Urban power grid real-time load collaborative peak regulation method based on multi-energy complementation and AI scheduling

Abstract

The invention discloses an urban power grid real-time load collaborative peak regulation method based on multi-energy complementation and AI scheduling. The method comprises the following steps of multi-source data access and high-dimensional feature space construction, embedded entropy calculation and interactive network construction, domain knowledge and data-driven model fusion, hierarchical scheduling and dual-stage optimization, and real-time decision and closed-loop feedback. According to the method, the real-time performance and hierarchical scheduling thought are emphasized, and an organic closed loop is formed on the three aspects of intra-day scheduling, hour-level rolling correction and minute-level or second-level emergency response. Meanwhile, by means of a multi-stage optimizer switching mechanism, the model can complete rapid convergence of high-dimensional parameters in a short time, finer strategy fine adjustment is carried out in the later period, and the accuracy and reliability of a peak regulation scheme are guaranteed; the method can be applied to advanced power grid systems such as intelligent power grid dispatching, a multi-energy collaborative optimization platform and demand side response management, and has the characteristics of high real-time performance, strong adaptability and good expandability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems and smart grid, in particular to a city power grid real-time load collaborative peak shaving method based on multi-energy complementation and AI scheduling. BACKGROUND

[0002] With the continuous acceleration of global urbanization process, the power demand of city power grid in peak period is becoming increasingly severe. The traditional thermal power unit scheduling mode is difficult to balance the real-time and environmental friendliness of the system, which increases the power generation cost and leads to additional carbon emissions. In recent years, the proportion of renewable energy such as wind power and photovoltaic power in the city power structure has gradually increased, making the new power supply mode of "multi-energy complementation and source-load collaboration" an important direction for future development. However, the fluctuation characteristics of wind power and photovoltaic power bring greater uncertainty to system operation. When the wind speed or irradiance changes dramatically, the load side of the power grid needs to make quick and accurate adjustments. If the traditional peak shaving means is still used, it is often difficult to respond efficiently to the instantaneous peak load in a short time, which may lead to risks such as frequency fluctuation and line loss growth in the power consumption peak period.

[0003] In this context, more and more urban grid dispatching studies are turning their attention to the "multi-energy complementary" strategy that combines energy storage systems and renewable energy. On the one hand, energy storage systems such as battery storage and pumped storage can store electricity during low load periods and release it during peak demand periods, thereby smoothing the load curve and reducing the peak-to-valley difference. On the other hand, with the increasing scale of distributed photovoltaic and distributed wind power, grid operation and maintenance must cope with more complex tidal changes and random fluctuations. To address this complexity, a series of dispatch optimization methods based on artificial intelligence (AI) or machine learning have emerged in recent years. For example, the Energy Systems Laboratory at Stanford University proposed a reinforcement learning-based method to coordinate wind farm and energy storage output [Anand et al., IEEE Transactions on Smart Grid, 2023, Transient Stability Constrained Social Welfare Maximization with Demand Response in Smart Electrical Grid]; the Department of Electrical Engineering at Tsinghua University implemented a multi-stage rolling optimization in its "Multi-source Coupling Dispatching Platform" project to dynamically control the output of different types of renewable energy [Zhang et al., Applied Energy, 2022, Renewable energy systems for building heating, cooling and electricity production with thermal energy storage]. These studies have improved the flexibility and economy of urban power grids to some extent.

[0004] However, most existing multi-energy complementary dispatching schemes still remain at the single time scale or offline optimization level. Traditional linear programming or mixed integer programming (MILP) can usually obtain a set of static intraday dispatching plans under given constraints, but when sudden situations such as sudden drop in renewable energy output or sudden surge in load occur in actual operation, the system often cannot quickly correct the peak shaving strategy in time. In addition, many data-driven models lack a full understanding of the physical characteristics of the power system, such as tidal distribution, frequency and voltage safety constraints, and device start-stop characteristics, resulting in their ability to function only when short-term prediction accuracy is high; once there are abnormalities in the input data, or the system size is large enough, their prediction and dispatching effect will be significantly degraded.

[0005] In the peak regulation link, another key problem is how to effectively train and deploy artificial intelligence models. When the power grid is large and there are many types of units and energy storage devices, the optimization variables grow exponentially, and the traditional gradient descent method is prone to slow convergence speed or local optimal problems. Therefore, some foreign teams try to use different stage optimizer switching techniques in the training process, such as first using Adam with adaptive learning rate to update the network parameters in large steps, and then using L-BFGS with second-order information for fine adjustment. This multi-stage optimization idea helps to significantly reduce the training time, and can reduce gradient oscillation while improving the final accuracy of the model [Wang et al., Energy AI, 2024, Does artificial intelligence promote energy transition and curb carbon emissions?The role of trade openness]. However, most of these works are limited to laboratory or small and medium-sized power grid scenarios, and there is no mature productization solution to realize landing in complex urban power grids.

[0006] The current multi-energy complementary and AI scheduling technology mainly faces the following bottlenecks when dealing with peak load of large-scale urban power grids:

[0007] Lack of real-time performance: It is difficult to link and control the load and distributed new energy fluctuations within minutes or even seconds, resulting in a lag in the scheduling scheme.

[0008] Limited scalability: When the power grid expands and the number of device types increases, the difficulty of model training and optimization increases nonlinearly.

[0009] Insufficient mechanism integration: Data-driven predictions often ignore physical constraints when devices start and stop; while linear or nonlinear programming based on mechanism models alone, it is easy to produce excessive computational overhead and local optimal solutions.

[0010] It is difficult to balance the cost and safety of peak regulation in a high-penetration renewable environment: The coupling degree of energy storage, wind power and thermal power is high, and single-dimensional or single-time-scale optimization cannot maximize the overall benefit. SUMMARY

[0011] In order to effectively solve these problems and fully exert the potential of multi-energy complementation, the present application proposes a city power grid real-time load collaborative peak shaving method based on multi-energy complementation and AI scheduling. Compared with the traditional method, the present application emphasizes real-time and hierarchical scheduling ideas, and forms an organic closed loop at three levels of intraday scheduling, hour-level rolling correction and minute or second-level emergency response. At the same time, with the help of multi-stage optimizer switching mechanism, the model can complete the rapid convergence of high-dimensional parameters in a short time, and carry out more fine strategy fine-tuning in the later period, guaranteeing the accuracy and reliability of the peak shaving scheme. Through a series of innovative means, the present application not only can effectively reduce the peak value of city power grid load in peak period, but also can flexibly call energy storage and conventional unit resources in the case of random fluctuation of renewable energy, realizing the maximization of green energy utilization. It can be applied to advanced power grid systems such as smart grid scheduling, multi-energy collaborative optimization platform and demand side response management,

[0012] The technical scheme of the present application is specifically introduced as follows.

