Power system distributed data anomaly diagnosis method based on deep neural network

Abstract

The invention provides a power system distributed data anomaly diagnosis method based on a deep neural network, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining the multi-modal data of each distributed energy of a power system, and obtaining the processed distributed data; modeling each node of the power system and the relationship thereof into a graph structure to obtain a multi-time-scale spatio-temporal feature vector; pre-training is carried out by constructing a self-supervised learning task, and effective features between different time scales and modal data are extracted; generating a normal data distribution model by using a generative adversarial network, and summarizing abnormal data samples into a comprehensive data set; obtaining a multi-task optimization model through sharing characteristics among tasks; a final optimized anomaly response system is obtained, and then the task of distributed data anomaly diagnosis of the power system is completed through the anomaly response system. According to the method, efficient anomaly diagnosis and response processing of the distributed data of the power system are realized, and the method has remarkable technical advantages and practical application value.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring and diagnosis of power systems, and is a method for diagnosing abnormal distributed data of power systems based on a deep neural network. This method is particularly applicable to monitoring and anomaly detection of data in application scenarios such as large power systems, microgrids, and intelligent distribution networks in the context of smart grids and the energy Internet. With the rapid development of the energy Internet, power systems are evolving towards a highly integrated, intelligent, and distributed direction. The present invention aims to address the challenges faced by new power systems in data analysis and safe operation. Background Art

[0002] In the field of power system anomaly diagnosis, technology is developing rapidly, especially driven by deep learning technology. To better understand the industry background in which the present invention is located and highlight its innovation and practicality, this chapter will be elaborated in detail from the following aspects: First, the characteristics and challenges of power system transformation under the background of carbon neutrality will be introduced; second, the existing power system anomaly detection methods will be classified and reviewed in detail, including rule-based methods, statistical methods, and machine learning methods; then, specific examples and data support for the deficiencies of the existing technology will be supplemented; next, a review of the latest progress in the application of deep learning technology in power systems will be added; finally, the problem orientation will be strengthened, and the specific pain points and challenges of the current technology will be clearly pointed out.

[0003] Characteristics and Challenges of Power System Transformation under the Background of Carbon Neutrality

[0004] With the increasing global attention to climate change issues, achieving carbon neutrality has become the consensus and goal of all countries. As an important area of energy consumption, the transformation of power systems is crucial for achieving the carbon neutrality goal. New power systems exhibit significant characteristics such as a high proportion of renewable energy, multi-interaction of power sources, grids, loads, and energy storage, and wide application of power electronic devices. However, these characteristics also bring new challenges to the safe and stable operation of power systems.

[0005] On the one hand, the access of a high proportion of renewable energy increases the uncertainty and volatility of power systems. The power generation of renewable energy such as wind power and photovoltaic power is greatly affected by meteorological conditions, with intermittency and volatility, which brings difficulties to the balanced dispatching of power systems. On the other hand, the multi-interaction of power sources, grids, loads, and energy storage makes the structure of power systems more complex. The access of new types of loads such as distributed power sources, energy storage devices, and electric vehicles makes the operation mode of power systems more diverse, but also increases the difficulty of system control and management. In addition, the wide application of power electronic devices also brings new challenges to the stability and reliability of power systems. The fast switching characteristics of power electronic devices may cause problems such as harmonics and oscillations, affecting the power quality and operation safety of power systems.

[0006] Existing Power System Anomaly Detection Methods

[0007] Power system anomaly detection is a key link to ensure the safe and stable operation of the power system. Traditional anomaly detection methods mainly include rule-based methods, statistical methods, and machine learning methods.

[0008] 1) Rule-based methods

[0009] Rule-based methods rely on human experience and preset rules to judge whether the system is abnormal by setting thresholds or logical conditions. For example, when parameters such as voltage and current exceed the set thresholds, the system will trigger an alarm. The advantage of this method is that it is simple and easy to understand and implement, but the disadvantage is that it is difficult to adapt to complex power grid structures and operating environments, and false alarms or missed alarms are likely to occur.

[0010] 2) Statistical methods

[0011] Statistical methods perform statistical analysis on historical data to establish a statistical model of the normal operating state, and then compare the real-time data with the model to judge whether there is an anomaly. Commonly used statistical methods include:

[0012] a) Parameter estimation methods: For example, Kalman filtering, particle filtering, etc., which estimate the system state to judge whether there is an anomaly.

[0013] b) Non-parametric methods: For example, kernel density estimation, support vector machines, etc., which estimate the data distribution to judge whether there is an anomaly.

[0014] The advantage of statistical methods is that they do not require preset rules and can automatically learn the distribution characteristics of data, but the disadvantage is that the computational complexity is relatively high and the quality requirements for data are relatively high.

[0015] 3) Machine learning methods

[0016] Machine learning methods learn from historical data to establish an anomaly detection model, and then input the real-time data into the model to judge whether there is an anomaly. Commonly used machine learning methods include:

[0017] a) Supervised learning methods: For example, support vector machines, decision trees, etc., which require a large amount of labeled data for training.

[0018] b) Unsupervised learning methods: For example, clustering algorithms, autoencoders, etc., which do not require labeled data and can automatically learn the characteristics of data.

[0019] The advantage of machine learning methods is that they can handle high-dimensional and non-linear data and have strong adaptability, but the disadvantage is that the quality and quantity requirements for data are relatively high, and the training and optimization of the model require a long time.

[0020] Deficiencies of the prior art

[0021] Although certain progress has been made in existing power system anomaly detection methods, there are still some deficiencies.

