A multi-region power energy intelligent scheduling method based on big data

Abstract

This invention provides a multi-regional intelligent power energy dispatching method based on big data, belonging to the field of power energy dispatching technology. The method includes: collecting multimodal data, performing data cleaning and spatiotemporal alignment, and constructing a structured feature dataset through standardization; constructing a power grid topology model consistent with the physical power grid system parameters, and integrating a multi-physics model to build a dynamic simulation system; identifying typical energy consumption patterns of the system operation, constructing a multi-dimensional anomaly pattern detection system, and performing core feature extraction and dimensionality reduction; designing a multi-agent architecture, establishing a cooperative-competitive game mechanism to guide agent collaborative optimization; dynamically generating dispatching decisions, establishing a safety constraint verification system, and introducing a human intervention mechanism. This invention achieves accurate identification of typical energy consumption patterns, improves anomaly detection accuracy, increases the renewable energy absorption rate compared to traditional dispatching methods, significantly improves dispatching real-time performance, and enables rapid response to emergencies.

Description

Technical Field

[0001] This invention relates to the field of power energy dispatching technology, and more specifically to a multi-regional intelligent power energy dispatching method based on big data. Background Technology

[0002] With the advancement of "dual-carbon" goals and the deepening of energy transition, multi-regional power systems and green computing centers are deeply integrated, and renewable energy is being connected to the grid on a large scale, forming a complex energy system with multi-dimensional coupling of "electricity, computing power, and environment." Big data technology provides data support for multi-regional energy dispatch, integrating multi-modal data such as load, power output, and environmental parameters from various regions to achieve refined and intelligent dispatch decisions. Currently, multi-regional power energy dispatch needs to balance renewable energy consumption, computing power load assurance, and system security and stability. Traditional dispatch methods are no longer suitable for the needs of heterogeneous data processing and multi-entity collaboration. Big data-based intelligent dispatch has become a key path to solve energy optimization problems and improve energy utilization efficiency, possessing significant engineering application value and promising industrial promotion prospects.

[0003] Existing technologies for multi-regional power energy dispatching still have many shortcomings: First, the multimodal data processing capability is weak, making it difficult to efficiently process heterogeneous data such as power, environment, and equipment operation. It is easily affected by data noise and missing values, and cannot accurately uncover hidden energy consumption patterns and abnormal states, resulting in a lack of reliable data support for dispatching decisions. Second, the level of multi-regional and multi-energy collaborative dispatching is insufficient, making it difficult to cope with the superimposed impact of renewable energy volatility and computing load fluctuations. Dispatch conflicts occur frequently between regions, and dispatching robustness and collaborative optimization effects are poor. Third, the real-time optimization and automated dispatching capabilities are lacking, relying on manual intervention. Decision response is lagging, and there is a lack of continuous iterative optimization mechanisms, making it impossible to dynamically adapt to changes in the operation of the power system and balance the multiple objectives of system energy efficiency, safety and stability, and cost control. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-regional intelligent power energy dispatching method based on big data, in order to solve the technical problems of existing power energy dispatching technologies, such as difficulty in accurately mining energy consumption patterns and detecting anomalies, difficulty in multi-regional and multi-energy coordinated dispatching, and difficulty in responding to dynamic changes in the system due to reliance on manual intervention.

[0005] To achieve the above objectives, this invention provides a multi-regional intelligent power energy dispatching method based on big data. The method includes: collecting multimodal data from a sensor network deployed in the target power region; cleaning and spatiotemporally aligning the data; and constructing a structured feature dataset through standardization. Based on the feature dataset, a power grid topology model consistent with the parameters of the physical power grid system is constructed, and a multi-physics model is integrated to build a dynamic simulation system, establishing a data interaction channel between the digital twin environment and the physical power grid system. A preset clustering algorithm is used to identify typical energy consumption patterns of the system operation, and a preset deep learning algorithm is combined to construct a multi-dimensional abnormal pattern detection system to detect abnormal operating states of the system, and core features are extracted and dimensionality reduced. Based on the dimensionality-reduced core features, a multi-agent architecture is designed, a cooperative-competitive game mechanism is established to guide the agents in collaborative optimization, and transfer learning technology is used to accelerate model training, resulting in a trained multi-agent reinforcement learning game model. The trained multi-agent reinforcement learning game model is deployed to the digital twin environment to dynamically generate dispatching decisions, establish a safety constraint verification system to comprehensively verify the dispatching decisions, and introduce a human intervention mechanism to achieve an organic combination of machine decision-making and human experience.

[0006] Optionally, the acquisition of multimodal data includes constructing a three-dimensional data acquisition system for power, environment, and operation. The data cleaning and spatiotemporal alignment include: using the isolated forest algorithm in unsupervised learning to identify outliers in the multimodal data; using the sliding window mean method for filling and repairing; and using linear interpolation to supplement missing data; resampling the multimodal data with a preset time benchmark; adjusting multimodal data with different sampling frequencies to a uniform time interval; establishing a unified coding system; and associating multimodal data according to the coding; and using the Kalman filter algorithm to fuse multi-source homogeneous data of the multimodal data, establish state equations and observation equations, and dynamically estimate the true value of the data.

[0007] Optionally, the construction of a power grid topology model consistent with the parameters of the physical power grid system, and the integration of a multiphysics model to build a dynamic simulation system, includes: importing wiring diagram data conforming to the general information model standard from the power grid GIS system and storing topology relationships using a preset graph database; constructing a power grid topology visualization scene using a preset professional 3D engine, and designing differentiated power grid topology models according to equipment type and functional characteristics; verifying the rationality and accuracy of the power grid topology model through power flow calculation, and checking whether the power grid topology model meets the operating rules of the physical power grid system; integrating an electrical simulation physical field model into the power grid topology model, using the Newton-Raphson method for power flow calculation, and performing steady-state simulation of a large-scale power grid system; integrating a renewable energy output model into the power grid topology model, dynamically associating environmental data with energy output, and connecting environmental data to the simulation system through an interface to dynamically adjust renewable energy output; integrating a green computing center equipment energy consumption model and a power grid equipment operation model into the power grid topology model, and constructing a dynamic simulation system to simulate the electrical characteristics, environmental impact, and equipment response of the physical power grid system under different operating scenarios.

[0008] Optionally, the construction of a multi-dimensional anomaly pattern detection system to detect abnormal system operation states and perform core feature extraction and dimensionality reduction includes: using a pre-defined statistical method combined with a deep learning joint detection strategy to screen extreme anomaly patterns and identify complex anomaly patterns; classifying the detected anomaly patterns and extracting corresponding anomaly features for different types of anomaly patterns; when an anomaly pattern is detected, initiating a root cause analysis process to trace back information in the digital twin environment and determine the root cause of the anomaly pattern through correlation analysis and causal reasoning; and using the t-SNE algorithm to reduce the dimensionality of high-dimensional features in the anomaly features while retaining the core information of the high-dimensional features.

[0009] Optionally, the establishment of a cooperative-competitive game mechanism to guide agent collaborative optimization includes: establishing a two-layer reward mechanism combining global and local rewards to balance global and local optimization; employing the Nash-Q learning algorithm to conduct multi-agent cooperative-competitive game, where the global coordinating agent dynamically adjusts the reward weights of agents in each region by evaluating global rewards, guiding agents in each region to collaborate towards the global optimum; and establishing a regional decision conflict coordination mechanism, whereby when the actions of agents in different regions conflict, the global coordinating agent formulates conflict solutions based on the severity of the conflict and the priority of each region.

