Intelligent power grid edge device adaptive decision-making method and system based on distributed architecture

Abstract

The invention relates to an intelligent power grid edge device self-adaptive decision-making method and system based on a distributed architecture, and the method comprises the following steps: deploying edge computing devices at different positions of a power grid, and achieving the real-time collection and local preprocessing of the operation data of the power grid; performing fault detection and diagnosis at the edge end by using a multi-scale convolutional neural network (MSCNN); an adaptive control strategy is formulated based on an improved multi-agent depth deterministic strategy gradient (MADDPG) algorithm, and by adopting the method provided by the invention, the operation efficiency and reliability of an intelligent power grid are effectively improved, the fault response time is shortened, and the access and management level of distributed energy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart grid, in particular to a smart grid edge device adaptive decision-making method and system based on distributed architecture. BACKGROUND

[0002] With the large number of distributed energy access and the continuous growth of electricity demand, the traditional centralized smart grid control architecture is facing problems such as high data transmission delay, heavy burden on the center server, and difficulty in adapting to real-time changes. Distributed architecture of smart grid edge computing devices emerged as the times require, but there are still deficiencies in the adaptive decision-making of edge devices, which cannot fully exert its advantages and realize the efficient and stable operation of the power grid.

[0003] In the traditional smart grid control architecture, data needs to be transmitted to the cloud or control center for processing and analysis, which causes a series of problems. First, the data transmission delay is high, and the power grid fault has suddenness and rapidity. For example, a short-circuit fault can cause a sudden drop in grid voltage and a sharp increase in current within a few milliseconds. Under the traditional architecture, data is transmitted to the cloud and then returns to the control instruction, with a delay of tens of milliseconds or even hundreds of milliseconds, missing the best control opportunity and making the fault larger, affecting the service life of the equipment and the stability of the power grid. Second, the center server is heavily burdened. With the expansion of the smart grid scale and the increase of monitoring points, the data volume grows exponentially, and the center server needs to process massive data, facing the algorithmic bottleneck, making it difficult to efficiently process and analyze in real time, leading to control instruction update delay, and difficulty in adapting to real-time changes in the power grid. Third, poor adaptability to real-time changes. The traditional centralized control is based on preset rules and models, and it is difficult to quickly adjust the control strategy in the face of real-time changes such as power grid equipment failure, weather change, and load fluctuation. For example, the output power of a photovoltaic power station suddenly changes due to cloud cover, and the traditional system cannot adjust in time, which easily causes local voltage fluctuation to affect the stability of the entire power grid.

[0004] Although distributed architecture has been introduced into smart grid, early edge devices only bear the task of simple data collection and forwarding, lack intelligent decision-making and autonomous control capabilities, and have not fully exerted their potential to improve power grid efficiency and reliability. In view of this, the present application aims to break through the limitations of traditional architecture and proposes a smart grid edge device adaptive decision-making method and system based on distributed architecture, which realizes local data collection and preprocessing, intelligent fault diagnosis and precise control by deploying edge computing devices everywhere in the power grid, to improve the efficiency, reliability and management level of distributed energy of the power grid. SUMMARY

[0005] The present application aims to solve the shortcomings of existing smart grids in edge device adaptive decision-making, and provides a smart grid edge device adaptive decision-making method and system based on a distributed architecture. The main purpose is to improve the operation efficiency and reliability of the smart grid, reduce the fault response time, improve the access and management level of distributed energy, fully utilize the advantages of distributed architecture and edge computing, and realize fast and intelligent adaptive decision-making.

[0006] By deploying edge computing devices in a distributed manner, real-time collection and local preprocessing of power grid operation data are realized, data transmission delay is reduced, data processing speed is accelerated, the power grid can respond to various operating conditions more quickly, and overall operation efficiency is improved. Using a multi-scale convolutional neural network (MSCNN) for fault detection and diagnosis at the edge, the power grid fault can be quickly and accurately detected, and the fault type and location can be determined, providing a basis for timely and effective control measures, thereby enhancing the reliability and fault resistance of the power grid. An improved multi-agent deep deterministic policy gradient (MADDPG) algorithm is used to develop an adaptive control strategy, so that the power grid can quickly adjust the control strategy according to the real-time operating state, isolate faults and restore power supply in time, reduce fault response time, and reduce the impact of faults on the power grid and users. Efficient access and management of distributed energy are realized, operation control of distributed power sources and energy storage systems is optimized, energy utilization efficiency is improved, and widespread application of distributed energy in smart grids is promoted, which promotes the optimization and upgrading of energy structure. The distributed characteristics and data processing capabilities of edge computing devices are embodied, local analysis and decision-making of data are realized, dependence on cloud computing resources is reduced, system communication burden is reduced, and real-time and autonomy of the system are improved.

[0007] To achieve the above purpose, the present application adopts the following technical solutions:

[0008] In a first aspect, embodiments of the present application provide a smart grid edge device adaptive decision-making method and system based on a distributed architecture, comprising the following steps:

[0009] Step 1: Distribute edge computing devices in a distributed manner at different locations of the smart grid, including substations, distribution lines, and user sides. These edge computing devices can collect real-time power grid operation data such as voltage, current, power, and frequency, and perform local preprocessing on the collected data, including data cleaning, filtering, feature extraction, etc., to improve data quality and processing efficiency and ensure data accuracy and reliability. In the data preprocessing stage, the present application uses wavelet packet transform algorithm for signal filtering. Compared with traditional wavelet transform, wavelet packet transform can divide the signal into multiple frequency bands more meticulously, more accurately retain useful signal features, and remove noise interference, which is suitable for the complex and variable signal characteristics in the smart grid and provides a higher quality data basis for subsequent fault detection and diagnosis.

[0010] Step 2: Use the multi-scale convolutional neural network (MSCNN) algorithm to perform fault detection and diagnosis on the preprocessed data in the edge computing device. The network structure of the MSCNN algorithm includes an input layer, multiple convolutional layers of different scales, a pooling layer, and an output layer. Different scale convolutional kernels can extract features of different scales, capture the changes of fault features at different scales, avoid the feature omission problem caused by single scale convolutional kernels, improve the accuracy of fault detection and the precision of fault type identification, especially suitable for complex and diverse fault features in smart grids, can quickly and accurately detect power grid faults and determine fault types and locations, providing reliable basis for subsequent control strategy formulation.