[0022] 1) Difficulty in handling multimodal data

[0023] The data of the power system exhibits multimodal characteristics, including electrical quantity data such as voltage, current, and power, as well as non-electrical quantity data such as meteorology and equipment status. Traditional anomaly detection methods often can only handle single-type data and are difficult to comprehensively utilize the information of multimodal data.

[0024] 2) Difficulty in adapting to dynamic changes

[0025] The operating state of the power system has the characteristic of dynamic change. The load, power source, network structure, etc. will all change over time. Traditional anomaly detection methods are often based on static models and are difficult to adapt to the dynamic change environment.

[0026] 3) Difficulty in identifying complex anomalies

[0027] Various complex anomaly situations may occur in the power system, such as multiple device failures and the coupling effect of multiple factors. Traditional anomaly detection methods often can only identify simple anomalies and are difficult to identify complex anomaly situations.

[0028] Applications of deep learning technology in power systems

[0029] As a machine learning technology that has developed rapidly in recent years, deep learning has achieved remarkable results in fields such as image recognition, speech recognition, and natural language processing. With the continuous improvement of the intelligent level of power systems, deep learning technology has gradually been applied to various fields of power systems, including load forecasting, fault diagnosis, and security assessment.

[0030] 1) Load forecasting

[0031] Deep learning models can automatically learn the temporal characteristics and non-linear relationships of load data, improving the accuracy of load forecasting. Commonly used deep learning models include recurrent neural network (RNN), long short-term memory network (LSTM), convolutional neural network (CNN), etc.

[0032] 2) Fault diagnosis

[0033] Deep learning models can automatically extract the characteristics of fault data, identify the fault type and location, and improve the efficiency and accuracy of fault diagnosis. Commonly used deep learning models include convolutional neural network (CNN), autoencoder (AE), etc.

[0034] 3) Security assessment

[0035] Deep learning models can automatically learn the operating status and safety constraints of power systems, evaluate the safety risks of the systems, and improve the accuracy and real-time performance of safety assessments. Commonly used deep learning models include deep neural networks (DNN), graph neural networks (GNN), etc.

[0036] Specific pain points and challenges of the prior art

[0037] Although deep learning technology has achieved preliminary applications in power systems, it still faces some challenges.

[0038] 1) Data acquisition and annotation

[0039] Training deep learning models requires a large amount of data, and the data of power systems are often difficult to acquire and annotate. Especially for fault data, due to its low occurrence probability, it is even more difficult to obtain.

[0040] 2) Interpretability of the model

[0041] Deep learning models often have a high degree of complexity, and their internal mechanisms are difficult to understand, resulting in poor interpretability of the models. This brings difficulties to the debugging and optimization of the models.

[0042] 3) Robustness of the model

[0043] The operating environment of power systems is complex and changeable. Deep learning models need to have strong robustness to adapt to different operating conditions.

[0044] All in all, the prior art faces problems such as high data complexity, high real-time requirements, and low detection accuracy in the abnormal diagnosis of power systems. The present invention precisely aims at these pain points and challenges, and proposes a distributed data abnormal diagnosis method for power systems based on deep neural networks, aiming to improve the intelligent level of power systems, reduce operation and maintenance costs, and ensure the safe and stable operation of power systems.

[0045] Summary of the invention

[0046] The purpose of the present invention is to provide a distributed data abnormal diagnosis method for power systems based on deep neural networks to solve the problems of high data complexity, high real-time requirements, and low detection accuracy existing in the prior art. This method realizes intelligent and automatic abnormal detection of multi-modal and massive data in power systems through the comprehensive application of advanced algorithms such as deep neural networks, graph neural networks (GNN), generative adversarial networks (GAN), and reinforcement learning, and optimizes the response strategy in real time.

[0047] To achieve the above purpose, the technical solutions proposed by the present invention are as follows:

[0048] In the first aspect, a distributed data anomaly diagnosis method for power systems based on deep neural networks is proposed, which specifically includes the following steps:

[0049] 1. Acquisition and preprocessing of multi-modal data

[0050] In the distributed data anomaly diagnosis method for power systems, the acquisition of multi-modal data is the first step. This involves establishing stable data acquisition channels with each distributed energy node (such as wind farms, photovoltaic power plants, energy storage devices, etc.) in the power system. Each node generates a large amount of data, including power output, environmental parameters (such as temperature, humidity, wind speed, etc.), equipment status, etc. To ensure the comprehensiveness and accuracy of the data, we need to adopt a variety of sensors and monitoring devices, as well as high-speed data transmission technologies.

[0051] After the data is acquired, the preprocessing step is crucial. First, we perform data cleaning, which includes identifying and filling in missing values, removing outliers, and correcting incorrect data. The purpose of data cleaning is to ensure the quality of the data for subsequent analysis. Next, we perform data normalization, converting data with different dimensions into the same scale for easy analysis and comparison. In addition, we also need to standardize the data to eliminate the bias and dimensional effects in the dataset, laying a solid foundation for subsequent feature extraction and modeling. During the preprocessing process, we must also pay attention to the time synchronization and spatial consistency of the data. Since the data in the power system is often time-series data, the data at different time points need to be precisely aligned to capture the real changes in the system state. At the same time, the data from different nodes and different sources need to be in a unified format for subsequent modeling and analysis.