[0010] Optionally, the establishment of a safety constraint verification system to comprehensively verify scheduling decisions includes: constructing a safety constraint system covering three major categories: power grid safety constraints, equipment operation constraints, and energy supply constraints. Each constraint has a limit range and verification standard based on industry standards and equipment parameters; using fast power flow calculation and constraint satisfaction checking algorithms to verify safety constraints. If a scheduling decision violates a constraint, the system automatically marks the type and degree of constraint violation and triggers a decision correction mechanism; and establishing an automatic scheduling decision correction mechanism. For scheduling decisions that violate constraints, a heuristic algorithm is used to adjust the parameters of the scheduling decision based on the importance and scope of the constraint.

[0011] Optionally, the introduction of a human intervention mechanism to organically combine machine decision-making with human experience includes: designing a scheduling decision visualization dashboard to visually display generated scheduling decisions, safety verification results, system operating status, and abnormal warning information, and triggering audible and visual warnings for high-risk scheduling decisions; designing a human intervention interactive interface, allowing operators to adjust, reject, or regenerate generated scheduling decisions based on actual operating conditions, experience, and external instructions, and supporting operation recording and backtracking; and establishing a human-machine collaborative decision-making iterative optimization mechanism, where operator intervention actions and decision feedback serve as incremental training data, and model parameters are updated through incremental learning technology.

[0012] Optionally, the intelligent power energy dispatching method further includes: establishing a multi-dimensional performance evaluation system based on the execution results of dispatching decisions, quantifying the implementation effect of the execution results and identifying problems, updating the multi-agent reinforcement learning game model through incremental learning technology, and dynamically calibrating the parameters of the digital twin environment by comparing the simulation data of the digital twin environment with the operation data of the physical power grid system.

[0013] Optionally, the establishment of a multi-dimensional performance evaluation system to quantify the implementation effect of the execution results and identify problems includes: formulating multi-dimensional performance evaluation indicators based on industry standards and project objectives, including at least economic, technical, environmental, and safety indicators; establishing an evaluation data collection mechanism to collect operational data after the execution of scheduling decisions, with the data collection frequency consistent with the scheduling decision cycle, and conducting real-time, periodic, and special evaluations; and using preset analysis methods to conduct in-depth analysis of the evaluation results, identify the causes of differences and weaknesses, and display the evaluation results through visual charts.

[0014] Optionally, the intelligent power energy dispatching method further includes: adapting and adjusting the intelligent power energy dispatching method to different application scenarios, planning a promotion and application path, and reducing the risk of promotion and application through compliance verification.

[0015] Through the above technical solutions, this invention accurately identifies typical energy consumption patterns in different regions and time periods by using unsupervised learning-driven energy consumption pattern mining and data analysis empowered by a digital twin environment. It quantifies regional energy consumption differences, improves anomaly detection accuracy, and enables early warning of energy consumption anomalies. By adapting a multi-agent reinforcement learning game model with a transfer learning acceleration model, the renewable energy absorption rate is improved compared to traditional scheduling methods, grid operation stability is significantly enhanced, the risk of line power flow exceeding limits is reduced, and a globally coordinated agent resolves inter-regional scheduling conflicts, enhancing the adaptability of the scheduling model. Through efficient integration of the model and the digital twin environment, full-process security verification, and human-machine collaborative decision-making, the real-time performance of scheduling is significantly improved, enabling rapid response to emergencies such as extreme weather and load surges. The level of scheduling automation is improved, requiring no manual intervention in daily scenarios, gradually reducing scheduling errors, and continuously improving system energy efficiency and stability.

[0016] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a flowchart illustrating the multi-regional intelligent power energy dispatching method based on big data according to the present invention. Figure 1 ;

[0019] Figure 2 This is a schematic diagram of the data cleaning and spatiotemporal alignment process in this invention;

[0020] Figure 3 This is a schematic diagram of the process of constructing the dynamic simulation system in this invention;

[0021] Figure 4 This is a schematic diagram of the process for core feature extraction and dimensionality reduction in this invention;

[0022] Figure 5 This is a schematic diagram of the process of establishing a cooperative-competitive game mechanism to guide the intelligent agents to cooperate and optimize in this invention;

[0023] Figure 6 This is a flowchart illustrating the comprehensive verification of scheduling decisions in this invention.

[0024] Figure 7 This is a flowchart illustrating the organic integration of machine decision-making and human experience in this invention.

[0025] Figure 8This is a flowchart illustrating the implementation effect of quantifying execution results and identifying problems in this invention;

[0026] Figure 9 This is a flowchart illustrating the multi-regional intelligent power energy dispatching method based on big data according to the present invention. Figure 2 . Detailed Implementation

[0027] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0028] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0029] Please refer to Figure 1 and Figure 9 This invention provides a multi-regional intelligent power energy dispatching method based on big data. The intelligent power energy dispatching method may include:

[0030] Step S100: Based on the sensor network deployed in the target power area, collect multimodal data, perform data cleaning and spatiotemporal alignment, and construct a structured feature dataset through standardization processing.

[0031] In this embodiment of the invention, the collection of multimodal data includes constructing a three-dimensional data acquisition system encompassing power, environment, and operation. During sensor network deployment, a comprehensive IoT sensor network is deployed in the target area (e.g., a green computing center, associated power grid nodes, and renewable energy plants). This network covers smart meters (e.g., for collecting real-time load data and power consumption data), high-precision meteorological monitoring equipment (e.g., for collecting environmental parameters such as temperature, humidity, wind speed, and light intensity), and server status monitoring agents (e.g., for collecting equipment status data such as CPU utilization, memory usage, device power consumption, and cooling system operating parameters). All sensor terminals establish real-time communication with the central data platform via 5G or fiber optic high-speed networks. The sampling frequency is dynamically adjusted according to the data type. The sampling frequency for core operational data ensures the continuity of time-series data, while the sampling frequency for environmental data meets the requirements for renewable energy output forecasting.

[0032] In this embodiment of the invention, power data is obtained from a dedicated interface of the power grid dispatch center, including time-series data such as multi-regional load curves, output data of various power sources, power flow of transmission lines, and substation operating parameters. Each data entry contains key attributes such as a unique timestamp, regional identifier, data type, and numerical value to ensure data traceability. Environmental data is accessed through a professional meteorological data service interface to obtain wind and solar energy resource assessment parameters and real-time meteorological indicators. Time-series alignment with power data is achieved through a timestamp synchronization mechanism to avoid data misalignment affecting subsequent analysis. Operational data is collected from the real-time status of server clusters and auxiliary systems through a computing center monitoring protocol (e.g., SNMP protocol). The data granularity is accurate to individual devices or functional modules, ensuring that energy consumption fluctuations at the device level can be captured.

[0033] Please refer to Figure 2 In a preferred embodiment of the present invention, data cleaning and spatiotemporal alignment may include:

[0034] Step S110: The Isolation Forest algorithm in unsupervised learning is used to identify outliers in multimodal data, the sliding window mean method is used for filling and repairing, and the linear interpolation method is used to supplement missing data.

