AI-driven intelligent power monitoring method and system

Abstract

The invention relates to the field of power monitoring, and discloses an AI-driven intelligent power monitoring method and system, and the method comprises the steps: S1, collecting the operation parameters of a multi-energy system in an intelligent park, and the energy consumption data, environment parameters and equipment operation state data of a user side in real time through a multi-type sensor network; s2, preprocessing the collected data by adopting an edge computing node; s3, inputting the preprocessed data into an AI analysis engine, wherein the AI analysis engine comprises a multi-modal deep learning model and a hierarchical reinforcement learning model which are operated in parallel; s4, virtual mapping of the intelligent park integrated energy system is constructed through a digital twin system, and system operation conditions under different energy scheduling strategies are simulated in real time in combination with user energy demand prediction and external environment changes; and S5, when potential energy supply abnormity or equipment failure is detected, triggering a multi-stage early warning mechanism, and generating a decision report.

Description

Technical Field

[0001] This invention relates to the field of power monitoring, specifically to an AI-driven intelligent power monitoring method and system. Background Technology

[0002] In the current field of power monitoring, especially in the scenario of multi-energy system management in smart parks, traditional monitoring methods face many challenges.

[0003] For example, 1. Data acquisition and processing challenges: In smart parks, multiple energy systems such as electricity, heat, and gas operate in parallel with diverse operating parameters. User-side energy consumption data, environmental parameters, and equipment operating status data are also complex and variable. Traditional sensor networks have low acquisition efficiency and poor accuracy, making it difficult to acquire comprehensive and real-time data. Furthermore, different types of data have different formats and frequencies, and there is a lack of unified preprocessing methods, resulting in inconsistent data quality and an inability to provide reliable support for subsequent analysis.

[0004] 2. Insufficient analytical model effectiveness: Existing monitoring systems lack models that can effectively integrate and analyze multi-source heterogeneous data; it is difficult to extract spatiotemporal data and topological features of multi-energy systems, making it difficult to uncover complex relationships between energy nodes; in terms of energy dispatching strategy formulation, it is unable to balance macro cost control and micro user experience, resulting in low energy utilization efficiency and failing to meet the needs of sustainable development in the park.

[0005] 3. Lagging Early Warning and Decision-Making Mechanisms: Early warnings for energy supply anomalies and equipment failures are mostly based on simple threshold judgments, lacking a tiered mechanism and making it difficult to differentiate the severity of anomalies. Anomaly cause analysis relies on manual experience, lacking scientific rigor and accuracy, resulting in slow and untargeted development of response plans. In emergencies, it is impossible to quickly guarantee the energy needs of critical users, impacting the normal operation of the park. Summary of the Invention

[0006] The purpose of this invention is to provide an AI-driven intelligent power monitoring method and system to solve at least one of the above-mentioned technical problems.

[0007] The objective of this invention can be achieved through the following technical solutions: An AI-driven smart power monitoring method includes the following steps: S1. Real-time collection of operating parameters of the multi-energy system consisting of electricity, heat and gas in the smart park, as well as energy consumption data, environmental parameters and equipment operating status data on the user side, through a multi-type sensor network; S2. Use edge computing nodes to preprocess the collected data, including multi-source data fusion, noise filtering, data alignment and outlier removal; S3. Input the preprocessed data into the AI ​​analysis engine, which includes a multimodal deep learning model and a hierarchical reinforcement learning model running in parallel. S4. Construct a virtual mapping of the intelligent park's integrated energy system through a digital twin system, and simulate the system operation under different energy dispatch strategies in real time by combining user energy demand forecasts and changes in the external environment. S5. When a potential energy supply anomaly or equipment failure is detected, a multi-level early warning mechanism is triggered, and a decision report is generated that includes anomaly location, cause analysis, energy dispatch optimization, and equipment maintenance and disposal recommendations.

[0008] As a further technical solution, the multimodal deep learning model adopts a spatiotemporal graph convolutional network to extract and fuse features from the spatiotemporal data and topology of the multi-energy system. The spatiotemporal graph convolutional network introduces an attention mechanism into the graph structure to adaptively learn the spatiotemporal relationships between different energy nodes, thereby enhancing the feature extraction capability of complex energy networks. The hierarchical reinforcement learning model is based on the hierarchical Q-learning algorithm. The upper-layer strategy is used to formulate a macro-allocation plan for energy production and storage, and the lower-layer strategy is used to dynamically adjust the energy consumption of each terminal device. The upper-layer strategy of the hierarchical reinforcement learning model takes the overall energy cost and carbon emission indicators of the park as optimization objectives, while the lower-layer strategy takes the operational stability of terminal equipment and user comfort as optimization objectives. The upper and lower-layer strategies are jointly optimized through a reward function. The hierarchical reinforcement learning model uses an ε-greedy strategy to explore new strategies in the early stages of training, and gradually reduces the exploration rate as training progresses; when new energy equipment is introduced into the park or the energy supply mode changes, the exploration rate is reset to explore again, so as to quickly adapt to system changes.