[0052] 2. Spatiotemporal feature extraction and modeling

[0053] Based on the preprocessed data, we use a graph neural network (GNN) to extract spatiotemporal features and perform modeling. GNN can model the various nodes and their connection relationships in the power system as a graph structure, where the nodes represent distributed energy sources, loads, and substations, and the edges represent the physical connections and power flows between the nodes. In this way, GNN can capture the spatiotemporal dependence relationships between the nodes, providing a basis for subsequent feature extraction.

[0054] During the feature extraction process, we divide the data into three time scales: short-term, medium-term, and long-term, according to the time dimension. Short-term data is used to capture rapid dynamic changes in the system, such as power fluctuations from minutes to hours; medium-term data is used to reflect periodic fluctuations, such as load changes during the day and on weekends; long-term data is used to analyze long-term trends and seasonal changes from months to years. Through the graph convolution operation of the GNN, we can extract spatio-temporal feature vectors that reflect these time-scale changes. These feature vectors not only contain the time series information of individual nodes but also the interaction information between nodes, thus providing a rich information basis for the spatio-temporal modeling of the model.

[0055] 3. Self-Supervised Learning and Contrastive Learning

[0056] After feature extraction, we use self-supervised learning and contrastive learning to further optimize the feature representation. Self-supervised learning encodes and decodes the input data by constructing an autoencoder, minimizing the reconstruction error, and thus extracting the deep features in the data. This method does not require labeled data and can utilize a large amount of unlabeled data to improve the generalization ability of the model.

[0057] Contrastive learning further optimizes the feature representation based on self-supervised learning. It extracts features that can effectively reflect the internal correlations in the data by comparing the similarities between data of different time scales and modalities. Contrastive learning can help the model learn more complex and abstract feature representations, thus improving the accuracy and robustness of anomaly detection.

[0058] 4. Generative Adversarial Networks and Anomaly Detection

[0059] Based on the extracted effective features, we use a generative adversarial network (GAN) to generate a distribution model of normal data. The GAN consists of a generator and a discriminator. The generator is responsible for generating data samples that look like normal data, while the discriminator is responsible for distinguishing between the data generated by the generator and the real data. Through this adversarial process, the generator can generate samples that are increasingly close to the real data.

[0060] In practical applications, we compare the normal data samples generated by the generator with the actual samples to detect the outliers that deviate from the normal distribution. These outliers may be caused by equipment failures, external interferences, or other abnormal situations. After detecting the outliers, we can generate abnormal data samples corresponding to these outliers to further optimize the anomaly detection model and improve its performance in practical scenarios.

[0061] 5. Multi-Task Learning Framework

[0062] To improve the practicality and efficiency of the model, we constructed a multi-task learning framework that can simultaneously perform multiple tasks such as anomaly detection, load forecasting, and fault diagnosis. In this framework, feature representations are shared among different tasks, which means that the features learned by one task can be used for other tasks, thereby enhancing the learning effects of each task.

[0063] Through feature sharing and task collaboration, the multi-task learning framework enables the model to comprehensively understand the operating state of the power system, improving the generalization ability and robustness of the model. In addition, this framework can also reduce the training time of the model because it can utilize the correlation between different tasks to improve the learning efficiency.

[0064] 6. Reinforcement Learning and Optimization of Anomaly Response Strategies

[0065] Based on multi-task learning, we adopt reinforcement learning to design the optimal anomaly response strategy. Reinforcement learning establishes the relationship between states, actions, and rewards, enabling the model to select the best action, that is, the anomaly response strategy, according to the current state of the system during the real-time online learning process.

[0066] The state in the reinforcement learning environment represents the current state of the system, the action corresponds to the anomaly response strategy, and the system response effect is optimized through the reward mechanism. During this process, the model will continuously update and optimize the response strategy to adapt to the dynamic changes of the power system. Through online learning and adaptive adjustment, the reinforcement learning model can ensure that the anomaly diagnosis system can efficiently handle problems in actual scenarios, thereby enhancing the operating stability and security of the power system.

[0067] Through the above steps, the deep neural network diagnosis method of the present invention can automatically and accurately perform anomaly detection and response in a multi-modal complex data environment, significantly enhancing the operating stability and security of the power system.

[0068] Preferably, the processed distributed data is divided according to short-term, medium-term, and long-term time scales, where short-term data is used to capture the rapid dynamic changes in the system, such as power fluctuations; medium-term data reflects periodic fluctuations, such as daily and weekend load changes; and long-term data reflects annual trends and seasonal changes.

[0069] Preferably, the graph structure modeling of the power system nodes includes distributed energy sources, load nodes, and substation nodes. The connections between nodes are based on factors such as physical relationships, power transmission capacity, and geographical location to generate a complete power system network topology structure.

[0070] Preferably, the spatio-temporal dependence relationship is captured through graph convolution and temporal convolution operations of the graph neural network, and the multi-time scale feature vectors of each node are processed through feature fusion to obtain a more accurate spatio-temporal feature representation.

[0071] Preferably, the normal data samples generated by the generative adversarial network are compared with the actual samples to detect abnormal points, and the generator generates abnormal data samples corresponding to the abnormal points to further optimize the anomaly detection model.

[0072] Preferably, the state in the reinforcement learning environment represents the current state of the system, the action corresponds to the abnormal response strategy, and the system response effect is optimized through the reward mechanism to ensure the optimization of the abnormal response strategy. Finally, through online learning and adaptive adjustment, the reinforcement learning model can adapt to the dynamic changes of the power system, ensuring that the anomaly diagnosis system can efficiently handle problems in actual scenarios.