[0013] The present application provides a city power grid real-time load collaborative peak shaving method based on multi-energy complementation and AI scheduling, comprising the following steps:

[0014] (1) Multi-source data access and high-dimensional feature space construction

[0015] Real-time renewable energy generation data, energy storage device state data, power grid load data and environmental data are collected, and preprocessing operations including cleaning, missing data completion and standardization are performed on the accessed multi-source heterogeneous data; after the preprocessing is completed, based on the obtained clean and dimensionally unified data, a high-dimensional time-space feature matrix containing static features and dynamic features is constructed to realize the unified representation of heterogeneous data while retaining the inherent characteristics of various data;

[0016] (2) Embedding entropy calculation and interaction network construction

[0017] Based on the preprocessed multi-source time series data, the embedding entropy of each node of the system is calculated to quantify the complexity and uncertainty of the node; based on the embedding entropy matrix, the interaction influence network between the nodes of the power grid is constructed to quantify the information flow and causal relationship;

[0018] (3) Domain knowledge and data-driven model fusion

[0019] Integrating power system domain knowledge and data-driven model, converting power system professional knowledge into prior information and constraint conditions available for data-driven model, constructing hybrid intelligent model; data-driven model includes supergraph neural network layer and space-time attention layer, supergraph neural network layer is responsible for modeling high-order relationship between multiple nodes, each superedge connects multiple nodes, representing complex group interaction; the space-time attention layer enhances the model's ability to perceive key space-time patterns and dynamically adjusts the importance weights of different features and time steps; the model fusion adopts an adaptive ensemble learning framework that dynamically adjusts the weights of different models based on real-time system state, balancing the accuracy and computational efficiency of the model;

[0020] (4) Hierarchical scheduling and two-stage optimization

[0021] A two-stage optimizer is used to train the hybrid intelligent model. Based on the trained hybrid intelligent model, hierarchical scheduling is implemented in both time and space dimensions to generate scheduling strategies. The two-stage optimizer is used to optimize model parameters in the training stage and optimize control / resource allocation variables in the scheduling stage.

[0022] (5) Real-time decision-making and closed-loop feedback

[0023] The generated peak shaving instructions are sent to each execution unit through the SCADA system or energy management system, and the peak shaving effect is monitored and evaluated in real time. Based on the monitoring and evaluation results, the system implements closed-loop feedback optimization to continuously improve the peak shaving effect.

[0024] In summary, the present application proposes an intelligent peak shaving method with high real-time performance, strong adaptability, and good scalability for multi-energy complementation and load real-time prediction, to solve the technical bottleneck of large-scale urban power grids in achieving fast and accurate scheduling under peak load and random fluctuations of renewable energy. Specifically, the present application implements a multi-time scale and hierarchical intelligent scheduling control method, significantly improving the training speed and execution accuracy of AI prediction models, optimizing source-load collaboration and energy storage management capabilities, and ensuring the safe and economic operation of power grids under high proportions of renewable energy. The main technical advantages and innovations of the present application include:

[0025] Unified representation and fusion of multi-source heterogeneous data: innovatively mapping heterogeneous data of renewable energy, energy storage devices, and power grid load side to a high-dimensional feature space, achieving seamless fusion and unified analysis of data, overcoming the limitations of traditional methods in handling multi-source heterogeneous data. Compared to traditional parallel data processing methods, the unified representation method of the present application can more effectively capture the internal relationships between data, improving the prediction accuracy and generalization ability of the model.

[0026] Embedding Entropy-Driven Dynamic System Analysis: For the first time, embedding entropy theory is applied to power grid interaction network analysis, providing a completely new perspective to understand and quantify the complex interaction between nodes in the grid. This method not only identifies explicit direct effects, but also finds implicit indirect effects and cascading effects, making system analysis more comprehensive and in-depth.

[0027] Deep Fusion of Domain Knowledge and Data-Driven Models: Breaks through the limitations of traditional models that rely too much on physical knowledge or rely too much on data, realizing the complementary advantages of the two paradigms. By encoding power system expert knowledge into network structure and training constraints, the model not only has data-driven adaptability and learning ability, but also maintains physical interpretability and robustness.

[0028] High-order relationship modeling based on hypergraph neural network: Using hypergraph neural network model to represent the multi-node high-order relationship in the power grid, breaking through the limitation of traditional graph neural network that can only handle pairwise relationships. Each hyperedge can connect multiple nodes, accurately describing the group interaction mode, so that the model can capture more complex system dynamics.

[0029] Efficient training method of two-stage optimizer: The innovative two-stage optimization strategy significantly improves the efficiency and quality of model training. The global exploration stage quickly locates the feasible solution region, and the local refinement stage fine-tunes the key parameters. The two work together to realize the efficient optimization process of "coarse tuning + fine tuning". Compared with traditional single optimization method, the two-stage optimizer shortens the training time by about 50%, while improving the quality and stability of the solution.

[0030] Real-time adaptability of hierarchical scheduling architecture: The hierarchical scheduling architecture in time and space dimensions enables the system to not only plan for the long term, but also quickly respond to real-time changes, greatly improving the flexibility and adaptability of scheduling. The coordination mechanism between different levels ensures the consistency and continuity of decisions, avoiding the conflicts and oscillation problems commonly seen in traditional multi-level scheduling. The response speed of the real-time layer reaches minutes, meeting the demand for fast peak shaving of urban power grids.

[0031] End-to-end closed-loop feedback optimization mechanism: The complete closed-loop feedback design makes the entire system form a self-optimizing agent that can continuously learn and improve from execution results. This mechanism enables the system to have the ability to continuously evolve, adapting to changing power grid environments and load characteristics, maintaining long-term efficient operation. The closed-loop feedback mechanism improves the average performance of the system in long-term operation by about 15-20%. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The overall architecture diagram of the urban power grid real-time load coordinated peak shaving method based on multi-energy complementation and AI scheduling described in the present invention.

[0033] Figure 2 Flow chart of multi-source data access and high-dimensional feature space construction of the application.

[0034] Figure 3 Schematic diagram of embedded entropy dynamic analysis and node interaction quantization of the application.

[0035] Figure 4 Domain knowledge and data-driven model fusion framework diagram of the application.

[0036] Figure 5 Hierarchical scheduling and two-stage optimizer workflow diagram of the application.

[0037] Figure 6 Real-time decision-making and closed-loop feedback mechanism schematic diagram of the application. DETAILED DESCRIPTION

[0038] The application is committed to building a real-time peak shaving solution for city-level power systems, using integrated technical means such as hierarchical scheduling, data and mechanism fusion, and multi-stage optimizer switching to overcome the real-time and scalability bottlenecks encountered by multi-energy complementation and AI scheduling in large-scale scenarios, providing new ideas and methods for the safe, stable and efficient operation of city power grids.

[0039] In order to better explain the implementation method of the application, the following will combine specific steps and key parameters in the load peak shaving scenario of a super large city power grid to describe the implementation process of the application in detail. Through the integrated operation of multiple modules, the system can realize real-time load collaborative peak shaving based on multi-energy complementation and AI scheduling.

[0040] The application designs a five-function module including data acquisition and preprocessing, load and output prediction, multi-energy complementation modeling and power flow analysis, multi-stage optimization scheduling and execution feedback monitoring:

[0041] I. Multi-source data access and high-dimensional feature space construction module

[0042] This module is responsible for real-time acquisition and processing of wind power, photovoltaic, energy storage, load and weather data in city power grids. Specifically, it includes acquisition of frequency data above minutes, data cleaning, interpolation, normalization processing, and automatic detection and correction of outliers. Multi-scale wavelet transform and adaptive filtering algorithm are used to process missing values and outliers to improve data quality. Nonlinear manifold learning method is used for feature space construction, which realizes unified representation of heterogeneous data through adaptive weight mapping while preserving the inherent characteristics of various data.

