set is generated based on all the training samples.

[0026] As described above, highly correlated historical data with a linear correlation higher than the preset value are selected from the preprocessed historical data to reduce the input dimension and improve training efficiency. Then, a sliding time window mechanism is introduced to solve the problem that existing anomaly detection methods rely on static samples and cannot perceive changes in system state in real time. This enables the model to dynamically construct inputs based on the data of the current time period, thereby improving the response capability to power anomalies under non-stationary operating conditions.

[0027] Furthermore, the latest data is input into the trained multi-layer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection, and the anomaly detection results include: Calculate the variance of the latest data; Based on the variance of the historical data and the variance of the latest data, it is determined whether there is an anomaly in the volatility of the latest data. If not, the latest data is compared with the historical data of the same time period to determine whether the data change of the latest data and the historical data of the same time period is within a preset range. If so, the latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism for prediction to obtain the first predicted value and the second predicted value. Obtain the actual observation value corresponding to the latest data; Calculate the first error between the first predicted value and the actual observed value, and calculate the second error between the second predicted value and the actual observed value; If both the first error and the second error exceed a preset percentage, and the first predicted value and the second predicted value show a preset number of abnormal samples consecutively, and the abnormal duration of the abnormal samples exceeds a preset time, then an anomaly is determined to exist, and the anomaly level is a severe anomaly. If the first error or the second error exceeds a preset percentage, and the duration of the abnormal sample in the first predicted value or the second predicted value does not exceed a preset time, then an anomaly is determined to exist, and the anomaly level is moderate. If neither the first error nor the second error exceeds a preset percentage, then various other detection methods are used to detect anomalies in the latest data. If at least two other detection methods mark the latest data as abnormal, then an anomaly is determined to exist, and the anomaly level is minor; otherwise, no anomaly is determined to exist.

[0028] As described above, combining the BP neural network model and the Transformer model for anomaly detection is effective. The BP neural network model is suitable for short-term power prediction (10 minutes to 24 hours), optimizing network weights through backpropagation. The Transformer model is more suitable for long-term time-series modeling, capable of identifying periodic trends and long-term changes, and relying on self-attention mechanisms to improve its ability to learn complex relationships between multiple variables. By combining the BP neural network model's ability to quickly model short-term disturbances with the Transformer model's advantages in long-sequence dependency modeling, a dual-channel parallel structure is constructed to achieve collaborative modeling of local changes and global trends in the photovoltaic system. This effectively improves the accuracy and coverage of anomaly detection in the photovoltaic system. Furthermore, during anomaly detection, not only is the existence of anomalies determined based on the prediction error and predicted value, but the anomaly level can also be determined, improving the reliability of anomaly detection.

[0029] Further, the latest data is input into the trained multi-layer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection. After obtaining the anomaly detection result, the method further includes: If the anomaly detection result indicates the presence of an anomaly and the anomaly level is minor, a first response strategy is generated. The first response strategy is to automatically remove the anomaly label if the data returns to normal after observation. If the anomaly detection result indicates the presence of an anomaly and the anomaly level is moderate, a second response strategy is generated. The second response strategy is to retain the anomaly record and increase the response strategy level if the anomaly recurs. If the anomaly detection result indicates the presence of an anomaly, and the anomaly level is severe, a third response strategy is generated. The third response strategy is to immediately trigger an alarm and initiate a maintenance and troubleshooting process.

[0030] As can be seen from the above description, there are corresponding response strategies for different abnormal detection results of photovoltaic systems, which can address different abnormal situations in a targeted manner and have stronger adaptability and deployment flexibility.

[0031] Further, the latest data is input into the trained multi-layer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection. After obtaining the anomaly detection result, the method further includes: Store the abnormal data based on the abnormality detection results; Determine whether the anomaly detection result triggers a preset model update rule. If so, retrain the trained multilayer feedforward neural network model and / or the trained deep learning model based on the anomaly data.