[0009] As a further technical solution, the multi-type sensor network in step S1 includes: Power monitoring layer: smart meters, harmonic sensors, insulation monitoring devices; Thermal monitoring layer: distributed heat meters, pipeline temperature sensors; Environmental sensing layer: temperature and humidity sensor, light intensity sensor, Concentration detector; Sensors at each layer transmit data in a standardized manner via the OPCUA protocol; The edge computing nodes and the AI ​​analysis engine use a 5G+industrial IoT hybrid network for data transmission. Before transmission, the data is losslessly compressed and encrypted and traced using blockchain technology to ensure efficient and secure data transmission.

[0010] As a further technical solution, the edge computing node adopts a multi-source data fusion algorithm based on deep learning to map different types of data such as electricity, heat and gas to the same feature space; The data fusion algorithm is based on an architecture combining a Long Short-Term Memory (LSTM) network and an autoencoder. First, the LSTM network is used to learn the long-term dependencies of time series data, and then the autoencoder is used to reduce the dimensionality and extract features from the data, thereby achieving efficient fusion of data from different modalities.

[0011] As a further technical solution, the digital twin system combines weather forecast data and user historical energy consumption data, and uses a long short-term memory network (LSTM) to predict users' future energy demand, providing a basis for energy dispatch simulation. The digital twin system uses the particle swarm optimization algorithm (PSO) to quantitatively evaluate different energy dispatch strategies based on the energy dispatch simulation results. By setting objective functions for energy cost, carbon emissions, and user satisfaction, it calculates the comprehensive score of each strategy and selects the optimal energy dispatch scheme.

[0012] As a further technical solution, the multi-level early warning mechanism divides the early warning level according to the degree of impact of the anomaly on the stability of the park's energy supply and user experience, performs probabilistic reasoning on the cause of the anomaly through a Bayesian network, and generates energy dispatch optimization and equipment maintenance plans in combination with an expert knowledge base. The warning levels are divided into three levels: general warning, severe warning, and emergency warning, each corresponding to different abnormal handling procedures. When an emergency warning is triggered, the system automatically activates the backup energy supply system and prioritizes the energy needs of critical users.

[0013] As a further technical solution, the AI ​​analysis engine adopts a federated learning framework during operation. The federated learning framework is used to achieve cross-regional and cross-departmental data collaborative training while ensuring the data privacy of each energy subsystem, so as to dynamically update the multimodal deep learning model and the hierarchical reinforcement learning model.

[0014] An AI-driven smart power monitoring system includes: The data acquisition unit collects in real time the operating parameters of the multi-energy system consisting of electricity, heat and gas in the smart park, as well as energy consumption data, environmental parameters and equipment operating status data on the user side through a multi-type sensor network; The data preprocessing unit uses edge computing nodes to preprocess the collected data, including multi-source data fusion, noise filtering, data alignment, and outlier removal. The data analysis unit inputs the preprocessed data into the AI ​​analysis engine, which includes a multimodal deep learning model and a hierarchical reinforcement learning model running in parallel. The simulation monitoring unit constructs a virtual mapping of the intelligent park's integrated energy system through a digital twin system. Combining user energy demand forecasts and changes in the external environment, it simulates the system operation under different energy dispatch strategies in real time. The early warning and reporting unit triggers a multi-level early warning mechanism when it detects potential energy supply anomalies or equipment failures, and generates a decision report that includes anomaly location, cause analysis, energy dispatch optimization, and equipment maintenance and handling suggestions.