[0035] In a preferred embodiment of the present invention, the Isolation Forest algorithm in unsupervised learning is used to identify data outliers. This algorithm can efficiently process high-dimensional data. By constructing multiple isolated trees to isolate the data, outliers are more easily isolated and identified. For outliers such as numerical mutations caused by sensor malfunctions or deviations from the normal range caused by data acquisition errors, a dynamic threshold based on data distribution characteristics is set. Outliers exceeding the threshold are filled and repaired using a sliding window mean method. The window size is adaptively adjusted according to the data sampling frequency and fluctuation characteristics. For short-term missing data, linear interpolation or interpolation based on the correlation of neighboring data is used to supplement the data to ensure data continuity. For long-term missing data, it is marked as invalid data and the reason for the missing data is recorded to avoid interference with subsequent analysis.

[0036] Step S120: Resample the multimodal data using a preset time base, adjust the multimodal data with different sampling frequencies to a uniform time interval, establish a unified coding system, and associate the multimodal data according to the coding.

[0037] In a preferred embodiment of the present invention, multi-source data is resampled using a unified time reference (e.g., UTC time) to adjust data with different sampling frequencies to a unified time interval, thus solving the clock deviation problem of different sensors. Spatial alignment is achieved through regional identifier mapping, establishing a unified coding system for geographical or functional areas, and associating meteorological data, power data, and operational data by code to ensure spatial consistency of different types of data under the same timestamp. For example, the sunlight data of a certain area is accurately bound to the output data of photovoltaic power stations in that area.

[0038] Step S130: Using the Kalman filter algorithm, multi-source homogeneous data of multimodal data are fused to establish state equations and observation equations, and the true value of the data is dynamically estimated.

[0039] In a preferred embodiment of the present invention, a Kalman filter algorithm is used to fuse homogeneous data from multiple sources. By establishing state equations and observation equations, the true values ​​of the data are dynamically estimated, reducing the impact of sensor noise and measurement errors on data accuracy. For example, by fusing wind speed data collected from multiple wind speed monitoring devices and combining equipment calibration parameters and environmental factors, the accuracy of wind energy resource assessment is optimized; and by fusing temperature data from different monitoring points, the temperature distribution in the computing center's computer room is accurately reflected.

[0040] In a preferred embodiment of the present invention, during feature extraction, based on unsupervised learning theory and the operating characteristics of energy systems, multi-dimensional statistical features are extracted, covering three main categories: time series features (e.g., load volatility, peak percentage, valley percentage, load duration, etc., used to characterize the load change pattern over time), cross-regional correlation features (e.g., inter-regional load correlation coefficient, wind and solar power output intermittency index, inter-regional energy exchange intensity, etc., used to reflect the collaborative operation characteristics of multi-regional systems), and dynamic trend features (e.g., slope of energy consumption change trend within a sliding window, energy output prediction error rate, load growth rate, etc., used to capture the dynamic changes in system operation).

[0041] In a preferred embodiment of the present invention, during data standardization, numerical features are processed using the Z-score standardization method. By calculating the mean and standard deviation of the data, features with different dimensions and numerical ranges are converted into a unified scale, eliminating the impact of dimensional differences on model training. Categorical features (e.g., region type, equipment operating status, weather type, etc.) are converted using one-hot encoding, transforming discrete features into binary vectors to ensure the model can effectively identify and process them. During the standardization process, the distribution characteristics of the original data are preserved to avoid data distortion.

[0042] Step S200: Based on the feature dataset, construct a power grid topology model consistent with the parameters of the physical power grid system, integrate a multi-physics model, construct a dynamic simulation system, and establish a data interaction channel between the digital twin environment and the physical power grid system.

[0043] Please refer to Figure 3 In this embodiment of the invention, a power grid topology model consistent with the parameters of the physical power grid system is constructed, and a multiphysics model is integrated to build a dynamic simulation system, which may include:

[0044] Step S210: Import wiring diagram data conforming to the general information model standard from the power grid GIS system, and store the topology relationship using a preset graph database.

[0045] In a preferred embodiment of the present invention, the wiring diagram data covers node entities such as substations, green computing centers, photovoltaic power stations, and wind farms, as well as edge entities such as transmission lines, distribution lines, and transformers. Node attributes include key parameters such as equipment name, rated capacity, voltage level, geographical coordinates, and equipment type; edge attributes include electrical parameters such as line length, impedance, resistance, reactance, rated current, and insulation class. A graph database (e.g., Neo4j) is used to store the topology relationships, and the node-edge-attribute structural model supports fast traversal, querying, and modification of the topology relationships.

[0046] Step S220: Using a preset professional 3D engine, construct a power grid topology visualization scene and design differentiated power grid topology models according to equipment type and functional characteristics.

[0047] In a preferred embodiment of the present invention, in the power grid topology model, substations are represented by a cube model, computing centers by a composite model, transmission lines by a cylindrical model, and photovoltaic panels and wind turbines by realistic 3D models. Different colors are used to encode the operating status of equipment; for example, green indicates normal operation, yellow indicates a warning state, and red indicates an overload or fault state. The scene supports interactive operations such as zooming, rotating, and panning, allowing operators to control the viewpoint via mouse and keyboard and intuitively observe the power grid topology and equipment operating status.

[0048] Step S230: Verify the rationality and accuracy of the power grid topology model through power flow calculation, and check whether the power grid topology model meets the operating rules of the physical power grid system.

[0049] In a preferred embodiment of the present invention, the rationality and accuracy of the topology model are verified through power flow calculation. Based on the principle of node power balance, the voltage phasor, power distribution, and line power flow of each node are calculated to check whether the topology model meets the operating rules of the physical power grid. Verification indicators include node voltage deviation, line power flow deviation, and power balance error. If the voltage deviation exceeds the allowable range, the line power flow exceeds the rated capacity, or the power balance error is too large, it is necessary to backtrack and check the imported topology data and parameter settings, correct the error information, and ensure that the power balance error of the topology model is less than 1% and the voltage deviation is within the allowable range.

[0050] Step S240: Integrate the electrical simulation physical field model into the power grid topology model, use the Newton-Raphson method to perform power flow calculations, and conduct steady-state simulation of the large-scale power grid system.

[0051] In a preferred embodiment of the invention, the Newton-Raphson method is used for power flow calculation. This method has fast convergence speed and high calculation accuracy, and is suitable for steady-state simulation of large-scale power grids. Based on the nodal power balance equations, the voltage magnitude and phase angle of each node are solved iteratively to obtain the line power flow, power distribution, and equipment operating parameters. The core principle of the nodal power balance equations is that the injected power at each node is equal to the sum of the power exchanges between that node and all other nodes. By continuously adjusting the nodal voltage phasors, the calculation results satisfy the power balance condition. Simultaneously, the influence of factors such as transformer turns ratio and reactive power compensation equipment on power flow distribution is considered to improve simulation accuracy.

[0052] Step S250: Integrate the renewable energy output model into the power grid topology model, dynamically link environmental data with energy output, and connect the environmental data to the simulation system through an interface to dynamically adjust the renewable energy output.