[0011] Step 3: According to the fault diagnosis results and the real-time operation state of the power grid, an improved multi-agent deep deterministic policy gradient (MADDPG) algorithm is used to formulate an adaptive control strategy. The MADDPG algorithm is a reinforcement learning algorithm suitable for multi-agent environments, which has been improved by introducing an attention mechanism, allowing agents to better focus on the interaction with other agents and more accurately estimate the impact of other agents' actions on the current agent's rewards, thereby improving the effectiveness of multi-agent collaborative control. At the same time, the experience replay mechanism is added, which learns by storing and randomly sampling historical experience samples, breaking the correlation between samples and improving the stability and convergence speed of the algorithm. In the smart grid, each agent can control key devices or regions in the power grid, such as distributed power sources, energy storage systems, and key load nodes. Through the collaborative work of agents, precise control of the entire power grid is achieved, improving the stability and reliability of the power grid and enhancing its ability to cope with complex conditions and faults.

[0012] Step 4: Real-time control of power grid equipment based on the formulated control strategy, and real-time monitoring and feedback of control effect. By monitoring the operation state and control effect of power grid equipment, such as whether the voltage and current have returned to normal, whether the fault has been effectively isolated, and whether the power grid frequency is stable, etc., the deficiencies in the control strategy can be discovered in time, and the control strategy can be dynamically adjusted based on feedback information to adapt to the real-time changes of the power grid and ensure the stable operation of the power grid. For example, after control execution, if the voltage in a certain area is still fluctuating, the system will pass this feedback information to the adaptive control strategy module, which will re-evaluate the power grid state and adjust the control strategy based on the MADDPG algorithm, such as further adjusting the output of distributed power sources or optimizing the charge and discharge power distribution of energy storage systems, until the power grid voltage stabilizes within the set range.

[0013] An adaptive decision-making method for smart grid edge devices based on a distributed architecture is as follows:

[0014] Step 1 Distributed data collection and preprocessing

[0015] Edge computing devices, including smart sensors, smart meters, distributed power interface devices, and energy storage system interface devices, are distributed at different locations of the smart grid (substations, distribution lines, customer side, etc.). These devices collect real-time grid operation data, such as voltage (V), current (I), power (P), and frequency (f), and perform local preprocessing on the collected data, including data cleaning, wavelet packet transform filtering, feature extraction, etc., to improve data quality and processing efficiency, ensuring data accuracy and reliability. The preprocessed data is stored locally on the edge computing device, providing a foundation for subsequent analysis. During data cleaning, the normal voltage range is set to [V min , V max ], the normal current range is set to [I min , I max ], the normal power range is set to [P min, P max ], and the normal frequency range is set to [f min , f max ]. Data outside the range is considered invalid and is removed. The wavelet packet transform filtering formula is:

[0016]

[0017] where ψ w,b (t) is the wavelet basis function, w is the scale factor, b is the translation factor, and X(w, b) is the wavelet packet transform coefficient. By selecting an appropriate threshold, the wavelet packet transform coefficients are thresholded to remove noise and achieve signal filtering. Taking voltage amplitude extraction as an example, the feature extraction calculation formula is:

[0018]

[0019] where N is the number of sampling points, and Vi is the voltage value of the i-th sampling point.

[0020] Step 2 Fault detection and diagnosis based on MSCNN

[0021] A multi-scale convolutional neural network (MSCNN) algorithm is deployed in the edge computing device, which includes an input layer, multiple convolutional layers of different scales, pooling layers, and an output layer. The input data is where H is the height, W is the width, and C is the number of channels. Different scale convolution kernels can extract features of different scales and capture the changes of fault features at different scales. Let the convolution kernel size be k x k and the step size be s, then the size of the feature map after convolution operation is:

[0022] (H-k) / s+1 and (W-k) / s+1

[0023] Different scales of convolution kernels correspond to different k values, so as to extract features of different scales. The preprocessed data is used to train MSCNN, so that it can identify data features in normal operation state and fault state. The training process updates the network parameters θ by minimizing the loss function L, which can use cross-entropy loss:

[0024]

[0025] Where N is the number of samples, yi is the true label, is the predicted label. Real-time analysis of preprocessed data can quickly detect power grid faults and determine fault type and location. By setting the fault probability threshold τ, when the predicted fault probability P fault > τ, it is determined that a fault has occurred, and the fault location is determined according to the fault location probability distribution output by the network.

[0026] Step 3: Formulate adaptive control strategy based on improved MADDPG

[0027] According to the fault diagnosis results and the real-time operation state of the power grid, an improved multi-agent deep deterministic policy gradient (MADDPG) algorithm is started. Let the agent set be Each agent A i corresponds to an edge computing device. The state s i of each agent A i includes local power grid operation data and state information of adjacent agents, the action a i is the control instruction to the power grid equipment, and the reward function r i is designed according to the power grid operation state, fault condition and control effect, and the formula is:

[0028] r i = α · stability index + β · fault handling index + Y · energy utilization index

[0029] Where α, β, Y are weight coefficients, respectively measuring the contribution of stability, fault handling capability and energy utilization efficiency to the reward. The attention mechanism is introduced to calculate the attention weight w i of agent A j to other agents A ij , and the formula is:

[0030]

[0031] Where Attention(s i , s j ) measures the attention of agent A i to A jThe correlation of states is calculated by a multi-layer perception (MLP). An experience replay mechanism is added to store historical experience samples (s i , a i , r i , s i ′ ), and a small batch of samples is randomly drawn for learning, updating the policy network parameters θ i and the value network parameters φi of the agent. The update formula is:

[0032]

[0033] where α is the learning rate, is the experience replay buffer, is the target value function. The agent formulates an adaptive control strategy based on local information and information of neighboring agents, such as adjusting the output power P source of the distributed power supply, controlling the charge and discharge power P storage of the energy storage system, adjusting the power P load of key load input or cut-off, etc. The update formula of the control strategy is:

[0034]

[0035] where π is the strategy, η is the strategy update step, and J(π) is the strategy performance indicator.

