A multi-source energy system fault detection and root cause analysis system

Abstract

This invention discloses a fault detection and root cause analysis system for multi-source energy systems, relating to the fields of energy management and fault diagnosis technology. The system acquires various operational and image data from the multi-source energy system in real time; preprocesses the operational and image data; inputs the preprocessed data into a hybrid algorithm combining wavelet transform, CNN, and LSTM to detect abnormal patterns and potential faults in the multi-source energy system; a digital twin model is dynamically updated based on the real-time acquired data; combining the feature extraction capabilities of the digital twin model and the hybrid algorithm, the system locates the key causes of faults, simulates the system's operating state, analyzes the potential causes of faults, and outputs the analysis results. This invention effectively improves the system's comprehensive analysis capabilities for multiple data sources, providing a comprehensive and accurate solution for the intelligent operation and maintenance of energy systems.

Description

Technical Field

[0001] This invention relates to the field of energy management and fault diagnosis technology, and in particular to a fault detection and root cause analysis system for multi-source energy systems. Background Art

[0002] With the continuous optimization of industrial sectors, especially renewable energy, energy storage devices, and traditional power generation technologies, factories and other industrial sites are increasingly relying on the coordinated operation of multiple energy sources, including solar power, grid power, lithium battery energy storage devices, and traditional power generation, to improve energy efficiency, reduce costs, and achieve green production. Intelligent and efficient management and fault prediction of multi-source energy systems are crucial for ensuring system stability and improving the intelligence of energy management. However, current fault detection technologies and diagnostic methods in this field still face some problems and challenges.

[0003] 1. Current Status of Fault Detection Technology for Lithium-ion Battery Energy Storage Systems

[0004] Currently, fault detection in lithium battery energy storage systems primarily relies on traditional hardware monitoring methods, such as sensors for temperature, voltage, and current, and data analysis-based models, such as statistical analysis and time series analysis. For example, by monitoring lithium battery voltage, current, and temperature data in real time, traditional fault detection methods identify equipment anomalies through threshold judgment or deviation detection. However, this approach has several drawbacks.

[0005] Problems with existing technology:

[0006] The data is too limited to fully reflect the system's operational status.

[0007] Most existing systems focus solely on the physical parameters of lithium batteries, such as voltage and current, as the basis for fault diagnosis, neglecting many crucial factors. Environmental factors, such as temperature and humidity, have a significant impact on lithium battery performance. Aging and wear on the device's exterior may indicate underlying internal problems, and there is a lack of monitoring of external factors such as grid load and power fluctuations. For example, lithium battery performance may degrade in high-temperature environments, but this is difficult to detect using only voltage and current data. Furthermore, grid load fluctuations may lead to frequent charging and discharging of lithium batteries, accelerating battery aging, but existing systems cannot effectively capture this correlation information, potentially missing the opportunity to detect some latent faults early.

[0008] Lack of dynamic model update mechanism

[0009] Traditional fault detection methods are typically based on fixed models, which cannot adapt to changes in system state and dynamic fluctuations in the external environment in real time. For example, changes in temperature and humidity, as well as load fluctuations and changes in grid load, can all affect the performance and operating status of lithium batteries. However, existing monitoring models lack effective dynamic adjustment and optimization mechanisms and cannot be updated in a timely manner according to actual conditions, resulting in a significant reduction in the accuracy and timeliness of fault detection.

[0010] Insufficient fault location capability

[0011] Existing technologies generally only provide preliminary fault detection results and struggle to accurately identify the root cause of faults. This is particularly true when dealing with the impact of external power grid load fluctuations on lithium battery equipment, making it impossible to precisely pinpoint the fault location. This limitation severely restricts the accuracy and efficiency of equipment fault diagnosis, leaving subsequent system optimization and maintenance work lacking clear direction and specific basis, and unable to effectively solve the fundamental problems.

[0012] 2. Application of Digital Twin Technology in Multi-Source Energy Systems

[0013] Digital twin technology, as an innovative monitoring and analysis tool, has been widely applied in various fields, particularly demonstrating great potential in industrial manufacturing and energy management. It enables real-time monitoring and prediction of system states by constructing virtual models that correspond one-to-one with the actual system in a virtual environment, thus providing strong support for the optimization and adjustment of the actual system. In multi-source energy systems, digital twin technology can simulate the operating state of the energy system and update and correct the virtual model based on real-time data, providing important support for fault detection and prediction. However, the application of digital twin technology in multi-source energy systems still faces some challenges.

[0014] Problems with existing digital twin technology:

[0015] Although digital twins enable system monitoring through virtual models, most existing digital twin systems lack automated update mechanisms and cannot update the virtual model in real time based on multimodal data, such as sensor data, image data, and power grid load data. This means that the model often lags behind changes in the actual system state and cannot reflect system anomalies and faults in a timely manner, thus affecting the timeliness and accuracy of fault detection.

[0016] Weak data fusion capabilities

[0017] Traditional digital twin technology often lacks effective fusion strategies when processing data from different data sources, such as batteries, images, environment, and grid connection. In multi-source energy systems, there is a large amount of data from lithium battery management systems, various sensors, cameras, traditional power generation equipment, and the power grid. Effective fusion of this data is crucial for accurate fault detection. However, existing digital twin systems typically cannot fully integrate these different types of data, resulting in underutilization of data resources and failing to fully leverage the advantages of digital twin technology in fault detection.

[0018] 3. Application of multimodal data fusion technology

[0019] Multimodal data fusion technology aims to extract information from multiple different types of energy sources and obtain a more comprehensive and accurate understanding of the system status through integrated analysis. In multi-source energy systems, in addition to collecting parameters such as voltage, current, and temperature of lithium batteries through sensors, image recognition technology can also be used to obtain information on changes in the appearance of equipment, thereby achieving a more comprehensive monitoring of the operating status of energy equipment. By fusing these different types of data, it is expected to improve the accuracy of fault detection and diagnostic capabilities. However, this technology still has some shortcomings in its current application.

[0020] Problems with existing multimodal data fusion technologies:

[0021] The fusion algorithm is not mature enough.

[0022] Existing multimodal data fusion technologies have not yet been widely applied in the field of lithium battery energy storage systems, and most focus on basic feature extraction and fusion algorithms, such as weighted averaging and decision trees. Technical bottlenecks remain for the deep fusion and automated analysis of high-dimensional, complex data, including grid load data. Especially in energy systems, where data types are diverse and the volume is enormous, more complex and efficient algorithms are needed to fully extract the potential information within the data.

[0023] Insufficient feature extraction and pattern recognition capabilities

[0024] While traditional multimodal data fusion techniques can handle different types of data, they typically lack sufficiently powerful feature extraction and pattern recognition capabilities. The operating states of multi-source energy systems are complex, and methods based solely on data fusion often struggle to extract deep-seated characteristics of system faults. This is especially true under the influence of external factors such as grid load fluctuations, making accurate fault prediction and root cause analysis difficult, thus impacting system reliability and stability.

[0025] 4. Application of Hybrid Algorithms

[0026] To overcome these problems, researchers have gradually begun to adopt more advanced algorithms, such as CNNs and LSTMs from deep learning, for the analysis and fusion of multi-source data. CNNs are convolutional neural networks, and LSTMs are long short-term memory networks. These algorithms have powerful feature extraction capabilities when processing time-series and image data, and can more effectively mine useful information from various types of data. However, in practical applications, hybrid algorithms still have some problems.

[0027] Application problems of existing hybrid algorithms

[0028] Feature extraction that relies too heavily on a single data source

[0029] While algorithms such as CNN and LSTM have achieved some success in their respective fields, many studies in practical applications rely too heavily on a single data source, such as using only sensor or image data, neglecting the potential of data fusion. This is especially true in multi-source energy systems, where faults can be caused by a combination of factors. A single data source often cannot accurately describe the system state, particularly under the influence of grid-connected data. Considering only a single data source can lead to an inability to comprehensively account for the impact of external factors on the system, thereby reducing the accuracy of fault detection.