[0053] In a preferred embodiment of the present invention, a renewable energy output model is constructed to realize the dynamic correlation between environmental factors and energy output. The photovoltaic output model is constructed based on irradiance, photovoltaic panel area, conversion efficiency, and temperature correction coefficient. The core logic is that photovoltaic output is positively correlated with irradiance, and is also affected by ambient temperature; an increase in temperature will lead to a decrease in conversion efficiency. The wind power output model is constructed by fitting the wind speed-output power curve. Different wind speed ranges correspond to different output levels. When the wind speed is lower than the cut-in wind speed or higher than the cut-out wind speed, the wind turbine stops outputting power, and the maximum output is achieved near the rated wind speed.

[0054] Step S260: Integrate the energy consumption model of the green computing center equipment and the operation model of the power grid equipment into the power grid topology model, and build a dynamic simulation system to simulate the electrical characteristics, environmental impact and equipment response of the physical power grid system under different operating scenarios.

[0055] In a preferred embodiment of the present invention, the energy consumption model of the computing center server includes two parts: idle power consumption and load power consumption. The total power consumption of the server is equal to the idle power consumption plus the product of the load factor and the maximum load power consumption. The load factor is related to operating parameters such as CPU utilization and memory utilization. The cooling system energy consumption model is constructed based on the computer room temperature, the rated power of the cooling equipment, and the energy efficiency ratio. The higher the computer room temperature, the greater the energy consumption of the cooling system. The power grid equipment operation model includes transformer loss model, line loss model, etc. Transformer loss includes idle loss and load loss. Line loss is proportional to the square of the line current. Equipment loss is calculated based on electrical simulation results.

[0056] In a preferred embodiment of the present invention, an API gateway can be constructed using the RESTful interface specification, providing a dedicated data transmission endpoint (e.g., / api / real-time-data) for receiving standardized multimodal data. The API gateway supports high-concurrency data transmission and can simultaneously process real-time data pushes from multiple data sources. It allocates processing resources through a load balancing mechanism to avoid interface congestion. An event-driven architecture is used to achieve real-time data synchronization. When the physical system's operating data is updated, a Webhook is triggered to notify the API gateway of the digital twin environment. After receiving the data, the gateway quickly forwards it to the simulation system, with synchronization latency controlled within 1 second. A heartbeat detection mechanism maintains the stability of the data transmission connection. If a connection interruption is detected, a reconnection mechanism is automatically triggered to ensure uninterrupted data synchronization. A precise mapping relationship between real-time data and topology model nodes is established. Based on regional identifiers and device codes, real-time data such as load data, power output data, and equipment operating parameters of the physical system are bound to the corresponding nodes and device models in the digital twin environment. For example, the real-time load data of a green computing center is bound to the computing center node in the topology model, and the real-time power output data of a wind farm is bound to the corresponding wind farm node.

[0057] For example, a digital twin topology model of a regional power grid was constructed. This model contains over 200 nodes and over 300 lines. For instance, photovoltaic power plants and wind farms are concentrated in the western region, while computing centers and industrial load nodes are concentrated in the eastern region. During model validation, power flow calculations revealed that the power flow of a certain transmission line exceeded its rated capacity by 5%. Upon investigation, it was found that the line impedance parameters were incorrectly imported. After correcting the parameters and recalculating, the power flow returned to within the rated capacity range, and the model validation passed. In the digital twin environment, an extreme weather scenario was simulated: at 14:00 on a certain day, due to cloud cover, the solar irradiance plummeted from 800W / m² to 200W / m², and the ambient temperature rose by 5°C. Calculations using the photovoltaic output model showed that the photovoltaic power plant output decreased by 30%. Combined with electrical simulations, the western renewable energy-rich area... A 0.2 pu decrease in node voltage in the domain leads to a redistribution of power flow on some transmission lines, with a 10% increase in power flow on one line. Simultaneously, the temperature in the computing center's computer room rises, causing a 15% increase in cooling system energy consumption. When a green computing center in the physical power grid experiences a 20% increase in load from its baseline due to adjustments in computing tasks, the center's monitoring system collects load change data in real time and pushes it to the digital twin environment via the API gateway. The digital twin environment receives and processes this data within 0.5 seconds, updating the load change to the corresponding computing center node through data mapping. The visualization interface simultaneously refreshes the node's load value and status color (e.g., changing from green to yellow alert status). Operators can observe the impact of load changes on power flow distribution and node voltage in real time, providing timely scenario feedback for scheduling decisions.

[0058] Step S300: Using a preset clustering algorithm, identify typical energy consumption patterns of the system operation. Combined with a preset deep learning algorithm, construct a multi-dimensional abnormal pattern detection system to detect abnormal operating states of the system and perform core feature extraction and dimensionality reduction processing.

[0059] In this embodiment of the invention, an improved K-shapes algorithm is used for temporal clustering. This algorithm is specifically designed for time-series data and uses dynamic time warping distance as a similarity metric. It can effectively handle the translation, scaling, and other deformations of time-series data and accurately measure the similarity of different time-series data. The number of clusters K is determined by combining the elbow rule and the silhouette coefficient. The elbow rule is used to find the point where the error decrease rate changes abruptly by plotting the curve of clustering error with the number of clusters, as a reference for the optimal number of clusters. The silhouette coefficient is used to evaluate the rationality of the clustering results by calculating the similarity between a sample and samples in the same cluster and samples in different clusters, and the number of clusters with the largest silhouette coefficient is selected as the final value.

[0060] In this embodiment of the invention, cluster analysis is performed on load time-series data over a period of time (e.g., 3 months) to identify different types of typical energy consumption patterns. Common patterns include peak load concentration patterns (e.g., bi-peak pattern in the morning and evening, corresponding to peak residential electricity consumption and industrial production), stable base load patterns (e.g., small load fluctuations, corresponding to continuously operating base loads), intermittent fluctuation patterns (e.g., load fluctuates frequently with changes in renewable energy output or computing tasks), and off-peak load patterns (e.g., load is at a low level, corresponding to nighttime or holidays). The center curve of each cluster represents the core characteristics of that type of pattern. By analyzing parameters such as the peak time, load amplitude, and fluctuation range of the center curve, the operating characteristics of each type of pattern are clarified.

[0061] Please refer to Figure 4 In this embodiment of the invention, a multi-dimensional anomaly pattern detection system is constructed to detect abnormal operating states of the system and to perform core feature extraction and dimensionality reduction processing, which may include:

[0062] Step S310: Use a pre-defined statistical method combined with a deep learning joint detection strategy to screen extreme anomaly patterns and identify complex anomaly patterns.

[0063] In a preferred embodiment of the present invention, the statistical method adopts the 3σ criterion, which, based on the normal distribution characteristics of the data, determines data points that deviate from the mean by more than 3 times the standard deviation as abnormal, and is suitable for detecting extreme outliers that significantly deviate from the normal range; the deep learning method adopts an autoencoder model, which compresses high-dimensional energy-consuming data into a low-dimensional feature space through an encoder, and then reconstructs the original data through a decoder. If the reconstruction error exceeds a set threshold, it is determined to be abnormal, and is suitable for detecting complex abnormal patterns hidden in normal data.

[0064] Step S320: Classify the detected abnormal patterns and extract the corresponding abnormal features for different types of abnormal patterns.

