Intelligent application method of power saving technology in power system operation and maintenance

Abstract

This invention discloses an intelligent application method of energy-saving technology in power system operation and maintenance, including the following steps: S1: Real-time operation data is collected through IoT sensors and monitoring devices deployed in the power system; The intelligent application method of energy-saving technology in power system operation and maintenance of this invention significantly improves the energy utilization efficiency of the power system by combining multi-source real-time data acquisition, multi-modal fusion modeling and reinforcement learning adaptive optimization. In the traditional operation and maintenance mode, power supply distribution often relies on manual experience or fixed rules, resulting in excessively high peak load and large transmission loss. This method uses a reinforcement learning algorithm to dynamically generate power supply distribution adjustment strategies with the goal of maximizing cumulative discount rewards, which can reduce peak load and minimize line transmission loss in real time.

Description

Technical Field

[0001] This invention relates to the field of power operation and maintenance technology, and more specifically to a method for the intelligent application of energy-saving technology in power system operation and maintenance. Background Technology

[0002] A power system is a unified whole that consists of power generation, transmission, transformation, distribution, and consumption. It is one of the most important infrastructures in modern society. The main function of the power system is to convert various energy sources such as coal, natural gas, hydropower, nuclear power, wind power, and solar energy into electrical energy, and then transmit it to cities and villages over long distances through high-voltage transmission lines. After being stepped down by substations, it is then delivered to factories, offices, and homes through the distribution network for everyone's use.

[0003] In practice, some problems still exist:

[0004] In current power system operation and maintenance, energy-saving technologies mostly adopt passive management, such as experience-based load balancing or simple threshold control, which cannot achieve dynamic optimization. In recent years, although the concept of smart grids and some AI applications have been introduced, such as demand-side response, fault diagnosis, and predictive maintenance, these technologies often exist in isolation, lacking deep fusion of multi-source data and end-to-end closed-loop optimization. For example, some systems use machine learning for load forecasting or equipment monitoring, but the prediction accuracy is limited by data heterogeneity and real-time requirements, making it difficult to cope with peak load reduction and transmission loss minimization in complex scenarios. At the same time, anomaly detection often relies on single indicator analysis, easily overlooking the risk of cascading failures; knowledge-assisted decision-making still mainly relies on manual queries, lacking intelligent recommendations; adjustment command generation does not fully consider the propagation of influence between nodes, resulting in poor distributed grid balance distribution. In addition, although existing edge computing deployments and photovoltaic integration applications have made progress, real-time response latency and privacy protection issues are prominent, and the self-learning mechanism is insufficient, making it impossible to accumulate optimization rules over a long period of time. Summary of the Invention

[0005] To address this, the present invention provides an intelligent application method for energy-saving technology in power system operation and maintenance, in order to solve the problems of passive, lack of deep integration and closed-loop optimization in existing power system operation and maintenance energy-saving technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The intelligent application method of energy-saving technology in power system operation and maintenance includes the following steps:

[0008] S1: Real-time operating data is collected through IoT sensors and monitoring devices deployed in the power system. The data specifically includes load current and power data, voltage amplitude and phase data, equipment surface temperature and ambient temperature data, equipment operating status, and external environmental factors. The collection frequency is set to at least once per minute.

[0009] S2: Preprocess the collected data, including noise filtering, missing value imputation and standardization, and then perform multimodal fusion to integrate heterogeneous data sources into a unified feature vector to build a dynamic energy consumption model.

[0010] S3: The dynamic energy consumption model is adaptively optimized using a reinforcement learning algorithm to generate a power supply distribution adjustment strategy, which achieves optimization by minimizing peak load and transmission loss.

[0011] S4: Perform fluctuation analysis and potential fault identification on the processed data, and integrate an anomaly detection module;

[0012] S5: Extract entity relationships from historical operation and maintenance records and ERP systems, and combine knowledge graph technology to support intelligent recommendation of optimization strategies;

[0013] S6: Generate adjustment instructions and output them to the power system operation and maintenance control terminal, calculate the propagation of power supply impact between nodes, and realize the balanced distribution of the distributed power grid;

[0014] S7: Monitor and adjust execution results in real time, perform self-learning feedback, update model parameters, and evaluate power-saving effects.

[0015] Furthermore, the multimodal fusion in step S2 specifically includes taking IoT sensor data, historical operation and maintenance records, and weather data as inputs, and extracting spatiotemporal fusion features through a convolutional neural network to improve the accuracy of energy demand forecasting.