[0030] Lack of integration and automated update mechanism

[0031] Most existing hybrid algorithms focus on processing specific data sources, lacking mechanisms for integrating multimodal data and automating updates. Especially in fault detection, hybrid algorithms exhibit weak dynamic adaptability, failing to respond promptly to changes in system operating status. Furthermore, under the influence of factors such as power grid load fluctuations, system state changes become more complex, and existing hybrid algorithms cannot adapt quickly enough. Stronger dynamic adaptability is needed to ensure stable system operation and effective fault detection.

[0032] By analyzing the background of existing technologies, we can clearly see that current fault detection technologies in multi-source energy systems have not yet achieved efficient multimodal data fusion, dynamic digital twin model updates, and accurate root cause analysis. Therefore, the fault detection and root cause analysis method for multi-source energy systems based on a hybrid algorithm of multimodal data fusion and digital twin updates proposed in this invention is an innovative solution addressing the aforementioned technical bottlenecks. Summary of the Invention

[0033] The purpose of this invention is to propose a fault detection and root cause analysis system for multi-source energy systems. This invention utilizes multimodal data fusion to extract frequency domain features through wavelet transform, extract spatial features through convolutional neural networks, and predict temporal features through long short-term memory networks. This enables the detection of abnormal patterns and potential faults in multi-source energy systems. Furthermore, by combining digital twin technology, it further determines the occurrence of faults and performs root cause analysis, effectively improving the system's comprehensive analysis capabilities across multiple data sources and solving the problem of accurate root cause analysis in multimodal data fusion in the prior art.

[0034] The technical solution adopted in this invention is as follows:

[0035] This invention is a fault detection and root cause analysis system for multi-source energy systems, comprising a data acquisition module, a data preprocessing module, a data feature extraction and dimensionality reduction module, a digital twin module, and a fault detection and root cause analysis module, wherein:

[0036] The data acquisition module acquires various operational and image data of the multi-source energy system in real time and uploads them to the data preprocessing module.

[0037] The data preprocessing module preprocesses the various operational and image data collected by the data acquisition module, including outlier detection, missing value imputation, noise reduction, and data standardization.

[0038] The data feature extraction and dimensionality reduction module performs feature extraction and dimensionality reduction on the preprocessed operational data and image data. The preprocessed operational data and image data are then input into a hybrid algorithm combining wavelet transform, CNN, and LSTM. Wavelet transform and CNN extract features from the preprocessed data, and the extracted features are fused to obtain a fused feature set. The high-dimensional data in the fused feature set is then dimensionality reduced, and LSTM is used for time series modeling to detect abnormal patterns and potential faults in multi-source energy systems.

[0039] The digital twin module builds a virtual model of a multi-source energy system based on initial data, namely the digital twin model. The digital twin model is dynamically updated according to real-time data collection, maintaining consistency with the physical equipment in the multi-source energy system, reflecting the status of the physical equipment, and providing a more accurate basis for fault detection and root cause analysis.

[0040] The fault detection and root cause analysis module uses a hybrid algorithm to analyze real-time collected data, detect abnormal patterns and potential faults in multi-source energy systems, compare the prediction results of the real-time data input digital twin model to determine whether there are any deviations, and further confirm the occurrence of the fault. Combining the feature extraction capabilities of the digital twin model and the hybrid algorithm, it locates the key causes of the fault, simulates the system operating state, analyzes the potential causes of the fault, and outputs the analysis results.

[0041] Furthermore, the various operational data of the multi-source energy system include photovoltaic power generation data, energy storage device data, traditional power generation device data, power load data, environmental data, and grid connection data. These operational data are acquired through various sensors installed in the multi-source energy system. The image data of the multi-source energy system, i.e., the changes in the appearance of the energy equipment, is collected by cameras and transmitted in real time to the data preprocessing module.

[0042] Furthermore, the data preprocessing module preprocesses the various operational and image data collected by the data acquisition module, specifically:

[0043] Outlier detection removes obviously abnormal data through statistical methods or threshold detection.

[0044] Missing value imputation uses interpolation, mean averaging, or machine learning prediction methods to fill in missing values.

[0045] For noise reduction, the collected operational data are treated as time-series data, and wavelet transform is used for noise reduction to extract frequency domain features related to the fault. The image data is processed by CNN to extract spatial features and remove irrelevant background noise. For grid-connected data, filtering is performed first, and then wavelet transform is used for noise reduction to extract frequency domain features related to the fault, so as to ensure data quality.

[0046] Data standardization involves normalizing data from different sources to ensure data format consistency.

[0047] Furthermore, the data feature extraction and dimensionality reduction module uses wavelet transform and CNN to extract features from the preprocessed data, specifically as follows:

[0048] The voltage and current data in the energy storage device data are non-stationary signals. Wavelet transform is used to decompose the non-stationary signals, extract frequency domain features of different frequency bands, and reveal potential fault modes.

[0049] The image data was used to extract key spatial features using CNN;

[0050] The traditional power generation equipment data is used to extract fuel consumption trends through data fitting and difference methods or time series decomposition methods, and to extract power output fluctuation characteristics through wavelet transform methods or rolling standard deviation methods.

[0051] The grid-connected data is extracted using wavelet transform to extract frequency domain features, and the spectral components and energy distribution in the frequency domain are analyzed to determine the matching between the grid load status and energy supply.

[0052] The feature extraction in the data feature extraction and dimensionality reduction module also includes:

[0053] The photovoltaic power generation data is used to extract characteristics of power generation stability, correlation between light intensity and power generation, and the influence of temperature on power generation efficiency.

[0054] The environmental data is used to extract the trends in humidity, wind speed, and air pressure. The mean and variance of these trends are then determined using time series analysis.

[0055] Furthermore, the preprocessed data features are extracted and then fused to obtain a fused feature set, specifically:

[0056] The grid-connected data is integrated with photovoltaic power generation data, energy storage equipment data, and traditional power generation equipment data. Through a multi-head attention mechanism, the features extracted from different data sources are weighted and integrated to provide more comprehensive energy supply information.

[0057] The environmental data, combined with features extracted from power load data, energy storage equipment data, traditional power generation equipment data, and photovoltaic power generation data, provides more multi-dimensional information for fault prediction and system scheduling.

[0058] Furthermore, the dimensionality reduction processing of the high-dimensional data in the fusion feature set specifically involves:

[0059] Principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional data in the fused feature set, retaining key features and reducing redundant data. The fused feature set includes:

[0060] The characteristics of power generation stability, correlation between light intensity and power generation, and influence of temperature on power generation efficiency were extracted from photovoltaic power generation data.

[0061] Voltage and current frequency domain characteristics extracted from energy storage device data;

[0062] Fuel consumption trends and power output fluctuation characteristics extracted from data of traditional power generation equipment;

[0063] Spatial features extracted from image data;

[0064] Frequency domain characteristics of grid stability and load variation extracted from grid-connected data;

[0065] Environmental data were used to extract trends in humidity, wind speed, and air pressure.

[0066] The principal component analysis method calculates the covariance matrix of the data to find eigenvalues ​​and eigenvectors, selects the main eigenvectors based on the magnitude of the eigenvalues, thereby projecting high-dimensional data into a low-dimensional space, preserving key features, and combining them with spatial features extracted from image data to form a unified low-dimensional feature set, simplifying the subsequent analysis process.

[0067] Furthermore, the LSTM models time series data based on a low-dimensional feature set, captures the long-term dependencies of system operating states, and predicts potential faults and abnormal behaviors.

[0068] Furthermore, the digital twin model is dynamically updated based on real-time collected data, specifically through the following steps:

[0069] The triggering mechanism automatically triggers the update of the digital twin model when the data acquisition module detects that one or more data points in the image data or various operational data have changed beyond the normal range.

[0070] Incremental learning uses incremental learning methods to dynamically update the parameters of the digital twin model, simulate the operation of a multi-source energy system, and ensure that the virtual model can reflect the operating status of the equipment in the actual multi-source energy system in a timely manner.

[0071] Real-time adjustments: The digital twin model provides real-time early warnings of system failures and makes adjustments based on updated data.

[0072] Furthermore, the results output by the fault detection and root cause analysis module are specifically shown as follows:

[0073] Fault warning information is issued to users or management systems through real-time monitoring and analysis.