[0065] In a preferred embodiment of the present invention, the anomalies are classified according to their causes into equipment failure anomalies, meteorological change anomalies, load fluctuation anomalies, and energy supply interruption anomalies. Equipment failure anomalies are characterized by a sudden surge or drop in the energy consumption of a certain type of equipment, such as a cooling system failure causing energy consumption to double. Meteorological change anomalies are characterized by a sudden change in the output of renewable energy due to extreme weather, which in turn causes load or energy consumption fluctuations. Load fluctuation anomalies are characterized by a sudden increase or decrease in computing load without a reasonable cause. Energy supply interruption anomalies are characterized by a sudden interruption in the supply of a certain type of energy, leading to abnormal changes in the consumption of other energy sources.

[0066] Step S330: When an abnormal pattern is detected, initiate the root cause analysis process, trace back information in the digital twin environment (e.g., event logs, equipment operation records, meteorological data changes, load adjustment records, etc.), and determine the root cause of the abnormal pattern through correlation analysis and causal reasoning.

[0067] For example, an abnormal surge in energy consumption of the cooling system in a computing center was detected. By reviewing environmental data, a sudden 5°C increase in the server room temperature was found during the same period. Further investigation of equipment operation records revealed a decrease in air conditioning efficiency, leading to the conclusion that the root cause of the anomaly was an air conditioning malfunction. Simultaneously, an anomaly early warning mechanism was established, classifying warning levels (e.g., general warning, important warning, emergency warning) based on the severity of the anomaly (e.g., scope of impact, duration, and losses caused). Warning information was sent to relevant personnel via digital twin visualization interfaces, SMS, email, and other means to remind them to take timely action.

[0068] Step S340: Use the t-SNE algorithm to reduce the dimensionality of high-dimensional features in the abnormal features, while retaining the core information of the high-dimensional features.

[0069] In a preferred embodiment of the present invention, the t-SNE algorithm can map high-dimensional features to two-dimensional or three-dimensional space while preserving the local structure and global distribution characteristics of the data, making it suitable for dimensionality reduction visualization of high-dimensional time series data. In the algorithm parameter settings, the perplexity parameter is adjusted according to the number of data samples to ensure that the local similarity of the data can be captured. The learning rate parameter controls the convergence speed of the dimensionality reduction process to avoid the occurrence of local optima. The number of iterations is set to a sufficiently large number to ensure that the algorithm converges.

[0070] In a preferred embodiment of the present invention, various visualization charts can be embedded in the digital twin environment to intuitively display energy consumption patterns, anomalies, and feature correlations. These visualization charts may include: first, scatter plot visualization, which displays the dimensionality-reduced low-dimensional features as a scatter plot, with different colors representing different clustering labels (typical energy consumption patterns) and different sizes representing anomaly levels, allowing operators to intuitively observe the distribution of various energy consumption patterns and the location of anomalies; second, time-series curve visualization, which plots the center curves of various typical energy consumption patterns and the time-series curves of anomaly patterns, comparing and displaying the differences between normal and anomaly patterns; third, heat map visualization, which displays the frequency of occurrence of various energy consumption patterns in different regions and time periods, intuitively reflecting the spatiotemporal distribution characteristics of energy consumption patterns; and fourth, feature correlation network diagrams, which display the correlation between various energy consumption features, using line thickness to represent the magnitude of the correlation coefficient, helping researchers understand the intrinsic connections between features.

[0071] Step S400: Based on the core features after dimensionality reduction, design a multi-agent architecture, establish a cooperative-competitive game mechanism to guide agents to optimize collaboratively, and use transfer learning technology to accelerate model training, thereby obtaining a trained multi-agent reinforcement learning game model.

[0072] In a preferred embodiment of the present invention, when designing a multi-agent architecture, a hierarchical architecture of "regional agents + global coordinating agents" can be adopted. Each region deploys a regional agent responsible for local energy scheduling decisions; a global coordinating agent is set up to coordinate the decision-making behavior of each regional agent, ensuring the achievement of the global optimization goal. The core responsibilities of the regional agents include perceiving the real-time operating status of their region (e.g., load, energy output, equipment status, etc.), formulating energy allocation strategies for their region, and interacting with other regional agents and the global coordinating agent. The core responsibilities of the global coordinating agent include collecting decision information from each regional agent and the overall system operating status, evaluating the global optimization effect, and sending coordination instructions to each regional agent to avoid inter-regional decision-making conflicts.

[0073] In a preferred embodiment of the present invention, the state space of the regional agent includes multi-dimensional features of the region, covering load features (e.g., real-time load, load forecast, load volatility, etc.), energy supply features (e.g., real-time output of renewable energy, available capacity of traditional energy, energy prices, etc.), equipment operation features (e.g., equipment utilization, fault status, energy consumption level, etc.), and environmental features (e.g., meteorological forecast data, extreme weather warnings, etc.). The dimensions of the state space are adjusted according to the scale and complexity of the region to ensure a comprehensive reflection of the region's operating status. The action space is defined as a set of scheduling operations that the regional agent can execute, including load adjustment strategies (e.g., load transfer, reduction of non-critical load, increase of delayable load, etc.), energy allocation strategies (e.g., adjusting the proportion of renewable energy consumption, traditional energy output plan, energy exchange between regions, etc.), and equipment control strategies (e.g., adjusting cooling system operating parameters, server cluster operating mode, etc.). The action space is represented by discretization or continuous representation, and an appropriate representation method is selected according to the characteristics of the scheduling operations.

[0074] Please refer to Figure 5 In this embodiment of the invention, establishing a cooperative-competitive game mechanism to guide the collaborative optimization of intelligent agents may include:

[0075] Step S410: Establish a two-tiered reward mechanism that combines global rewards with local rewards to balance global optimization and local optimization.

[0076] In a preferred embodiment of the present invention, the global reward comprehensively considers macro-level objectives such as system energy efficiency improvement, carbon emission reduction, and safe and stable operation of the power grid. The calculation method is as follows: the global reward equals the energy efficiency improvement weight multiplied by the energy efficiency improvement rate, plus the carbon emission reduction weight multiplied by the carbon emission reduction amount, minus the safety risk weight multiplied by the safety risk value. The weight coefficients are determined by the analytic hierarchy process to ensure that the importance of each objective matches the actual needs. The local reward focuses on local objectives such as reducing regional operating costs, ensuring supply and demand balance, and reducing equipment losses. The calculation method is as follows: the local reward equals the cost saving weight multiplied by the cost saving amount, plus the supply and demand balance weight multiplied by the supply and demand balance degree, minus the equipment loss weight multiplied by the equipment loss value.

[0077] Step S420: Using the Nash-Q learning algorithm, a multi-agent cooperative-competitive game is conducted. The global coordinating agent dynamically adjusts the reward weights of each regional agent by evaluating the global reward, guiding each regional agent to cooperate towards the global optimum.

[0078] In a preferred embodiment of the present invention, the Nash-Q learning algorithm is based on Nash equilibrium theory. It assumes that each agent believes that the policies of other agents are fixed and learns the optimal action under the current policies of other agents by iteratively updating its own Q-value function. During the game, each regional agent selects an action according to its own Q-value function, and at the same time receives coordination instructions from the global coordinating agent to adjust its own strategy and avoid decision-making conflicts between regions. The global coordinating agent dynamically adjusts the reward weights of each regional agent by evaluating the global reward, guiding each regional agent to cooperate towards the global optimum.