[0073] In a second aspect, the present application provides a power system distributed data anomaly diagnosis device based on a deep neural network, which is constituted as follows:

[0074] Memory: used to store computer programs required to execute the power system distributed data anomaly diagnosis method based on a deep neural network, including but not limited to software implementations of algorithms such as data acquisition, preprocessing, feature extraction, anomaly detection, multi-task learning, and reinforcement learning.

[0075] Processor: responsible for executing the computer programs stored in the memory to implement each step of the anomaly diagnosis method. The processor should have sufficient computing power to process the massive multi-modal data in the power system and be able to complete complex data analysis and model training tasks within a specified time.

[0076] Data interface module: responsible for communicating with distributed energy nodes in the power system to obtain multi-modal data such as power output, meteorological data, and device operating status in real time. The data interface module should support multiple data transmission protocols and have data verification and error handling mechanisms to ensure the stability and reliability of data transmission.

[0077] Output module: used to output the anomaly diagnosis results and optimized response strategies to the monitoring interface or control center of the power system. The output module should have a visualization function and be able to display the anomaly detection and diagnosis results in the form of charts, reports, etc., providing intuitive information support for operators.

[0078] In a third aspect, the present application further provides a readable storage medium, on which a computer program for executing the above-mentioned power system distributed data anomaly diagnosis method based on a deep neural network is stored. The program includes but is not limited to the following functions:

[0079] Data acquisition and preprocessing: The program can automatically obtain multi-modal data from distributed energy nodes in the power system and perform cleaning, normalization, and standardization processing on it to provide a high-quality data basis for subsequent model training and analysis.

[0080] Feature extraction and modeling: The program uses a graph neural network (GNN) to model the power system and extracts feature vectors reflecting the spatio-temporal dependence relationships between nodes through graph convolution operations. These feature vectors can effectively capture the dynamic changes and abnormal patterns of the power system.

[0081] Anomaly detection and multi-task learning: The program optimizes the feature representation through self-supervised learning and contrastive learning, and uses a generative adversarial network (GAN) to generate a distribution model of normal data, thereby achieving high-precision anomaly detection. At the same time, the program also supports a multi-task learning framework, which can perform load forecasting and fault diagnosis while detecting anomalies, improving the overall performance of the system.

[0082] Reinforcement learning and response strategy optimization: The program adopts a reinforcement learning algorithm to design and optimize anomaly response strategies according to the real-time state and anomalies of the power system. Through online learning and adaptive adjustment, the program can ensure that the anomaly diagnosis system can efficiently handle problems in actual scenarios, improving the operational stability and security of the power system.

[0083] The readable storage medium can be a hard disk drive, a solid-state drive, an optical disc, or any other medium capable of storing and reading computer programs. When the program is executed on a processor, it can implement the steps of the above-mentioned distributed data anomaly diagnosis method for power systems based on deep neural networks, providing intelligent support and services for the operation of power systems.

[0084] The beneficial effects of the present invention are as follows:

[0085] Advantages of data preprocessing and fusion: The present invention effectively solves the problems encountered by traditional methods in processing complex data through the preprocessing and fusion of multi-modal data. The preprocessing process includes cleaning, normalization, and standardization processing, which ensure the high quality of data input and provide a solid data foundation for subsequent anomaly detection and model training. In addition, the present invention adopts a spatio-temporal data modeling method with multiple time scales, models each node and its relationships in the power system as a graph structure, and uses a graph neural network (GNN) to capture the spatio-temporal dependence relationships between nodes, thereby realizing the effective extraction of short-term, medium-term, and long-term spatio-temporal features. This innovative data processing method significantly enhances the model's ability to capture complex abnormal patterns and improves the accuracy of anomaly detection. For example, short-term features can capture rapid fluctuations in the power system, medium-term features can reflect periodic fluctuation patterns, and long-term features can reveal seasonal and trend changes. Through such feature extraction, the model can more comprehensively understand the operating state of the power system, thus achieving higher accuracy in anomaly detection.

[0086] Advantages of Self-Supervised Learning and Feature Extraction: By constructing self-supervised learning tasks such as future data point prediction and data reconstruction, the present invention realizes the effective extraction of deep features. Self-supervised learning does not require a large amount of labeled data and can utilize unlabeled data to improve the generalization ability of the model. Through self-supervised learning, the model can learn the internal structure and rules of the data, thereby extracting more meaningful features. In addition, the present invention uses the method of contrast learning to further optimize the feature representation and extract effective features between data of different time scales and modalities. By comparing the similarities between data of different time scales and modalities, contrast learning enables the model to learn deeper associations and patterns. This process significantly improves the feature representation ability of the model and its adaptability to different data patterns. For example, contrast learning can help the model learn the correlations between data at different time scales, thereby better identifying abnormal patterns in anomaly detection.

[0087] Advantages of Multi-Task Learning and Transfer Learning: The present invention constructs a multi-task learning framework that can simultaneously perform related tasks such as anomaly detection, load prediction, and fault diagnosis. Through the shared features between tasks, the present invention not only improves the detection effect but also enhances the data utilization efficiency and the comprehensive performance of the model. The multi-task learning framework enables the model to learn from each other between different tasks through shared feature representations, thereby improving the generalization ability and robustness of the model. In addition, the present invention designs an optimal anomaly response strategy based on reinforcement learning. Through online learning and adaptive adjustment, the anomaly detection model is optimized in real time to adapt to the dynamically changing system environment. Reinforcement learning can design and optimize the anomaly response strategy according to the real-time state and anomaly situation of the power system, thereby ensuring the stable operation of the power system. Through online learning and adaptive adjustment, the model can continuously update and optimize the response strategy according to the actual operation situation, improving the accuracy and response speed of anomaly detection.