[0074] Root cause analysis report provides a detailed analysis report, listing the type of failure, possible causes, and recommended maintenance measures;

[0075] Data visualization displays system operating status, fault prediction, and diagnostic results in chart form, making it easier for users to understand and make decisions.

[0076] This invention is applicable to multi-source energy management systems in factories or other industrial settings, particularly for the coordinated optimization and management of solar power generation, grid power supply, energy storage devices, and traditional backup power generation equipment. By employing multimodal data fusion and hybrid algorithms, combined with digital twin technology, this invention can monitor and analyze the operating status of multi-source energy systems in real time, enabling immediate fault detection, prediction, and root cause analysis. Digital twin technology constructs virtual models to reflect the real-time operating status of equipment, while the hybrid algorithm improves the accuracy and efficiency of fault diagnosis through in-depth mining of multimodal data. This invention provides a comprehensive and precise solution for the intelligent operation and maintenance of multi-source energy systems, significantly improving the reliability, operating efficiency, and emergency response capabilities of these systems.

[0077] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0078] This invention is a fault detection and root cause analysis system for multi-source energy systems. It uses cameras to acquire real-time image data from energy storage devices within the system, detecting potential anomalies such as cracks, damage, and stains on the device surfaces. The image data provides direct visual evidence of external faults, enabling timely identification of external fault sources, and offering strong support for detecting external faults in energy storage devices. Combining image data with sensor data through multimodal data fusion enhances the accuracy of equipment fault diagnosis, enabling more comprehensive identification and analysis of fault root causes. Furthermore, the fusion of grid-connected data, power load data, and image data further enhances the fault prediction capabilities of multi-source energy systems in grid-connected environments.

[0079] This invention is a fault detection and root cause analysis system for multi-source energy systems. It employs a hybrid algorithm combining wavelet transform, CNN, and LSTM to overcome the limitations of traditional methods in fault diagnosis. Specifically, wavelet transform is used for signal denoising and decomposition, combined with the advantages of CNN in local feature extraction, and LSTM is used to model time-series data. This combination effectively improves the accuracy and real-time performance of fault detection, enabling it to cope with the complex and ever-changing operating environment of multi-source energy systems, especially adapting to their dynamic characteristics. Furthermore, grid-connected data, such as grid load fluctuations and electricity price fluctuations, can also effectively improve the accuracy of predictions and the system's response speed.

[0080] This invention is a fault detection and root cause analysis system for multi-source energy systems. The hybrid algorithm combines wavelet transform, convolutional neural networks, and long short-term memory networks, fully leveraging their respective advantages to achieve high-precision and high-efficiency fault detection in multi-source energy systems. This algorithm integrates traditional signal processing and deep learning techniques, enabling it to fully mine potential information from multimodal data and provide more comprehensive fault detection and root cause analysis capabilities. Through frequency domain characteristic analysis of wavelet transform, image feature extraction using CNN, and time series data modeling using LSTM, the three work synergistically to significantly improve the accuracy of fault diagnosis and enhance the system's robustness in complex and dynamic environments, allowing it to maintain efficient operation under varying operating conditions and respond to faults in real time.

[0081] This invention is a fault detection and root cause analysis system for multi-source energy systems. It utilizes digital twin technology to create virtual models of equipment and simulates the equipment's state and behavior through real-time updates and optimization. The digital twin model can quickly adapt to equipment changes based on actual operating data and provide early warnings and diagnoses before or during fault occurrence. Through incremental learning and online update mechanisms, the digital twin model maintains consistency with the physical equipment, adjusts in real time, enhances the accuracy of fault detection, significantly improves equipment reliability, reduces the risk of fault occurrence, and optimizes system operation and maintenance efficiency. Under grid-connected conditions, the digital twin model can simultaneously consider the interaction effects between energy storage devices and the power grid, reflecting the overall operating status of the system in real time.

[0082] This invention is a fault detection and root cause analysis system for multi-source energy systems. By integrating multiple data sources and combining digital twin technology, it creates a digital twin model of the multi-source energy system. Through real-time updates and optimization of the digital twin model, it achieves dynamic monitoring of equipment status and fault detection. This invention provides comprehensive and accurate fault detection and root cause analysis, enhancing system reliability and real-time response capabilities. Especially in complex and dynamic environments, it can reflect changes in equipment status in real time and provide accurate fault warnings. Grid-connected data provides real-time feedback on the external environment, further enhancing fault detection and scheduling optimization capabilities.

[0083] This invention is a multi-source energy system fault detection and root cause analysis system capable of real-time detection and diagnosis of equipment faults, rapidly generating fault warnings and providing detailed root cause analysis reports. The reports are presented through a graphical interface, helping maintenance personnel quickly identify problems and take effective measures. This function significantly reduces equipment downtime, optimizes system operating efficiency, and improves overall system reliability. The generation of fault warnings and analysis reports provides precise decision support, helping maintenance personnel take timely maintenance actions to avoid system downtime. The introduction of grid connection data helps to better assess the impact of grid load fluctuations on equipment and provides timely maintenance recommendations.

[0084] This invention is a fault detection and root cause analysis system for multi-source energy systems. It possesses strong adaptability and scalability, enabling wide application in multi-source energy systems of varying scales and types. It can be flexibly adjusted to meet different needs, supporting fault detection and analysis in different equipment and environments. Furthermore, it supports future equipment expansion and upgrades, ensuring greater application potential in future development. The system's scalability is not only reflected in its support for multiple types of energy equipment but also in its ability to integrate grid data, providing management and optimization solutions for larger-scale energy systems. Attached Figure Description

[0085] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort, wherein:

[0086] Figure 1 This is a system block diagram of the present invention;

[0087] Figure 2 This is a flowchart of the data processing of the feature extraction and dimensionality reduction module in this invention;

[0088] Figure 3 This is a data processing flowchart of Embodiment 1 of the present invention. Detailed Implementation

[0089] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0090] It should be noted that the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0091] Example 1

[0092] like Figure 1 As shown, this invention proposes a fault detection and root cause analysis system for multi-source energy systems. The system includes a data acquisition module, a data preprocessing module, a feature extraction and dimensionality reduction module, a digital twin module, and a fault detection and root cause analysis module. Details are as follows:

[0093] Data acquisition module

[0094] Various sensors installed in multi-source energy devices and their environment collect real-time operational data, including voltage, current, temperature, pressure, humidity, and wind speed. Cameras are used to acquire external data of the energy storage devices, such as whether there are cracks, stains, or damage on the device surface, to further analyze the device's condition. The system also collects operational data on the power grid, traditional power generation equipment, and photovoltaic power generation systems to provide comprehensive energy monitoring and system management.

[0095] The sensors in this embodiment include, but are not limited to, voltage sensors with a measurement accuracy of ±0.5mV, current sensors with an accuracy of ±0.1mA, and sensors with an accuracy of ±0.5℃, which can accurately acquire multi-dimensional operational data including voltage, current, temperature, pressure, humidity, and wind speed. Simultaneously, a high-resolution 1080p camera is used to monitor subtle changes in the appearance of the energy storage device in real time, such as whether cracks, stains, or damage appear on the device surface, providing intuitive visual information for a comprehensive assessment of the device's condition. Furthermore, specialized power monitoring equipment is used to collect grid status data, as well as operational data from traditional power generation equipment and photovoltaic power generation systems, for comprehensive energy monitoring and system management. The specific data collected includes:

[0096] Photovoltaic power generation data: power generation capacity, irradiance, temperature, etc., used to assess the photovoltaic power supply capability;

[0097] Energy storage device data includes: voltage, current, temperature, state of health (SOH), charge / discharge cycle count, etc. In this embodiment, the energy storage device is a lithium battery, and the data from the energy storage device can monitor the battery status in real time.

[0098] Traditional power generation equipment data includes operating parameters such as fuel consumption rate, power generation, and start-up frequency, which are used to optimize power generation efficiency.

[0099] Electricity load data: Electricity price fluctuations and load demand data, used together with other sensor data for fault detection and prediction;

[0100] Image data: Capture changes in the appearance of energy equipment, such as cracks, damage, and stains, through cameras to provide visual information about the equipment's status;

[0101] Environmental data, such as humidity, wind speed, and air pressure, are used for fault prediction and optimized scheduling.