[0079] Step S430: Establish an inter-regional decision-making conflict coordination mechanism. When the actions of agents in different regions conflict (for example, simultaneously vying for the capacity of an energy transmission channel), the global coordinating agent formulates conflict solutions based on the severity of the conflict and the priority of each region.

[0080] In a preferred embodiment of the present invention, the priority is determined based on factors such as the load importance of the region, energy supply characteristics, and system security requirements. For example, the power supply priority of regions with concentrated computing load is higher than that of ordinary load regions, and the energy transmission priority of regions rich in renewable energy is higher than that of other regions under the new energy consumption target. Through the conflict coordination mechanism, the actions of the intelligent agents in each region are coordinated and consistent, thereby achieving global optimization.

[0081] In a preferred embodiment of the present invention, when designing the transfer learning strategy, parameter transfer involves using the parameters of the well-trained source region agent network (such as the weights and biases of the Actor and Critic networks) as the initialization parameters of the target region agent network, and only fine-tuning the last 1-2 layers of the network, thereby accelerating the convergence of the target region model by utilizing the training experience of the source region; knowledge transfer involves transferring the knowledge such as energy consumption patterns, scheduling rules, and game strategies learned during the training process of the source region to the target region in the form of a rule base or feature mapping, thereby helping the target region agent quickly understand its own operating environment and scheduling requirements.

[0082] In a preferred embodiment of the present invention, during course learning and sample transfer, the target region agent first learns to independently schedule tasks within a single region, becoming familiar with its own state space and action space. Then, it is introduced to perform simple interaction tasks with the source region, learning basic inter-regional coordination rules. Finally, it learns complex multi-regional interaction and conflict coordination tasks, comprehensively improving the target region agent's decision-making capabilities. Simultaneously, high-quality training samples from the source region are transferred to the target region to supplement insufficient training data. However, the transferred samples need adaptive adjustments, such as modifying relevant features based on the target region's load characteristics and energy supply parameters to ensure the samples match the target region's operating environment.

[0083] In a preferred embodiment of the present invention, an evaluation index system for transfer learning effectiveness is established, including the proportion of training time reduction, model convergence speed, and the effect of scheduling strategy optimization. The proportion of training time reduction is used to measure the improvement of training efficiency by transfer learning; the model convergence speed is measured by the number of convergence iterations of the Q-value function; the effect of scheduling strategy optimization is measured by indicators such as energy utilization efficiency, cost saving rate, and new energy consumption rate. If the transfer effect does not meet expectations, the transfer strategy needs to be adjusted, such as increasing the number of layers for parameter fine-tuning, optimizing the proportion of sample transfer, and adjusting the task difficulty of course learning, to ensure that the model performance of the target region agent meets the requirements.

[0084] Step S500: Deploy the trained multi-agent reinforcement learning game model into the digital twin environment, dynamically generate scheduling decisions, establish a security constraint verification system, comprehensively verify the scheduling decisions, and introduce a human intervention mechanism to achieve an organic combination of machine decision-making and human experience.

[0085] In a preferred embodiment of the present invention, a multi-agent reinforcement learning game model is encapsulated using Docker containerization technology. The model's code, dependency libraries, configuration files, etc., are packaged into a standardized container image to ensure the model's portability and consistency in different environments. Multiple simulation instances of different scenarios are run simultaneously in the digital twin environment, including normal operation scenarios, extreme weather scenarios, equipment failure scenarios, and load change scenarios. The model independently generates scheduling decisions for each scenario, evaluates the adaptability and optimization effect of the decisions under different scenarios, and predicts scheduling risks under extreme scenarios in advance through parallel simulation to optimize decision-making strategies. A data interaction feedback loop between the model and the digital twin environment is established. The digital twin environment pushes real-time status data to the model. After the model generates scheduling decisions, it feeds them back to the digital twin environment. The digital twin environment executes the decisions and simulates the decision effects, feeding back the simulation results (e.g., energy utilization efficiency, system stability indicators, cost changes, etc.) to the model again. The model dynamically adjusts its decision-making strategy based on the feedback results.

[0086] Please refer to Figure 6 In this embodiment of the invention, establishing a security constraint verification system to comprehensively verify scheduling decisions may include:

[0087] Step S510: Construct a safety constraint system covering three major categories: power grid safety constraints, equipment operation constraints, and energy supply constraints. Each constraint is based on industry standards and equipment parameters to formulate limit ranges and verification standards.

[0088] In a preferred embodiment of the present invention, power grid safety constraints include node voltage limits, line power flow limits, frequency stability limits, power angle stability limits, etc., to ensure that the power grid maintains safe and stable operation after the execution of dispatch decisions; equipment operation constraints include equipment rated capacity limits, temperature rise limits, start-stop frequency limits, operating time limits, etc., to ensure that the equipment does not exceed the safe operating range; energy supply constraints include renewable energy output fluctuation limits, traditional energy minimum output limits, energy storage equipment charging and discharging power limits, etc., to ensure the stability and continuity of energy supply.

[0089] Step S520: Use fast power flow calculation and constraint satisfaction check algorithm to check safety constraints. If the scheduling decision violates a constraint, the system automatically marks the type and degree of constraint violation and triggers the decision correction mechanism.

[0090] In a preferred embodiment of the present invention, fast power flow calculation is based on simplified power flow equations to quickly calculate the grid operation parameters such as node voltage and line power flow after the execution of scheduling decisions, and compare them with the grid safety constraint limits; the constraint satisfaction check algorithm verifies one by one whether the scheduling decision meets the equipment operation constraints and energy supply constraints, for example, checking whether the equipment operation parameters exceed the rated capacity and whether the energy supply can meet the load demand.

[0091] Step S530: Establish an automatic correction mechanism for scheduling decisions. For scheduling decisions that violate constraints, adjust the parameters of the scheduling decisions using heuristic algorithms based on the importance and scope of the constraints (e.g., reduce the amount of energy transmitted beyond the line power flow limit, reduce the amount of load allocation beyond the rated capacity of the equipment, and ensure that the corrected decision meets all safety constraints).

[0092] In a preferred embodiment of the present invention, the Monte Carlo simulation method is used to analyze the robustness of the decision-making process. Uncertain factors such as load forecasting error and renewable energy output forecasting error are injected to test the stability and reliability of the scheduling decision under different disturbance conditions. If the robustness success rate of the decision within the error range exceeds 95%, it is determined to pass the verification; otherwise, the decision parameters are further optimized to improve the robustness.

[0093] Please refer to Figure 7 In this embodiment of the invention, a human intervention mechanism is introduced to achieve an organic combination of machine decision-making and human experience, which may include:

[0094] Step S501: Design a visual dashboard for scheduling decisions, which will visualize the generated scheduling decisions, safety verification results, system operating status, and abnormal warning information, and trigger audible and visual warnings for high-risk scheduling decisions.