[0015] The beneficial effects of this invention are: (1) Comprehensive data from multiple energy systems and users within the smart park are collected in real time at a specific frequency through a multi-type sensor network. Using edge computing nodes, preprocessing is performed using wavelet transform threshold denoising, timestamp synchronization, and the 3σ principle to improve data quality, provide a reliable basis for subsequent analysis, and ensure the accuracy and integrity of the basic data of the monitoring system; (2) The spatiotemporal graph convolutional network of the multimodal deep learning model, combined with the multi-head attention mechanism, accurately extracts spatiotemporal data and topological features of multi-energy systems, enhancing the understanding of complex energy networks. The hierarchical reinforcement learning model is based on the hierarchical Q-learning algorithm. By dynamically adjusting the reward function, it achieves the coordinated optimization of upper-level macro-energy management and lower-level terminal equipment energy consumption, thereby improving energy utilization efficiency, reducing costs, reducing carbon emissions, and enhancing user satisfaction. (3) The digital twin system combines weather forecasts and users' historical energy consumption data, and uses a Long Short-Term Memory (LSTM) network to accurately predict users' future energy needs. The Particle Swarm Optimization (PSO) algorithm is used to quantitatively evaluate energy dispatch strategies, select the optimal solution, make energy dispatch more scientific and reasonable, effectively reduce energy costs, reduce carbon emissions, and improve the stability of energy supply. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a diagram illustrating the method steps of the present invention; Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0019] Please see Figures 1-2 As shown, this invention is an AI-driven smart power monitoring method, comprising the following steps: S1. Real-time acquisition of operating parameters of multi-energy systems consisting of electricity, heat, and gas within the smart park, as well as energy consumption data, environmental parameters, and equipment operating status data from the user side, through a multi-type sensor network; for example, the acquisition frequency is once per second for electricity parameters, once every 5 minutes for heat and gas parameters, once per minute for environmental parameters, and equipment operating status data is adaptively acquired based on the vibration frequency of key equipment components. S2. Edge computing nodes are used to preprocess the collected data, including multi-source data fusion, noise filtering, data alignment, and outlier removal. Among them, wavelet transform threshold denoising method is used for noise filtering, data alignment is synchronized by timestamp, and outlier removal adopts a statistical method based on the 3σ principle. S3. Input the preprocessed data into the AI ​​analysis engine, which includes a multimodal deep learning model and a hierarchical reinforcement learning model running in parallel. S4. Construct a virtual mapping of the intelligent park's integrated energy system through a digital twin system. Combine user energy demand forecasts and changes in the external environment to simulate the system's operation under different energy dispatch strategies in real time. For example, the simulation time step is set to 15 minutes and calibrated with the actual system operation data once per hour. S5. When a potential energy supply anomaly or equipment failure is detected, a multi-level early warning mechanism is triggered, and a decision report is generated that includes anomaly location, cause analysis, energy scheduling optimization, and equipment maintenance and handling recommendations. For example, anomaly detection uses an unsupervised learning model based on the isolated forest algorithm, and the early warning response time does not exceed 30 seconds.

[0020] In this embodiment, to address the technical problems of low efficiency in collecting operating parameters and user-side related data from multiple energy systems within a smart park, lack of unified preprocessing leading to poor data quality and inability to effectively support subsequent analysis, and untimely detection of energy supply anomalies and equipment failures hindering rapid decision-making, the following measures are taken: First, the data collection frequency of multiple types of sensor networks is clearly defined to ensure the timeliness and effectiveness of data collection. Second, specific data preprocessing methods are employed at edge computing nodes, such as wavelet transform threshold denoising, timestamp synchronization, and the 3σ principle to remove outliers, improving data quality and providing reliable data for subsequent AI analysis. Third, the isolated forest algorithm is used for anomaly detection, ensuring that the early warning response time does not exceed 30 seconds, enabling timely detection of potential problems, buying time for energy scheduling optimization and equipment maintenance, and improving the stability and reliability of the energy system operation.

[0021] The multimodal deep learning model employs a spatiotemporal graph convolutional network to extract and fuse features from the spatiotemporal data and topology of multi-energy systems. The spatiotemporal graph convolutional network introduces an attention mechanism into the graph structure to adaptively learn the spatiotemporal relationships between different energy nodes, thereby enhancing the feature extraction capability of complex energy networks. The hierarchical reinforcement learning model is based on the hierarchical Q-learning algorithm. The upper-layer strategy is used to formulate a macro-allocation plan for energy production and storage, and the lower-layer strategy is used to dynamically adjust the energy consumption of each terminal device. The upper-layer strategy of the hierarchical reinforcement learning model takes the overall energy cost and carbon emission indicators of the park as optimization objectives, while the lower-layer strategy takes the operational stability of terminal equipment and user comfort as optimization objectives. The upper and lower-layer strategies are jointly optimized through a reward function. The hierarchical reinforcement learning model uses an ε-greedy strategy to explore new strategies in the early stages of training, and gradually reduces the exploration rate as training progresses; when new energy equipment is introduced into the park or the energy supply mode changes, the exploration rate is reset to explore again, so as to quickly adapt to system changes.