[0016] Furthermore, the reinforcement learning algorithm in step S3 uses the following optimization objective function to adaptively adjust the power supply distribution:

[0017]

[0018] in:

[0019] Let be the policy function, representing the mapping for selecting the power adjustment action in a given state;

[0020] Indicates the expected value;

[0021] This is a discount factor used to balance immediate rewards and long-term benefits;

[0022] For instant reward functions;

[0023] and These represent the total power consumption before and after the adjustment;

[0024] These are the weighting coefficients;

[0025] For line transmission loss, For the power grid edge collection, For branch current, Branch resistance;

[0026] By iteratively maximizing the cumulative discount reward, peak load and transmission loss are minimized.

[0027] Furthermore: the anomaly detection module in step S4 constructs a directed graph model based on the device relationship network, where nodes represent power equipment and edges represent the connection relationships and data dependencies between devices;

[0028] By calculating the volatility index of each node as the ratio of the standard deviation to the mean of the data deviation, and calculating the proportion index as the share of the volatility index of a single node in the sum of the volatility indices of all nodes, potential faulty nodes whose volatility index exceeds a preset threshold are identified.

[0029] When a potentially faulty node is identified, the system prioritizes generating power adjustment instructions for that node, including reducing the load on that node or isolating the power supply path to that node.

[0030] Furthermore: the dynamic energy consumption model in step S2 is represented by converting the collected data into a two-dimensional power supply distribution map, where the horizontal axis of the map represents the location of the power grid node and the vertical axis represents the energy consumption level;

[0031] Image processing techniques are applied to divide the distribution map into multiple grid patches. Gray-scale value normalization is performed on each patch to unify the pixel intensity range. Then, edge detection operators are used to extract the patch boundary features, and the sum of the squared differences between the supplied power and the demand power within each patch is calculated as a difference index.

[0032] Based on these differential indicators, a visual interface is generated, including color-coded heat maps and overlaid arrows indicating the direction of energy flow, which helps maintenance personnel to intuitively identify high-consumption areas and assist in decision-making.

[0033] Furthermore: In step S5, entity relationships are extracted from historical maintenance records, maintenance work orders, and ERP data systems, including equipment identifiers as entities, fault types as entities, and maintenance events as entities;

[0034] By using a positional translation embedding model, these entities and relations are mapped to a continuous vector space. The relation path representation is enhanced by minimizing the norm of the head entity vector plus the relation vector minus the tail entity vector.

[0035] A knowledge graph is constructed based on the enhanced path representation, which supports intelligent querying and strategy recommendation of power-saving strategies. The shortest path matching optimization scheme is found through graph traversal algorithm.

[0036] Furthermore: the adjustment instruction in step S6 calculates the propagation of the influence between parent and child nodes during the generation process, where the parent node represents the upper-level power supply node and the child node represents the lower-level load node;

[0037] The influence is quantified by using a weighted coefficient as the product of the inverse of the node distance and the power correlation coefficient. Specifically, the adjustment value is propagated step by step from the parent node by traversing the power grid topology, ensuring that the adjustment received by each child node is multiplied by the cumulative weight on the propagation path. This achieves balanced distribution of the distributed power grid, including uniform load distribution to avoid local overload, and the overall transmission loss is calculated as the sum of the product of the square of the line current and the resistance after simulated propagation.

[0038] Furthermore: Step S7 monitors the adjustment execution results in real time, including collecting energy consumption data, voltage stability indicators, and equipment status changes within one minute after the adjustment;

[0039] The energy-saving effect is calculated by comparing the difference in energy consumption before and after the adjustment as the ratio of the previous value minus the subsequent value divided by the previous value, and the gradient descent algorithm is used to update the model parameters, neural network weights, and biases based on this effect.

[0040] When the energy-saving effect reaches a preset threshold of 10%, the current optimization rule is expressed as a condition and automatically added to the dedicated rule database.

[0041] Furthermore, the method is applicable to photovoltaic power plant scenarios. It integrates image data and sensor data obtained from drone inspections, including images of the photovoltaic panel surface for detecting dust accumulation or damage, and real-time output power data for evaluating power generation efficiency.

[0042] Achieving smart access to renewable energy includes predicting photovoltaic output using regression models based on historical data and weather factors, and matching it with grid load by dynamically allocating excess photovoltaic power to high-demand nodes; energy-saving optimization involves prioritizing the use of photovoltaic power to replace traditional energy sources, and storing or transferring photovoltaic output when it exceeds demand, with the goal of reducing dependence on traditional energy sources to below 30%.

[0043] Furthermore, the method employs edge computing deployment, installing computing modules on edge nodes to localize data acquisition and processing, thereby reducing the transmission latency from the edge to the cloud to within 10ms.