[0036] Step 4 Real-time Control and Feedback Adjustment

[0037] According to the formulated control strategy, the edge computing device performs real-time control on the power grid equipment. The calculation formula of the control command is:

[0038]

[0039] where u(t) is the control command, e(t) is the control error, K p , K i , and K d are the proportional, integral, and derivative gains, respectively. The running state and control effect of the power grid equipment are monitored in real time, including whether the voltage V and current I return to normal, whether the fault is effectively isolated, and whether the power grid frequency f is stable, etc. The quantification formula of the monitoring indicators is:

[0040]

[0041] where f nominal is the rated frequency. According to the monitoring results, the control effect is fed back to the adaptive control strategy module. The feedback information includes the control error e control , the fault indication F indicator , and the energy efficiency E efficiencyEtc. The calculation formula is:

[0042] e control = target state - actual state

[0043]

[0044] The module re-evaluates the grid state based on the MADDPG algorithm and dynamically adjusts the control strategy to adapt to the real-time changes of the grid, ensuring stable operation of the grid. The adjusted control strategy parameter update formula is:

[0045]

[0046] where θ i is the policy network parameter of agent A i , α is the learning rate, and J(θ i ) is the policy performance index.

[0047] An intelligent power grid edge device adaptive decision system based on a distributed architecture consists of:

[0048] Module 1 edge computing device module

[0049] Distributed deployment at different locations in the smart grid for real-time collection of grid operation data and local preprocessing of collected data. Edge computing devices include intelligent sensors, smart meters, distributed power supply interface devices, energy storage system interface devices, etc. Each device is connected through wired or wireless communication networks. Each edge computing device has independent data processing and decision-making capabilities and can independently analyze and process local data. Let the set of edge computing devices be where M is the number of devices, the processing capacity of each device D i is C i , and the storage capacity is S i .

[0050] Module 2 fault detection and diagnosis module

[0051] Based on the multi-scale convolutional neural network (MSCNN) algorithm, the preprocessed data is analyzed to realize fast detection and accurate diagnosis of grid faults, and to determine the fault type and location. The network structure of the MSCNN algorithm includes an input layer, multiple convolutional layers of different scales, a pooling layer, and an output layer. Let the input data be The size of the convolution kernel is k x k, and the step is s. The size of the feature map after convolution operation is:

[0052] (H-k) / s+1 and (W-k) / s+1

[0053] Different scales of convolution kernels correspond to different k values, so as to extract features of different scales. By setting a fault probability threshold T, when the predicted fault probability P fault >τ, it is determined that a fault occurs, and the fault location is determined according to the fault location probability distribution output by the network.

[0054] Module 3 adaptive control strategy module

[0055] According to the fault diagnosis result and the real-time operation state of the power grid, an improved multi-agent deep deterministic policy gradient (MADDPG) algorithm is used to formulate an adaptive control strategy. Let the agent set be Each agent A i corresponds to an edge computing device. The state s i of the agent includes local power grid operation data and state information of adjacent agents, the action a i is a control instruction for the power grid equipment, and the reward function r i is designed according to the operation state of the power grid, the fault condition and the control effect, and the formula is:

[0056] r i =α·stability index+β·fault handling index+Y·energy utilization index, where α, β and Y are weight coefficients. The attention mechanism is introduced to calculate the attention weight W i of the agent A i to other agents A ij , and the formula is:

[0057]

[0058] where Attention(s i , s j ) is calculated by a multi-layer perception (MLP). The experience replay mechanism is added to store historical experience samples (s i , a i , r i , s i ′), and a small batch of samples are randomly extracted for learning, and the policy network parameter θ i and the value network parameter φi of the agent are updated. The update formula is:

[0059]

[0060] where α is the learning rate, is the experience replay buffer, is the target value function. The agent formulates an adaptive control strategy according to local information and information of adjacent agents, such as adjusting the output power P source of the distributed power supply and controlling the charge and discharge power P storage, adjust the key load input or cut off power P load The update formula of the control strategy is:

[0061]

[0062] Where π is the strategy, η is the strategy update step, and J(π) is the strategy performance index.

[0063] Module 4 controls the execution module

[0064] According to the formulated control strategy, real-time control is performed on the power grid equipment to realize stable operation of the power grid. The calculation formula of the control command is:

[0065]

[0066] Where u(t) is the control command, e(t) is the control error, K p , K i , and K d are the proportional, integral, and differential gains, respectively. Real-time monitoring of the operating state and control effect of the power grid equipment is performed, including whether the voltage V and current I have returned to normal, whether the fault has been effectively isolated, and whether the power grid frequency f is stable. The quantification formula of the monitoring indicators is:

[0067]

[0068] Where f nominal is the rated frequency. According to the monitoring results, the control effect is fed back to the adaptive control strategy module. The feedback information includes the control error e control , fault indication F indicator , and energy efficiency E efficiency , etc. The calculation formula is:

[0069] e control = target state - actual state

[0070]

[0071] The module re-evaluates the power grid state based on the MADDPG algorithm and dynamically adjusts the control strategy to adapt to real-time changes in the power grid, ensuring stable operation of the power grid. The updated control strategy parameter update formula is:

[0072]

[0073] Where θ i is the policy network parameter of agent A i , α is the learning rate, and J(θ i ) is the strategy performance index.

[0074] The present application aims to solve the shortcomings of existing smart grids in edge device adaptive decision-making, and proposes a smart grid edge device adaptive decision-making system and method based on a distributed architecture. By distributing edge computing devices, real-time collection and local preprocessing of power grid operation data are realized, fault detection and diagnosis are performed at the edge using the multi-scale convolutional neural network (MSCNN) algorithm, and adaptive control strategies are formulated based on the improved multi-agent deep deterministic policy gradient (MADDPG) algorithm to achieve precise control of power grid equipment. The system and method effectively improve the operation efficiency and reliability of the smart grid, reduce the fault response time, improve the access and management level of distributed energy, fully utilize the advantages of distributed architecture and edge computing, and achieve fast and intelligent adaptive decision-making.