[0032] As described above, after anomaly detection, when the anomaly detection result triggers the preset model update rule, the stored anomaly data is used to retrain the trained BP neural network model and / or the trained Transformer model. Through online updates, the continuous anomaly detection effect of the model is optimized, ensuring the stability and reliability of anomaly detection.

[0033] The above-described method and system for multi-channel anomaly detection in photovoltaic systems based on a sliding window mechanism are applicable to anomaly detection scenarios in photovoltaic systems, and are particularly suitable for anomaly monitoring scenarios in photovoltaic power plants with multiple strings operating in parallel. The following detailed implementation methods illustrate these methods: Please refer to Figure 1 One embodiment of the present invention is as follows: A method for multi-channel anomaly detection in a photovoltaic system based on a sliding window mechanism includes the following steps: S1. Obtain historical data of the photovoltaic system and preprocess the historical data to obtain preprocessed historical data, specifically including S11-S16: S11. Obtain key operating parameters of the photovoltaic system, such as historical voltage, current, output power, ambient temperature, relative humidity, and solar radiation intensity, to obtain historical data.

[0034] Specifically, by deploying a distributed sensor network at the photovoltaic power plant site, key operating parameters such as voltage, current, output power, ambient temperature, relative humidity, and solar radiation intensity of each string are collected in real time. The collected data is then aggregated by a remote monitoring system and transmitted to a data processing platform via an API interface. This method can be used to obtain historical data of the photovoltaic system.

[0035] S12. For data loss or anomalies caused by equipment failure or communication interruption in the historical data, an interpolation completion strategy is adopted to repair data breakpoints, remove data that deviates significantly from the physical reasonable range, and unify the timestamp format to obtain historical data with unified format, thereby ensuring the integrity and consistency of time series data.

[0036] S13. Based on the minimum and maximum value ranges of each physical quantity in the unified historical data, normalization processing is performed to obtain normalized historical data, thereby improving the uniformity of the model input data. The normalization formula is as follows: ; In the formula, D n This represents the normalized historical data. D This represents historical data after the format has been standardized. D min This represents the minimum value of each physical quantity. Dmax This represents the maximum value of each physical quantity.

[0037] S14. Filter the data that conforms to the preset time period from the normalized historical data to obtain the filtered historical data.

[0038] The preset time period includes periods of sufficient light intensity and active power output (typically from 08:00 to 16:00), plus one hour before sunrise and one hour after sunset. This is to assist in modeling edge response behavior. The preset time period can be dynamically adjusted according to seasonal changes.

[0039] S15. Perform continuity and uniqueness checks on the timestamp field of the filtered historical data, identify and remove duplicate or disordered records, and obtain the historical data after timestamp verification, thereby preventing time sequence errors from affecting model training.

[0040] S16. Calculate the standard deviation and variance of environmental variables such as irradiance, temperature and humidity in the historical data after the timestamp verification, identify abnormal fluctuation points that occur in a short period of time, and obtain pre-processed historical data to preliminarily determine the possibility of equipment failure or environmental interference.

[0041] S2. Based on the preprocessed historical data, a training set is generated using a sliding window, specifically including S21-S25: S21. Use the Pearson correlation coefficient to analyze the degree of linear correlation between the preprocessed historical data and the performance of the photovoltaic system.

[0042] In an alternative implementation, prior to S21, the following may also be included: The consistency of the preprocessed historical data is verified, and the stability of each variable in the preprocessed historical data at the physical level is obtained by combining the variance analysis.

[0043] S22. Filter highly correlated historical data (such as ambient temperature, relative humidity, solar radiation, and DC power) with a linear correlation higher than a preset value from the preprocessed historical data to reduce the input dimension and improve training efficiency.

[0044] S23. Set a fixed-length time window.

[0045] S24. Using the time window, continuously slide the highly relevant historical data in chronological order. Each time the data is slideped, extract the continuous data within the time window as a training sample.

[0046] For example, suppose the original highly correlated historical data sequence is: ; In the formula, X represents a highly correlated historical data sequence, x T Indicates the first T Highly correlated historical data at each point in time, x t Indicates the first t Highly correlated historical data at each point in time, including d Each feature dimension.