[0043] II. Embedded entropy dynamic analysis and node interaction quantization module

[0044] The concept of embedding entropy is introduced to dynamically analyze multi-source time series information and accurately quantify the high-order interaction between nodes in the urban power grid. This module first constructs a multivariate embedding entropy matrix to calculate the mutual information and conditional mutual information between different nodes, identify potential causal relationships and coupling patterns. Then, through time delay embedding technology and phase space reconstruction, the dynamic evolution characteristics of the system are analyzed, and nonlinear and non-stationary interaction patterns are captured. Finally, based on the concepts of transfer entropy and causal entropy, the information flow and influence transmission between different nodes are quantified, and a complete interaction influence network is constructed to provide accurate information support for subsequent peak shaving decisions.

[0045] III. Domain knowledge and data-driven model fusion module

[0046] The power system domain knowledge model (such as power flow analysis, node sensitivity, safety constraints) and data-driven model (such as hypergraph neural network, spatio-temporal attention network) are innovatively integrated to build a hybrid intelligent model with physical interpretation. This module uses knowledge distillation and transfer learning techniques to encode expert knowledge in the power system as prior constraints and structured biases, guiding the learning process of the neural network. Hypergraph neural networks effectively capture complex interactions between multiple nodes through high-order relationship modeling, while spatio-temporal attention mechanisms enhance the model's ability to perceive key spatio-temporal patterns. The model fusion uses an adaptive ensemble learning framework to dynamically adjust the weights of different models according to real-time system state, balancing model accuracy and computational efficiency.

[0047] A hierarchical scheduling architecture is designed in both time and space dimensions, and an innovative two-stage optimizer is equipped to realize fast training and optimization of the hypergraph neural network. In the time dimension, it is divided into day-ahead planning, hour-level adjustment, and real-time response; in the spatial dimension, it is divided into system-level, regional-level, and device-level. The two-stage optimizer includes a global exploration stage and a local refinement stage: the global exploration uses an improved particle swarm algorithm to quickly search the large-scale solution space; the local refinement is based on gradient descent and Lagrange multiplier method to fine-tune key parameters. The optimization objective is a multi-objective function, considering multiple dimensions such as minimum peak-valley difference, minimum peak shaving cost, and maximum system stability.

[0048] V. Real-time decision and closed-loop feedback module

[0049] Based on the analysis results of the preceding modules, a load transfer strategy and peak shaving plan are generated in real time, and the decision quality is continuously optimized through a closed-loop feedback mechanism. The decision generation adopts a model predictive control (MPC) framework to solve the optimal control sequence in a rolling time domain, while considering system constraints and uncertainties. The execution strategy includes direct control signals (such as energy storage charging and discharging instructions) and indirect incentive signals (such as real-time electricity price adjustments), forming a multi-level regulation system. The closed-loop feedback mechanism monitors the system response in real time, evaluates the decision effect, and dynamically adjusts the model parameters and decision strategy according to the deviation, achieving adaptive optimization of the system. In addition, the module also contains an emergency response mechanism that can quickly identify abnormal situations and trigger pre-set emergency peak shaving schemes.

[0050] The operation flow of the present application specifically comprises the following steps:

[0051] (1) System initialization and data access

[0052] Configure system parameters, establish connection channels for various data sources, initialize databases and computing engines. Real-time access to renewable energy generation data (such as photovoltaic power, wind power), energy storage device state data (such as state of charge SOC, charging and discharging power), grid load data (such as partition load, typical user load curve) and environmental data (such as meteorological data, temperature and humidity). Standardize the raw data accessed to ensure data format consistency and time synchronization.

[0053] (2) Data preprocessing and feature space construction

[0054] Clean, complete and standardize the multi-source heterogeneous data accessed. Use wavelet transform and Kalman filter methods to remove noise and outliers, and use interpolation algorithms and deep learning methods to process missing data. Construct a spatio-temporal feature matrix, including static features (such as device parameters, network topology) and dynamic features (such as load change rate, power fluctuation characteristics). Through nonlinear manifold learning methods, map different types of features to a unified high-dimensional feature space to form a standardized representation of data.

[0055] (3) Embedded entropy calculation and interaction network construction

[0056] Based on the preprocessed data, calculate the single-variable embedded entropy of each node to evaluate the complexity and uncertainty of the node. Construct a multivariate embedded entropy matrix to calculate the mutual information and conditional mutual information between node pairs, quantifying the degree of information coupling. Based on the principle of transfer entropy, analyze the direction and strength of information flow, identify causal relationships and influence paths. Construct a complete interaction influence network, mark key nodes and key connections, and provide a decision basis for subsequent load regulation. At the same time, based on the time sequence changes of the interaction network, analyze the dynamic evolution law of the system, and predict potential unstable factors and risk points.

[0057] (4) Hybrid intelligent model training and optimization

[0058] Integrate domain knowledge (e.g., power flow equations, security constraints) and data-driven models to build a hybrid intelligent architecture. Hypergraph neural network layers model high-order relationships between multiple nodes, with each hyperedge connecting multiple nodes to represent complex group interactions. Spatio-temporal attention layers enhance the model's ability to perceive key spatio-temporal patterns and dynamically adjust the importance weights of different features and time steps. A two-stage optimizer is used to train the model: the first stage uses an improved particle swarm optimization algorithm to perform global exploration in a large solution space, quickly locating potential optimal solution regions; the second stage uses a gradient-based optimization method for local refinement to fine-tune model parameters. Distributed computing framework is used for model training, supporting incremental learning and online updating to maintain the timeliness of the model.

[0059] (5) Hierarchical dispatching strategy generation

[0060] Based on the trained hybrid intelligent model, hierarchical dispatching is implemented in both time and space dimensions. In the time dimension, the dispatching is gradually refined from day-ahead planning to real-time response; in the space dimension, the dispatching is decomposed from system level to device level. For the day-ahead planning layer, load forecasting, renewable energy output forecasting, and energy storage status are considered to develop a preliminary 24-hour peak shaving plan. For the hourly adjustment layer, short-term forecasts and real-time monitoring data are used to optimize and adjust the dispatching plan for the next 4-6 hours. For the real-time response layer, based on the latest system state, precise control instructions are generated at a minute-level time granularity, including energy storage charging and discharging power, adjustable load adjustment amount, etc. All levels of dispatching decisions meet system operation constraints and safety requirements.

[0061] (6) Peak shaving instruction execution and effect monitoring

[0062] The generated peak shaving instructions are sent to various execution units through the SCADA system or energy management system, including energy storage system controllers, renewable energy controllers, and load-side response controllers. Real-time monitoring of instruction execution and system response is performed, and key performance indicators such as actual peak-valley difference, peak shaving cost, and system stability indicators are collected. A multi-dimensional evaluation system is established to comprehensively evaluate the peak shaving effect. In response to deviations and abnormalities during execution, appropriate corrective measures are triggered to ensure the effective implementation of the peak shaving plan.

[0063] (7) Closed-loop feedback and continuous optimization

[0064] Based on the peak shaving effect evaluation results, a closed-loop feedback mechanism is formed. The deviation between the expected peak shaving effect and the actual effect is compared, and the reasons for the deviation are analyzed, including model prediction error, execution delay, external interference, etc. According to the analysis results, the model parameters, optimizer configuration and decision strategy are dynamically adjusted to realize the adaptive optimization of the system. Periodically, the historical data and optimization effect are deeply mined to identify long-term trends and patterns, providing basis for model upgrade and strategy improvement. At the same time, expert knowledge and successful experience are accumulated to enrich the prior knowledge base, and the intelligent level and the ability to cope with complex scenarios of the system are improved.