[0075] The beneficial effects of the present application are:

[0076] By distributing edge computing devices at different locations in the power grid, real-time collection and local preprocessing of power grid operation data are realized, and fault detection and diagnosis are performed quickly at the edge using the multi-scale convolutional neural network (MSCNN). Compared with traditional centralized architecture, the delay of data transmission to the cloud and then returning control instructions is avoided, the fault type and location can be quickly determined, the fault response time is shortened from tens of milliseconds or even hundreds of milliseconds in traditional architecture to a few milliseconds, the impact of faults on power grid equipment and users is effectively reduced, and the stability and reliability of the power grid are improved.

[0077] Based on the improved multi-agent deep deterministic policy gradient (MADDPG) algorithm, adaptive control strategies are formulated, multiple agents work cooperatively, and the control strategy is adjusted in real time according to the real-time operation state of the power grid. The edge computing device controls the power grid equipment in real time according to the formulated control strategy, realizes precise regulation of the power grid, quickly restores the stability of the power grid voltage and frequency, reduces power grid fluctuations, improves power supply quality, reduces the risk of power outages, and ensures stable power supply for various power consumption scenarios such as industrial production and residential life.

[0078] Traditional centralized control architecture relies on high-performance central servers to process massive data, consumes a lot of computing resources, and has high cost. The present application utilizes the local data processing and decision-making capabilities of edge computing devices, only performs data preprocessing and preliminary analysis locally, and does not need to upload a large amount of raw data to the cloud, reducing data transmission volume and cloud computing burden, reducing the performance requirements of the central server, effectively saving computing cost and resource consumption, and improving the economic efficiency and scalability of the system.

[0079] Achieving efficient access and fine management of distributed energy, optimizing the operation control of distributed power and energy storage systems, and improving energy utilization efficiency. For example, through real-time monitoring and control, reasonable scheduling of distributed power output, optimization of energy storage system charging and discharging strategy, ensuring the coordinated operation of distributed energy and main grid, promoting clean energy consumption, reducing the dependence on traditional fossil energy, promoting the optimization and upgrading of energy structure, realizing the economic, efficient and green operation of smart grid.

[0080] The distributed architecture allows easy addition or removal of edge computing devices, facilitating flexible adjustment of system configuration according to changes in grid scale and demand. The addition of new devices does not cause significant disturbance to the operation of the existing system, reducing the difficulty and cost of system maintenance and upgrading, improving the adaptability and sustainable development capability of smart grid, and better meeting the changing and expanding needs of future power grid. BRIEF DESCRIPTION OF DRAWINGS

[0081] The drawings accompanying the specification of this application are used to provide further understanding of this application, the illustrative embodiments of this application and their descriptions serve to explain this application, and do not constitute a limitation on this application.

[0082] Figure 1 The flowchart of the adaptive decision-making method of the smart grid edge device based on the distributed architecture of the present application shows the complete process from data acquisition, fault detection and diagnosis, control strategy formulation to control execution, ensuring stable operation and efficient management of the power grid.

[0083] Figure 2 The system architecture diagram of the adaptive decision-making system of the smart grid edge device based on the distributed architecture of the present application includes edge computing device module, fault detection and diagnosis module, adaptive control strategy module and control execution module, each module cooperates with each other to realize adaptive decision-making and control of smart grid.

[0084] Figure 3 The network structure diagram of the multi-scale convolutional neural network (MSCNN) algorithm shows the connection relationship and data flow between the input layer, different scale convolutional layers, pooling layers and output layer, and reflects how the algorithm extracts features of different scales through multi-scale convolutional kernels to realize fault detection and diagnosis.

[0085] Figure 4 The agent interaction and policy update diagram of the improved multi-agent deep deterministic policy gradient (MADDPG) algorithm shows the distribution of multiple agents in the power grid environment, the information interaction mechanism between them and the policy update process based on attention mechanism and experience replay mechanism, reflecting how the algorithm realizes precise control of power grid equipment and multi-agent collaborative work.

[0086] Figure 5 PID control and feedback adjustment diagram, depicting the monitoring process of the real-time monitoring module on the operating state of the power grid equipment (such as voltage, current, frequency, etc.), and marking the calculation formula of the monitoring index (such as recovery rate, fault isolation rate, frequency stability, etc.). Reflect the feedback adjustment mechanism, mark how to dynamically adjust the control strategy according to the monitoring results, and the feedback control process of the adjusted control strategy on the power grid equipment. DETAILED DESCRIPTION

[0087] The intelligent power grid edge device adaptive decision-making method and system based on distributed architecture of the present application will be described in detail below in combination with the drawings and specific embodiments.

[0088] Take an intelligent power grid containing multiple distributed power sources and energy storage systems as an example, the specific scenario is as follows:

[0089] The intelligent power grid covers a certain industrial park, and there are multiple distributed power sources (such as photovoltaic power stations, small wind power stations), energy storage systems (such as lithium battery energy storage) and numerous industrial and commercial users in the park. The power grid is connected to the main power grid through the substation, and multiple distribution lines and user-side monitoring points are set up in the park.

[0090] Edge computing devices are deployed at key nodes of the substation, distribution lines, and each enterprise user side in the industrial park. These devices include intelligent sensors (for collecting voltage, current, power, frequency, etc.), smart meters (for measuring user-side power consumption), distributed power interface devices (for controlling the output power of photovoltaic power stations and wind power stations), energy storage system interface devices (for managing the charging and discharging process of lithium battery energy storage), and edge computing gateways (for coordinating communication and data transmission between devices).

[0091] Fault simulation: In the afternoon, a distribution line in the park was accidentally dug and broken due to external construction, causing a short circuit fault. After the short circuit fault occurred, the edge computing devices near the line responded quickly and collected abnormal data such as a sharp drop in voltage and a sharp rise in current.

[0092] Taking a smart grid containing multiple distributed power sources and energy storage systems as an example, first, edge computing devices are distributed at different locations such as substations, distribution lines, and user sides (step 1). These edge computing devices collect real-time grid operation data such as voltage, current, power, frequency, etc., and perform local preprocessing on the collected data, including data cleaning, filtering, feature extraction, etc. In the data preprocessing stage, the wavelet packet transform algorithm is used for signal filtering, which decomposes the signal into multiple frequency bands, more accurately retains useful signal features, and removes noise interference. For example, for a grid area containing distributed photovoltaic power and wind power, the edge computing device can effectively filter out the output fluctuations of the photovoltaic power caused by weather changes and the high-frequency noise interference of the wind power, and extract the characteristic signals reflecting the main operation state of the grid.