[0047] Assume the fixed length of the time window is w Step size is Then the first i The input data X corresponding to each training sample window (i) Defined as: ; S25. Generate a training set based on all the training samples.

[0048] The above-described sliding window-based sample generation method of the present invention uses a fixed-length window to slide sequentially over time series data. The data within the window is used to construct a single training sample, thereby achieving continuous rolling updates of data input and adaptive expansion of the training set, enhancing the model's ability to capture dynamic operating states.

[0049] S3. Input the training set into the multilayer feedforward neural network model and the deep learning model based on the self-attention mechanism for training, to obtain the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism.

[0050] Among them, the multilayer feedforward neural network model is suitable for short-term power prediction (10 minutes to 24 hours). It optimizes network weights through backpropagation, and the input data is usually selected from the historical records of the most recent n days (n is usually 10) to reduce the interference of short-term environmental fluctuations. Deep learning models based on self-attention mechanisms are more suitable for long-term time series modeling, and can identify periodic trends and long-term changes. They rely on self-attention mechanisms to improve the ability to learn complex relationships between multiple variables.

[0051] Specifically, for a multi-layer feedforward neural network model, environmental variables (such as temperature, humidity, and radiation) at the current time point are selected from the training set as input to predict the power output at that moment. Although this network lacks temporal memory capabilities, it can continuously generate samples through a sliding window mechanism, achieving short-cycle response and update capabilities. For a deep learning model based on a self-attention mechanism, a dual-channel sample system with an encoder-decoder structure is constructed based on the training set. Each sample consists of seven consecutive environmental data sequences as encoder input, while the previous six corresponding historical power data sequences are introduced as decoder input. Finally, the power value at the end of the predicted window is used as the supervision label. Therefore: No. i The encoder input X for each sample (i)enc Defined as: ; set up w =7 indicates that the sample uses the 7 most recent records as the input sequence.

[0052] Decoder input Y (i) dec For the encoder time interval before w Historical power values ​​at -1 time point: ; In the formula, y i Indicates the first i Output labels at each time point k Indicates the target output dimension.

[0053] Supervision label for each sample y (i) target For the last window i + w -1 point in time: actual output y i+w-1 : .

[0054] This structure, through the self-attention mechanism in the Transformer architecture, can improve the ability to model long-term dependent features such as periodic trends and slowly changing anomalies while maintaining the integrity of the time series context.

[0055] like Figure 3 As shown, Figure 3 This paper demonstrates the specific method for constructing Transformer model input samples based on the sliding window mechanism in this invention. A fixed-length sliding window is set along the time series dimension, and it slides continuously with a set step size. Each sliding window consists of three parts: encoder input, decoder input, and target output. The encoder input corresponds to the continuous environmental feature sequence within the current window, the decoder input is the historical power output sequence corresponding to the encoder's time interval, and the target output is the power value at the end of the prediction window. Figure 3 The paper further demonstrates the sliding process of the window along the time axis, including the definition of the sliding step size and the method of generating repeated windows for multiple time segments to form multiple training samples. Through this mechanism, the model can capture dynamic change patterns across time steps, enabling the modeling of long-term operating trends and anomaly detection of photovoltaic systems.

[0056] In one alternative implementation, after S3, the following may also be included: Calculate the prediction errors of the trained multilayer feedforward neural network model and the trained deep learning model based on self-attention mechanism; determine whether the prediction error exceeds a preset error; if so, optimize the trained multilayer feedforward neural network model and the trained deep learning model based on self-attention mechanism by adjusting hyperparameters, increasing the training dataset, or improving feature selection, thereby improving the accuracy and generalization ability of the model.

[0057] S4. Obtain the latest data of the photovoltaic system.

[0058] S5. Input the latest data into the trained multi-layer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection, and obtain the anomaly detection results, specifically including S51-S58: S51. Calculate the variance of the latest data, specifically as follows: ; ; In the formula, The standard deviation of the data n Indicates the number of data points. X i Representing data points i , This represents the mean of the data. Var ( X ) represents the variance of the data.