[0088] Improvement of System Comprehensive Performance: By constructing a distributed data anomaly diagnosis method for power systems based on deep neural networks, the present invention effectively solves the limitations of traditional methods in dealing with complex data of power systems. This method not only improves the accuracy and response speed of anomaly detection but also realizes the rapid deployment and application of the model through multi-task learning and transfer learning, significantly enhancing the operation efficiency and security of the power system. Through the multi-task learning framework, the model can share features and knowledge between multiple tasks, thereby improving the data utilization efficiency and enhancing the comprehensive performance of the model. In addition, the present invention also uses transfer learning technology to apply the trained model to new power system scenarios, thereby realizing the rapid deployment and application of the model. These advantages enable the present invention to achieve higher performance and efficiency in practical applications, providing intelligent support and services for the operation of power systems.

[0089] Other features and advantages of the present invention will be described in the subsequent specification, and in part will be obvious from the specification, or can be understood by implementing the embodiments of the present invention. The objectives and other advantages of the present invention can be achieved and obtained by the structures specifically pointed out in the written specification, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and thus should not be regarded as a limitation of the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can be obtained based on these drawings.

[0091] Figure 1 It is a schematic flowchart of the method for distributed data anomaly diagnosis of a power system based on a deep neural network described in the embodiments of the present invention;

[0092] Figure 2 It is a schematic structural diagram of the device for distributed data anomaly diagnosis of a power system based on a deep neural network described in the embodiments of the present invention.

[0093] In the figure: 800, a marker receiving device; 801, a processor; 802, a memory; 803, a multimedia component; 804, an I / O interface; 805, a communication component. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0094] The present invention provides a method for distributed data anomaly diagnosis of a power system based on a deep neural network, aiming to achieve comprehensive and intelligent monitoring and diagnosis of the operating state of the power system. To more clearly elaborate the technical solutions of the present invention, this section will describe a complete embodiment in detail, supplement implementation variations and optimization solutions in various application scenarios, provide parameter setting and tuning guidelines, as well as step-by-step instructions for actual deployment and application, and finally give performance evaluation and verification methods.

[0095] 1. Hardware Environment

[0096] The implementation of the present invention depends on a set of high-performance computing platforms. Data acquisition terminals are deployed at each distributed energy node (such as wind farms, photovoltaic power stations, energy storage devices, etc.) to be responsible for real-time collection of operation data. The data center is responsible for receiving, storing, and processing this data, and running deep learning models for anomaly diagnosis and response strategy optimization. The data center usually consists of multiple servers to achieve high availability and load balancing. Considering the massive amount and complexity of power system data, the following hardware configuration is recommended:

[0097] a) Data acquisition terminal:

[0098] Processor: ARM Cortex-A53 quad-core or above

[0099] Memory: 4GB RAM

[0100] Storage: 32GB Flash

[0101] Communication interface: Support power industry standard protocols such as Modbus, IEC 61850, etc.

[0102] b) Data center server:

[0103] Processor: Intel Xeon Gold 6248R or above

[0104] Memory: 256GB DDR4 ECC RAM

[0105] Storage: 1TB NVMe SSD + 10TB SATA HDD (RAID 5)

[0106] GPU: NVIDIA Tesla V100 (32GB) or equivalent performance GPU

[0107] Network: 10Gbps Ethernet

[0108] 2. Software architecture

[0109] The software architecture of the present invention mainly includes a data acquisition module, a data preprocessing module, a model training module, an anomaly detection module, a multi-task learning module, and a response strategy optimization module. The specific functions of each module are as follows:

[0110] a) Data acquisition module: Responsible for collecting multi-modal data from various distributed energy nodes, including power output, meteorological data, equipment operation status, etc. This module needs to support multiple data transmission protocols and have data verification and error handling mechanisms to ensure the stability and reliability of data transmission.

[0111] b) Data preprocessing module: Responsible for cleaning, normalizing, and standardizing the collected data. Data cleaning includes identifying and filling missing values, removing outliers, and correcting incorrect data. Data normalization converts data with different dimensions into the same scale for easy analysis and comparison. Data standardization eliminates the bias and dimensionality effects in the dataset.

[0112] c) Model training module: Responsible for training deep learning models, including graph neural networks (GNNs), autoencoders, generative adversarial networks (GANs), and reinforcement learning networks, etc. This module needs to support multiple deep learning frameworks, such as TensorFlow, PyTorch, etc., and provide model optimization and hyperparameter tuning tools to improve the performance of the model.

[0113] d) Anomaly Detection Module: Responsible for using the trained deep learning model for anomaly detection. This module receives real-time data and inputs it into the model to identify potential anomaly patterns. The detected anomaly information will be presented to the operators in the form of charts, reports, etc.

[0114] e) Multi-Task Learning Module: Responsible for constructing a multi-task learning framework to perform related tasks such as anomaly detection, load forecasting, and fault diagnosis simultaneously. This module improves the detection effect and enhances the comprehensive performance of the model through shared features between tasks.

[0115] f) Response Strategy Optimization Module: Responsible for designing the optimal anomaly response strategy based on reinforcement learning. This module establishes the relationship between states, actions, and rewards, enabling the model to select the best action, i.e., the anomaly response strategy, according to the current state of the system during real-time online learning.