[0044] Real-time processing includes running a simplified version of the dynamic energy consumption model and reinforcement learning algorithm on the node to ensure instant power-saving response in high-load scenarios, and generating adjustment instructions locally and sending them directly to the device.

[0045] The present invention has the following advantages:

[0046] 1. The intelligent application method of the energy-saving technology in power system operation and maintenance described in this invention significantly improves the energy utilization efficiency of the power system by combining multi-source real-time data acquisition, multi-modal fusion modeling, and reinforcement learning adaptive optimization. In the traditional operation and maintenance mode, power supply distribution often relies on manual experience or fixed rules, resulting in excessively high peak loads and large transmission losses. This method uses a reinforcement learning algorithm to dynamically generate power supply distribution adjustment strategies with the goal of maximizing cumulative discount rewards. It can reduce peak loads in real time and minimize line transmission losses. In practical applications, it can reduce the overall energy consumption of the system by 10%-20%, especially during high-load periods. At the same time, the introduction of multi-modal fusion technology of convolutional neural networks in step S2 greatly improves the accuracy of energy demand prediction, enabling the model to perceive load changes in advance and avoid unnecessary energy waste. This method not only reduces the operating costs of power companies but also extends the service life of equipment and reduces maintenance expenses caused by overload.

[0047] 2. This invention significantly improves the safety and reliability of power system operation and maintenance by integrating an anomaly detection module, intelligent recommendation based on knowledge graphs, and a self-learning feedback mechanism. Step S4, based on fluctuation analysis of the equipment relationship network, can promptly identify potential fault nodes and prioritize adjusting the power supply path before the fault spreads, effectively preventing additional energy losses and system collapse risks caused by cascading faults. Compared with traditional passive fault response, this method achieves proactive prevention, significantly reducing power outage time and energy waste caused by faults. Meanwhile, the knowledge graph constructed in step S5 extracts entity relationships from massive historical operation and maintenance records, supporting intelligent query and strategy recommendation, enabling operation and maintenance personnel to quickly obtain historical optimal solutions and avoid repeated trial and error. The self-learning closed loop in step S7 further accumulates and optimizes rules, and as the running time increases, the system decision accuracy and energy-saving effect continue to improve.

[0048] 3. This invention is particularly applicable to grid-connected renewable energy scenarios such as photovoltaic power plants. It achieves efficient real-time response through edge computing deployment, demonstrating significant environmental benefits and technological foresight. In photovoltaic power plant applications, this method integrates UAV inspection images and sensor data to achieve intelligent prediction and load matching of renewable energy. It prioritizes the use of photovoltaic power to replace traditional energy and dynamically transfers excess power when photovoltaic output is sufficient. The goal is to reduce the dependence on traditional energy to below 30%, effectively reducing fossil fuel consumption and carbon emissions. At the same time, edge computing deployment pushes data processing and model operation down to edge nodes such as substations, reducing transmission latency to within 10ms, ensuring real-time power-saving response in high-load scenarios, and avoiding the latency and privacy risks caused by cloud reliance.

[0049] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0050] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0051] Fig. 1 This is a flowchart illustrating the overall workflow of the intelligent application method of the energy-saving technology of this invention in power system operation and maintenance.

[0052] Fig. 2 This is a flowchart illustrating the intelligent power-saving closed-loop optimization method for power system operation and maintenance, which is the application method of the power-saving technology of the present invention in power system operation and maintenance. Detailed Implementation

[0053] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figs. 1-2As shown, the intelligent application method of energy-saving technology in power system operation and maintenance includes the following steps:

[0055] S1: Real-time operating data is collected through IoT sensors and monitoring devices deployed in the power system. The data specifically includes load current and power data, voltage amplitude and phase data, equipment surface temperature and ambient temperature data, equipment operating status, and external environmental factors. The collection frequency is set to at least once per minute.

[0056] S2: Preprocess the collected data, including noise filtering, missing value imputation and standardization, and then perform multimodal fusion to integrate heterogeneous data sources into a unified feature vector to build a dynamic energy consumption model.

[0057] S3: The dynamic energy consumption model is adaptively optimized using a reinforcement learning algorithm to generate a power supply distribution adjustment strategy, which achieves optimization by minimizing peak load and transmission loss.

[0058] S4: Perform fluctuation analysis and potential fault identification on the processed data, and integrate an anomaly detection module;

[0059] S5: Extract entity relationships from historical operation and maintenance records and ERP systems, and combine knowledge graph technology to support intelligent recommendation of optimization strategies;

[0060] S6: Generate adjustment instructions and output them to the power system operation and maintenance control terminal, calculate the propagation of power supply impact between nodes, and realize the balanced distribution of the distributed power grid;

[0061] S7: Monitor and adjust execution results in real time, perform self-learning feedback, update model parameters, and evaluate power-saving effects.