[0059] S52. Based on the variance of the historical data and the variance of the latest data, determine whether there is an anomaly in the volatility of the latest data. If not, compare the latest data with the historical data of the same time period to determine whether the data change of the latest data and the historical data of the same time period is within a preset range. If so, input the latest data into the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism for prediction to obtain a first predicted value and a second predicted value.

[0060] Determining whether the volatility of the latest data is abnormal based on the variance of the historical data and the variance of the latest data includes: Determine whether the difference between the variance of the historical data and the variance of the latest data is within a preset difference value. If so, determine that the volatility of the latest data is not abnormal; otherwise, determine that the volatility of the latest data is abnormal.

[0061] S53. Obtain the actual observation value corresponding to the latest data.

[0062] S54. Calculate the first error between the first predicted value and the actual observed value, and calculate the second error between the second predicted value and the actual observed value.

[0063] S55. If both the first error and the second error exceed a preset percentage, and the first predicted value and the second predicted value show a preset number of abnormal samples consecutively, and the abnormal duration of the abnormal samples exceeds a preset time, then it is determined that there is an anomaly, and the anomaly level is a severe anomaly.

[0064] Such anomalies may originate from hardware issues such as component aging, loose wiring, or inverter failure. In such cases, anomaly logs will be recorded, and the operation and maintenance system will be notified to conduct on-site verification.

[0065] In one optional implementation, the preset percentage is 10%; the preset number is 3.

[0066] S56. If the first error or the second error exceeds a preset percentage, and the abnormal duration of the abnormal sample in the first predicted value or the second predicted value does not exceed a preset time, then it is determined that there is an anomaly, and the anomaly level is moderate anomaly.

[0067] Possible causes of this type of anomaly include component contamination and temporary occlusion.

[0068] S57. If neither the first error nor the second error exceeds a preset percentage, then use a variety of other detection methods to perform anomaly detection on the latest data.

[0069] Among these, the various other detection methods include the standard deviation threshold method, the Z-score outlier detection method, the differential fluctuation detection method, and the smoothing comparison method based on simple moving average (SMA).

[0070] The standard deviation threshold method is specifically as follows: If the standard deviation of a data row in the latest data is greater than a preset threshold (e.g., 1000), it is marked as abnormal data; The Z-score outlier detection method is specifically as follows: For each data point in the latest data (such as temperature, humidity, radiation, etc.), calculate its mean and standard deviation; The Z-score for this data point is calculated based on the mean and standard deviation, specifically as follows: ; In the formula, Z i Representing data points X i Z-score; If the absolute value of the Z-score of a data point is greater than 3, it is considered a possible outlier and the data point is marked as an anomalous data point.

[0071] The differential fluctuation detection method is specifically as follows: Calculate the change between adjacent data points in the latest data. If the magnitude of the change exceeds a preset threshold, the point may have a sudden fluctuation, and the data point is marked as abnormal data.

[0072] Specifically, calculating the change between adjacent data points in the latest data involves: ; For example, if If the value is greater than 2, then the data point is considered to have experienced sudden fluctuations.

[0073] The smoothing comparison method based on simple moving average is as follows: The latest data is smoothed using a sliding window. The difference between the original latest data and the smoothed value is compared. If the absolute value of the difference is greater than 0.5, the data is marked as outlier. The smoothing calculation formula is as follows: ; In the formula, S i Indicates the smoothing value. W Indicates the length of the sliding window. W =3, X j This indicates the latest original data.

[0074] S58. If at least two other detection methods mark the latest data as abnormal, then an abnormality is determined to exist, and the abnormality level is minor; otherwise, no abnormality is determined to exist.

[0075] Such anomalies may be caused by environmental fluctuations or sampling noise.