[0022] In this embodiment, the multimodal deep learning model adopts a spatiotemporal graph convolutional network, which includes three temporal convolutional layers and two spatial convolutional layers. The temporal convolutional layers use a one-dimensional convolutional kernel of size 5, and the spatial convolutional layers use a two-dimensional convolutional kernel of size 3×3. Feature extraction and fusion are performed on the spatiotemporal data and topology of the multi-energy system. The power, heat, and gas systems are abstracted into a graph structure composed of nodes and edges. The node attributes contain the real-time parameters of each energy system, and the edge attributes represent the energy flow relationship between energy nodes. The spatiotemporal graph convolutional network introduces a multi-head attention mechanism into the graph structure, sets up 8 attention heads, adaptively learns the spatiotemporal correlation between different energy nodes, and enhances the feature extraction capability of complex energy networks by calculating the attention weights between nodes. The hierarchical reinforcement learning model is based on the hierarchical Q-learning algorithm. The decision cycle of the upper-level strategy is 1 day, which is used to formulate a macro-allocation plan for energy production and storage, including determining the start-up and shutdown times of each energy production device and the charging and discharging plan of the energy storage device. The decision cycle of the lower-level strategy is 15 minutes, which is used to dynamically adjust the energy consumption of each terminal device, including the power regulation of air conditioning and lighting equipment. The upper-layer strategy of the hierarchical reinforcement learning model optimizes the overall energy cost and carbon emission indicators of the park, with energy cost having a weight of 0.6 and carbon emission having a weight of 0.4. The lower-layer strategy optimizes the operational stability of terminal devices and user comfort, with operational stability having a weight of 0.7 and user comfort having a weight of 0.3. The upper and lower-layer strategies are collaboratively optimized through a reward function, which is dynamically adjusted according to the achievement of the objectives. The expression of the reward function is as follows: ; in, Total reward value; Weighted by energy cost; Energy cost-related reward value; Carbon emission weights; Carbon emission-related reward values; As a weight for equipment operational stability; Reward values ​​related to equipment operational stability; Weighting based on user comfort; Rewards related to user comfort.

[0023] ; The number of energy types, covering electricity, heat, and natural gas; The cost incentive weight for the j-th energy type is set based on the proportion and importance of that energy in the park's energy consumption; An adjustment coefficient is set based on the rate of change of energy cost for the j-th type, when the rate of change of energy cost in a certain period... hour, , Additional reward coefficient; otherwise ; To set a threshold; The expected cost of the j-th energy source can be determined based on historical data and market price forecasts. The actual cost of the j-th energy source is obtained through statistics from energy metering equipment; ; This is the carbon emission incentive coefficient, used to adjust the sensitivity of carbon emission incentives; The carbon emission target set for the park; The actual carbon emissions of the park are estimated through carbon emission monitoring equipment or based on energy consumption data. ; This is the reward coefficient for equipment operational stability. The number of samples for the equipment operating parameters; the first Sampled values ​​of operating parameters of each device; This represents the average value of the equipment's operating parameters; This represents the standard deviation of the equipment's operating parameters; the formula calculates the degree of fluctuation in the equipment's operating parameters, with smaller fluctuations resulting in higher reward values. ; The number of dimensions for evaluating user comfort, such as temperature, humidity, and light intensity; For the first The weights of each comfort evaluation indicator are set according to user preferences; For the first The scores for each comfort evaluation indicator are quantified by constructing a user comfort evaluation system and combining sensor monitoring data and user feedback data. For example, each parameter is compared with its respective reference area; if it falls within the reference area, the comparison score is 100; otherwise, the comparison score is 20. User feedback is then incorporated into the evaluation, ranging from 1 to 5 points. Finally, the score is calculated using the formula: =0.7 * comparison score + 0.3 * user feedback score * 20.