[0065] The following is a specific embodiment.

[0066] Embodiment 1: Large-scale urban core area power grid load peak shaving

[0067] Implementation environment setting

[0068] In this embodiment, the system is applied to the load peak shaving scenario of the core area power grid of a certain large city, relying on the following implementation environment.

[0069] Network: The scale of the power grid includes 30 110kV substations and 150 10kV distribution stations, with a total peak load of about 3.2GW and a daily peak-valley difference rate of 45%, belonging to a typical commercial and residential mixed power consumption area.

[0070] Energy type: Renewable energy includes distributed photovoltaic (total installed capacity 350MW), urban wind power (total installed capacity 120MW), and small-scale biomass power stations (60MW).

[0071] Energy storage equipment: Energy storage equipment includes electrochemical energy storage power stations (total capacity 200MWh), pumped storage (100MW), electric vehicle charging station networks (60 sites, about 15000 vehicles, V2G capability), and building air conditioning cold storage systems (equivalent capacity 150MWh).

[0072] Step 1: Multi-source data access and high-dimensional feature space construction

[0073] (1) Data source access and synchronization

[0074] This step first establishes a data collection network to access the following multi-source heterogeneous data:

[0075] Power grid SCADA system: Collect real-time data of substations and distribution stations, including bus voltage, line power, transformer load rate, etc., with a sampling period of 5 seconds;

[0076] Renewable energy management system: Collect real-time output data of distributed photovoltaic and wind power, including active power, reactive power, voltage, etc., with a sampling period of 10 seconds;

[0077] Energy storage management system: Collects status data of energy storage devices, including SOC, charging and discharging power, available capacity, etc., with a sampling period of 15 seconds;

[0078] Demand-side management system: Collects load data from various users, including industrial users, commercial buildings, and residential communities, with a sampling period of 1 minute;

[0079] Meteorological monitoring system: Collects meteorological data such as temperature, humidity, sunshine, and wind speed, with a sampling period of 5 minutes;

[0080] Electricity market system: Collects market information such as day-ahead market prices and real-time electricity prices, with an update cycle of 5 minutes.

[0081] All data is transmitted via a dedicated communication network, supporting multiple communication protocols such as OPC UA, IEC 61850, and Modbus. The data acquisition process employs a data quality marking mechanism to monitor data integrity, accuracy, and timeliness in real time. Time synchronization utilizes NTP service with microsecond-level precision to ensure time consistency across multiple data sources.

[0082] (2) Data preprocessing and quality control

[0083] Data preprocessing employs a multi-stage pipeline architecture, including the following steps:

[0084] Data cleaning: First, the raw data is standardized to remove the influence of units and dimensions. Then, an outlier detection algorithm based on Z-Score is used to identify outliers. Minor outliers are smoothed using median filtering, while severe outliers are marked as invalid data.

[0085] Missing value handling: For short-term missing values ​​(<5 minutes), cubic spline interpolation is used for imputation; for medium-term missing values ​​(5-30 minutes), similar day pattern matching is used for estimation; for long-term missing values ​​(>30 minutes), LSTM-based deep learning model is used for prediction imputation.

[0086] Multi-scale processing: Wavelet transform is used to decompose the data at multiple scales, separating the trend, periodic and random components. Different processing strategies are adopted for different components to improve the accuracy and efficiency of data processing.

[0087] Noise Suppression: A data smoothing algorithm based on an adaptive Kalman filter is designed to dynamically adjust the process noise covariance matrix and the measurement noise covariance matrix, balancing data smoothness and response speed.

[0088] Data quality control adopts a three-level index system, including completeness index (valid data proportion > 99.5%), accuracy index (abnormal value proportion < 0.5%) and timeliness index (average delay < 200ms). The system calculates these indicators in real time and triggers corresponding alarms and remedial measures when the indicators are abnormal.

[0089] High-dimensional feature space construction: it carries out feature engineering and unified representation on the preprocessed data, including feature classification, time series expansion and window splicing, nonlinear manifold learning / embedding, and splicing to form a tensor.

[0090] Step 2: Embedding entropy dynamic analysis and node interaction quantification

[0091] (1) Embedding entropy calculation

[0092] Based on the preprocessed multi-source time series data, the embedding entropy of each node of the system is calculated to quantify the complexity and uncertainty of the node:

[0093] Phase space reconstruction: the time series data {x(t)} of each node is reconstructed in phase space to determine the optimal embedding dimension m and time delay τ. The embedding dimension is determined by the false nearest neighbor method, with a value range of 3-10; the time delay is determined by the mutual information minimum method, usually 5-20 sampling periods. The reconstructed phase space vector is: X(t) = [x(t), x(t+τ), x(t+2τ),..., x(t+(m-1)τ)];

[0094] Single-variable embedding entropy calculation: based on the reconstructed phase space, the permutation entropy algorithm is used to calculate the embedding entropy value of each node. For the sequence {X(t)}, a set of permutation patterns with length L is constructed, the probability distribution of each permutation pattern p_i is calculated, and then the embedding entropy is calculated: H(X) = -Σ(p_i·log(p_i)). The higher the embedding entropy value, the more complex and uncertain the node behavior. The system sets the threshold Hthreshold = 0.75, and marks the nodes with embedding entropy higher than this value as high volatility nodes, which are given priority for regulation.

[0095] Multivariate embedding entropy matrix construction: for the N key candidate nodes (which can be all nodes or monitoring nodes screened by the running party) in the system, an N×N embedding entropy matrix EMij is constructed. The matrix element EMij represents the joint embedding entropy of node i and node j, and the calculation formula is: EMij = H([Xi,Xj]) - H(Xi) - H(Xj) + I(Xi;Xj) where H([Xi,Xj]) is the joint entropy and I(Xi;Xj) is the mutual information. The larger the EMij value, the more complex the interaction between nodes i and j.

[0096] In practice, the system updates the embedded entropy calculation every 15 minutes to capture the dynamic changes in node behavior patterns. To improve computational efficiency, the edge nodes are responsible for the preliminary calculation of local embedded entropy, and the central node is responsible for the integration and analysis of multivariate embedded entropy.

[0097] (2) Interactive network construction

[0098] Based on the embedded entropy matrix, an interactive influence network between grid nodes is constructed to quantify information flow and causal relationships:

[0099] Transfer entropy calculation: Transfer entropy (TE) is used to quantify the flow of information between nodes. For node i and node j, the transfer entropy TEij represents the amount of information transfer from i to j: TEij = H(Xj(t+1) | Xj(t)) - H(Xj(t+1) | Xj(t), Xi(t)) where H(· | ·) represents conditional entropy. TEij > TEji indicates that information mainly flows from node i to node j, i.e., i has a stronger causal influence on j.

[0100] Typical pattern recognition: Based on the transfer entropy matrix, a community detection algorithm is used to identify typical interaction patterns in the system. The Louvain method is used to cluster the transfer entropy matrix, dividing the system into several tightly interacting subgroups, with transfer entropy values within each subgroup significantly higher than between subgroups. These subgroups usually correspond to physically or functionally closely related device sets, such as devices in the same substation, load nodes in the same area, etc.