[0076] In an alternative implementation, a multi-dimensional joint anomaly detection strategy may also be introduced, including: Co-variable analysis: If multiple key operating variables (such as temperature, humidity, light intensity and power output) all show abnormal fluctuations within the same time period, the confidence level of the anomaly determination is significantly improved.

[0077] Comparison of prediction error trends: If the error change trends of the multi-layer feedforward neural network model and the deep learning model based on the self-attention mechanism are similar (such as rising or falling synchronously), it can be used as a basis for enhancing the reliability of anomaly detection.

[0078] Historical trend review: If similar anomalies occur frequently in the past week, the system tends to attribute them to potential equipment failure rather than occasional environmental interference.

[0079] In one alternative implementation, after S5, the following is also included: If the anomaly detection result indicates the presence of an anomaly and the anomaly level is minor, a first response strategy is generated. The first response strategy is to automatically remove the anomaly label if the data returns to normal after observation. If the anomaly detection result indicates the presence of an anomaly and the anomaly level is moderate, a second response strategy is generated. The second response strategy is to retain the anomaly record and increase the response strategy level if the anomaly recurs. If the anomaly detection result indicates the presence of an anomaly, and the anomaly level is severe, a third response strategy is generated. The third response strategy is to immediately trigger an alarm and initiate a maintenance and troubleshooting process.

[0080] In one optional implementation, the impact of the anomaly on long-term operating performance will be comprehensively assessed. If the anomaly has little impact on the overall power generation efficiency of the system, the data will be recorded and archived. If the prediction error continues to exceed the threshold, the sample set update and model retraining process will be automatically initiated to improve long-term stability.

[0081] In one alternative implementation, after S5, the following is also included: Store the abnormal data based on the abnormality detection results; Determine whether the anomaly detection result triggers a preset model update rule. If so, retrain the trained multilayer feedforward neural network model and / or the trained deep learning model based on the anomaly data.

[0082] The preset model update rules include: Rule 1: If the frequency of anomaly detection is high (≥10 times) within a short period of time (e.g., within 1 day), but most of them are not recognized by the model, then a small-scale fine-tuning process of the multi-layer feedforward neural network model is triggered. Rule 2: If the first occurrence of an abnormal event lasts for more than 30 minutes and the model's recognition accuracy is insufficient, then incremental learning processing should be performed on the deep learning model based on the self-attention mechanism. Rule 3: If environmental variables (such as light or temperature and humidity) continue to deviate from their original operating patterns within 30 days, a full retraining process will be performed on the multi-layer feedforward neural network model and the deep learning model based on the self-attention mechanism.

[0083] like Figure 4 , Figure 5 and Figure 6 As shown, Figure 4 , Figure 5 and Figure 6 The actual power output of 10 typical photovoltaic strings (PV1-PV15) in a photovoltaic power plant during the operating period (08:00-16:00), the prediction results of the BP neural network model, and the prediction results of the Transformer model are shown respectively.

[0084] like Figure 7 As shown, Figure 7 The Pearson correlation coefficient of the DC output power of the photovoltaic string is shown to change between adjacent time points. The correlation is close to 1 for most time periods, indicating that the string output has significant temporal continuity, which verifies the rationality and effectiveness of introducing a sliding window mechanism in the model design to capture time-series dependencies.

[0085] When modeling and detecting anomalies in photovoltaic system channels PV1-PV15, to demonstrate the predictive effect and diagnostic accuracy of the method of this invention, PV5 and PV6 were selected as typical representatives for comparative analysis. To ensure channel representativeness and stability, their operating data in the training set were statistically analyzed, and the channel stability was quantified using the standard deviation of the PV output mean. ; In the formula, CV i Indicates channel i The relative volatility, Indicates channel i The fluctuation range of the power value within the sampling period, expressed in W. Indicates channel i The average output power over the entire sampling period (e.g., one day), measured in W.

[0086] like Figure 8 As shown, Figure 8 The graph shows the fluctuation of DC output power of each photovoltaic channel in the training set data during the entire day's operation (08:00-16:00). It can be seen that channels PV5 and PV6 have small standard deviations and low volatility for most of the time period, indicating stable output and making them suitable for modeling and demonstration. In contrast, channels such as PV1 and PV13 exhibit larger power fluctuations or local anomalies.