[0116] 3. Algorithm Implementation

[0117] The core algorithms of the present invention include Graph Neural Network (GNN), Autoencoder, Generative Adversarial Network (GAN), and Reinforcement Learning Algorithm. The implementation details of these algorithms are described separately below:

[0118] a) Graph Neural Network (GNN):

[0119] The graph neural network is used to extract the spatio-temporal features of the power system. Specifically, when implementing, the various nodes and their connection relationships of the power system can be modeled as a graph structure, where the nodes represent distributed energy sources, loads, and substations, and the edges represent the physical connections and power flows between the nodes. Then, graph convolution operations are used to extract the spatio-temporal dependence relationships between the nodes.

[0120] The formula for graph convolution operation is as follows:

[0121] H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W (l) )

[0122] Where, H (l) represents the node feature matrix of the l-th layer, A represents the adjacency matrix of the power system, D represents the degree matrix, W (l) represents the weight matrix of the l-th layer, and σ represents the activation function.

[0123] b) Autoencoder:

[0124] Autoencoders are used to extract deep features of data. In specific implementation, a neural network consisting of an encoder and a decoder can be constructed. The encoder compresses the input data into a low-dimensional feature vector, and the decoder reconstructs the feature vector into the original data. By minimizing the reconstruction error, the deep features in the data can be extracted.

[0125] The objective function of the autoencoder is as follows:

[0126] L AE = ||X - Dec(Enc(X))|| 2

[0127] where X represents the input data, Enc represents the encoder, and Dec represents the decoder.

[0128] c) Generative Adversarial Network (GAN):

[0129] Generative Adversarial Networks are used to generate a distribution model of normal data and detect outliers. A GAN consists of a generator and a discriminator. The generator is responsible for generating new samples similar to the normal data distribution, and the discriminator is responsible for distinguishing between the data generated by the generator and the real data.

[0130] The objective function of the generator is as follows:

[0131] min G max D V(D,G) = E x～pdata(x) [log D(x)] + E z～pz(z) [log(1 - D(G(z)))]

[0132] where D represents the discriminator, G represents the generator, x represents the real data, and z represents the random noise.

[0133] d) Reinforcement learning algorithms:

[0134] Reinforcement learning algorithms are used to design optimal anomaly response strategies. In specific implementation, algorithms such as Q-learning or Deep Deterministic Policy Gradient (DDPG) can be adopted.

[0135] The goal of reinforcement learning is to maximize the cumulative reward:

[0136] R = ∑ t＝0 ∞ γ t r t

[0137] where γ represents the discount factor, and r t represents the reward at the t-th step.

[0138] 4. Implementation variations and optimization schemes in multiple application scenarios

[0139] The present invention is applicable to power systems of various scales and types, including large-scale power systems, microgrids, smart distribution grids, and new energy power generation, etc. For different application scenarios, the implementation scheme can be correspondingly deformed and optimized:

[0140] a) Large-scale power system:

[0141] In a large-scale power system, the data volume is huge and the network topology is complex. To improve the efficiency of the algorithm, a distributed computing framework such as Spark, Hadoop, etc. can be adopted to process the data in parallel. In addition, the network structure of the GNN can be optimized to reduce the number of parameters and the computational complexity.

[0142] b) Microgrid:

[0143] In a microgrid, the data volume is relatively small, but the real-time requirement is high. To meet the real-time requirement, lightweight deep learning models such as MobileNet, ShuffleNet, etc. can be adopted. In addition, the algorithm can be pruned and quantized to reduce the storage space and computational amount of the model.

[0144] c) Smart distribution grid:

[0145] In a smart distribution grid, there are various types of devices and diverse data types. To improve the robustness of the algorithm, multi-modal data fusion technology can be adopted to effectively integrate different types of data. In addition, domain knowledge such as power system operation regulations, equipment maintenance manuals, etc. can be introduced to improve the accuracy of anomaly detection.

[0146] d) New energy power generation:

[0147] In new energy power generation, the data fluctuates greatly and the prediction accuracy requirement is high. To improve the prediction accuracy, time series prediction models such as LSTM, Transformer, etc. can be adopted. In addition, meteorological data such as wind speed, light intensity, etc. can be combined to improve the prediction accuracy.

[0148] 5. Parameter setting and tuning guidelines

[0149] The performance of the present invention is affected by various parameters, such as the number of network layers of the GNN, the dimension of the hidden layer of the autoencoder, the learning rate of the GAN, the discount factor of reinforcement learning, etc. To obtain the best performance, these parameters need to be reasonably set and tuned. Some common parameter tuning methods and experiences are provided below:

[0150] a) The number of network layers of the GNN:

[0151] The number of network layers in the GNN determines the model's ability to perceive the graph structure. Generally speaking, the more network layers there are, the stronger the model's ability to perceive the graph structure, but overfitting is also likely to occur. It is recommended to start with a smaller number of network layers, gradually increase it, and evaluate the model's performance through the validation set.

[0152] b) Dimension of the hidden layer of the autoencoder:

[0153] The dimension of the hidden layer of the autoencoder determines the model's ability to compress data features. The smaller the dimension of the hidden layer, the stronger the model's ability to compress data features, but information is also likely to be lost. It is recommended to start with a larger dimension of the hidden layer, gradually decrease it, and evaluate the model's performance through the validation set.

[0154] c) Learning rate of the GAN:

[0155] The learning rate of the GAN determines the update speed of the model parameters. If the learning rate is too large, the model is likely to be unstable; if the learning rate is too small, the model is likely to converge slowly. It is recommended to use the Adam optimizer and set the learning rate to 0.0002.