[0101] Key node identification: The betweenness centrality (BC) and influence range of each node are calculated to identify key nodes in the system. The betweenness centrality BCi represents the importance of node i in network information transfer: BCi = ∑j≠i≠k(σjk(i) / σjk) where σjk is the number of shortest paths from node j to k, and σjk(i) is the number of shortest paths from j to k through node i. The system marks the top 10% of nodes in terms of betweenness centrality as key nodes, which usually correspond to important substations, large energy storage facilities, or key load centers.

[0102] Dynamic interaction atlas construction: Based on the above analysis, a complete grid dynamic interaction atlas is constructed. The atlas is represented by a directed and weighted graph, with nodes representing system components and edges representing information flow, and edge weights corresponding to transfer entropy values. To capture the dynamic characteristics of the system, the system maintains multiple interaction atlas snapshots within a sliding time window, analyzing the time-varying characteristics of interaction patterns. In actual implementation, the system generates a complete interaction atlas every hour to guide subsequent peak shaving decisions.

[0103] Step 3: Fusion of domain knowledge and data-driven model

[0104] (1) Power system domain knowledge encoding

[0105] Transforming power system expertise into model usable prior information and constraints:

[0106] Physical constraint encoding: Transform the basic physical constraints of the power system, such as the flow equation, node power balance, line capacity limit, etc., into hard constraint conditions of the model. For example, for node i, the power balance constraint is expressed as:

[0107] Pi+ΣjPij=0

[0108] Where Pi is the injected power of node i, and Pij is the line power from node i to j. These constraints are encoded as penalty terms in the loss function of the model to ensure that the model output meets the physical feasibility.

[0109] Expert rule extraction: Through in-depth interviews with power dispatching experts and historical dispatching case analysis, extract expert rules and heuristic strategies for peak shaving decisions. For example, "discharge energy storage during peak period, charge during low valley period", "reduce conventional unit generation when photovoltaic output is high" and so on. These rules are encoded as decision tree structures as the initial strategy and safety guarantee of the model.

[0110] System safety margin setting: Based on historical operation data and expert experience, set safety margins and alarm thresholds for each parameter of the system. For example, set the transformer load rate alarm value to 85%, and the lower limit of the energy storage SOC to 20%. These thresholds are integrated into the constraint conditions of the model to ensure that the peak shaving decision does not cause the system to exceed the safe operating range.

[0111] Knowledge graph construction: Establish the relationship knowledge graph between power grid components, including physical connection relationship, functional dependence relationship and historical interaction mode. The knowledge graph uses triplets (entity-relation-entity) to represent, such as (substation A-feeds-load B), (energy storage C-supports-load area D) and so on. The graph is mapped to a vector space through embedding learning to provide prior structural information for subsequent graph neural networks.

[0112] (2) Hypergraph neural network construction

[0113] Based on the aforementioned interactive network, construct a hypergraph neural network model:

[0114] Supergraph structure design: The system uses a supergraph G = (V, E) to represent the interaction relationships in the power grid, where V represents the node set (such as generators, energy storage devices, load centers, etc.), and E represents the superedge set. Unlike traditional graphs, each superedge e ∈ E in the supergraph can connect two or more nodes, representing the group interaction relationship between multiple nodes. For example, a superedge can simultaneously connect a photovoltaic power station, an energy storage device, and a group of load nodes, representing their collaborative relationship.

[0115] Supergraph construction rules: Based on the analysis of transfer entropy and embedding entropy, the system automatically constructs the supergraph structure. When the transfer entropy values between multiple nodes all exceed the threshold TEthreshold = 0.4, a superedge connecting these nodes is created. In addition, through clustering analysis, node groups with similar behavior patterns are identified, and shared superedges are created for them. The final constructed supergraph contains about 3000 nodes and 1200 superedges, with an average of 2.8 nodes connected to each superedge.

[0116] Supergraph neural network architecture: The model uses a three-layer supergraph convolutional network (HGCN) architecture. For the node feature matrix X, the supergraph convolution operation of the first layer is defined as: H^(l+1) = σ(D^(-1 / 2)HWD^(-1 / 2)H^(l)Θ^(l)) where H is the superedge-node association matrix, D is the degree diagonal matrix, Θ^(l) is the learnable parameter matrix of the first layer, and σ is the nonlinear activation function (using LeakyReLU with a slope of 0.2). The dimension settings of each layer of the network are: input layer 128 dimensions, hidden layer 256 dimensions, and output layer 64 dimensions.

[0117] Temporal-spatial attention enhancement: To capture the temporal and spatial dependencies of the system, a temporal-spatial attention mechanism is added on top of the supergraph convolutional layer. The temporal attention calculates the importance weights of different time steps through self-attention; the spatial attention adjusts the information contribution of different nodes through the graph attention network (GAT). The temporal-spatial attention mechanism significantly improves the model's ability to perceive key patterns, especially in scenarios of load mutations and renewable energy fluctuations.

[0118] Step 4: Hierarchical scheduling and two-stage optimization

[0119] (1) Time dimension hierarchical scheduling

[0120] In the time dimension, the system implements a three-layer scheduling architecture, forming a complete closed loop from long-term planning to real-time control:

[0121] Day-ahead planning layer: Based on 24-48 hours ahead load and renewable power forecasts, a preliminary day-ahead peak shaving plan is formulated. The goal is to minimize the total peak-valley difference while satisfying system security and economic constraints. The optimization problem is formulated as: min F = w1 · (Pmax - Pmin) + w2 · Cost + w3 · Stability s.t. g(x) = 0, h(x) ≤ 0 where Pmax, Pmin are the daily maximum and minimum loads, Cost is the peak shaving cost, Stability is the system stability index, g(x) and h(x) are equality and inequality constraints (e.g. power balance, device capacity limits, etc.). The weight coefficients w1, w2, w3 are dynamically set according to the scheduling objectives, which are 0.5, 0.3, 0.2 in this example.

[0122] Hourly adjustment layer: Based on short-term load forecasts (2-6 hours ahead), the day-ahead plan is dynamically adjusted. This layer mainly targets the deviations between the day-ahead plan and actual conditions, as well as foreseeable load and renewable power changes in the short term. The adjustment uses a model predictive control (MPC) framework to solve the optimization problem in a rolling horizon: min F = Σt=1..T(w1 · (L(t) - Ltarget(t))^2 + w2 · ΔE(t)^2) s.t. system operation constraints where L(t) is the system load at time t, Ltarget(t) is the target load curve, ΔE(t) is the control variable change (e.g. energy storage charge / discharge power change), T is the prediction horizon (usually 12 time steps, i.e. 6 hours). The MPC solution frequency is every 30 minutes, ensuring that the plan adapts to the latest system state and forecasts.

[0123] Real-time response layer: Real-time control of the system at a minute-level time granularity, mainly responding to unexpected situations such as prediction deviations and device failures. This layer uses a rule-based fast response strategy and a simplified optimization algorithm to ensure that the response speed meets the real-time control requirements (<30 seconds).