[0087] like Figure 9 As shown, Figure 9 This demonstrates a comparison between the actual power output curves of photovoltaic strings PV5 and PV6 and the prediction results of the BP neural network model; for example... Figure 10 As shown, Figure 10This chart compares the actual power output of photovoltaic strings PV5 and PV6 with the prediction results of the Transformer model. The horizontal axis represents the time step number (each step corresponds to a 5-minute sampling interval), and the vertical axis represents the DC power output (unit: watts, W). The power shown in the figure is the average output power of each photovoltaic module, which has been normalized by dividing by the number of modules in the string (e.g., PV5 contains 17 modules, PV6 contains 17 modules). The two channels are representative channels selected from the full channel set. Their average power levels in the training data are moderate, with coefficients of variation (CV) of 0.173 and 0.171, respectively, indicating stable power output without significant shading or fault effects, making them suitable for demonstrating the modeling results.

[0088] Furthermore, this invention has been deployed in the PyCharm Community environment for model inference and diagnostic processes, achieving rolling prediction and anomaly identification based on a sliding time window. The system continuously updates the input data in fixed steps (5 minutes) and outputs the power prediction result for the current moment and the corresponding anomaly judgment flag within each prediction cycle. Figure 11 As shown, this is the rolling prediction output for a photovoltaic power station branch (such as PV5) over a continuous time period after calling the model module in the PyCharm platform. It includes measured power, model prediction values, and anomaly diagnosis labels.

[0089] Regarding the output of diagnostic results, this invention uses an anomaly highlighting mechanism to emphasize (italicize) predicted samples judged to be abnormal, and outputs multi-channel results containing diagnostic information in Excel format, forming a structured and traceable diagnostic record document. For example... Figure 12 As shown, the generated file contains fields such as channel number, timestamp, environmental characteristics, measured value, predicted value, and diagnostic label. It can be directly connected to the operation and maintenance system or used for manual review, and has good readability and deployment versatility.

[0090] In summary, the proposed multi-channel anomaly detection method for photovoltaic systems based on a sliding window mechanism addresses the limitations of existing anomaly detection methods that rely on static samples and cannot perceive real-time system state changes. This allows the model to dynamically construct inputs based on current time-series data, improving its response to power anomalies under non-stationary operating conditions. Combining the rapid modeling capabilities of BP neural networks for short-term disturbances with the advantages of Transformer models in long-sequence dependency modeling, a dual-channel parallel structure is constructed to achieve collaborative modeling of local changes and global trends in the photovoltaic system. The multi-channel parallel modeling approach supports simultaneous diagnosis of multiple photovoltaic branches, enhancing the method's applicability in real-world power plant operation scenarios. The use of a relative residual dynamic threshold judgment mechanism and a joint judgment strategy across three consecutive time points effectively suppresses false alarms caused by instantaneous fluctuations, improving the accuracy and stability of anomaly identification. Furthermore, the model possesses online update and state adaptation capabilities, automatically adjusting the identification strategy based on factors such as component degradation and environmental changes, ensuring continuous detection effectiveness during long-term system operation. In summary, this invention has significant improvements over existing technologies in terms of state synchronization modeling, anomaly diagnosis accuracy, channel expansion capability, and deployment adaptability, and is suitable for intelligent anomaly monitoring needs under complex operating conditions of photovoltaic systems.

[0091] According to another aspect of the invention, Figure 2 This is a schematic diagram illustrating a multi-channel anomaly detection system for a photovoltaic system based on a sliding window mechanism according to an embodiment of the present invention. The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various steps of the multi-channel anomaly detection method for a photovoltaic system based on a sliding window mechanism as described above.