[0156] d) Discount factor of reinforcement learning:

[0157] The discount factor of reinforcement learning determines the degree to which the model values future rewards. The larger the discount factor, the more the model values future rewards, but the model is also likely to be unstable. It is recommended to set the discount factor to 0.99.

[0158] 6. Instructions on the steps of actual deployment and application

[0159] The actual deployment and application of the present invention mainly include the following steps:

[0160] a) Data preparation: Collect historical data of the power system, including power output, meteorological data, equipment operation status, etc. Clean, normalize, and standardize the data.

[0161] b) Model training: Use the processed data to train deep learning models, including GNN, autoencoder, GAN, and reinforcement learning network, etc.

[0162] c) Model deployment: Deploy the trained model to the data center server.

[0163] d) Real-time monitoring: Collect data of the power system in real time and input it into the model to identify potential abnormal patterns.

[0164] e) Abnormal response: According to the detected abnormal information, take corresponding countermeasures, such as adjusting the generator output, switching lines, etc.

[0165] 7. Performance evaluation and verification methods

[0166] To evaluate the performance of the present invention, the following metrics can be adopted:

[0167] a) Abnormal detection accuracy: It refers to the proportion of correctly detected abnormal samples in the total abnormal samples.

[0168] b) Fault power outage time: It refers to the time from the occurrence of a fault to the restoration of power supply.

[0169] c) Operation and maintenance cost: It refers to the expenses required for the operation and maintenance of the power system.

[0170] The superiority of the present invention can be verified by comparing it with traditional abnormal detection methods. For example, experiments can be conducted on actual power system data to compare the performance of the present invention with methods based on thresholds, statistics, and machine learning.

[0171] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0172] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or replacements, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A distributed data anomaly diagnosis method for power systems based on deep neural networks, characterized in that, It includes the following steps: S1: Multimodal data acquisition and preprocessing: Acquire multimodal data of each distributed energy node in the power system, and perform data cleaning, data normalization, and data standardization on the multimodal data to obtain processed distributed data. Among them, the multimodal data includes power output, meteorological data, and equipment operation status; the data cleaning includes identifying and filling missing values, removing outliers, and correcting incorrect data; the data normalization is to convert data with different dimensions to the same scale range; the data standardization is to eliminate the bias and dimension influence existing in the dataset; S2: Spatiotemporal feature extraction and modeling: Decompose the processed distributed data into data with three time scales of short-term, medium-term, and long-term. Model the nodes and their relationships in the power system as a graph structure, and use a graph neural network (GNN) to capture the spatiotemporal dependence relationships in the graph structure to obtain spatiotemporal feature vectors with multiple time scales. Among them, the nodes include wind farms, photovoltaic power plants, and energy storage devices; the short-term time scale is from minutes to hours, the medium-term time scale is from days to weeks, and the long-term time scale is from months to years; S3: Self-supervised learning and contrastive learning: According to the spatiotemporal feature vectors with multiple time scales, construct self-supervised learning tasks for pre-training and extract deep features; use the contrastive learning method to extract effective features between different time scales and modal data; S4: Generative adversarial network and outlier detection: Based on the effective features, use a generative adversarial network (GAN) to generate a normal data distribution model, and analyze new normal sample data from the normal data distribution model; by comparing the new normal sample data with the actual samples, detect outliers that deviate from the normal distribution, and generate abnormal data samples for enhancing model training according to the outliers, and summarize the abnormal data samples into a comprehensive dataset; S5: Multitask learning framework: Construct a multitask learning framework according to the comprehensive dataset, and simultaneously perform multitask learning of outlier detection, load forecasting, and fault diagnosis. Through the shared features between tasks, obtain a multitask optimization model; S6: Reinforcement learning and abnormal response strategy optimization: Based on the multitask optimization model, use reinforcement learning to design an optimal abnormal response strategy, and through online learning and adaptive adjustment, optimize the outlier detection model in real time to obtain a finally optimized abnormal response system adapted to dynamic changes, and then complete the task of abnormal diagnosis of distributed data in the power system through the abnormal response system.

2. The distributed data anomaly diagnosis method for power systems based on deep neural networks according to claim 1, characterized in that In step S2, decomposing the processed distributed data into data with three time scales of short-term, medium-term, and long-term includes: Short-term data, used to capture the rapidly changing dynamic information in the power system, and the dynamic information includes instantaneous power fluctuations and instantaneous changes in equipment status; Medium-term data, used to capture the daily and periodic changes in the power system, and the daily and periodic changes include daily load fluctuations and weekend electricity consumption patterns; Long-term data, used to capture the seasonal and annual trends in the power system, and the seasonal and annual trends include seasonal load changes and annual energy consumption patterns.

3. The distributed data anomaly diagnosis method for power systems based on a deep neural network according to claim 1, wherein In step S2, the nodes and their relationships in the power system are modeled as a graph structure, including: Modeling each energy node, load node, and substation node in the power system as a node in the graph; Modeling the physical connections, power flows, and geographical location relationships between nodes as edges in the graph, where the weights of the edges are set according to the power transmission capacity, distance, and transmission efficiency, thereby forming a network graph containing all nodes and edges. The network graph includes the topological structure and operating state of the power system, and its calculation formula is as follows: G = (V, E) In the formula, V is the set of nodes, and E is the set of edges.