[0124] (2) Spatial dimension hierarchical scheduling

[0125] In the spatial dimension, the system also adopts a three-layer architecture to achieve coordinated control from global to local:

[0126] System-level scheduling: Responsible for resource coordination and global optimization within the entire network, formulating the overall peak shaving strategy and target curve. The system-level scheduling considers global objective functions and system-level constraints, such as total peak-valley difference, system reserve capacity, main network section power flow, etc. Decision variables include regional target load curves, cross-regional resource allocation, system-level energy storage scheduling plans, etc.

[0127] Regional-level scheduling: Based on the target and resource quota issued by the system level, optimize the resource allocation and load regulation within the region. A region usually corresponds to a substation power supply range or a city functional area. The optimization goal of regional-level scheduling is to minimize the deviation of regional load from the target curve, while considering the economic and feasibility constraints of regional resources.

[0128] Device-level scheduling: Execute specific device control instructions, such as energy storage system charging and discharging power, controllable load adjustment, etc. Device-level scheduling needs to meet the physical constraints and operating conditions of the device itself, while responding to the control signals of the upper level. Through device's own closed-loop control and real-time feedback, ensure the safe and efficient operation of the device.

[0129] The three spatial levels communicate through a hierarchical protocol: the upper level issues instructions through target values and constraints, and the lower level uploads information through execution feedback and state reports. The system uses an incentive-based coordination mechanism, where the upper level can provide incentive signals (such as compensation fees, performance rewards, etc.) to guide the lower level to actively cooperate with the global optimization goal.

[0130] (3) Two-stage optimizer design

[0131] To efficiently solve large-scale optimization problems, the system designs an innovative two-stage optimizer:

[0132] Global exploration stage: Use the improved particle swarm optimization algorithm (IPSO) to conduct global search in the large-scale solution space. The improvements include:

[0133] Adaptive inertia weight: w = wmax - (wmax - wmin) · (iter / itermax)^2, initial wmax = 0.9, final wmin = 0.4;

[0134] Diversity preservation mechanism: Every 10 generations, evaluate the diversity of the population. When the diversity index is below the threshold, perturb and reinitialize 30% of the particles;

[0135] Constraint handling: Use the penalty function-based method to handle constraint conditions, impose a penalty on infeasible solutions, and increase the penalty factor with the number of iterations;

[0136] IPSO algorithm setting parameters: population size 100, maximum iteration number 200, convergence threshold 0.1% change in optimal value for 30 consecutive generations. The goal of global exploration is to quickly locate the feasible solution region, rather than accurately find the optimal solution, usually reaching the expected goal in about 100 generations.

[0137] Local refinement stage: Starting from the best solution of global exploration, the system applies gradient-based optimization methods for fine tuning. The specific algorithm is L-BFGS (Limited-memory BFGS), which is suitable for large-scale optimization problems due to its low memory requirement. L-BFGS parameters: memory length m = 10, convergence threshold is gradient norm <1e-6 or relative change of function value <1e-8. Local refinement usually requires 20-50 iterations to fine-tune key variables and improve the solution quality.

[0138] Two-stage collaborative strategy: The two stages work collaboratively through an intelligent switching strategy. The system sets multiple switching indicators, including global exploration convergence, solution feasibility, computation time limit, etc. When the switching conditions are met, the system transfers the best solution of global exploration to the local refinement algorithm and switches the optimization strategy. If the local refinement finds a better local optimal solution during the iteration process, the system will record the solution and continue to explore. Finally, the solution with the best global performance is selected as the output.

[0139] Computational resource allocation: The system dynamically allocates computational resources based on scheduling levels and time urgency. For day-ahead planning, longer computation time (5-10 minutes) is allowed, enabling complete execution of two-stage optimization; for hourly adjustments, computation time is limited to 2 minutes, which may terminate global exploration early; for real-time response, computation time is limited to 30 seconds, mainly relying on pre-computed simplified models and rule engines.

[0140] Step 5: Real-time decision-making and closed-loop feedback

[0141] (1) Real-time load shifting strategy generation

[0142] Based on the analysis results of the preceding modules, the system generates real-time load shifting strategies and peak shaving plans:

[0143] Peak target curve generation: The system generates a target load curve based on the peak-valley difference optimization objective. The target curve generation adopts the "peak clipping and valley filling" principle by solving the following optimization problem:

[0144] min F = Σt(Lt-Lavg)^2 s.t. ΣtLt = ΣtL0t Lt≥Lmin,t |Lt-L0t|≤ΔLmax,t

[0145] Where F is the peak-valley difference, Lt is the target load at time t, L0t is the predicted original load, Lavg is the average load, Lmin,t is the minimum load constraint, and ΔLmax,t is the maximum adjustment constraint. Through this optimization, the system generates a target curve with minimum peak-valley difference, total power conservation, and satisfaction of constraints in each period.

[0146] Resource allocation algorithm: To achieve the target curve, the system needs to allocate various types of peak-shaving resources. The resource allocation adopts a priority strategy based on cost-effectiveness:

[0147] Construct a resource capacity matrix Cij, which represents the adjustable capacity of resource i at time period j;

[0148] Construct a resource cost matrix Kij, which represents the unit control cost of resource i at time period j;

[0149] Construct a resource benefit index Eij = Cij / Kij, which represents the adjustment capacity per unit cost;

[0150] Sort by benefit index from high to low, and prefer to call high-benefit resources;

[0151] In actual implementation, the system also considers the synergistic effect and constraint conditions between resources, such as energy storage charging and discharging balance, load regulation continuity, etc., and solves the overall optimal configuration through a mixed integer linear programming (MILP) model.

[0152] Control instruction generation: Based on the resource allocation results, the system generates specific control instructions, including:

[0153] Energy storage charging and discharging instructions: specify the charging and discharging power and energy of each energy storage device at each time period;

[0154] Controllable load adjustment instructions: specify the power consumption plan adjustment amount of various controllable loads (such as air conditioners, electric vehicle charging piles);

[0155] Distributed power generation adjustment instructions: specify the output plan and adjustment range of distributed power sources;

[0156] Price signal: generate time-of-use electricity price or incentive signal to guide users to actively adjust their electricity consumption behavior. The control instructions adopt standard format and communication protocol, support automatic execution and manual confirmation modes, and ensure system safety and user friendliness.

[0157] Emergency response mechanism: For sudden situations (such as equipment failure, load mutation, extreme weather, etc.), the system prepares multiple sets of emergency response strategies. Emergency response adopts a rule-based rapid decision-making method, without complex optimization calculation, to ensure response speed. Emergency strategies focus on maintaining system safety and stable operation, which may temporarily sacrifice economic and comfort targets.

[0158] (2) Multi-level execution framework

[0159] The system designs a multi-level execution framework to ensure that the peak-shaving instructions can be effectively implemented:

[0160] Direct Control Channel: For devices directly controlled by the system (such as energy storage power stations, part of distributed power), the system sends control instructions through SCADA or energy management systems to achieve precise control. The direct control channel uses encryption communication and instruction verification mechanism to ensure control safety and reliability. Control instructions comply with standards such as IEC 61850, compatible with multiple devices and systems.