[0092] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention's specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A multi-channel anomaly detection method for photovoltaic systems based on a sliding window mechanism, characterized in that, Including the following steps: Historical data of the photovoltaic system is acquired, and the historical data is preprocessed to obtain preprocessed historical data; A training set is generated using a sliding window based on the preprocessed historical data. The training set is input into the multilayer feedforward neural network model and the deep learning model based on the self-attention mechanism for training, to obtain the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism. Obtain the latest data of the photovoltaic system; The latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection, and anomaly detection results are obtained.

2. The method for multi-channel anomaly detection in a photovoltaic system based on a sliding window mechanism according to claim 1, characterized in that, Generating a training set using a sliding window based on the preprocessed historical data includes: The Pearson correlation coefficient was used to analyze the degree of linear correlation between the preprocessed historical data and the performance of the photovoltaic system; Filter the highly correlated historical data whose linear correlation is higher than a preset value from the preprocessed historical data; Set a fixed-length time window; The time window is used to continuously slide on the highly relevant historical data in chronological order. Each time the data is slideped, the continuous data within the time window is extracted as a training sample. A training set is generated based on all the training samples.

3. The method for multi-channel anomaly detection in a photovoltaic system based on a sliding window mechanism according to claim 1, characterized in that, The latest data is input into the trained multilayer feedforward neural network model and the trained self-attention-based deep learning model for anomaly detection, and the anomaly detection results include: Calculate the variance of the latest data; Based on the variance of the historical data and the variance of the latest data, it is determined whether there is an anomaly in the volatility of the latest data. If not, the latest data is compared with the historical data of the same time period to determine whether the data change of the latest data and the historical data of the same time period is within a preset range. If so, the latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism for prediction to obtain the first predicted value and the second predicted value. Obtain the actual observation value corresponding to the latest data; Calculate the first error between the first predicted value and the actual observed value, and calculate the second error between the second predicted value and the actual observed value; If both the first error and the second error exceed a preset percentage, and the first predicted value and the second predicted value show a preset number of abnormal samples consecutively, and the abnormal duration of the abnormal samples exceeds a preset time, then an anomaly is determined to exist, and the anomaly level is a severe anomaly. If the first error or the second error exceeds a preset percentage, and the duration of the abnormal sample in the first predicted value or the second predicted value does not exceed a preset time, then an anomaly is determined to exist, and the anomaly level is moderate. If neither the first error nor the second error exceeds a preset percentage, then various other detection methods are used to detect anomalies in the latest data. If at least two other detection methods mark the latest data as abnormal, then an anomaly is determined to exist, and the anomaly level is minor; otherwise, no anomaly is determined to exist.

4. The method for multi-channel anomaly detection in a photovoltaic system based on a sliding window mechanism according to claim 3, characterized in that, The latest data is input into the trained multilayer feedforward neural network model and the trained self-attention-based deep learning model for anomaly detection. After obtaining the anomaly detection results, the method further includes: If the anomaly detection result indicates the presence of an anomaly and the anomaly level is minor, a first response strategy is generated. The first response strategy is to automatically remove the anomaly label if the data returns to normal after observation. If the anomaly detection result indicates the presence of an anomaly and the anomaly level is moderate, a second response strategy is generated. The second response strategy is to retain the anomaly record and increase the response strategy level if the anomaly recurs. If the anomaly detection result indicates the presence of an anomaly, and the anomaly level is severe, a third response strategy is generated. The third response strategy is to immediately trigger an alarm and initiate a maintenance and troubleshooting process.

5. The method for multi-channel anomaly detection in a photovoltaic system based on a sliding window mechanism according to claim 1, characterized in that, The latest data is input into the trained multilayer feedforward neural network model and the trained self-attention-based deep learning model for anomaly detection. After obtaining the anomaly detection results, the method further includes: Store the abnormal data based on the abnormality detection results; Determine whether the anomaly detection result triggers a preset model update rule. If so, retrain the trained multilayer feedforward neural network model and / or the trained deep learning model based on the anomaly data.