4. The distributed data anomaly diagnosis method for power systems based on deep neural networks according to claim 1, wherein In step S2, a GNN is used to capture the spatio-temporal dependence relationships in the graph structure to obtain spatio-temporal feature vectors of multiple time scales, including: Based on three time scales and the graph structure, spatio-temporal convolution operations are performed using a graph neural network. Graph convolution and time convolution processing are performed on the data of each node at different time scales. Combining with time convolution, the dependence relationships of nodes in the time dimension are captured to obtain graph feature vectors containing spatio-temporal features; Feature extraction and fusion processing are performed on the graph feature vectors, including performing feature fusion on the feature vectors of each node at different time scales to obtain the first fusion feature; based on the first fusion feature, the feature vectors of each node at different time scales are integrated to obtain the final spatio-temporal feature vectors of multiple time scales.

5. The distributed data anomaly diagnosis method for power systems based on a deep neural network according to claim 1, characterized in that In step S3, based on the spatio-temporal feature vectors of multiple time scales, a self-supervised learning task is constructed for pre-training to extract deep features. At the same time, a contrast learning method is used to extract effective features between different time scales and modal data, including: According to the spatio-temporal feature vectors of multiple time scales, a self-supervised learning task is constructed, including pre-training using an autoencoder for time scales and modal data, training the encoder and decoder for the feature vectors of each node, reconstructing the input data and minimizing the reconstruction error to obtain deep features; Based on the deep features, using the contrast learning method, learn to compare the feature vectors from different time scales and modal data. For each node, a siamese network is constructed, and the similarity loss between positive and negative samples is calculated to obtain optimized deep features; According to the optimized deep features, the feature vectors of each node at short-term, medium-term, and long-term time scales are integrated. Through feature fusion technology, the effective feature vectors of different time scales and modal data are integrated, and finally, a multi-time-scale feature representation for anomaly detection and prediction tasks is obtained.

6. The distributed data anomaly diagnosis method for power systems based on deep neural networks according to claim 1, characterized in that In step S4, based on the effective features, a normal data distribution model is generated using a generative adversarial network, and new normal sample data is analyzed from the normal data distribution model; by comparing the new normal sample data with the actual samples, abnormal points deviating from the normal distribution are detected, and abnormal data samples for enhancing model training are generated according to the abnormal points. The abnormal data samples are summarized into a comprehensive data set, including: Train the generator part of the generative adversarial network according to the effective features, where the generator receives random noise as input and generates new samples similar to the actual normal data distribution. During the training process, minimize the difference between the generated samples and the actual samples; Based on the trained generator, sample new normal samples from the noise distribution, compare the generated new samples with the actual normal samples, and use anomaly detection algorithms to detect the outliers deviating from the normal distribution; According to the detected outliers, use the generator of the generative adversarial network to generate abnormal data samples, where the abnormal data samples are abnormal noise inputs, and at the same time summarize the abnormal data samples into a comprehensive data set.

7. The distributed data anomaly diagnosis method for a power system based on a deep neural network according to claim 1, characterized in that In step S5, construct a multi-task learning framework according to the comprehensive data set, and simultaneously perform multi-task learning of anomaly detection, load prediction, and fault diagnosis. Through the shared features between tasks, obtain a multi-task optimization model, including: Based on the comprehensive data set, design a multi-task learning framework, and simultaneously train the anomaly detection, load prediction, and fault diagnosis tasks, where the framework is used to improve the overall performance by sharing the underlying feature extractor and utilizing the correlation between different tasks; Use the network structure to extract features from the input comprehensive data set. In the multi-task learning framework of the neural network, extract shared features by sharing hidden layers or shared convolution kernels, where the network structure includes convolutional neural networks, recurrent neural networks, or transformers; Train the multi-task learning of anomaly detection, load prediction, and fault diagnosis according to the loss function, and its calculation formula is as follows: L MTL = ∑ i＝1 T λ i L i where λ i is the task weight coefficient and L is the overall loss function; After multi-task learning training, based on the shared features, obtain a comprehensive multi-task optimization model, where the multi-task optimization model is used to process the anomaly detection, load prediction, and fault diagnosis tasks at the same time.

8. The distributed data anomaly diagnosis method for power systems based on a deep neural network according to claim 1, characterized in that In step S6, based on the multi-task optimization model, use reinforcement learning to design the optimal anomaly response strategy, and through online learning and adaptive adjustment, optimize the anomaly detection model in real time to obtain the final optimized anomaly response system adapted to dynamic changes, and then complete the task of distributed data anomaly diagnosis of the power system through the anomaly response system, including: Learn the multi-task optimization model according to the anomaly response strategy designed by reinforcement learning. In the reinforcement learning environment, define the state s as the current system state, the action a as the response strategy, and the reward r as the effect after the system response. Use the Q-learning algorithm to maximize the cumulative reward through optimizing the strategy to obtain the optimal anomaly response strategy; Based on the optimal anomaly response strategy, through real-time online learning, continuously update the reinforcement learning model to adapt to the dynamic changes of the power system, and use the deep deterministic policy gradient algorithm to optimize the policy parameters in the real-time data stream to obtain the finally updated reinforcement learning model; According to the updated reinforcement learning model, perform adaptive adjustment and optimization through the adaptive dynamic programming method, and feedback and adjust the model parameters in real time to obtain the optimized anomaly detection model; Use the optimized anomaly detection model to monitor and respond to the anomalies in the distributed data of the power system in real time, and complete the anomaly diagnosis task, and its calculation formula is as follows: Diagnosis = f(InputData) In the formula, InputData is the input data and Diagnosis is the diagnosis result.

Images