[0062] In the above, by deploying a large number of IoT sensors and monitoring devices at key nodes of the power system, comprehensive real-time perception of the operating status is achieved. Core electrical parameters such as load current, power, voltage amplitude and phase, equipment surface temperature, ambient temperature, and switch status are collected, as well as external environmental factors such as weather, humidity, and wind speed. High-frequency collection at least once per minute ensures data timeliness. These data serve as the basic input for all subsequent steps, providing multi-dimensional information that accurately reflects the real-time operating status of the system for the dynamic energy consumption model. This ensures that the entire energy-saving optimization process is based on reliable data support, avoiding optimization deviations caused by data lag or missing data.

[0063] The multimodal fusion in step S2 specifically involves taking IoT sensor data, historical operation and maintenance records, and weather data as inputs, and extracting spatiotemporal fusion features through a convolutional neural network to improve the accuracy of energy demand forecasting.

[0064] In step S2, multimodal fusion technology is used to deeply integrate heterogeneous sources such as electrical data collected in real time by IoT sensors, event sequences in historical operation and maintenance records, and weather data. Specifically, a convolutional neural network is used to perform multi-layer convolution and pooling operations on the input data to extract spatiotemporal fusion features, effectively capturing the implicit correlations between different modal data, improving the accuracy and robustness of energy demand forecasting. This fusion method enables the dynamic energy consumption model to more comprehensively reflect the actual operating rules of the power system, providing high-quality feature vectors for subsequent reinforcement learning optimization, thereby improving the accuracy and adaptability of the overall energy-saving strategy.

[0065] The reinforcement learning algorithm in step S3 uses the following optimization objective function to adaptively adjust the power supply distribution:

[0066]

[0067] in:

[0068] Let be the policy function, representing the mapping for selecting the power adjustment action in a given state;

[0069] Indicates the expected value;

[0070] This is a discount factor used to balance immediate rewards and long-term benefits;

[0071] For instant reward functions;

[0072] and These represent the total power consumption before and after the adjustment;

[0073] These are the weighting coefficients;

[0074] For line transmission loss, For the power grid edge collection, For branch current, Branch resistance;

[0075] By iteratively maximizing the cumulative discount reward, peak load and transmission loss are minimized.

[0076] In the above, step S3 uses a reinforcement learning algorithm to continuously optimize the power supply distribution with the goal of maximizing the long-term cumulative discount reward. The policy function adjusts the output power based on the current state, and the immediate reward function comprehensively considers the total power savings before and after the adjustment and the line transmission loss. The loss term is calculated by summing the square of the branch current and the resistance. The discount factor balances the immediate and future benefits. The algorithm iteratively updates the policy by continuously interacting with the environment, enabling the system to adaptively generate the optimal power supply distribution adjustment strategy under complex and ever-changing operating conditions. This achieves effective reduction of peak load and continuous minimization of transmission loss, thereby significantly improving the overall energy utilization efficiency of the power system.

[0077] The anomaly detection module in step S4 constructs a directed graph model based on the device relationship network, where nodes represent power equipment and edges represent the connection relationships and data dependencies between devices.

[0078] By calculating the volatility index of each node as the ratio of the standard deviation to the mean of the data deviation, and calculating the proportion index as the share of the volatility index of a single node in the sum of the volatility indices of all nodes, potential faulty nodes whose volatility index exceeds a preset threshold are identified.

[0079] When a potentially faulty node is identified, the system prioritizes generating power adjustment instructions for that node, including reducing the load on that node or isolating the power supply path to that node.

[0080] In the above, in step S4, the anomaly detection module first constructs a directed graph model based on the physical connection and data dependency of the devices. Then, it calculates the fluctuation index for each node in real time as the ratio of the standard deviation of the data deviation to the mean, and calculates the proportion index to reflect the degree of anomaly of the node in the whole network. When the fluctuation index exceeds the preset threshold, it is immediately identified as a potential fault node. The system then prioritizes generating power supply adjustment instructions for the node, such as reducing the load or isolating the power supply path, so as to actively intervene before the fault spreads and avoid large-scale energy waste and system instability caused by chain reaction.

[0081] The dynamic energy consumption model in step S2 is represented by converting the collected data into a two-dimensional power supply distribution map, where the horizontal axis of the map represents the location of the power grid node and the vertical axis represents the energy consumption level.

[0082] Image processing techniques are applied to divide the distribution map into multiple grid patches. Gray-scale value normalization is performed on each patch to unify the pixel intensity range. Then, edge detection operators are used to extract the patch boundary features, and the sum of the squared differences between the supplied power and the demand power within each patch is calculated as a difference index.