[0093] Then, the multi-scale convolutional neural network (MSCNN) algorithm is used to detect and diagnose faults in the preprocessed data in the edge computing device (step 2). The MSCNN algorithm extracts features of different scales through different scale convolution kernels, capturing the changes of fault features at different scales. For example, in a certain distribution line, the edge computing device detects abnormal fluctuations in voltage and current signals, and the MSCNN algorithm captures transient fault features in the signal through small-scale convolution kernels and extracts steady-state fault features in the signal through large-scale convolution kernels, and after comprehensive analysis, determines that it is a line short-circuit fault, and accurately locates the fault position at a distance of 3.5 kilometers from the substation, providing accurate information for subsequent fault handling.

[0094] Next, based on the fault diagnosis results and the real-time operation state of the grid, an improved multi-agent deep deterministic policy gradient (MADDPG) algorithm is used to develop an adaptive control strategy (step 3). In the smart grid, each agent controls a key device or area in the grid, such as distributed power sources, energy storage systems, key load nodes, etc. By introducing an attention mechanism, the agent can better focus on the interaction with other agents and more accurately estimate the impact of other agents' actions on the current agent's reward. For example, when a short-circuit fault occurs, the agent controlling the energy storage system will evaluate the weight of its charge-discharge action on the stability of the grid based on its distance from the fault area, energy storage state, and grid frequency changes, etc., while combining the experience replay mechanism to learn the optimal charge-discharge power adjustment strategy in similar fault conditions from historical experience. Through the collaborative work of agents, precise control of the entire grid is achieved, such as adjusting the output power of distributed power sources, controlling the charge and discharge of energy storage systems, adjusting the input or removal of key loads, etc., to restore the normal operation of the grid.

[0095] Finally, according to the control strategy, the power grid equipment is controlled in real time, and the control effect is monitored and fed back in real time (step 4). The edge computing device sends control instructions to the power grid equipment according to the control strategy, realizing real-time control of the power grid equipment. At the same time, the running state and control effect of the power grid equipment are monitored in real time, such as whether the voltage and current are restored to normal, whether the fault is effectively isolated, whether the power grid frequency is stable, and other indicators. For example, after the energy storage system adjusts the charging and discharging power according to the control strategy, the edge computing device monitors the output current, voltage of the energy storage system and the voltage change of the corresponding area of the power grid in real time. If it is monitored that the power grid voltage has not completely recovered to stable, the system will feed back this information to the adaptive control strategy module, and the module will re-evaluate the power grid state and adjust the control strategy based on the MADDPG algorithm, such as further adjusting the output of the distributed power supply or optimizing the charging and discharging power distribution of the energy storage system, until the power grid voltage is stable within the set range, ensuring the stable operation of the power grid.

[0096] Each edge computing device pre-processes the collected data locally.

[0097] First, data cleaning is started. The normal range of voltage is set to V min = 220V-10%, V max = 220V+10%, the normal range of current is I min = 0A, I max = 100A, the normal range of power is P min = 0W, P max = 22kW, and the normal range of frequency is f min = 49.5Hz, f max = 50.5Hz. Data outside the range is determined as invalid data and is removed.

[0098] Second, the wavelet packet transform filter is used to decompose the signal x(t) into multiple frequency bands, remove noise interference, and extract useful signal features. The wavelet packet transform formula is:

[0099]

[0100] where ψ w,b (t) is the wavelet basis function, w=2 is the scale factor, b=1 is the translation factor, and X(w,b) is the coefficient after wavelet packet transform. By selecting an appropriate threshold, the threshold of the wavelet packet transform coefficient is processed to remove noise and achieve signal filtering.

[0101] Through feature extraction, key features such as voltage amplitude V amp、 , current waveform feature I feature , and power change trend P trend are extracted. Taking voltage amplitude extraction as an example, the calculation formula is:

[0102]

[0103] where N = 1000 is the number of sampling points, V i is the voltage value of the i-th sampling point. Assuming that the collected voltage data is calculated over a certain period of time, V amp = 210 V.

[0104] Next, based on the MSCNN-based fault detection and diagnosis, a multi-scale convolutional neural network (MSCNN) algorithm is deployed in the edge computing device, which has a network structure including an input layer, multiple convolutional layers of different scales, a pooling layer, and an output layer. Let the input data be where height H = 224, width W = 224, and channel number C = 3.

[0105] Different scale convolution kernels can extract features of different scales and capture the changes of fault features at different scales. Assuming that the convolution kernel size is k = 3 and the step size is s = 1, the size of the feature map after convolution operation is:

[0106] (224-3) / 1+1=222 and (224-3) / 1+1=222

[0107] Different scale convolution kernels correspond to different k values, thereby extracting features of different scales.

[0108] The MSCNN is trained using the preprocessed data to enable it to recognize data features in normal and fault states. The training process updates the network parameters θ by minimizing the loss function L, which can use cross-entropy loss:

[0109]

[0110] where N = 1000 is the number of samples, yi is the true label (0 or 1, indicating normal or fault), is the predicted label. Assuming that after training, the model achieves an accuracy of 98% on the test set.

[0111] Real-time analysis of preprocessed data quickly detects power grid faults and determines the fault type and location. By setting a fault probability threshold T = 0.8, when the predicted fault probability P fault = 0.9 > T, it is determined that a fault has occurred, and according to the fault location probability distribution output by the network, the fault location is determined to be in a certain specific area of the distribution line, for example, 2.5 kilometers away from the substation.

[0112] Finally, based on the improved MADDPG adaptive control strategy, according to the fault diagnosis results and the real-time operation state of the power grid, the improved multi-agent deep deterministic policy gradient (MADDPG) algorithm is started. Set the agent set as Each agent A i corresponds to an edge computing device.