[0102] Grid connection data: grid voltage, current, power, latitude, longitude, etc., used to monitor the stability of grid power supply and grid load status, affecting the switching decisions of energy sources;

[0103] The collected multimodal data is transmitted to the data preprocessing module using either a wireless sensor network (WSN) or wired communication to ensure efficient data transmission and real-time performance. In this embodiment, a wireless sensor network with a ZigBee communication protocol, a frequency band of 2.4 GHz, a transmission rate of 250 kbps, and a transmit power of 20 dBm is used for data transmission.

[0104] Data preprocessing module

[0105] The collected multimodal data undergoes preprocessing, including outlier detection, missing value imputation, noise reduction, and data normalization, to ensure data quality and usability and provide a reliable data foundation for subsequent analysis.

[0106] Outlier detection: For all collected operational and image data, outlier detection is performed first, using statistical methods or threshold detection to remove obviously abnormal data. For example, for voltage data of energy storage devices, the normal range is determined based on historical data or equipment specifications. If the voltage value at a certain moment exceeds the normal range, it is identified as an outlier and removed.

[0107] Missing value imputation: Various flexible and effective imputation methods are employed for missing values ​​in a dataset. For time series data, if data is missing at a certain point in time, it can be imputed using methods such as linear interpolation and spline interpolation, based on the trends of data before and after that point, to maintain data continuity. Simultaneously, for some complex data, machine learning prediction methods can be used, such as algorithms based on support vector machines (SVM) regression and neural networks for prediction and imputation. When using machine learning methods, historical data similar to the time period of the missing data are carefully selected as the training set. The data undergoes rigorous cleaning and normalization, and the model is optimized using evaluation metrics such as root mean square error (RMSE) and coefficient of determination (R²) to ensure the accuracy and reliability of the imputed values.

[0108] Noise Reduction: All collected operational data are treated as time-series data. Wavelet transform is used for noise reduction to effectively extract frequency domain features related to faults. For image data, the powerful image processing capabilities of CNNs are utilized to extract key spatial features and remove irrelevant background noise, highlighting crucial information in the images. For grid-connected data, an efficient filtering algorithm is first used for preliminary filtering to remove high-frequency noise and other interference. Then, wavelet transform is applied for further noise reduction to ensure accurate extraction of frequency domain features, thereby improving data quality and providing clear and reliable data signals for subsequent analysis.

[0109] In this embodiment, the wavelet function Daubechies, specifically the Db4 wavelet from the wavelet series, is selected. It has good time-frequency resolution based on the frequency characteristics of the signal and the fault characteristic frequency. The number of decomposition layers is determined according to the complexity of the signal and the noise level. An empirical formula determines it to be 5 layers. A soft thresholding method is used for noise reduction, thereby accurately extracting the frequency domain features of different frequency bands and effectively revealing potential fault modes.

[0110] For image data, the powerful image processing capabilities of CNNs are leveraged to extract key spatial features and remove irrelevant background noise, highlighting crucial information within the image. The CNN structure comprises three convolutional layers with 3×3 kernels and strides of 1 and 2 pooling layers, and two fully connected layers with 128 and 64 neurons respectively. The ReLU activation function is used to extract spatial features from the image, remove background noise, and identify damage or anomalies in the appearance of energy storage devices, such as minor cracks or dirt.

[0111] For grid-connected data, an efficient filtering algorithm, such as a Butterworth filter with a filter order of 4 and a cutoff frequency of 50Hz, is first used for preliminary filtering to remove high-frequency noise and other interference. Then, wavelet transform is used for further noise reduction to ensure accurate extraction of frequency domain features, thereby improving data quality and providing clear and reliable data signals for subsequent analysis.

[0112] Data standardization: To ensure the consistency and comparability of data from different sources, advanced data cleaning and normalization algorithms, such as the Z-score standardization method, are used to convert the data collected by various sensors into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating differences in dimensions and orders of magnitude between data, and creating favorable conditions for subsequent data fusion and analysis.

[0113] Data feature extraction and dimensionality reduction module

[0114] This module uses a hybrid algorithm of wavelet transform, CNN, and LSTM to perform in-depth data analysis and fault diagnosis.

[0115] Wavelet transform is a commonly used signal processing method that effectively extracts local features from signals, and is particularly suitable for the analysis of non-stationary signals. Wavelet transform plays a crucial role in fault detection, especially in multi-source energy systems, where it can process time-series signals such as battery voltage and current, removing noise and enhancing the visibility of fault characteristics. Therefore, wavelet transform is often combined with other deep learning algorithms such as CNN and LSTM to improve the accuracy of fault detection and prediction.

[0116] By combining wavelet transform with CNN and LSTM, time-series signals and image data of multi-source energy systems can be effectively processed and analyzed. This hybrid algorithm not only improves feature extraction capabilities but also enables efficient fault prediction and root cause analysis. Wavelet transform is used to denoise and decompose the signal, extracting local features, which are then learned through CNN. Finally, LSTM is used to model and predict the time-series data, significantly improving the accuracy and real-time performance of fault detection.

[0117] Wavelet transform: used to process time series data from sensors, extract frequency domain features, and perform noise reduction to obtain fault characteristic information.

[0118] Specifically, wavelet transform processes time-series data from sensors. By carefully selecting appropriate wavelet functions, non-stationary signals are decomposed, such as the Db4 wavelet from the Daubechies wavelet series, which offers good time-frequency resolution based on the signal's frequency characteristics and fault feature frequencies. A reasonable number of decomposition levels is determined, based on signal complexity and noise levels, through experiments or empirical formulas. Soft or hard thresholding methods are then used to accurately extract frequency domain features from different frequency bands, effectively revealing potential fault modes. The soft thresholding formula is as follows:

[0119] ,

[0120] in, For the first frequency band Wavelet coefficients of position, Used to determine the location of wavelet functions in the time or spatial domain, different The value represents the different translation positions of the wavelet function in time or space. , For the threshold, At that time, wavelet coefficients According to Perform shrinkage treatment; when At that time, wavelet coefficients It was set directly to 0. For symbolic functions, .

[0121] By conducting in-depth analysis of the energy distribution in different frequency bands, such as calculating the energy of each frequency band... ,in, Let K be the wavelet coefficient at position k in the j-th frequency band. For the first The energy of the frequency band can accurately determine whether there are faults such as short circuits or plate aging inside the battery, such as a sudden increase in low-frequency energy or a gradual decrease in high-frequency energy.

[0122] In this invention, the voltage and current data in the energy storage device are non-stationary signals. Wavelet transform is used to decompose the non-stationary signals, extract the frequency domain features of different frequency bands, and reveal potential fault modes. By analyzing the energy distribution of different frequency bands, it is possible to determine whether there are faults such as short circuits or plate aging inside the battery.

[0123] Traditional power generation equipment data is used to extract fuel consumption trends through data fitting and difference methods or time series decomposition methods, and to extract power output fluctuation characteristics through wavelet transform methods or rolling standard deviation methods.

[0124] Fuel consumption trend feature extraction:

[0125] Data fitting and difference method: First, collect time series data of fuel consumption of traditional power generation equipment, perform polynomial fitting based on the time series, and then obtain the trend of fuel consumption by calculating the derivative or difference of the fitting function.

[0126] Time series decomposition method: This method decomposes fuel consumption data into trend components, seasonal components, periodic components, and stochastic components, focusing on the trend component after decomposition. For example, the Holt-Winters method can be used to separate the long-term trend of fuel consumption. The Holt-Winters method includes three smoothing equations and level equations. ,

[0127] Trend equation ,

[0128] Seasonal equation ,

[0129] in, In time The level value, In time The level value, In time The trend value, In time Seasonal factors, In time The actual observed value, , , For smoothing parameters, In time Seasonal factors, For seasonal cycles, the optimal decomposition results are obtained by optimizing these parameters, for example... , , .