[0083] Based on these differential indicators, a visual interface is generated, including color-coded heat maps and overlaid arrows indicating the direction of energy flow, which helps maintenance personnel to intuitively identify high-consumption areas and assist in decision-making.

[0084] In the above, in step S2, the dynamic energy consumption model maps multi-dimensional operational data into a two-dimensional power supply distribution map. The horizontal axis corresponds to the location of power grid nodes, and the vertical axis corresponds to the energy consumption level. Image processing technology is used to divide the distribution map into grids and perform grayscale normalization to unify the pixel intensity range. Then, edge detection operators are used to extract boundary features, and the sum of the squared differences between the power supplied and the power demanded in each map block is calculated as a difference index. Based on this, a color-coded heat map is generated and an energy flow direction arrow is superimposed to form an intuitive visualization interface. This allows operation and maintenance personnel to quickly locate high-consumption areas and understand energy flow trends, providing a clear visual basis for manual assisted decision-making and automatic system optimization.

[0085] In step S5, entity relationships are extracted from historical maintenance records, maintenance work orders, and ERP data systems, including device identifiers as entities, fault types as entities, and maintenance events as entities.

[0086] By using a positional translation embedding model, these entities and relations are mapped to a continuous vector space. The relation path representation is enhanced by minimizing the norm of the head entity vector plus the relation vector minus the tail entity vector.

[0087] A knowledge graph is constructed based on the enhanced path representation, which supports intelligent querying and strategy recommendation of power-saving strategies. The shortest path matching optimization scheme is found through graph traversal algorithm.

[0088] In the above, step S5 extracts entities such as equipment identifiers, fault types, and maintenance events, as well as their interrelationships, from historical maintenance records, maintenance work orders, and the ERP system. The entities and relationships are mapped to a low-dimensional continuous vector space using a location translation embedding model. The embedding quality is optimized by minimizing the norm of the head entity vector plus the relationship vector minus the tail entity vector, thus enhancing the ability to represent multi-hop relationship paths. The completed knowledge graph supports intelligent querying and strategy recommendation based on the current operating status. The system quickly finds the historically optimal power-saving solution with the shortest path matching through a graph traversal algorithm, thereby significantly improving the intelligence level and decision-making efficiency of optimization strategy selection.

[0089] The adjustment instructions in step S6 calculate the propagation of the influence between parent and child nodes during the generation process, where the parent node represents the upper-level power supply node and the child node represents the lower-level load node.

[0090] The influence is quantified by using a weighted coefficient as the product of the inverse of the node distance and the power correlation coefficient. Specifically, the adjustment value is propagated step by step from the parent node by traversing the power grid topology, ensuring that the adjustment received by each child node is multiplied by the cumulative weight on the propagation path. This achieves balanced distribution of the distributed power grid, including uniform load distribution to avoid local overload, and the overall transmission loss is calculated as the sum of the product of the square of the line current and the resistance after simulated propagation.

[0091] In the above, when generating the adjustment command in step S6, the hierarchical relationship between parent and child nodes in the power grid topology is fully considered. The parent node is the upper-level power supply unit, and the child node is the lower-level load unit. The product of the inverse of the node distance and the power correlation coefficient is used as a weighting coefficient to quantify the influence propagation intensity. By propagating the adjustment value level by level along the topology starting from the parent node, the adjustment amplitude finally received by the child node is matched with the cumulative path weight, so as to achieve uniform load distribution of the distributed power grid and avoid local overload. At the same time, the overall transmission loss is evaluated by recalculating the sum of the square of the line current and the resistance after simulated propagation, so as to ensure that the loss after balanced distribution is kept at the lowest level.

[0092] The S7 step monitors and adjusts the results in real time, including collecting energy consumption data, voltage stability indicators, and equipment status changes within one minute after the adjustment.

[0093] The energy-saving effect is calculated by comparing the difference in energy consumption before and after the adjustment as the ratio of the previous value minus the subsequent value divided by the previous value, and the gradient descent algorithm is used to update the model parameters, neural network weights, and biases based on this effect.

[0094] When the energy-saving effect reaches a preset threshold of 10%, the current optimization rule is expressed as a condition and automatically added to the dedicated rule database.

[0095] In step S7, after each adjustment command is executed, the system immediately collects data on energy consumption, voltage stability, and equipment status changes within the following minute. By comparing the difference in energy consumption before and after the adjustment, the system calculates the energy-saving effect ratio and uses this as a feedback signal to drive the gradient descent algorithm to update the neural network weights and biases. When the energy-saving effect reaches a preset threshold, such as 10%, the current successful optimization conditions and corresponding actions are expressed as rule pairs and automatically stored in a dedicated rule database. This allows the system to directly call the accumulated rules for rapid decision-making in similar scenarios, thereby forming a continuous self-learning closed loop and continuously improving the energy-saving performance and adaptability in long-term operation and maintenance.