[0024] Energy cost incentives Calculate and sum the cost rewards for each energy type separately, and give additional rewards considering cost change trends; carbon emission rewards. The reward is calculated based on the difference between actual carbon emissions and the target value; the larger the difference, the higher the reward. (Equipment operation stability reward) By evaluating the fluctuations in equipment operating parameters, it was found that smaller fluctuations result in higher rewards; user comfort rewards... The total reward is calculated by combining comfort scores from multiple dimensions; the total reward value is obtained by multiplying each reward by its corresponding weight and then summing the results. This function is used to update the Q-value table and guide the optimization of subsequent energy dispatch strategies. It can more comprehensively and dynamically reflect the operation goals of the intelligent park's integrated energy system, adapt to complex and ever-changing operating environments, improve the effect of hierarchical reinforcement learning models in energy dispatch strategy optimization, and achieve efficient, economical, and green operation of the energy system.

[0025] The hierarchical reinforcement learning model sets ε to 0.8 in the early stage of training and uses an ε-greedy strategy to explore new strategies. As training progresses, the exploration rate is gradually reduced at a rate of ε = ε - 0.001. When new energy equipment is introduced into the park or the energy supply mode changes, ε is reset to 0.8 and the exploration is restarted to quickly adapt to system changes.

[0026] Through the above technical solutions, the spatiotemporal graph convolutional network, combined with a multi-head attention mechanism, can adaptively learn the spatiotemporal relationships between different energy nodes by setting specific network layers, convolutional kernel size, and number of attention heads. This accurately extracts and integrates features of multiple energy systems, providing precise data support for energy scheduling. The hierarchical reinforcement learning model, based on the hierarchical Q-learning algorithm, clarifies the decision-making cycle, optimization objectives, and weights of upper and lower layer strategies. Through collaborative optimization of the reward function, it can simultaneously formulate macro-level plans for energy production and storage while taking into account the operational stability of terminal equipment and user comfort. This achieves refined hierarchical optimization of energy scheduling strategies, improving energy utilization efficiency and user satisfaction.

[0027] The multi-type sensor network in step S1 includes: Power monitoring layer: smart meters, harmonic sensors, insulation monitoring devices; Thermal monitoring layer: distributed heat meters, pipeline temperature sensors; Environmental sensing layer: temperature and humidity sensor, light intensity sensor, Concentration detector; Sensors at each layer transmit data in a standardized manner via the OPCUA protocol; The edge computing nodes and the AI ​​analysis engine use a 5G+industrial IoT hybrid network for data transmission. Before transmission, the data is losslessly compressed and encrypted and traced using blockchain technology to ensure efficient and secure data transmission.

[0028] In this embodiment, the edge computing node and the AI ​​analysis engine use a 5G + Industrial IoT hybrid network for data transmission. The 5G network is used for data transmission with high real-time requirements, while the Industrial IoT network is used for non-real-time data transmission. Before transmission, the data undergoes lossless compression using the LZ4 compression algorithm, achieving a compression ratio of 2:1-5:1. Data encryption and traceability are achieved through blockchain technology, employing an Ethereum consortium blockchain. Each data block includes a timestamp, hash value, and digital signature. This technical solution uniformly adopts the OPCUA protocol to achieve standardized data transmission and supports encryption and authentication, ensuring the accuracy and security of data during transmission. The use of a 5G + Industrial IoT hybrid network, combined with the LZ4 compression algorithm and blockchain technology, allows for reasonable data transmission planning, lossless compression, and encrypted traceability. This satisfies the real-time transmission needs of different data types while improving transmission efficiency, ensuring high-efficiency and secure data transmission, and preventing data leakage and tampering.

[0029] The edge computing node uses a deep learning-based multi-source data fusion algorithm to map different types of data, such as electricity, heat, and gas, to the same feature space. The data fusion algorithm is based on an architecture combining a Long Short-Term Memory (LSTM) network and an autoencoder. First, the LSTM network is used to learn the long-term dependencies of time series data, and then the autoencoder is used to reduce the dimensionality and extract features from the data, thereby achieving efficient fusion of data from different modalities.