[0095] In a preferred embodiment of the present invention, the dashboard is divided into a decision details area, a risk assessment area, and a status monitoring area. The decision details area displays the scheduling actions, execution basis, and expected effects of each area. The risk assessment area uses red, yellow, and green to mark the decision risk level. Red represents high-risk decisions (e.g., involving overload of critical equipment or power outage risk of important loads), which require mandatory manual review. Yellow represents medium-risk decisions (e.g., some constraints are close to the limit), which can be selectively reviewed. Green represents low-risk decisions (e.g., all constraints are fully met, and the optimization effect is significant), which can be executed automatically. The status monitoring area displays the key operating parameters and abnormal situations of the system in real time.

[0096] Step S502: Design a manual intervention interface. Operators can adjust, reject, or regenerate the generated scheduling decisions based on actual operating conditions, experience, and external instructions. Operation records and backtracking are also supported.

[0097] In a preferred embodiment of the present invention, the intervention methods include parameter adjustment, action modification, and target weight adjustment. Parameter adjustment allows operators to modify specific values ​​in the decision (e.g., energy export volume, load adjustment ratio, etc.). Action modification allows operators to add, delete, or replace scheduling actions. Target weight adjustment allows operators to adjust the weights of targets such as energy efficiency, cost, and safety. The model recalculates the scheduling strategy in real time based on the adjusted weights.

[0098] Step S503: Establish a human-machine collaborative decision-making iterative optimization mechanism, using the operator's intervention actions and decision feedback as incremental training data, and update the model parameters through incremental learning technology.

[0099] For example, if an operator adjusts the safety target weights of a model in an extreme weather scenario, the model can incorporate this adjustment experience into subsequent training and automatically optimize decision-making strategies in similar scenarios. Through human-machine collaborative iteration, machine decision-making and human experience can mutually promote each other and continuously improve the level of scheduling decisions.

[0100] Step S600: Based on the execution results of scheduling decisions, establish a multi-dimensional performance evaluation system, quantify the implementation effect of the execution results and identify problems, update the multi-agent reinforcement learning game model through incremental learning technology, and compare the simulation data of the digital twin environment with the operation data of the physical power grid system to dynamically calibrate the parameters of the digital twin environment.

[0101] Please refer to Figure 8 In this embodiment of the invention, establishing a multi-dimensional performance evaluation system to quantify the implementation effect of the execution results and identify problems may include:

[0102] Step S610: Based on industry standards and project objectives, formulate multi-dimensional performance evaluation indicators, including at least economic indicators (e.g., unit computing power energy consumption cost, total regional energy procurement cost, equipment operation and maintenance cost saving rate, return on investment, etc., to measure the economics of scheduling decisions), technical indicators (e.g., energy utilization efficiency, renewable energy absorption rate, voltage qualification rate, line power flow qualification rate, load power supply reliability, equipment utilization rate, etc., to measure the impact of scheduling decisions on system operation technical performance), environmental indicators (e.g., carbon emissions, fossil energy consumption reduction, pollutant emission reduction rate, etc., to evaluate the low-carbon and environmental protection effects of scheduling decisions), and safety indicators (e.g., number of safety constraint violations, fault recovery time, system stability margin, etc., to measure the safety of scheduling decisions).

[0103] Step S620: Establish an evaluation data collection mechanism to collect operational data (e.g., energy consumption data, cost data, equipment operation data, environmental data, safety event records, etc.) after the execution of scheduling decisions. The data collection frequency should be consistent with the scheduling decision cycle, and real-time evaluation, periodic evaluation, and special evaluation should be carried out.

[0104] In a preferred embodiment of the present invention, real-time evaluation is performed after each scheduling decision is executed to quickly calculate key indicators (e.g., voltage qualification rate, renewable energy absorption rate) and promptly identify serious problems; periodic evaluation (e.g., daily, weekly) comprehensively evaluates the scheduling effect over a period of time and analyzes the trend of indicator changes; special evaluation is conducted for special scenarios (e.g., extreme weather, equipment failure) to summarize lessons learned.

[0105] Step S630: Using a preset analysis method, conduct an in-depth analysis of the evaluation results, identify the causes of the differences and the weak points, and display the evaluation results through visualization charts.

[0106] In a preferred embodiment of the present invention, a dedicated incremental data pool is established to store recent (e.g., one month) operational data, scheduling decision execution results, performance evaluation data, and records of manual intervention. An incremental learning algorithm based on gradient descent is employed, utilizing new data from the incremental data pool to fine-tune the model's network parameters. A model update and verification mechanism is established, where the updated model is simulated and verified in a digital twin environment across multiple scenarios. The performance metrics (e.g., optimization effect, robustness, response speed) of the model under different scenarios are evaluated to ensure that the updated model outperforms the original model. After successful verification, the model is deployed using a rolling update approach to avoid system interruptions caused by model replacement.

[0107] In a preferred embodiment of the present invention, actual operating data of the physical system (e.g., node voltage, line power flow, equipment energy consumption, renewable energy output, etc.) are periodically collected and compared with simulation data from the digital twin environment during the same period. The simulation error of key parameters is calculated, and the error indicators include absolute error, relative error, root mean square error, etc. For parameters and models with large errors, differentiated calibration strategies are formulated. For equipment parameters (e.g., line impedance, equipment energy consumption parameters), the least squares method is used to adjust the parameter values ​​in reverse to minimize the error between the simulation data and the actual data. For renewable energy output models (e.g., photovoltaic and wind power models), the model parameters are refitted by combining new meteorological data and actual output data to optimize the model's prediction accuracy. For environmental coupling models, new environmental influencing factors (e.g., the impact of extreme temperatures on equipment efficiency) are added to improve the model structure. After parameter calibration, multi-scenario simulation verification is performed in the digital twin environment. The simulation errors before and after calibration are compared to evaluate the calibration effect. The verification indicators include the percentage reduction in the error of key parameters, the improvement in the overall system simulation accuracy, and the consistency between the scheduling decision verification results and the actual execution results.

[0108] Step S700: Adapt and adjust the intelligent power dispatching method to different application scenarios, plan the promotion and application path, and reduce the risk of promotion and application through compliance verification.

[0109] In a preferred embodiment of the present invention, the structural characteristics, operational requirements, and constraints of different application scenarios are analyzed in depth, and the core modules of the technical solution are adapted and adjusted accordingly to ensure that the technical solution can accurately match the needs of the scenario and achieve optimal results. A clear and feasible technology promotion path is formulated, covering four stages: technical document preparation, training system construction, pilot application verification, and large-scale promotion, to ensure that the technical solution can be successfully implemented and its promotion scope can be gradually expanded. The technical solution is ensured to comply with relevant industry standards and specifications, and compliance verification is conducted to reduce the risks of promotion and application and enhance the industry recognition and credibility of the technical solution.