[0096] The method is applicable to photovoltaic power plant scenarios. It integrates image data and sensor data obtained by drone inspection, including photovoltaic panel surface images for detecting dust accumulation or damage, and real-time output power data for evaluating power generation efficiency.

[0097] Achieving smart access to renewable energy includes predicting photovoltaic output using regression models based on historical data and weather factors, and matching it with grid load by dynamically allocating excess photovoltaic power to high-demand nodes; energy-saving optimization involves prioritizing the use of photovoltaic power to replace traditional energy sources, and storing or transferring photovoltaic output when it exceeds demand, with the goal of reducing dependence on traditional energy sources to below 30%.

[0098] In the above-mentioned photovoltaic power plant application scenario, the system integrates high-resolution photovoltaic panel surface images obtained by drone inspections with real-time sensor output power data to accurately detect factors affecting power generation efficiency, such as dust accumulation and panel damage. At the same time, it combines historical data and weather factors to build a regression model to predict photovoltaic output, match it with the grid load in real time, dynamically allocate excess photovoltaic power to nodes with high demand, prioritize the use of clean photovoltaic power to replace traditional thermal power, and automatically trigger energy storage or transfer mechanisms when photovoltaic output exceeds immediate demand, so as to reduce the proportion of traditional energy dependence to below 30%, thereby achieving deeper energy saving optimization and carbon emission reduction goals under the condition of high renewable energy penetration.

[0099] The method employs edge computing deployment, installing computing modules on edge nodes to localize data acquisition and processing, thereby reducing the transmission latency from the edge to the cloud to within 10ms.

[0100] Real-time processing includes running a simplified version of the dynamic energy consumption model and reinforcement learning algorithm on the node to ensure instant power-saving response in high-load scenarios, and generating adjustment instructions locally and sending them directly to the device.

[0101] In the above, by deploying computing modules at edge nodes such as substations and smart gateways, data preprocessing, multimodal fusion, simplified dynamic energy consumption models, and reinforcement learning algorithms are executed locally. This significantly reduces the transmission latency from the edge to the cloud to within 10ms, ensuring that the system can generate and directly issue adjustment instructions to control equipment in real time under high load or sudden scenarios. This avoids the impact of latency and network fluctuations caused by cloud computing. At the same time, localized processing reduces the long-distance transmission of raw sensitive data, enhances data privacy protection, and enables the entire power-saving optimization process to maintain high responsiveness and reliability under a distributed architecture.

[0102] Working principle: This invention discloses an intelligent application method of energy-saving technology in power system operation and maintenance. Its working principle is to form a complete closed-loop intelligent control system through seven steps: multi-source data perception, fusion modeling, intelligent optimization, anomaly protection, knowledge assistance, instruction execution and self-learning feedback, so as to realize dynamic energy-saving optimization in the power system operation and maintenance process.

[0103] The method first utilizes IoT sensors and monitoring devices deployed on substations, transmission lines, power distribution equipment, and terminal loads in step S1 to collect high-frequency real-time data on the power system's operating status. The collected data includes load current and power data, voltage amplitude and phase data, equipment surface temperature and ambient temperature data, equipment operating status (such as switch position and relay protection action signals), and external environmental factors (such as weather temperature, humidity, wind speed, and holiday load patterns). The data collection frequency is set to at least once per minute to ensure that the system can quickly respond to sudden load changes, environmental changes, or equipment anomalies, providing a timely and accurate data foundation for subsequent analysis.

[0104] Step S2 preprocesses the collected multi-source heterogeneous data, including noise filtering, missing value imputation, and standardization, to eliminate data quality issues. Subsequently, multimodal fusion is performed to integrate real-time IoT sensor data, historical operation and maintenance records, and weather data into a unified feature vector. This process uses a convolutional neural network to extract spatiotemporal fusion features, which can effectively capture the correlation between different data modalities, thereby constructing a high-precision dynamic energy consumption model. This model can not only predict energy demand in future periods, but also convert real-time data into a two-dimensional power supply distribution map. Through image processing technology, grid division, grayscale normalization, and edge feature extraction are performed to calculate the sum of squared differences between supply and demand, forming a color-coded heat map and a visual interface with energy flow arrows to help operation and maintenance personnel intuitively identify high-energy-consuming areas.