[0030] In this example, to address the technical problem of significant differences in modalities among different data types such as electricity, heat, and gas, which hinders effective data fusion and affects the accuracy and efficiency of data analysis, a deep learning-based multi-source data fusion algorithm is adopted for the edge computing node. This algorithm consists of one input layer, two hidden layers, and one output layer. The number of neurons in the input layer is determined based on the number of data types, the number of neurons in the hidden layers are 128 and 64 respectively, and the number of neurons in the output layer represents the feature dimension of the fused data. This maps different types of data, such as electricity, heat, and gas, to the same feature space. The data fusion algorithm described is based on an architecture combining a Long Short-Term Memory (LSTM) network and an autoencoder. The LSTM network contains two hidden layers, each with 128 neurons, used to learn long-term dependencies in time-series data. The autoencoder's encoder part contains three fully connected layers with 256, 128, and 64 neurons respectively, while the decoder part has a symmetrical structure used for dimensionality reduction and feature extraction, achieving efficient fusion of different modalities. Through the above technical solution, this deep learning-based multi-source data fusion algorithm constructs a network architecture with a specific number of layers and neurons, combining LSTM and an autoencoder. First, LSTM is used to learn long-term dependencies in time-series data, and then the autoencoder is used for dimensionality reduction and feature extraction, mapping different types of data to the same feature space. This achieves efficient fusion of multi-modal data, improves the accuracy and efficiency of data processing, provides high-quality fused data for AI analysis engines, and enhances the analytical capabilities of energy systems.

[0031] The digital twin system combines weather forecast data and users' historical energy consumption data, and uses a Long Short-Term Memory (LSTM) network to predict users' future energy needs, providing a basis for energy dispatch simulation. The digital twin system uses the particle swarm optimization algorithm (PSO) to quantitatively evaluate different energy dispatch strategies based on the energy dispatch simulation results. By setting objective functions for energy cost, carbon emissions, and user satisfaction, it calculates the comprehensive score of each strategy and selects the optimal energy dispatch scheme.

[0032] In this example, to address the technical problems of inaccurate user energy demand forecasting, lack of quantitative evaluation methods for energy dispatch strategies, and difficulty in formulating optimal dispatch schemes, the digital twin system is used in conjunction with weather forecast data and user historical energy consumption data. A Long Short-Term Memory (LSTM) network is employed to predict users' future energy demand. The LSTM network has three hidden layers, each with 256 neurons, and the prediction period is 24 hours in the future, with a 1-hour time interval. This provides a basis for energy dispatch simulation. The digital twin system uses the particle swarm optimization algorithm (PSO) to quantitatively evaluate different energy dispatch strategies based on the energy dispatch simulation results. The particle swarm size is set to 50, and the maximum number of iterations is 100. By setting objective functions for energy cost, carbon emissions, and user satisfaction, with energy cost weighted at 0.5, carbon emissions weighted at 0.3, and user satisfaction weighted at 0.2, the system calculates the comprehensive score of each strategy and selects the optimal energy dispatch scheme. The above technical solution utilizes a Long Short-Term Memory (LSTM) network with specific network layers and parameters. By combining weather forecasts and historical energy consumption data, it achieves accurate prediction of users' energy demand for the next 24 hours, providing a reliable basis for energy dispatch simulation. Furthermore, it employs a Particle Swarm Optimization (PSO) algorithm, setting the particle swarm size and maximum number of iterations. By defining objective functions that include energy costs, carbon emissions, and user satisfaction, along with corresponding weights, it quantitatively evaluates different energy dispatch strategies, selects the optimal energy dispatch scheme, improves the scientific rigor and rationality of energy dispatch, reduces energy costs, decreases carbon emissions, and enhances user satisfaction.

[0033] The multi-level early warning mechanism classifies the early warning level according to the degree of impact of anomalies on the stability of park energy supply and user experience. It uses Bayesian networks to perform probabilistic reasoning on the causes of anomalies and combines expert knowledge base to generate energy dispatch optimization and equipment maintenance plans. The warning levels are divided into three levels: general warning, severe warning, and emergency warning, each corresponding to different abnormal handling procedures. When an emergency warning is triggered, the system automatically activates the backup energy supply system and prioritizes the energy needs of critical users.

[0034] In this embodiment, to address the technical problems of lacking a graded early warning mechanism for energy supply anomalies and equipment failures, unscientific anomaly cause analysis, and difficulty in quickly formulating effective handling solutions, a multi-level early warning mechanism is used to classify early warning levels based on the degree of impact of anomalies on the stability of the park's energy supply and user experience. A Bayesian network is used to perform probabilistic reasoning on the causes of anomalies. The Bayesian network contains 10 nodes, each representing a different anomaly type and influencing factor. Combined with an expert knowledge base, energy scheduling optimization and equipment maintenance solutions are generated. The warning levels are divided into three levels: general warning, severe warning, and emergency warning. General warning corresponds to situations where the anomaly has a minor impact on energy supply and user experience, and recording and continuous monitoring measures are taken. Severe warning corresponds to situations where the anomaly may affect the energy supply of some areas, and local energy dispatch optimization is initiated. Emergency warning corresponds to situations that may lead to large-scale energy outages. When an emergency warning is triggered, the system automatically starts the backup energy supply system and prioritizes the energy needs of key users, who are determined through a preset user priority list. Through the above technical solutions, the multi-level early warning mechanism divides the warning levels into three levels based on the impact of anomalies on the stability of the park's energy supply and user experience, and clarifies the anomaly handling procedures corresponding to different levels; by using Bayesian networks to perform probabilistic reasoning on the causes of anomalies and combining them with an expert knowledge base, it can scientifically and accurately analyze the causes of anomalies and quickly generate energy dispatch optimization and equipment maintenance plans; in the event of an emergency warning, it automatically activates the backup energy supply system and prioritizes key users, improving the energy system's ability to cope with emergencies and ensuring the stability of the park's energy supply and the normal energy use of key users.