[0130] Power output fluctuation feature extraction:

[0131] Wavelet transform method: For power output signals from power generation equipment, which may be non-stationary, wavelet transform is used to decompose them into sub-signals of different scales or frequencies, and the energy changes and fluctuations of each sub-signal are analyzed. In this embodiment, the feature extraction step selects the Mort wavelet, and the wavelet coefficient formula is:

[0132] ,

[0133] in, These are wavelet coefficients. The power output signal is related to time. The function, is the scaling parameter used to scale the width of the wavelet function. These are translation parameters used to shift the wavelet function along the time axis. For wavelet functions, This represents the wavelet function after scaling and translation transformations, and is obtained by analyzing the squared modulus of the wavelet coefficients. The fluctuation characteristics of power output are described by how they change over time and scale.

[0134] Rolling standard deviation method: Set a rolling window, such as a window size of 10, and calculate the standard deviation of the power output data within the window. As the window rolls, a series of standard deviation values ​​are obtained to describe the change in the degree of power output fluctuation over time.

[0135] The photovoltaic power generation data is used to extract characteristics of power generation stability, correlation between light intensity and power generation, and the influence of temperature on power generation efficiency.

[0136] Extraction of power generation stability features:

[0137] First, a time period is selected to examine the stability of power generation. Within this time period, the power generation value of the photovoltaic system is recorded at each moment. Then, the average level of these power values ​​is calculated. ,in, This represents the average power level. For the first Power generation at any given moment This refers to the number of data points within a given time period. For example, if 240 data points were collected in one day, then... Then calculate the deviation between the power value and the average value. The formula is: By comprehensively calculating these deviations, such as by calculating the variance or standard deviation, a value that reflects the magnitude of power fluctuations is obtained. This value is the power generation stability characteristic. If this value is small, it means that the power generation is relatively stable during that period; conversely, it indicates that the power fluctuations are large.

[0138] For example, by observing the power generation over the past day, calculating the average power for that day, and then looking at the differences between the power at each time point and the average value, and summarizing these differences, we can know how stable the power generation is for that day, and thus understand the reliability of the output power of the photovoltaic power generation system during that period. This is crucial for grid connection and energy dispatch. Stable power output helps maintain the balance and stable operation of the grid, and avoids the impact on the grid caused by large power fluctuations.

[0139] Feature extraction of the correlation between light intensity and power generation:

[0140] Select a time range and record the light intensity at each moment within that time range. and power generation Calculate the average light intensity. and the average power generation ,in, This represents the average light intensity. Average power generation for each entity If 10,080 data points were collected within a week, then... Then analyze the deviation between the light intensity and the average value. and the deviation of power generation from the average value The study investigates the correlation between these two sets of deviations to obtain a numerical value that reflects the close relationship between light intensity and power generation, namely the correlation characteristic.

[0141] Calculate the correlation coefficient, such as using the Pearson correlation coefficient: ,

[0142] in, The Pearson correlation coefficient is used. For the first The deviation of the light intensity at any given time from the average value. For the first The deviation between the power generation at any given time and the average value. The number of data points;

[0143] If power generation increases with rising light intensity, and the correlation coefficient is close to 1, the correlation is strong. Conversely, if power generation does not change significantly with changes in light intensity, and the correlation coefficient is close to 0, the correlation is weak. Understanding this correlation helps predict changes in power generation when the trend of light intensity changes is known, thereby optimizing energy allocation and improving energy efficiency. For example, it allows for the preparation of energy storage devices to receive more electricity when light intensity is about to increase.

[0144] For example, record relevant data over a week, calculate the average values ​​of light intensity and power generation, and then examine the correlation between their deviations from the average. If power generation increases along with light intensity, and this relationship is significant, the correlation is strong; conversely, if power generation does not change significantly with changes in light intensity, the correlation is weak. Understanding this correlation helps predict changes in power generation when the trend of light intensity changes, thereby optimizing energy allocation and improving energy efficiency. For instance, when light intensity is about to increase, prepare energy storage devices in advance to receive more electricity.

[0145] Feature extraction of the effect of temperature on power generation efficiency:

[0146] Continuously track the temperature and power generation of photovoltaic (PV) power generation equipment over a period of time, accumulating data. Based on temperature, attempt to identify the intrinsic relationship between it and power generation, constructing a model that describes this relationship. Use the relevant parameters in this model to measure the degree of temperature's impact on power generation efficiency.

[0147] For example, by continuously collecting temperature and power generation data over a month and observing how power generation changes with temperature, a suitable mathematical model can be constructed to describe this relationship. A certain coefficient in the model can serve as a characteristic of how temperature affects power generation efficiency. If this coefficient shows a significant decrease in power generation as temperature rises, then attention needs to be paid to the equipment's heat dissipation, as this may lead to reduced power generation efficiency, affecting the performance and power output of the entire photovoltaic power generation system. Conversely, if the coefficient is abnormal, further analysis is needed to determine whether the equipment is operating within a suitable temperature range to ensure normal operation and efficient power generation.

[0148] The environmental data is used to extract the trends in humidity, wind speed, and air pressure. The mean and variance of these trends are then determined using time series analysis.

[0149] Humidity change trend feature extraction:

[0150] Record ambient humidity data at regular time intervals, such as once per hour. Calculate the average humidity data over a longer period, such as a week or a month. Then observe how each humidity data point changes compared to the average—whether it is higher or lower than the average, and the trend of this difference over time—whether it continues to rise, fall, or fluctuates. Further analysis can be conducted on the dispersion of the humidity data, i.e., the magnitude of the fluctuation.

[0151] For example, by recording humidity hourly throughout a summer month and calculating the average humidity for that month, and then observing the difference between the humidity at different times of the day and the average, it can be found that if the humidity is consistently higher than the average for a few days with small fluctuations, it may indicate that rainy weather is coming soon. This can be helpful for protecting outdoor energy equipment from moisture. However, when the humidity fluctuates greatly, it may increase the risk of equipment getting damp, affecting the electrical performance and lifespan of the equipment. Therefore, it is necessary to take corresponding measures in advance based on the characteristics of humidity change trends, such as strengthening equipment sealing or adding dehumidification equipment.

[0152] Wind speed change trend feature extraction:

[0153] Wind speed data is collected at fixed time intervals. The average wind speed over a certain period, such as a day or a week, is calculated. The maximum and minimum wind speeds during this period and the times they occur are determined to understand extreme wind conditions. Simultaneously, the distribution of wind speed data around the average value is analyzed, i.e., the degree of fluctuation.

[0154] For example, in a region rich in wind resources, wind speed data can be recorded continuously for a week to calculate the average wind speed and identify the times and values ​​of maximum and minimum wind speed. If the average wind speed is high and fluctuates little, it may mean relatively stable power generation conditions for wind power equipment. However, if the maximum wind speed exceeds the equipment's tolerance limit, wind protection measures are needed to protect the equipment. On the other hand, large wind speed fluctuations may affect the stability of wind power generation, requiring the use of energy storage devices and other means to balance power output and ensure the stability of energy supply. At the same time, it is also necessary to reinforce transmission lines to prevent damage caused by strong winds.

[0155] Extraction of air pressure change trend features:

[0156] Record air pressure data regularly, observe the trajectory of air pressure changes over time, and determine whether the air pressure is rising, falling, or remaining relatively stable. Calculate the rate of change of air pressure over a period of time, and the distribution of air pressure across different value ranges.

[0157] For example, long-term monitoring of air pressure near energy facilities in mountainous areas can reveal a sustained drop in pressure, potentially indicating impending weather changes such as rain or storms. This could impact solar power equipment, as weather changes can alter sunlight intensity, affecting power generation efficiency. Understanding air pressure variation patterns helps predict the impact of weather on energy equipment operation, allowing for timely adjustments to operating strategies or the implementation of protective measures. For instance, when air pressure changes foreshadow severe weather, equipment can be reinforced or certain outdoor maintenance work can be suspended to ensure the safe and stable operation of the energy system.

[0158] It can also extract frequency features under extreme environmental conditions:

[0159] Based on the operational requirements of energy equipment and local environmental characteristics, judgment criteria or thresholds for extreme environmental conditions such as high temperature, high humidity, and strong winds are established. For each recorded temperature, humidity, and wind speed data point, it is compared with the corresponding threshold; if the threshold is exceeded, an extreme event of that type is recorded. Within a specific observation period, such as a quarter, the number of occurrences of various extreme events is counted, and then the frequency of these extreme events is calculated.