[0113] Each agent A i 's state s i includes local power grid operation data and state information of adjacent agents, such as voltage, current, power, frequency, etc. Action a i is the control instruction for the power grid equipment, such as adjusting the output power of distributed power supply, controlling the charge and discharge of energy storage system, adjusting the input or removal of key load, etc.

[0114] The reward function r i is designed according to the operation state of the power grid, the fault condition and the control effect, and the formula is:

[0115] r i = α · stability index + β · fault handling index + γ · energy utilization index, where α = 0.3, β = 0.3, γ = 0.4 are weight coefficients. Assuming that in a certain period, the stability index is 0.8, the fault handling index is 0.9, and the energy utilization index is 0.85, then the reward value is:

[0116] r i = 0.3 × 0.8 + 0.3 × 0.9 + 0.4 × 0.85 = 0.83

[0117] The attention mechanism is introduced to calculate the attention weight w j of agent A i to other agents A ij , and the formula is:

[0118]

[0119] where Attention(s i , s j ) measures the correlation of the states of agents A i and A j , which is calculated by a multi-layer perception (MLP). Assuming that the state correlation of agent A1 and A2 is the highest, then the value of w 12 is larger.

[0120] The experience replay mechanism is added to store historical experience samples (s i , a i , r i , s i ′), and a small batch of samples are randomly selected for learning to update the policy network parameters θ of the agent.i and the value network parameter φi. The update formula is:

[0121]

[0122] where α = 0.01 is the learning rate, is the experience replay buffer, is the target value function. Assuming after several iterations, the agent's policy network parameters θ i get effective updates, and the control policy performance improves by 15%.

[0123] The agent formulates an adaptive control policy based on local information and information from neighboring agents, such as adjusting the output power P source of distributed power sources, storage controlling the charge and discharge power P load of energy storage systems, old and adjusting the key load input or cut-off power P new , etc. The update formula for the control policy is:

[0124]

[0125] where π is the policy, η = 0.01 is the policy update step size, and J(π) is the policy performance indicator. Assuming the initial policy π old , the power grid frequency stability is 85%, and after updating, the new policy π new , the power grid frequency stability improves to 92%.

[0126] Finally, real-time control and feedback adjustment are performed. According to the formulated control policy, the edge computing device performs real-time control on the power grid equipment:

[0127] First, the control command is calculated through a proportional-integral-derivative (PID) controller, with the formula:

[0128]

[0129] where K p = 0.5, K i = 0.1, and K d = 0.05 are the gain parameters of the PID controller, and e(t) is the control error, such as the difference between the target voltage and the actual voltage. Assuming the target voltage is 220V and the actual voltage is 210V, the initial control error is 10V, and the control command u(t) is calculated according to the formula to adjust the operating state of the power grid equipment.

[0130] Next, the operating state and control effect of the power grid equipment are monitored in real time, including whether the voltage and current have returned to normal, whether the fault has been effectively isolated, whether the power grid frequency is stable, and other indicators. The quantification formula for the monitoring indicators is:

[0131]

[0132] where f nominal = 50 Hz is the rated frequency. Assuming that after the control strategy adjustment, the power grid returns to normal operation time of 180 seconds, the total operation time is 200 seconds, then the recovery rate is:

[0133]

[0134] The number of successful fault isolation is 1, and the total number of faults is 1, so the fault isolation rate is:

[0135]

[0136] The maximum value of the grid frequency after control f max = 50.2 Hz, the minimum value f min = 49.8 Hz, then the frequency stability is:

[0137]

[0138] According to the monitoring results, the control effect is fed back to the adaptive control strategy module. The feedback information includes control error e control , fault indication F indicator , energy efficiency E efficiency , etc. The calculation formula is:

[0139] e control = target state - actual state

[0140] For example, the target voltage is 220V, and the actual voltage after control is 218V, then the control error is:

[0141] e control = 220-218 = 2V

[0142]

[0143] After fault isolation, this value changes from 1 to 0, indicating that the fault has been eliminated.

[0144]

[0145] Assuming that in a certain period, the useful energy output is 1800kWh, and the total energy input is 2000kWh, then the energy efficiency is:

[0146]

[0147] The module re-evaluates the grid state based on the MADDPG algorithm and dynamically adjusts the control strategy to adapt to the real-time changes of the grid, ensuring stable operation of the grid. The updated control strategy parameter update formula is:

[0148]

[0149] where θ i is the policy network parameter of agent A i , α=0.01 is the learning rate, and J(θ i ) is the policy performance indicator. Assuming that after several adjustments, the performance indicator J(θ i ) of the control policy is improved by 10%, indicating that the effectiveness of the control policy is further improved.

[0150] Through the above application examples of specific scenarios, the application can effectively deal with sudden failures in smart grids, improve the operation efficiency and reliability of the power grid, reduce the fault response time, improve the access and management level of distributed energy, fully utilize the advantages of distributed architecture and edge computing, and realize fast and intelligent adaptive decision-making.

[0151] Application fields include but are not limited to the following:

[0152] Smart grid full-link monitoring and control: can be applied in different positions of smart grid substations, distribution lines, and user sides. In substations, the running state of devices such as transformers and circuit breakers can be monitored in real time, and potential faults can be found in time; in distribution lines, the voltage, current, and other parameters of the lines can be monitored to realize fast positioning and isolation of faults; at the user side, devices such as smart meters, distributed power sources, and energy storage systems can be managed to optimize user power consumption experience and energy utilization efficiency.

[0153] Distributed energy access and management: suitable for smart grid scenarios containing multiple distributed power sources (such as photovoltaic power stations and small wind power stations) and energy storage systems (such as lithium battery energy storage). Through the operation control of distributed power sources and energy storage systems, efficient access and fine management of distributed energy can be realized, the output and charge-discharge strategies can be optimized, the coordinated operation with the main grid can be ensured, and clean energy consumption can be promoted.

[0154] Local grid scenarios such as industrial parks: taking a smart grid covering an industrial park as an example, there are multiple distributed power sources, energy storage systems, and numerous industrial and commercial users in the scenario, and through the deployment of edge computing devices, real-time monitoring, fault diagnosis, and adaptive control of the power grid in the park can be realized to ensure stable power supply for various power consumption scenarios in the park.