[0035] The AI ​​analysis engine employs a federated learning framework during operation. This framework enables cross-regional and cross-departmental collaborative training of data while ensuring the data privacy of each energy subsystem, thereby dynamically updating the multimodal deep learning model and the hierarchical reinforcement learning model.

[0036] In this embodiment, to address the technical challenges of ensuring data privacy across energy subsystems, difficulties in cross-regional and cross-departmental collaborative training, and untimely model updates, the AI ​​analysis engine employs a federated learning framework during operation. This framework, based on a horizontal federated learning model, includes a central server and multiple participating nodes, which are edge computing nodes of each energy subsystem. It is used to achieve cross-regional and cross-departmental collaborative training while ensuring data privacy across energy subsystems, thereby dynamically updating the multimodal deep learning model and the hierarchical reinforcement learning model. During federated learning training, the central server is responsible for aggregating and distributing model parameters, using the FedAvg algorithm for model aggregation, with 10 iterations per training round; participating nodes train the model locally, with training data remaining on the local node, only uploading updated model parameter values; when the rate of change of the updated model parameter values ​​is less than 0.01, the model is considered to have converged, and training stops. The above technical solution employs a federated learning framework based on a horizontal federated learning model, clearly defining the responsibilities of the central server and participating nodes. This enables cross-regional and cross-departmental collaborative data training while ensuring data privacy across energy subsystems. The FedAvg algorithm is used for model aggregation, setting the number of training iterations and model convergence conditions for each round. This allows for dynamic updates of the multimodal deep learning model and hierarchical reinforcement learning model, enabling the model to adapt to changes in the energy system in a timely manner, improving the model's accuracy and applicability, and enhancing the intelligence level and long-term effectiveness of smart power monitoring methods.

[0037] An AI-driven smart power monitoring system includes: The data acquisition unit collects in real time the operating parameters of the multi-energy system consisting of electricity, heat and gas in the smart park, as well as energy consumption data, environmental parameters and equipment operating status data on the user side through a multi-type sensor network; The data preprocessing unit uses edge computing nodes to preprocess the collected data, including multi-source data fusion, noise filtering, data alignment, and outlier removal. The data analysis unit inputs the preprocessed data into the AI ​​analysis engine, which includes a multimodal deep learning model and a hierarchical reinforcement learning model running in parallel. The simulation monitoring unit constructs a virtual mapping of the intelligent park's integrated energy system through a digital twin system. Combining user energy demand forecasts and changes in the external environment, it simulates the system operation under different energy dispatch strategies in real time. The early warning and reporting unit triggers a multi-level early warning mechanism when it detects potential energy supply anomalies or equipment failures, and generates a decision report that includes anomaly location, cause analysis, energy dispatch optimization, and equipment maintenance and handling suggestions.

[0038] It should be noted that the calculation formulas and all parameters involved in the calculations in this invention have been dimensionless beforehand. The process of dimensionless processing is well known in the industry and will not be described here.

[0039] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An AI-driven intelligent power monitoring method, characterized in that, Includes the following steps: S1. Real-time collection of operating parameters of the multi-energy system consisting of electricity, heat and gas in the smart park, as well as energy consumption data, environmental parameters and equipment operating status data on the user side, through a multi-type sensor network; S2. Use edge computing nodes to preprocess the collected data, including multi-source data fusion, noise filtering, data alignment and outlier removal; S3. Input the preprocessed data into the AI ​​analysis engine, which includes a multimodal deep learning model and a hierarchical reinforcement learning model running in parallel. S4. Construct a virtual mapping of the intelligent park's integrated energy system through a digital twin system, and simulate the system operation under different energy dispatch strategies in real time by combining user energy demand forecasts and changes in the external environment. S5. When a potential energy supply anomaly or equipment failure is detected, a multi-level early warning mechanism is triggered, and a decision report is generated that includes anomaly location, cause analysis, energy dispatch optimization, and equipment maintenance and disposal recommendations.