[0160] For example, in coastal energy systems, high-temperature thresholds are set at 35°C, high-humidity thresholds at 80%, and strong wind thresholds at level 10. Within a quarter, data is recorded each time to determine if these extreme conditions have been met. If statistics show a high frequency of high-temperature events, it may be necessary to optimize equipment cooling or add cooling measures to prevent damage from high temperatures; if high humidity is frequent, moisture-proof and rust-proof treatments for equipment should be strengthened; if strong winds are frequent, windproof design and reinforcement measures for equipment need to be enhanced to ensure the normal operation of energy equipment in extreme environments, reduce equipment failures and downtime caused by extreme environments, and improve the reliability and stability of the energy system.

[0161] CNN: Used to analyze spatial features in image data to identify damage or anomalies in the appearance of energy storage devices, such as cracks and stains.

[0162] Specifically, CNNs, with their powerful convolutional and pooling layer structures, perform deep analysis of image data to extract key spatial features. By appropriately setting the convolutional kernel size and selecting suitable activation functions, they can effectively identify damage or anomalies in the appearance of energy storage devices, such as minor cracks and stains, providing intuitive visual evidence for device condition assessment.

[0163] In this embodiment, the CNN structure includes three convolutional layers with a kernel size of 3×3 and strides of 1 and 2 pooling layers and two fully connected layers. The number of neurons in the fully connected layers are 128 and 64, respectively, and the activation function is the ReLU function.

[0164] Multimodal data fusion and dimensionality reduction: The features extracted from various operational data and the spatial features extracted by CNN are combined and dimensionality reduced by PCA to form a unified low-dimensional feature set, which is then used by LSTM for further modeling and prediction.

[0165] The frequency domain features extracted by wavelet transform are combined with the spatial features extracted by CNN. Specifically, during data fusion, a multi-head attention mechanism is used to assign appropriate weights to different features based on their importance and relevance. The number of attention heads is determined based on factors such as the number of key performance indicators of energy storage devices in a multi-source energy system, including energy storage capacity and charging / discharging efficiency, the complexity of interactions with other energy sources, and the accuracy requirements for monitoring the status of energy storage devices. The weight of each attention head is determined by calculating indicators such as the dynamic correlation coefficient and information entropy between features, and different initial weights are assigned according to the importance of the energy storage device in the system. The fused feature set is then subjected to dimensionality reduction using PCA to calculate the covariance matrix of the data, solve for eigenvalues ​​and eigenvectors, select the main eigenvectors based on the magnitude of the eigenvalues, and project the high-dimensional data into a low-dimensional space to form a unified low-dimensional feature set, providing optimized data input for LSTM modeling.

[0166] The number of attention heads is determined based on factors such as the number of key performance indicators of energy storage devices in a multi-source energy system, the complexity of their interactions with other energy sources, and the accuracy requirements of the system for monitoring the status of energy storage devices. Specifically:

[0167] Energy storage devices in multi-source energy systems have several key performance indicators, such as energy storage capacity, charge / discharge efficiency, remaining capacity, and state of health (SOH). Each key performance indicator has a significant impact on the system's operating status and fault detection; therefore, an attention head is needed for each indicator. For example, when considering the two key performance indicators of energy storage capacity and charge / discharge efficiency, at least two attention heads are required. If the system further considers indicators such as remaining capacity and state of health, depending on actual needs and computing resources, this may increase to four or more attention heads to ensure a comprehensive assessment of the impact of different aspects of the energy storage device's performance on system operation.

[0168] In multi-source energy systems, energy storage devices have complex interactions with energy sources such as photovoltaic power generation, traditional power generation equipment, and the power grid. If energy exchange between energy storage devices and other energy sources is frequent—for example, when photovoltaic power generation is unstable, energy storage devices need to frequently charge and discharge to balance energy supply and demand—the interactions become complex. In this case, more attention heads are needed to capture the characteristics of the interaction data between different energy sources. In a simple multi-source energy system, where energy storage devices only exchange a small amount of energy with the power grid, perhaps only 2-3 attention heads are needed to process the data interactions related to the power grid. However, in a complex system where energy storage devices have close energy scheduling and collaborative working relationships with photovoltaic power generation, traditional power generation equipment, and the power grid, 4-6 attention heads may be needed to comprehensively analyze the interaction characteristics between various energy sources in order to accurately determine the system's operating status and potential faults.

[0169] If the system requires high precision in monitoring the status of energy storage devices, such as accurately assessing the health status of energy storage devices to prevent failures in advance, more detailed analysis of various data characteristics is needed. In this case, to improve monitoring accuracy, the number of attention heads will be increased. Taking the health status monitoring of energy storage devices as an example, if only a general understanding of their health status is needed, 2-3 attention heads may be sufficient for preliminary data analysis; however, if the goal is to pinpoint minute potential faults within the battery, such as the early stages of plate aging, 4-8 attention heads may be required. By focusing on and analyzing more data characteristics, the accuracy of energy storage device status monitoring is improved, leading to more accurate fault prediction and diagnosis. In this invention, considering the above factors and based on the actual operational needs of multi-source energy systems and the accuracy requirements for energy storage device status monitoring, the number of attention heads is determined to be 4. This achieves a good balance between computational resources and monitoring accuracy, meeting the system's requirements for fault detection and root cause analysis.

[0170] LSTM: LSTM is used to model time series data, learn the temporal characteristics of the data, capture the long-term dependencies of system operating states, and perform fault prediction and abnormal behavior identification. LSTM utilizes its unique gating structure, by reasonably setting the number of hidden layers and the learning rate, and combining external information such as grid voltage and load fluctuations, to effectively improve the accuracy and real-time performance of fault prediction, identify potential faults and abnormal behaviors in advance, and provide reliable protection for the stable operation of the system.

[0171] The LSTM model structure in this embodiment includes an input layer, hidden layers, and an output layer. By carefully adjusting parameters such as the number of hidden layers and the learning rate (e.g., a learning rate of 0.001), model performance is optimized. During training, the mean squared error (MSE) loss function and the Adam optimization algorithm are used to iteratively train the model, enabling it to accurately learn temporal patterns in the data, predict potential faults and abnormal behaviors in advance, and provide reliable assurance for the stable operation of the system.

[0172] like Figure 2 As shown, the data undergoes processing through the data feature extraction and dimensionality reduction modules.

[0173] 1. Multimodal data input: The input data is divided into two parts: time series data, such as voltage and current from sensors, and image data, such as the appearance of the device from the camera.

[0174] 2. Wavelet transform: Extract frequency domain features from time series data and perform noise reduction processing.

[0175] 3. CNN: Extracts spatial features from image data to identify damage to the appearance of equipment, such as cracks and stains.

[0176] 4. Multimodal data fusion: Combining the frequency domain features extracted by wavelet transform with the spatial features extracted by CNN to perform data dimensionality reduction and obtain a low-dimensional feature set.

[0177] 5. LSTM Modeling: Based on the dimensionality-reduced low-dimensional feature set data, LSTM is used to model time series data and predict faults and abnormal behaviors.

[0178] 6. Output Results: Finally, the fault prediction and anomaly diagnosis results are output through a hybrid algorithm.

[0179] Digital twin module

[0180] A digital twin model is a virtual model of a multi-source energy system built based on initial data. It establishes a one-to-one correspondence between the physical system and the virtual model, enabling real-time monitoring of the system's operational status. The digital twin model can be dynamically updated based on real-time data collection, such as changes in voltage and current, and detection results of external damage, generating corresponding virtual models that maintain consistency with the physical equipment. The digital twin model integrates data from energy storage devices, photovoltaic power generation, traditional power generation equipment, and the power grid, ensuring that the virtual model accurately reflects the coordination and interaction of all energy sources.

[0181] Updates to the digital twin model

[0182] Triggering Mechanism: When the image data collected by the data acquisition module shows significant changes in the appearance of the energy equipment, such as cracks or damage, or when one or more parameters in the operational data, such as voltage, current, and temperature, exceed the preset normal range, the system automatically triggers the digital twin model update process to ensure that the virtual model can respond promptly to changes in the physical equipment's state. Taking a lithium battery as an example, when the collected voltage value is less than 2.5V or greater than 4.2V, it is determined to be outside the normal range, triggering a digital twin model update.