[0105] Step S3 is the core optimization step of the entire method. It uses reinforcement learning algorithm to adaptively optimize the dynamic energy consumption model. To avoid additional energy waste caused by equipment failure, step S4 integrates an anomaly detection module. This module constructs a directed graph model based on the device relationship network. With devices as nodes and connection relationships as edges, it calculates the fluctuation index (the ratio of the standard deviation of the data deviation to the mean) and the proportional index of each node. When the fluctuation index exceeds the preset threshold, it is identified as a potential fault node. The system immediately generates power supply adjustment instructions for the node, such as reducing the load or isolating the power supply path, to prevent the fault from spreading and causing cascading energy losses.

[0106] Step S5 introduces knowledge graph technology to enhance intelligent decision-making. It extracts entities and their relationships, such as equipment, fault types, and maintenance events, from historical operation and maintenance records, operation and maintenance work orders, and ERP systems. It uses a location translation embedding model to map entities and relationships to a continuous vector space and enhances the relationship path representation by minimizing the scoring function. The completed knowledge graph supports intelligent query and strategy recommendation functions, such as quickly retrieving the historical best solution based on the current operating status or recommending matching power-saving strategies through graph traversal algorithms.

[0107] Step S6 is responsible for converting the optimization results into executable instructions. When generating adjustment instructions, the system calculates the propagation of the influence between parent and child nodes. The parent node is the upper-level power supply unit, and the child node is the lower-level load unit. The product of the inverse of the node distance and the power correlation coefficient is used as the weighting coefficient. The adjustment value is propagated level by level through topology traversal to achieve load balancing distribution of the distributed power grid, avoid local overload and reduce overall transmission loss.

[0108] Step S7 forms a closed-loop self-learning mechanism. The system monitors and adjusts the execution results in real time, collects energy consumption, voltage stability and equipment status data in a short period after adjustment, calculates the power saving effect ratio, and updates the model parameters through the gradient descent algorithm. When the power saving effect reaches the preset threshold (e.g., 10%), the current optimization rule is automatically stored in the dedicated rule database in the form of condition-action pairs to achieve knowledge accumulation and long-term adaptability improvement.

[0109] This method constitutes a data-driven intelligent energy-saving closed loop, applicable to traditional power grids and photovoltaic power plant scenarios. Edge computing deployment further ensures real-time performance and reliability, ultimately achieving efficient energy saving and stable operation of the power system.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for the intelligent application of energy-saving technology in power system operation and maintenance, characterized in that, Includes the following steps: S1: Real-time operating data is collected through IoT sensors and monitoring devices deployed in the power system. The data specifically includes load current and power data, voltage amplitude and phase data, equipment surface temperature and ambient temperature data, equipment operating status, and external environmental factors. The collection frequency is set to at least once per minute. S2: Preprocess the collected data, including noise filtering, missing value imputation and standardization, and then perform multimodal fusion to integrate heterogeneous data sources into a unified feature vector to build a dynamic energy consumption model. S3: The dynamic energy consumption model is adaptively optimized using a reinforcement learning algorithm to generate a power supply distribution adjustment strategy, which achieves optimization by minimizing peak load and transmission loss. S4: Perform fluctuation analysis and potential fault identification on the processed data, and integrate an anomaly detection module; S5: Extract entity relationships from historical operation and maintenance records and ERP systems, and combine knowledge graph technology to support intelligent recommendation of optimization strategies; S6: Generate adjustment instructions and output them to the power system operation and maintenance control terminal, calculate the propagation of power supply impact between nodes, and realize the balanced distribution of the distributed power grid; S7: Monitor and adjust execution results in real time, perform self-learning feedback, update model parameters, and evaluate power-saving effects.

2. The intelligent application method of the energy-saving technology in power system operation and maintenance according to claim 1, characterized in that, The multimodal fusion in step S2 specifically includes taking IoT sensor data, historical operation and maintenance records, and weather data as inputs, and extracting spatiotemporal fusion features through a convolutional neural network to improve the accuracy of energy demand forecasting.

3. The intelligent application method of the energy-saving technology in power system operation and maintenance according to claim 1, characterized in that, The reinforcement learning algorithm in step S3 uses the following optimization objective function to adaptively adjust the power supply distribution: in: Let be the policy function, representing the mapping for selecting the power adjustment action in a given state; Indicates the expected value; This is a discount factor used to balance immediate rewards and long-term benefits; For instant reward functions; and These represent the total power consumption before and after the adjustment; These are the weighting coefficients; For line transmission loss, For the power grid edge collection, For branch current, Branch resistance; By iteratively maximizing the cumulative discount reward, peak load and transmission loss are minimized.