6. A multi-channel anomaly detection system for a photovoltaic system based on a sliding window mechanism, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it performs the following steps: Historical data of the photovoltaic system is acquired, and the historical data is preprocessed to obtain preprocessed historical data; A training set is generated using a sliding window based on the preprocessed historical data. The training set is input into the multilayer feedforward neural network model and the deep learning model based on the self-attention mechanism for training, to obtain the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism. Obtain the latest data of the photovoltaic system; The latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on self-attention mechanism for anomaly detection, and anomaly detection results are obtained.

7. The photovoltaic system multi-channel anomaly detection system based on a sliding window mechanism according to claim 6, characterized in that, Generating a training set using a sliding window based on the preprocessed historical data includes: The Pearson correlation coefficient was used to analyze the degree of linear correlation between the preprocessed historical data and the performance of the photovoltaic system; Filter the highly correlated historical data whose linear correlation is higher than a preset value from the preprocessed historical data; Set a fixed-length time window; The time window is used to continuously slide on the highly relevant historical data in chronological order. Each time the data is slideped, the continuous data within the time window is extracted as a training sample. A training set is generated based on all the training samples.

8. The photovoltaic system multi-channel anomaly detection system based on a sliding window mechanism according to claim 6, characterized in that, The latest data is input into the trained multilayer feedforward neural network model and the trained self-attention-based deep learning model for anomaly detection, and the anomaly detection results include: Calculate the variance of the latest data; Based on the variance of the historical data and the variance of the latest data, it is determined whether there is an anomaly in the volatility of the latest data. If not, the latest data is compared with the historical data of the same time period to determine whether the data change of the latest data and the historical data of the same time period is within a preset range. If so, the latest data is input into the trained multilayer feedforward neural network model and the trained deep learning model based on the self-attention mechanism for prediction to obtain the first predicted value and the second predicted value. Obtain the actual observation value corresponding to the latest data; Calculate the first error between the first predicted value and the actual observed value, and calculate the second error between the second predicted value and the actual observed value; If both the first error and the second error exceed a preset percentage, and the first predicted value and the second predicted value show a preset number of abnormal samples consecutively, and the abnormal duration of the abnormal samples exceeds a preset time, then an anomaly is determined to exist, and the anomaly level is a severe anomaly. If the first error or the second error exceeds a preset percentage, and the duration of the abnormal sample in the first predicted value or the second predicted value does not exceed a preset time, then an anomaly is determined to exist, and the anomaly level is moderate. If neither the first error nor the second error exceeds a preset percentage, then various other detection methods are used to detect anomalies in the latest data. If at least two other detection methods mark the latest data as abnormal, then an anomaly is determined to exist, and the anomaly level is minor; otherwise, no anomaly is determined to exist.

9. A multi-channel anomaly detection system for a photovoltaic system based on a sliding window mechanism according to claim 8, characterized in that, The latest data is input into the trained multilayer feedforward neural network model and the trained self-attention-based deep learning model for anomaly detection. After obtaining the anomaly detection results, the method further includes: If the anomaly detection result indicates the presence of an anomaly and the anomaly level is minor, a first response strategy is generated. The first response strategy is to automatically remove the anomaly label if the data returns to normal after observation. If the anomaly detection result indicates the presence of an anomaly and the anomaly level is moderate, a second response strategy is generated. The second response strategy is to retain the anomaly record and increase the response strategy level if the anomaly recurs. If the anomaly detection result indicates the presence of an anomaly, and the anomaly level is severe, a third response strategy is generated. The third response strategy is to immediately trigger an alarm and initiate a maintenance and troubleshooting process.

10. A multi-channel anomaly detection system for a photovoltaic system based on a sliding window mechanism according to claim 6, characterized in that, The latest data is input into the trained multilayer feedforward neural network model and the trained self-attention-based deep learning model for anomaly detection. After obtaining the anomaly detection results, the method further includes: Store the abnormal data based on the abnormality detection results; Determine whether the anomaly detection result triggers a preset model update rule. If so, retrain the trained multilayer feedforward neural network model and / or the trained deep learning model based on the anomaly data.

Images