[0183] Incremental learning: Utilizing incremental learning methods, such as gradient-based online learning algorithms, the parameters of the digital twin model are dynamically optimized. By introducing newly acquired data in real time, the model can quickly adapt to changes in the operating conditions of the multi-source energy system, ensuring that the virtual model always maintains a high degree of consistency with the operating status of equipment in the actual system, providing an accurate simulation environment for fault detection and prediction.

[0184] Real-time Adjustment: The fault detection algorithm is optimized online by incorporating newly acquired data, improving its adaptability and real-time performance in dynamic environments. Based on updated data, the digital twin model immediately provides real-time warnings of system faults and adjusts the model accordingly. Simultaneously, the fault detection algorithm is optimized online using new data, improving the model's adaptability and response speed to changes in the dynamic environment, ensuring that the system can promptly detect and handle potential faults under complex and changing operating conditions. For example, when significant grid voltage fluctuations are detected, the digital twin model promptly adjusts the simulation of the energy storage device's charging and discharging strategy, predicts the optimal operating state of the energy storage device under current grid conditions, and feeds relevant information back to the fault detection and root cause analysis module to assist in more accurate fault diagnosis and analysis.

[0185] Fault Detection and Root Cause Analysis Module

[0186] This system performs real-time, comprehensive monitoring of various devices in multi-source energy systems. It uses hybrid algorithms to deeply analyze collected multimodal data and combines this with digital twin models to conduct root cause analysis, pinpointing the faulty devices and their causes. The multimodal data includes various data sources such as energy storage equipment data, grid connection data, power load data, photovoltaic power generation data, traditional power generation equipment data, image data, and environmental data, enabling comprehensive identification and analysis of potential fault sources.

[0187] For example, if abnormal parameters such as voltage, current, and temperature are detected, the system will use data processing and model analysis, combined with the simulation and prediction results of the digital twin model of the equipment's operating status, to locate possible causes of failure, such as equipment overheating, battery damage, load fluctuations, etc.

[0188] The results of fault diagnosis and root cause analysis are displayed through a visual interface to support the decision-making of operation and maintenance personnel. This module will comprehensively display data from different data sources, including real-time operating data of energy storage devices, power output of photovoltaic power generation, operating status of traditional power generation equipment, grid fluctuations, and environmental data trends. It will also provide accurate fault warning information, offering comprehensive decision-making support for operation and maintenance personnel.

[0189] By monitoring and analyzing in real time, the results of fault diagnosis and root cause analysis are sent to users or management systems to issue fault warnings. Combined with front-end display, such as HTML5 and JavaScript, this invention can generate detailed fault diagnosis reports, intuitive root cause analysis charts, and forward-looking trend prediction charts in real time. This helps maintenance personnel to quickly and accurately understand the system status and take timely and effective fault troubleshooting measures or preventive maintenance strategies to ensure the stable and efficient operation of multi-source energy systems.

[0190] The following example illustrates this:

[0191] In a multi-source energy system, multiple lithium battery modules provide power through charging and discharging processes. During the operation of the multi-source energy system, if one of the battery modules overheats, a temperature sensor will detect abnormal temperature data, such as... Figure 3 As shown, the system of the present invention performs fault diagnosis and prediction through the following steps:

[0192] (1) Temperature anomaly detection

[0193] First, wavelet transform is used to perform frequency domain analysis on the temperature signal at different frequency bands, extracting the high-frequency components of the signal to identify potential fault characteristics. Wavelet transform helps to denoise and extract subtle anomalies in battery temperature changes, thus enabling earlier problem detection. This signal data is then processed together with other relevant data from the power grid and traditional power generation equipment, such as load fluctuations and grid voltage, to provide the system with more comprehensive operational status information.

[0194] (2) Image Data Analysis

[0195] Simultaneously, convolutional neural networks are used to analyze battery surface image data to extract damage features such as cracks, stains, and other damage, aiding in determining the cause of the malfunction. The image data, combined with sensor data, helps the system comprehensively assess the battery's health status, particularly potential surface-related faults.

[0196] (3) Time series data prediction

[0197] Long Short-Term Memory (LSTM) networks are used to model time-series temperature signals, capturing the long-term dependencies in battery module temperature fluctuations. LSTM can identify temperature trends and sudden anomalies, providing accurate early warning signals for subsequent fault prediction. At this point, power load data, such as electricity price fluctuations, and grid connection data, such as grid load fluctuations, will jointly participate in time-series prediction to optimize prediction accuracy.

[0198] (4) Grid connection and energy coordination

[0199] Joint processing of grid status data and battery energy storage system data helps improve the system's adaptability to external load fluctuations. Real-time load changes and electricity price fluctuations in the grid provide important external environmental data for fault prediction, making the system more real-time and accurate in fault diagnosis.

[0200] (5) Digital twin model update

[0201] With real-time updates to temperature and image data, the digital twin model receives timely data from sensors, automatically triggering model updates for the battery module. By simulating performance changes caused by battery overheating, the digital twin model reflects the battery's state changes under abnormal environments. The digital twin model ensures that the virtual model always remains consistent with the physical device, thereby enabling real-time monitoring of the device's operating status and simulating the interaction between the battery and the grid under grid-connected conditions.

[0202] (6) Root cause analysis

[0203] Supported by digital twin models and hybrid algorithms, further root cause analysis was conducted. Through comprehensive analysis of multiple data sources, including data from the power grid, energy storage devices, and load data, the root cause of the failure was ultimately confirmed to be an internal short circuit in the battery module. This failure not only caused abnormal temperatures but also affected the battery module's discharge capacity and the overall stability of the system.

[0204] (7) Fault warning and maintenance suggestions

[0205] Fault information is generated into reports through a visual interface, allowing maintenance personnel to view the fault type of the battery module, possible causes, and suggested repair measures. The system intelligently recommends repair steps to help maintenance personnel take quick action, ensuring stable system operation and collaboratively managing other energy sources, such as grid load and photovoltaic power generation, to optimize energy dispatch.