[0155] Data involved

[0156] Power grid operation basic data: including analog signals such as voltage (V), current (I), power (P), and frequency (f), which are collected by intelligent sensors and are the most basic parameters reflecting the operation state of the power grid.

[0157] Device status data: such as the output power of distributed power interface devices, the charge and discharge status of energy storage system interface devices, etc., to understand the working conditions of various devices in the power grid.

[0158] User electricity consumption data: electricity consumption data measured by smart meters, which can help analyze user electricity consumption behavior and load characteristics.

[0159] Fault feature data: through multi-scale convolutional neural network (MSCNN) algorithm to analyze the preprocessed data, extract key features such as voltage amplitude (V_amp), current waveform feature (I_feature), power trend (P_trend), etc., for fault detection and diagnosis.

[0160] Control strategy related data: including the state of the agent (s_i), action (a_i), reward function (r_i), etc., as well as control instructions (u(t)), control error (e(t)), etc., for formulating and adjusting adaptive control strategies.

[0161] Monitoring and feedback data: such as recovery rate, fault isolation rate, frequency stability, etc. Monitoring indicators, as well as control error (e_control), fault indication (F_indicator), energy efficiency (E_efficiency), etc. Feedback information, for evaluating control effect and dynamically adjusting strategy.

Claims

1. An adaptive decision-making method for smart grid edge devices based on a distributed architecture, characterized in that, Includes the following steps: Step 1: Distribute edge computing devices in different locations of the smart grid to collect grid operation data in real time and perform local preprocessing; Step 2: Use the Multi-Scale Convolutional Neural Network (MSCNN) algorithm to perform fault detection and diagnosis on the preprocessed data in the edge computing device to determine the fault type and location; Step 3: Based on the fault diagnosis results and the real-time operating status of the power grid, formulate an adaptive control strategy using the improved Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm; Step 4: Perform real-time control of power grid equipment according to the established control strategy, monitor and provide feedback on the control effect in real time, and dynamically adjust the control strategy.

2. The adaptive decision-making method for smart grid edge devices based on a distributed architecture according to claim 1, characterized in that: Step 1.1: Distribute edge computing devices, including smart sensors, smart meters, distributed power interface devices, and energy storage system interface devices, in different locations of the smart grid (substations, distribution lines, user side, etc.). Step 1.2: Edge computing devices collect real-time power grid operation data, including voltage V, current I, power P, frequency f, etc. Step 1.3: Perform local preprocessing on the collected data, including data cleaning, wavelet packet transform filtering, feature extraction, etc., to improve data quality and processing efficiency, and ensure data accuracy and reliability. Specifically: Data cleaning removes missing values, outliers, and other invalid data. The normal voltage range is set to [V]. min V max The normal range of current is [I]. min I max The normal power range is [P]. min P max The normal frequency range is [f] min f max Data outside the specified range is considered invalid and removed. Wavelet packet transform filtering is used to decompose the signal x(t) into multiple frequency bands, remove noise interference, and extract useful signal features. The wavelet packet transform formula is: Where, ψ w,b Let (t) be the wavelet basis function, w be the scaling factor, b be the translation factor, and X(w, b) be the coefficients after wavelet packet transform. By selecting an appropriate threshold, thresholding is performed on the coefficients after wavelet packet transform to remove noise and achieve signal filtering. Key features are extracted through feature extraction, such as voltage amplitude V. amp Current waveform characteristics I feature Power variation trend P trend Etc. Taking voltage amplitude extraction as an example, the calculation formula is: Where N is the number of sampling points, V i Let be the voltage value at the i-th sampling point. Step 1.4: The preprocessed data is stored locally on the edge computing device, providing a foundation for subsequent analysis.

3. The adaptive decision-making method for smart grid edge devices based on a distributed architecture according to claim 1, characterized in that: Step 2.1: Deploy a multi-scale convolutional neural network (MSCNN) algorithm in an edge computing device. Its network structure includes an input layer, multiple convolutional layers of different scales, pooling layers, and an output layer. Let the input data be... Where H is the height, W is the width, and C is the number of channels. MSCNN is an improved convolutional neural network (CNN) designed to extract features at different scales through multi-scale convolution operations, thereby improving the ability to identify complex patterns (such as fault features in smart grids). For input data Output feature map of the l-th layer of the convolutional layer Calculated by the following formula: in, k l It is the size of the kernel of the l-th layer. W l These are the convolution kernel weights. b 1 It is a bias term. σ is an activation function, such as ReLU. MSCNN uses multiple convolutional kernel sizes {k1, k2, ..., k}. m Extract features at different scales. The feature maps corresponding to each convolutional kernel are as follows: Then these feature maps are stitched together: Output feature map of pooling layer The calculation method is as follows: Max pooling: Average pooling: The fully connected layer maps the convolutional and pooled features to the classification result. The output layer uses the softmax function to calculate the fault probability. Among them, o c It is the score of the c-th type of fault output by the fully connected layer. Fault location can be predicted by adding a location regression branch to the output layer, using the mean squared error loss function: Where, N pos It is the number of location samples. and y i These are the predicted and actual location coordinates, respectively. IV. Optimization The network is trained using backpropagation and gradient descent, with the cross-entropy loss function: Among them, y i and These are the true label and the predicted probability, respectively. Step 2.2: Multi-scale convolutional layers consist of multiple parallel convolutional layers, each using a different sized kernel. Kernels of different sizes can extract features at different scales, capturing the changes in fault features at different scales. Let the kernel size be k×k and the stride be s, then the size of the feature map after the convolution operation is... and Different scales of convolutional kernels correspond to different k values, thereby extracting features at different scales. Step 2.3: Train the MSCNN using the preprocessed data to enable it to recognize data features under normal operating and fault conditions. The training process updates the network parameters θ by minimizing the loss function L, which can be a cross-entropy loss function. Where N is the number of samples, y i For real labels, For predicted labels. Step 2.4: Real-time analysis of preprocessed data to quickly detect power grid faults and determine fault type and location. This is achieved by setting a fault probability threshold τ, where the predicted fault probability P... fault When the value is greater than τ, a fault is determined, and the fault location is determined based on the fault location probability distribution output by the network.