[0161] Indirect Incentive Mechanism: For user-side resources (such as residential appliances, commercial building loads), the system guides users to adjust their electricity consumption behavior through price signals and incentive measures. Indirect incentives include:

[0162] Real-time Electricity Price: Dynamically adjust electricity prices according to system load conditions, increase prices during peak periods, and reduce prices during off-peak periods;

[0163] Demand Response Subsidy: Provide economic subsidies to users participating in peak load shifting, with subsidy amounts related to response quantity and time period;

[0164] Green Energy Points: Users participating in renewable energy consumption can earn points, which can be exchanged for discounts or services;

[0165] Hybrid Collaborative Regulation: For some large commercial users and industrial parks, the system uses a hybrid regulation mode of "direct control + incentive mechanism". The system interfaces with the user's energy management system, sends load target recommendations, and the user's system optimizes and adjusts the electricity plan on its own premise of ensuring internal demand. Such users usually enjoy more preferential electricity prices, but need to commit to a certain proportion of controllable load capacity.

[0166] Manual Intervention Interface: Although the system has high automation capabilities, it still retains a manual intervention interface to allow dispatchers to manually adjust decisions when necessary. Manual intervention is mainly used to handle complex situations or safety-critical decisions that the system cannot automatically handle. The interface design uses a "suggestion + confirmation" mode, where the system provides decision suggestions, and the dispatcher confirms and executes them to reduce the risk of human error.

[0167] (3) Real-time Monitoring and Evaluation

[0168] The system monitors the peak shaving execution effect in real time and conducts multi-dimensional evaluation:

[0169] Key Performance Indicators: The system monitors the following key performance indicators in real time:

[0170] Load Curve Flatness: Evaluate the peak-valley difference by calculating the standard deviation and mean ratio σ / μ of the daily load curve;

[0171] Peak Shaving Resource Utilization: The actual utilization of various peak shaving resources accounts for the proportion of available capacity;

[0172] Peak shaving cost: including energy storage wear and tear cost, demand response subsidy, backup startup cost, etc.

[0173] System stability index: voltage / frequency qualification rate, line load rate, backup capacity adequacy, etc.

[0174] User satisfaction: indirectly assessed through feedback surveys and complaint rates.

[0175] Execution deviation analysis: compare the actual execution result with the planned instruction deviation, analyze the deviation reason:

[0176] Prediction deviation: the difference between load prediction or renewable energy output prediction and actual value;

[0177] Response deviation: the difference between the actual response of the regulation resource and the requirement of the instruction;

[0178] Communication delay: time delay in the process of instruction transmission and execution;

[0179] External interference: unforeseen system events or external condition changes.

[0180] Benefit evaluation analysis: comprehensive evaluation of the comprehensive benefits of peak shaving measures, including:

[0181] Economic benefits: peak-valley price difference income, reduced backup capacity cost, reduced network investment;

[0182] Technical benefits: improve equipment utilization, reduce line loss, improve system stability;

[0183] Environmental benefits: improve renewable energy consumption ratio, reduce carbon emissions;

[0184] Social benefits: improve power supply reliability, reduce power outage risk, promote user participation;

[0185] Trend analysis and early warning: based on historical data analysis of load characteristics and long-term trends of peak shaving effect, identify potential problems and give early warning:

[0186] Seasonal changes: identify seasonal patterns of load changes, adjust peak shaving strategies in advance;

[0187] Load growth: monitor load growth trends, assess future peak shaving resource needs;

[0188] Changes in user behavior: analyze the evolution of user response patterns, optimize incentive mechanisms;

[0189] Equipment aging: monitor the performance degradation of energy storage and other equipment, arrange maintenance or update.

[0190] (4) Closed-loop feedback and self-optimization

[0191] Based on the monitoring and evaluation results, the system implements closed-loop feedback optimization to continuously improve the peak shaving effect:

[0192] Online model updating: Based on actual operation data and feedback, the system continuously updates various models and parameters:

[0193] Prediction model update: Use the latest data to retrain the load prediction model and renewable energy prediction model to adapt to changes in load characteristics;

[0194] Response model calibration: According to the actual response data of users, calibrate the load response model parameters to improve the accuracy of response prediction;

[0195] Equipment model update: Update key parameters such as energy storage efficiency and capacity attenuation based on equipment operating status;

[0196] Strategy optimization iteration: Based on the evaluation of execution effect, the system continuously optimizes the peak shaving strategy:

[0197] Resource allocation optimization: Adjust the priority and allocation ratio of various resources to improve resource utilization efficiency;

[0198] Time period division optimization: Adjust the division standard of peak, valley and flat periods according to the change of load characteristics;

[0199] Price signal optimization: Based on user response elasticity, adjust the price gradient and time distribution;

[0200] Safety margin optimization: Adjust the margin setting of various safety constraints according to the stability performance of the system;

[0201] Knowledge base accumulation and expert rule extraction: The system continuously accumulates operation experience and knowledge:

[0202] Typical scenario library: Records and classifies various peak shaving typical scenarios and their optimal processing strategies;

[0203] Abnormal event library: Collects system abnormal events and successful processing methods as an abnormal handling knowledge base;

[0204] Expert experience extraction: Extract expert decision rules by recording and analyzing manual intervention operations and integrate them into the automatic decision system;

[0205] Long-term evolution mechanism: The system has the ability of long-term self-evolution to cope with environmental and demand changes:

[0206] Architecture adaptation: The system architecture can dynamically adjust according to actual needs, such as adding new data sources, expanding model types, and upgrading algorithms;

[0207] Dynamic balance of goals: According to the changes of policy guidance and market environment, dynamically adjust the weights of economy, reliability, environmental protection and other multi-objectives;

[0208] Frontier technology integration: The system design reserves interfaces for integrating new artificial intelligence technologies and optimization algorithms, maintaining technological advancement;

[0209] Long-term evolution is driven by quarterly evaluations and annual planning, combined with technological development trends and changes in business needs, to develop a system upgrade roadmap, ensuring the system's continuous competitiveness and adaptability.

[0210] The innovation of this step is to build a complete real-time decision-making and closed-loop feedback mechanism, enabling the system to continuously learn and optimize from actual operation, forming a self-evolving intelligent system. Compared with traditional static systems, the closed-loop feedback mechanism improves the average performance of the system in long-term operation by about 15-20%, significantly enhances the adaptability to abnormal situations, and gradually accumulates knowledge and experience, reducing the dependence on manual intervention.

[0211] The above describes the preferred embodiments of the application in detail. It should be understood that those skilled in the art can make many modifications and changes without creative labor according to the concept of the application. For example, the data source configuration, optimization algorithm selection and parameter setting can be adjusted according to the specific regional characteristics, or the model architecture and decision strategy can be adjusted according to different types of power systems. Therefore, any technical solution obtained by logical analysis, reasoning or limited experiment based on the existing technology according to the concept of the application shall be within the protection scope determined by the claims.