[0206] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be conceived by those skilled in the art within the technical scope disclosed in the present invention without creative effort should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A fault detection and root cause analysis system for a multi-source energy system, characterized in that, It includes a data acquisition module, a data preprocessing module, a data feature extraction and dimensionality reduction module, a digital twin module, and a fault detection and root cause analysis module, among which: The data acquisition module acquires various operational and image data of the multi-source energy system in real time and uploads them to the data preprocessing module. The data preprocessing module preprocesses the various operational and image data collected by the data acquisition module, including outlier detection, missing value imputation, noise reduction, and data standardization. The data feature extraction and dimensionality reduction module performs feature extraction and dimensionality reduction on the preprocessed operational data and image data. The preprocessed operational data and image data are then input into a hybrid algorithm combining wavelet transform, CNN, and LSTM. Wavelet transform and CNN extract features from the preprocessed data, and the extracted features are fused to obtain a fused feature set. The high-dimensional data in the fused feature set is then dimensionality reduced, and LSTM is used for time series modeling to detect abnormal patterns and potential faults in multi-source energy systems. The digital twin module builds a virtual model of a multi-source energy system based on initial data, namely the digital twin model. The digital twin model is dynamically updated according to real-time data collection, maintaining consistency with the physical equipment in the multi-source energy system, reflecting the status of the physical equipment, and providing a more accurate basis for fault detection and root cause analysis. The fault detection and root cause analysis module uses a hybrid algorithm to analyze real-time acquired data, detect abnormal patterns and potential faults in multi-source energy systems, compare the prediction results of the real-time data input digital twin model to determine whether there are any deviations, and further confirm the occurrence of the fault; combining the feature extraction capabilities of the digital twin model and the hybrid algorithm, it locates the key causes of the fault, simulates the system operating state, analyzes the potential causes of the fault, and outputs the analysis results. The various operational data of the multi-source energy system include photovoltaic power generation data, energy storage device data, traditional power generation equipment data, power load data, environmental data, and grid connection data. These operational data are acquired through various sensors installed in the multi-source energy system. The image data of the multi-source energy system, i.e., the changes in the appearance of the energy equipment, is collected by cameras and transmitted in real time to the data preprocessing module. The data preprocessing module preprocesses the various operational and image data collected by the data acquisition module. Specifically: Outlier detection removes obviously abnormal data through statistical methods or threshold detection. Missing value imputation uses interpolation, mean averaging, or machine learning prediction methods to fill in missing values. For noise reduction, the collected operational data are treated as time-series data, and wavelet transform is used for noise reduction to extract frequency domain features related to the fault. The image data is processed by CNN to extract spatial features and remove irrelevant background noise. For grid-connected data, filtering is performed first, and then wavelet transform is used for noise reduction to extract frequency domain features related to the fault, so as to ensure data quality. Data standardization involves normalizing data from different sources to ensure data format consistency. The data feature extraction and dimensionality reduction module uses wavelet transform and CNN to extract features from the preprocessed data, specifically: The voltage and current data in the energy storage device data are non-stationary signals. Wavelet transform is used to decompose the non-stationary signals, extract frequency domain features of different frequency bands, and reveal potential fault modes. The image data was used to extract key spatial features using CNN; The traditional power generation equipment data is used to extract fuel consumption trends through data fitting and difference methods or time series decomposition methods, and to extract power output fluctuation characteristics through wavelet transform methods or rolling standard deviation methods. The grid-connected data is extracted using wavelet transform to extract frequency domain features, and the spectral components and energy distribution in the frequency domain are analyzed to determine the matching between the grid load status and energy supply. The various frequency domain features extracted by wavelet transform and the spatial features extracted by CNN are combined. Specifically, during the data fusion process, a multi-head attention mechanism is used to assign reasonable weights to different features based on the importance and relevance of the data. The number of attention heads is determined based on the complexity of the interaction between the key performance indicators of the energy storage device in the multi-source energy system and other energy sources, as well as the accuracy requirements of the system for monitoring the status of the energy storage device. The weight of each attention head is determined by calculating the dynamic correlation coefficient and information entropy index between features, and different initial weights are assigned according to the importance of the energy storage device in the system. The fused feature set is then reduced using the PCA dimensionality reduction method to calculate the covariance matrix of the data, solve for the eigenvalues ​​and eigenvectors, select the main eigenvectors based on the magnitude of the eigenvalues, and project the high-dimensional data into a low-dimensional space to form a unified low-dimensional feature set, providing optimized data input for LSTM modeling. The number of attention points is determined based on the number of key performance indicators of energy storage devices in a multi-source energy system, the complexity of their interaction with other energy sources, and the accuracy requirements of the system for monitoring the status of energy storage devices. The feature extraction in the data feature extraction and dimensionality reduction module also includes: The photovoltaic power generation data is used to extract characteristics of power generation stability, correlation between light intensity and power generation, and the influence of temperature on power generation efficiency. The environmental data is used to extract the trends in humidity, wind speed, and air pressure. The mean and variance of these trends are then determined using time series analysis. After feature extraction from the preprocessed data, the features are fused to obtain a fused feature set, specifically: The grid-connected data is integrated with photovoltaic power generation data, energy storage equipment data, and traditional power generation equipment data. Through a multi-head attention mechanism, the features extracted from different data sources are weighted and integrated to provide more comprehensive energy supply information. The environmental data is combined with features extracted from power load data, energy storage equipment data, traditional power generation equipment data, and photovoltaic power generation data to provide more multi-dimensional information for fault prediction and system scheduling. The dimensionality reduction processing of the high-dimensional data in the fusion feature set specifically involves: Principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional data in the fusion feature set, retaining key features and reducing redundant data. The fusion feature set includes: The characteristics of power generation stability, correlation between light intensity and power generation, and influence of temperature on power generation efficiency were extracted from photovoltaic power generation data. Voltage and current frequency domain characteristics extracted from energy storage device data; Fuel consumption trends and power output fluctuation characteristics extracted from data of traditional power generation equipment; Spatial features extracted from image data; Frequency domain characteristics of grid stability and load variation extracted from grid-connected data; Environmental data were used to extract trends in humidity, wind speed, and air pressure. The principal component analysis method calculates the covariance matrix of the data to find eigenvalues ​​and eigenvectors, selects the main eigenvectors based on the magnitude of the eigenvalues, thereby projecting high-dimensional data into a low-dimensional space, retaining key features, and combining them with spatial features extracted from image data to form a unified low-dimensional feature set, simplifying the subsequent analysis process. The digital twin model is dynamically updated based on real-time collected data. The specific steps are as follows: The triggering mechanism automatically triggers the update of the digital twin model when the data acquisition module detects that one or more data points in the image data or various operational data have changed beyond the normal range. Incremental learning uses incremental learning methods to dynamically update the parameters of the digital twin model, simulate the operation of a multi-source energy system, and ensure that the virtual model can reflect the operating status of the equipment in the actual multi-source energy system in a timely manner. Real-time adjustments: Based on updated data, the digital twin model immediately provides real-time warnings of system faults and makes corresponding adjustments to the model according to the warning information. At the same time, it combines new data to optimize the fault detection algorithm online, improves the model's adaptability and response speed to dynamic environmental changes, and ensures that the system can promptly detect and handle potential faults under complex and ever-changing operating conditions. By conducting in-depth analysis of the energy distribution in different frequency bands, the energy of each frequency band is calculated. ,in, For the first j frequency band k Wavelet coefficients of position, For the first j The energy of the frequency band is used to accurately determine whether there is a short circuit or plate aging fault inside the battery. The judgment condition is a sudden increase in low-frequency energy or a gradual decrease in high-frequency energy. The LSTM model time series data based on a low-dimensional feature set, captures the long-term dependencies of system operating states, and predicts potential faults and abnormal behaviors. The system performs fault diagnosis and prediction through the following steps: (1) Temperature anomaly detection: First, wavelet transform is used to perform frequency domain analysis on the temperature signal at different frequency bands to extract the high-frequency components of the signal and identify potential fault characteristics. This signal data is then compared with relevant data from the power grid and traditional power generation equipment. (2) Image data analysis: By analyzing battery surface image data through convolutional neural networks, damage features of the battery appearance can be extracted; the image data combined with sensor data helps the system to comprehensively assess the health status of the battery, especially potential failures on the battery surface. (3) Time series data prediction: Long Short-Term Memory (LSTM) networks are used to model time-series data of temperature signals to capture the long-term dependence of battery module temperature fluctuations; LSTM can identify temperature trends and sudden anomalies, providing accurate early warning signals for subsequent fault prediction; at this time, power load data will also participate in time-series prediction to optimize the accuracy of prediction. (4) Grid connection and energy coordination: Joint processing of grid status and battery energy storage system data helps improve the system's adaptability to external load fluctuations; real-time load changes and electricity price fluctuations of the grid provide important external environmental data for fault prediction, making the system more real-time and accurate in the fault diagnosis process. (5) Digital twin model update: With real-time updates of temperature and image data, the digital twin model receives real-time data from sensors and automatically triggers model updates for the battery module. The digital twin model reflects the battery's state changes under abnormal environments by simulating performance changes caused by battery overheating. The digital twin model ensures that the virtual model always remains consistent with the physical device, thereby monitoring the device's operating status in real time and simulating the interaction between the battery and the grid under grid-connected conditions. (6) Root cause analysis of the failure: With the support of digital twin models and hybrid algorithms, root cause analysis is further conducted. By comprehensively analyzing multiple data sources, including data from the power grid, energy storage devices, and load data, the root cause of the failure is finally identified. (7) Fault warning and maintenance suggestions: Fault information is generated into reports through a visual interface, allowing maintenance personnel to view the fault type of the battery module, possible causes, and suggested maintenance measures. The system helps maintenance personnel take quick action by intelligently recommending maintenance steps to ensure stable system operation and to coordinate the management of other energy sources, grid load, and photovoltaic power generation to optimize energy dispatch.

2. The multi-source energy system fault detection and root cause analysis system according to claim 1, characterized in that, The results output by the fault detection and root cause analysis module are specifically shown as follows: Fault warning information is issued to users or management systems through real-time monitoring and analysis. Root cause analysis report provides a detailed analysis report, listing the type of failure, possible causes, and recommended maintenance measures; Data visualization displays system operating status, fault prediction, and diagnostic results in chart form, making it easier for users to understand and make decisions.

Images