4. The adaptive decision-making method for smart grid edge devices based on a distributed architecture according to claim 1, characterized in that: Step 3.1: Based on the fault diagnosis results and the real-time operating status of the power grid, initiate the improved Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. Let the set of agents be... Each intelligent agent (Ai) corresponds to one edge computing device. Step 3.2: Each agent A i state s i This includes local power grid operation data and the status information of neighboring intelligent agents, action a i For control commands to power grid equipment, the reward function r i The design is based on the power grid's operating status, fault conditions, and control effectiveness. Specifically: r i = α·Stability index + β·Fault handling index + γ·Energy utilization index Among them, α, β, and γ are weighting coefficients that measure the contribution of stability, fault handling capability, and energy utilization efficiency to the reward, respectively. Step 3.3: Introduce an attention mechanism to calculate agent A. i For other intelligent agents A j Attention weight w ij The formula is: Among them, Attention(s) i s j ) Measuring agent A i and A j The correlation of states is calculated using a multilayer perceptron (MLP). Step 3.4: Add an experience replay mechanism to store historical experience samples (s) i a i r i s i The agent's policy network parameters θ are updated by randomly sampling a small batch of samples for learning. i and value network parameters φ i The updated formula is: Where α is the learning rate. For experience replay buffer, The objective value function. Step 3.5: Based on local information and information from neighboring agents, the agent formulates an adaptive control strategy, such as adjusting the output power P of the distributed power source. source Controlling the charging and discharging power P of the energy storage system storage Adjusting the power input or output of critical loads (P) load The update formula for the control strategy is as follows: Where π is the policy, η is the policy update step size, and J(π) is the policy performance index.

5. The adaptive decision-making method for smart grid edge devices based on a distributed architecture according to claim 1, characterized in that: Step 4.1: Based on the established control strategy, the edge computing device performs real-time control of the power grid equipment. The formula for calculating the control command is: Where u(t) is the control command, e(t) is the control error, and K p K i K d These are proportional, integral, and differential gains, respectively. Step 4.2: Real-time monitoring of the operating status and control effectiveness of power grid equipment, including whether voltage V and current I have returned to normal, whether faults have been effectively isolated, and whether the power grid frequency f is stable. The quantitative formula for the monitoring indicators is: Among them, f nominal This is the rated frequency. Step 4.3: Based on the monitoring results, the control effect is fed back to the adaptive control strategy module. The feedback information includes the control error e. control Fault indication F indicator Energy efficiency E efficiency The calculation formula is as follows: e control = Target state - Actual state Step 4.4: The module reassesses the power grid state based on the MADDPG algorithm and dynamically adjusts the control strategy to adapt to real-time changes in the power grid and ensure stable grid operation. The updated control strategy parameters are updated using the following formula: in, and For agent A i The network parameters for the new and old strategies are α, where α is the learning rate, and J(θ) is the network parameter for the new and old strategies. i () is a performance indicator for the strategy.

6. An adaptive decision-making system for smart grid edge devices based on a distributed architecture, characterized in that, include: Edge computing device modules are used for distributed deployment in different locations within the smart grid to collect real-time grid operation data and perform local preprocessing on the collected data. Edge computing devices include smart sensors, smart meters, distributed power supply interface devices, energy storage system interface devices, etc., and are connected via wired or wireless communication networks. Each edge computing device has independent data processing and decision-making capabilities, and can autonomously analyze and process local data. Let the set of edge computing devices be denoted as... Where M is the number of devices, and D is the number of devices. i Its processing power is C i Storage capacity is S i . Fault Detection and Diagnosis Module: This module analyzes preprocessed data using a multi-scale convolutional neural network (MSCNN) algorithm to achieve rapid detection and accurate diagnosis of power grid faults, and to determine the fault type and location. The MSCNN algorithm's network structure includes an input layer, multiple convolutional layers of different scales, pooling layers, and an output layer. Let the input data be... If the kernel size is k×k and the stride is s, then the size of the feature map after the convolution operation is: (Hk) / s+1 and (Wk) / s+1 Different scales of convolutional kernels correspond to different k values, thereby extracting features at different scales. By setting a fault probability threshold τ, when the fault probability P predicted by MSCNN... fault When the value is greater than τ, a fault is determined, and the fault location is determined based on the fault location probability distribution output by the network. The adaptive control strategy module is used to formulate an adaptive control strategy based on fault diagnosis results and the real-time operating status of the power grid, employing an improved Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. Let the set of agents be... Each agent A i This corresponds to one edge computing device. The agent's state s i This includes local power grid operation data and the status information of neighboring intelligent agents, action a i For control commands to power grid equipment, the reward function r i The design is based on the power grid operating status, fault conditions, and control effectiveness, and the formula is: r i = α·Stability index + β·Fault handling index + γ·Energy utilization index Where α, β, and γ are weighting coefficients. Control execution module: Used to perform real-time control of power grid equipment according to the established control strategy to achieve stable operation of the power grid.

7. The adaptive decision-making system for smart grid edge devices based on a distributed architecture according to claim 6, characterized in that: The smart sensors in the edge computing device module are used to collect analog signals such as voltage V, current I, power P, and frequency f from the power grid and convert them into digital signals. Smart meters are used to measure electricity consumption; distributed power source interface devices are used to connect and control the output of distributed power sources; energy storage system interface devices are used to manage the charging and discharging process of energy storage systems; and edge computing gateways are used to coordinate communication and data transmission between these devices. Each edge computing device has an independent data processing unit and storage unit, capable of autonomously completing data preprocessing, feature extraction, and local decision-making tasks. Let edge computing device D... i The data processing latency is T proc,i The communication delay is T comm,i Then the total delay T of the system total for: T totaql =max(T proc,i +T comm,i )i=1,2,...,M Where M represents the number of edge computing devices.

8. The adaptive decision-making system for smart grid edge devices based on a distributed architecture according to claim 6, characterized in that: The multi-scale convolutional neural network (MSCNN) algorithm in the fault detection and diagnosis module runs independently on each edge computing device, using only local data for training and inference, ensuring data real-time performance and privacy. Let the training dataset be... in For the j-th input sample of the i-th edge computing device, For the corresponding tag.

Images