2. The AI-driven smart power monitoring method according to claim 1, characterized in that, The multimodal deep learning model employs a spatiotemporal graph convolutional network to extract and fuse features from the spatiotemporal data and topology of multi-energy systems. The spatiotemporal graph convolutional network introduces an attention mechanism into the graph structure to adaptively learn the spatiotemporal relationships between different energy nodes, thereby enhancing the feature extraction capability of complex energy networks. The hierarchical reinforcement learning model is based on the hierarchical Q-learning algorithm. The upper-layer strategy is used to formulate a macro-allocation plan for energy production and storage, and the lower-layer strategy is used to dynamically adjust the energy consumption of each terminal device. The hierarchical reinforcement learning model has an upper-layer strategy that optimizes the overall energy cost and carbon emission indicators of the park, and a lower-layer strategy that optimizes the operational stability of terminal devices and user comfort. The upper and lower-layer strategies are jointly optimized through a reward function.

3. The AI-driven smart power monitoring method according to claim 2, characterized in that, The multi-type sensor network in step S1 includes: Power monitoring layer: smart meters, harmonic sensors, insulation monitoring devices; Thermal monitoring layer: distributed heat meters, pipeline temperature sensors; Environmental sensing layer: temperature and humidity sensor, light intensity sensor, Concentration detector; Each layer of sensors uses the OPCUA protocol to achieve standardized data transmission.

4. The AI-driven smart power monitoring method according to claim 3, characterized in that, The edge computing node employs a deep learning-based multi-source data fusion algorithm to map different types of data, such as electricity, heat, and gas, to the same feature space.

5. The AI-driven smart power monitoring method according to claim 4, characterized in that, The digital twin system combines weather forecast data and users' historical energy consumption data, and uses a Long Short-Term Memory (LSTM) network to predict users' future energy needs, providing a basis for energy dispatch simulation. The digital twin system uses the particle swarm optimization algorithm (PSO) to quantitatively evaluate different energy dispatch strategies based on the energy dispatch simulation results. By setting objective functions for energy cost, carbon emissions, and user satisfaction, it calculates the comprehensive score of each strategy and selects the optimal energy dispatch scheme.

6. The AI-driven smart power monitoring method according to claim 1, characterized in that, The multi-level early warning mechanism classifies the early warning level according to the degree of impact of anomalies on the stability of park energy supply and user experience. It uses Bayesian networks to perform probabilistic reasoning on the causes of anomalies and combines expert knowledge base to generate energy dispatch optimization and equipment maintenance plans. The warning levels are divided into three levels: general warning, severe warning, and emergency warning, each corresponding to a different abnormal handling procedure.

7. The AI-driven smart power monitoring method according to claim 1, characterized in that, The AI ​​analysis engine employs a federated learning framework during operation. This framework enables cross-regional and cross-departmental collaborative training of data while ensuring the data privacy of each energy subsystem, thereby dynamically updating the multimodal deep learning model and the hierarchical reinforcement learning model.

8. An AI-driven intelligent power monitoring system, characterized in that, The system is used to execute the AI-driven smart power monitoring method according to any one of claims 1-8, comprising: The data acquisition unit collects in real time the operating parameters of the multi-energy system consisting of electricity, heat and gas in the smart park, as well as energy consumption data, environmental parameters and equipment operating status data on the user side through a multi-type sensor network; The data preprocessing unit uses edge computing nodes to preprocess the collected data, including multi-source data fusion, noise filtering, data alignment, and outlier removal. The data analysis unit inputs the preprocessed data into the AI ​​analysis engine, which includes a multimodal deep learning model and a hierarchical reinforcement learning model running in parallel. The simulation monitoring unit constructs a virtual mapping of the intelligent park's integrated energy system through a digital twin system. Combining user energy demand forecasts and changes in the external environment, it simulates the system operation under different energy dispatch strategies in real time. The early warning and reporting unit triggers a multi-level early warning mechanism when it detects potential energy supply anomalies or equipment failures, and generates a decision report that includes anomaly location, cause analysis, energy dispatch optimization, and equipment maintenance and handling suggestions.

Images