[0110] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0111] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0112] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0113] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0114] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0115] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0116] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0117] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0118] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A multi-regional intelligent power energy dispatching method based on big data, characterized in that, The intelligent power energy dispatching method includes: Based on the sensor network deployed in the target power area, multimodal data is collected, data is cleaned and spatiotemporally aligned, and a structured feature dataset is constructed through standardization processing. Based on the feature dataset, a power grid topology model consistent with the parameters of the physical power grid system is constructed, and a multi-physics model is integrated to build a dynamic simulation system and establish a data interaction channel between the digital twin environment and the physical power grid system. Using a pre-defined clustering algorithm, typical energy consumption patterns of the system are identified. Combined with a pre-defined deep learning algorithm, a multi-dimensional anomaly detection system is constructed to detect abnormal operating states of the system and perform core feature extraction and dimensionality reduction. Based on the core features after dimensionality reduction, a multi-agent architecture is designed, a cooperative-competitive game mechanism is established to guide the agents to optimize collaboratively, and transfer learning technology is used to accelerate model training, resulting in a trained multi-agent reinforcement learning game model. The trained multi-agent reinforcement learning game model is deployed in a digital twin environment to dynamically generate scheduling decisions, establish a security constraint verification system, conduct comprehensive verification of scheduling decisions, and introduce a human intervention mechanism to achieve an organic combination of machine decision-making and human experience.

2. The multi-regional intelligent power energy dispatching method based on big data according to claim 1, characterized in that, The acquisition of multimodal data includes constructing a three-dimensional data acquisition system encompassing power, environment, and operation; the data cleaning and spatiotemporal alignment include: The isolated forest algorithm in unsupervised learning is used to identify outliers in multimodal data. The sliding window mean method is used for imputation and repair, and linear interpolation is used to supplement missing data. Multimodal data is resampled using a preset time base, and multimodal data with different sampling frequencies are adjusted to a uniform time interval. A unified coding system is established, and multimodal data are associated according to the coding. The Kalman filter algorithm is used to fuse multi-source homogeneous data from multiple modalities, establish state equations and observation equations, and dynamically estimate the true value of the data.

3. The multi-regional intelligent power energy dispatching method based on big data according to claim 1, characterized in that, The construction of a power grid topology model consistent with the parameters of the physical power grid system, and the integration of a multiphysics model to build a dynamic simulation system, includes: Import wiring diagram data conforming to the general information model standard from the power grid GIS system, and store the topology relationships using a pre-set graph database; Using a pre-set professional 3D engine, a power grid topology visualization scene is constructed, and differentiated power grid topology models are designed according to equipment type and functional characteristics; The rationality and accuracy of the power grid topology model are verified by power flow calculation, and it is checked whether the power grid topology model meets the operating rules of the physical power grid system. To integrate the electrical simulation physical field model into the power grid topology model, the Newton-Raphson method is used for power flow calculation, and steady-state simulation of large-scale power grid systems is performed. The renewable energy output model is integrated into the power grid topology model, and environmental data is dynamically correlated with energy output. Environmental data is accessed into the simulation system through an interface, and the renewable energy output is dynamically adjusted. To integrate the energy consumption model of the green computing center equipment and the operation model of the power grid equipment into the power grid topology model, a dynamic simulation system is constructed to simulate the electrical characteristics, environmental impact and equipment response of the physical power grid system under different operating scenarios.

4. The multi-regional intelligent power energy dispatching method based on big data according to claim 1, characterized in that, The construction of a multi-dimensional anomaly pattern detection system to detect abnormal system operation states and perform core feature extraction and dimensionality reduction includes: A joint detection strategy combining pre-defined statistical methods and deep learning is used to screen extreme anomaly patterns and identify complex anomaly patterns. The detected abnormal patterns are classified, and corresponding abnormal features are extracted for different types of abnormal patterns. When an anomalous pattern is detected, the root cause analysis process is initiated to trace back information in the digital twin environment and determine the root cause of the anomalous pattern through correlation analysis and causal reasoning. The t-SNE algorithm is used to reduce the dimensionality of high-dimensional features in the outlier features while retaining the core information of the high-dimensional features.

5. The multi-regional intelligent power energy dispatching method based on big data according to claim 1, characterized in that, The establishment of a cooperative-competitive game mechanism to guide the collaborative optimization of intelligent agents includes: Establish a two-tiered reward mechanism that combines global rewards with local rewards to balance global optimization and local optimization; The Nash-Q learning algorithm is used to conduct multi-agent cooperative-competitive game. The global coordinating agent dynamically adjusts the reward weight of each regional agent by evaluating the global reward, thereby guiding the regional agents to cooperate towards the global optimum. Establish a regional decision-making conflict coordination mechanism. When there is a conflict between the actions of agents in different regions, the global coordinating agent formulates a conflict resolution solution based on the severity of the conflict and the priority of each region.

6. The multi-regional intelligent power energy dispatching method based on big data according to claim 1, characterized in that, The establishment of a security constraint verification system to comprehensively verify scheduling decisions includes: A safety constraint system covering three major categories—power grid security constraints, equipment operation constraints, and energy supply constraints—is constructed, with each constraint having its limit range and verification standard based on industry standards and equipment parameters. Fast power flow calculation and constraint satisfaction checking algorithms are used to verify safety constraints. If a scheduling decision violates a constraint, the system automatically marks the type and degree of constraint violation and triggers a decision correction mechanism. Establish an automatic correction mechanism for scheduling decisions. For scheduling decisions that violate constraints, adjust the parameters of the scheduling decision using a heuristic algorithm based on the importance and scope of the constraints.

7. The multi-regional intelligent power energy dispatching method based on big data according to claim 6, characterized in that, The introduction of a human intervention mechanism to organically combine machine decision-making with human experience includes: Design a visual dashboard for scheduling decisions, which will visually display the generated scheduling decisions, safety verification results, system operating status, and abnormal warning information, and trigger audible and visual warnings for high-risk scheduling decisions; The interface for manual intervention is designed so that operators can adjust, reject, or regenerate the generated scheduling decisions based on actual operating conditions, experience, and external instructions. Operation records and backtracking are also supported. Establish a human-machine collaborative decision-making iterative optimization mechanism, using the operator's intervention actions and decision feedback as incremental training data, and update the model parameters through incremental learning technology.

8. The multi-regional intelligent power energy dispatching method based on big data according to claim 1, characterized in that, The intelligent power energy dispatching method also includes: Based on the execution results of scheduling decisions, a multi-dimensional performance evaluation system is established to quantify the implementation effect of the execution results and identify problems. Through incremental learning technology, the multi-agent reinforcement learning game model is updated, and the simulation data of the digital twin environment is compared with the operation data of the physical power grid system to dynamically calibrate the parameters of the digital twin environment.

9. The multi-regional intelligent power energy dispatching method based on big data according to claim 8, characterized in that, The establishment of a multi-dimensional performance evaluation system to quantify the implementation effect of the execution results and identify problems includes: Develop multi-dimensional performance evaluation indicators based on industry standards and project objectives, including at least economic, technical, environmental and safety indicators. Establish an evaluation data collection mechanism to collect operational data after the execution of scheduling decisions. The data collection frequency should be consistent with the scheduling decision cycle, and real-time evaluation, periodic evaluation, and special evaluation should be carried out. Using a pre-defined analysis method, the evaluation results are analyzed in depth to identify the causes of the differences and the weak points, and the evaluation results are displayed through visual charts.

10. The multi-regional intelligent power energy dispatching method based on big data according to claim 8, characterized in that, The intelligent power energy dispatching method also includes: For different application scenarios of intelligent power dispatch, the intelligent power dispatch method is adapted and adjusted, the promotion and application path is planned, and the risks of promotion and application are reduced through compliance verification.

Images