Claims

1. A real-time load collaborative peak-shaving method for urban power grids based on multi-energy complementarity and AI scheduling, characterized in that, Includes the following steps: (1) Multi-source data access and high-dimensional feature space construction Real-time acquisition of renewable energy power generation data, energy storage device status data, grid load data, and environmental data; and preprocessing operations including cleaning, missing data completion, and standardization are performed on the multi-source heterogeneous data. After preprocessing, based on the clean and dimensionally uniform data obtained, a spatiotemporal high-dimensional feature matrix containing both static and dynamic features is constructed to achieve a unified representation of heterogeneous data while preserving the inherent characteristics of various types of data. (2) Embedded entropy calculation and interactive network construction Based on preprocessed multi-source time series data, the embedding entropy of each node in the system is calculated, and the complexity and uncertainty of the nodes are quantified. Based on the embedded entropy matrix, an interactive influence network among power grid nodes is constructed to quantify information flow and causal relationships. (3) Integration of domain knowledge and data-driven models By integrating power system domain knowledge with data-driven models, power system expertise is transformed into prior information and constraints usable by the data-driven model, thus constructing a hybrid intelligent model. The data-driven model includes a hypergraph neural network layer and a spatiotemporal attention layer. The hypergraph neural network layer is responsible for modeling high-order relationships between multiple nodes, with each hyperedge connecting multiple nodes and representing complex group interactions. The spatiotemporal attention layer enhances the model's ability to perceive key spatiotemporal patterns and dynamically adjusts the importance weights of different features and time steps. The model fusion adopts an adaptive ensemble learning framework, which dynamically adjusts the weights of different models according to the real-time system state to balance the accuracy and computational efficiency of the models. (4) Hierarchical scheduling and two-stage optimization A two-stage optimizer is applied to train a hybrid intelligent model. Based on the trained hybrid intelligent model, hierarchical scheduling is implemented in both time and space dimensions to generate scheduling strategies in a hierarchical manner. (5) Real-time decision-making and closed-loop feedback The generated peak-shaving instructions are sent to each execution unit through the SCADA system or energy management system to monitor and evaluate the peak-shaving effect in real time. Based on the monitoring and evaluation results, the system implements closed-loop feedback optimization to continuously improve the peak-shaving effect.

2. The urban power grid real-time load collaborative peak-shaving method based on multi-energy complementarity and AI scheduling according to claim 1, characterized in that, In step (2), based on the preprocessed multi-source time-series data, the embedding entropy of each node in the system is calculated, and the complexity and uncertainty of the nodes are quantified, including: First, the phase space of the time-series data of each node is reconstructed to determine the optimal embedding dimension and time delay; Then, based on the reconstructed phase space, the embedding entropy value of each node is calculated using the permutation entropy algorithm; Next, the multivariate embedding entropy matrix is ​​constructed: for the N key candidate nodes in the system, an N×N embedding entropy matrix is ​​constructed; the matrix element EMij represents the joint embedding entropy of node i and node j, and the calculation formula is: EMij=H([Xi,Xj])-H(Xi)-H(Xj)+I(Xi;Xj) where H([Xi,Xj]) is the joint entropy, H(Xi) and H(Xj) are the embedding entropies of node i and node j respectively, and I(Xi;Xj) is the mutual information.

3. The real-time load coordination peak-shaving method for urban power grids based on multi-energy complementarity and AI scheduling according to claim 1, characterized in that, In step (2), based on the embedded entropy matrix, an interaction network between power grid nodes is constructed to quantify information flow and causal relationships, as follows: First, the information flow between nodes is quantified using transfer entropy; Then, based on the transfer entropy matrix, a community detection algorithm is used to identify typical interaction patterns in the system; Next, the betweenness centrality and influence range of each node are calculated to identify the key nodes in the system; Finally, based on the above analysis, a complete dynamic interaction graph of the power grid is constructed. The graph is represented by a directed weighted graph, where nodes represent system components, edges represent information flow, and edge weights correspond to the entropy values.

4. The method for real-time load coordination and peak shaving of urban power grids based on multi-energy complementarity and AI scheduling according to claim 1, characterized in that, In step (3), power system expertise is transformed into prior information and constraints usable in the data-driven model, specifically including: Physical constraint encoding: Transforming the basic physical constraints of the power system into hard constraints for the model; Expert rule extraction: Through in-depth interviews with power dispatching experts and analysis of historical dispatching cases, expert rules and heuristic strategies for peak shaving decisions are extracted. System safety margin setting: Based on historical operating data and expert experience, safety margins and alarm thresholds are set for each system parameter; Knowledge graph construction: Establish a knowledge graph of relationships between power grid components, including physical connection relationships, functional dependencies, and historical interaction patterns. The knowledge graph is represented by triples.

5. The method for real-time load coordination and peak shaving of urban power grids based on multi-energy complementarity and AI scheduling according to claim 1, characterized in that, In step (3), the hypergraph neural network layer construction method is as follows: Hypergraph Structure Design: The system uses a hypergraph G = (V, E) to represent the power grid interaction relationship, where V represents the set of nodes, including generators, energy storage devices, and load centers, and E represents the set of hyperedges; each hyperedge e ∈ E in the hypergraph can connect two or more nodes, representing the group interaction relationship between multiple nodes; Hypergraph construction rules: Based on transit entropy and embedding entropy analysis, the system automatically constructs the hypergraph structure; when the transit entropy values ​​between multiple nodes all exceed the threshold TEthreshold = 0.4, a hyperedge connecting these nodes is created; in addition, cluster analysis is used to identify groups of nodes with similar behavioral patterns, and shared hyperedges are created for them. Hypergraph Neural Network Architecture: The model adopts a three-layer hypergraph convolutional network (HGCN) architecture.

6. The urban power grid real-time load collaborative peak shaving method based on multi-energy complementarity and AI scheduling according to claim 1 is characterized in that, in step (4), the dual-stage optimizer is used to train the hybrid intelligent model and to solve the actual decision problem at each layer in the scheduling stage; the dual-stage optimizer includes a global exploration stage and a local refinement stage: the global exploration adopts an improved particle swarm algorithm to quickly search the large-scale solution space; the local refinement is based on gradient descent and the Lagrange multiplier method to finely adjust key parameters; the two stages work together through an intelligent switching strategy; the optimization objective is a multi-objective function, which simultaneously considers multiple dimensions such as minimizing peak-valley difference, minimizing peak shaving cost, and maximizing system stability.

7. The method for real-time load coordination and peak shaving of urban power grids based on multi-energy complementarity and AI scheduling according to claim 1, characterized in that, In step (4), the time dimension is divided into three levels: day-ahead planning, hour-level adjustment and real-time response; the spatial dimension is divided into three levels: system level, regional level and equipment level.

8. The method for real-time load coordination and peak shaving of urban power grids based on multi-energy complementarity and AI scheduling according to claim 1, characterized in that, In step (4), the system adopts a cost-effective priority strategy to allocate various peak-shaving resources, generates a target load curve based on the peak-valley difference optimization objective, and considers the synergistic effect and constraints between resources. The overall optimal configuration is solved by the mixed integer linear programming (MILP) model.

9. The method for real-time load coordination and peak shaving of urban power grids based on multi-energy complementarity and AI scheduling according to claim 1, characterized in that, In step (5), the execution unit includes an energy storage system controller, a renewable energy controller, and a load-side response controller; based on a multi-level execution framework, it ensures that peak-shaving instructions can be effectively implemented.

10. The method for real-time load coordination and peak shaving of urban power grids based on multi-energy complementarity and AI scheduling according to claim 1, characterized in that, In step (5), the deviation between the expected peak-shaving effect and the actual effect is analyzed to identify the reasons for the deviation. Based on the analysis results, the model parameters, optimizer configuration and decision-making strategy are dynamically adjusted to achieve adaptive optimization of the system. Historical data and optimization effects are regularly mined to identify long-term trends and patterns, providing a basis for model upgrades and strategy improvements. At the same time, expert knowledge and successful experience are accumulated to enrich the prior knowledge base and improve the system's intelligence level and ability to cope with complex scenarios.

Images