4. The intelligent application method of the energy-saving technology in power system operation and maintenance according to claim 1, characterized in that, The anomaly detection module in step S4 constructs a directed graph model based on the device relationship network, where nodes represent power equipment and edges represent the connection relationships and data dependencies between devices. By calculating the volatility index of each node as the ratio of the standard deviation to the mean of the data deviation, and calculating the proportion index as the share of the volatility index of a single node in the sum of the volatility indices of all nodes, potential faulty nodes whose volatility index exceeds a preset threshold are identified. When a potentially faulty node is identified, the system prioritizes generating power adjustment instructions for that node, including reducing the load on that node or isolating the power supply path to that node.

5. The intelligent application method of the energy-saving technology in power system operation and maintenance according to claim 1, characterized in that, The dynamic energy consumption model in step S2 is represented by converting the collected data into a two-dimensional power supply distribution map, where the horizontal axis of the map represents the location of the power grid node and the vertical axis represents the energy consumption level. Image processing techniques are applied to divide the distribution map into multiple grid patches. Gray-scale value normalization is performed on each patch to unify the pixel intensity range. Then, edge detection operators are used to extract the patch boundary features, and the sum of the squared differences between the supplied power and the demand power within each patch is calculated as a difference index. Based on these differential indicators, a visual interface is generated, including color-coded heat maps and overlaid arrows indicating the direction of energy flow, which helps maintenance personnel to intuitively identify high-consumption areas and assist in decision-making.

6. The intelligent application method of the energy-saving technology in power system operation and maintenance according to claim 1, characterized in that, In step S5, entity relationships are extracted from historical maintenance records, maintenance work orders, and ERP data systems, including device identifiers as entities, fault types as entities, and maintenance events as entities. By using a positional translation embedding model, these entities and relations are mapped to a continuous vector space. The relation path representation is enhanced by minimizing the norm of the head entity vector plus the relation vector minus the tail entity vector. A knowledge graph is constructed based on the enhanced path representation, which supports intelligent querying and strategy recommendation of power-saving strategies. The shortest path matching optimization scheme is found through graph traversal algorithm.

7. The intelligent application method of the energy-saving technology in power system operation and maintenance according to claim 1, characterized in that, The adjustment instruction in step S6 calculates the propagation of the influence between parent and child nodes during the generation process, where the parent node represents the upper-level power supply node and the child node represents the lower-level load node. The influence is quantified by using a weighted coefficient as the product of the inverse of the node distance and the power correlation coefficient. Specifically, the adjustment value is propagated step by step from the parent node by traversing the power grid topology, ensuring that the adjustment received by each child node is multiplied by the cumulative weight on the propagation path. This achieves balanced distribution of the distributed power grid, including uniform load distribution to avoid local overload, and the overall transmission loss is calculated as the sum of the product of the square of the line current and the resistance after simulated propagation.

8. The intelligent application method of the energy-saving technology in power system operation and maintenance according to claim 1, characterized in that, The S7 step monitors the adjustment execution results in real time, including collecting energy consumption data, voltage stability indicators and equipment status changes within one minute after the adjustment. The energy-saving effect is calculated by comparing the difference in energy consumption before and after the adjustment as the ratio of the previous value minus the subsequent value divided by the previous value, and the gradient descent algorithm is used to update the model parameters, neural network weights, and biases based on this effect. When the energy-saving effect reaches a preset threshold of 10%, the current optimization rule is expressed as a condition and automatically added to the dedicated rule database.

9. The intelligent application method of the energy-saving technology in power system operation and maintenance according to claim 1, characterized in that, The method is applicable to photovoltaic power plant scenarios. It integrates image data and sensor data obtained by drone inspection, including photovoltaic panel surface images for detecting dust accumulation or damage, and real-time output power data for evaluating power generation efficiency. Achieving smart access to renewable energy includes predicting photovoltaic output using regression models based on historical data and weather factors, and matching it with grid load by dynamically allocating excess photovoltaic power to high-demand nodes; energy-saving optimization involves prioritizing the use of photovoltaic power to replace traditional energy sources, and storing or transferring photovoltaic output when it exceeds demand, with the goal of reducing dependence on traditional energy sources to below 30%.

10. The intelligent application method of the energy-saving technology in power system operation and maintenance according to claim 1, characterized in that, The method employs edge computing deployment, installing computing modules on edge nodes to localize data acquisition and processing, thereby reducing the transmission latency from the edge to the cloud to within 10ms. Real-time processing includes running a simplified version of the dynamic energy consumption model and reinforcement learning algorithm on the node to ensure instant power-saving response in high-load scenarios, and generating adjustment instructions locally and sending them directly to the device.

Images