Cooperative regulation and control method and system for source network load storage system

Abstract

The invention discloses a source network load storage system cooperative regulation and control method and system, and the method comprises the steps: employing a high-precision sensor and multi-protocol communication to obtain system multi-dimensional data through global data collection and preprocessing, and carrying out the noise reduction; constructing a dynamic association model based on a graph neural network, and accurately capturing a system node relationship in combination with a multi-head attention mechanism and topological constraints; layered multi-objective decision, reinforcement learning real-time regulation and control, layered control architecture and closed-loop feedback correction are adopted, and economic optimization, safety guarantee and strategy iteration are considered. The system is composed of a global data sensing unit, a system dynamic modeling unit and the like, and all the units are in close cooperation. According to the method and the system, the defects of insufficient data processing, low model precision, single regulation and control and the like in the prior art are effectively overcome, the system fault prediction accuracy can be improved, the carbon emission is reduced, the power fluctuation coping capacity is enhanced, and the operation efficiency and the stability of the source network load storage system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of power system automation and smart grid, and in particular to a method and system for coordinated control of a source-grid-load-storage system. Background Art

[0002] As the global energy transition deepens, the energy system is accelerating its transformation toward clean, low-carbon, safe, and efficient energy. The large-scale access of renewable energy sources, such as wind power and photovoltaics, to the grid has resulted in intermittent and fluctuating power generation on the source side. Consumer electricity demand is becoming increasingly diverse and personalized, significantly increasing the uncertainty of loads on the load side. While energy storage technology continues to advance, its collaborative application within the system still needs to be deepened. Against this backdrop, the traditional power system's one-way power supply and decentralized operation model is difficult to adapt to the new situation. There is an urgent need to build a new power system that integrates the synergistic interaction of sources, grids, loads, and storage, achieving multi-energy complementarity and coordinated optimization, and improving the comprehensive energy utilization efficiency and overall system stability.

[0003] While extensive research and practice have been conducted in the field of coordinated control of power generation, grid, load, and storage, significant technical bottlenecks remain. At the data level, existing system data collection suffers from incomplete coverage and insufficient accuracy. Inconsistent data communication protocols between different devices and systems make data fusion difficult, hindering a comprehensive and accurate reflection of system operational status. Furthermore, traditional modeling methods, often based on static parameters and simplified assumptions, are unable to effectively capture the complex dynamic interactions and spatiotemporal coupling characteristics of various links within the power generation, grid, load, and storage system. Consequently, the prediction accuracy and real-time performance struggle to meet actual control requirements.

[0004] From the perspective of control strategies, existing control methods often focus on single objectives or local optimization, lacking a comprehensive consideration of the system's overall economic efficiency, safety, and low-carbon nature. When dealing with fluctuations in renewable energy output and sudden load changes, control responses are slow and have limited regulatory capacity, making it difficult to quickly balance power fluctuations. Furthermore, most control strategies lack flexibility and are unable to adapt to system operating modes and external environmental changes. This results in high system operating costs and low resource allocation efficiency, making it difficult to fully leverage the advantages of coordinated operation of power generation, grid, load, and storage. Innovative coordinated control methods and systems are urgently needed to address these issues. Summary of the Invention

[0005] In order to overcome the shortcomings and deficiencies of the prior art, the present invention aims to provide a method and system for coordinated control of a source-grid-load-storage system.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for coordinated control of a source-grid-load-storage system, comprising the following steps:

[0008] S1: Build a global data acquisition system covering source-side power generation units, grid-side transmission networks, load-side power terminals, and storage-side energy storage devices. This system synchronously samples real-time power, voltage, frequency, temperature, and state of charge, and eliminates data noise through adaptive filtering algorithms.

[0009] S2: Utilizes graph neural networks to construct a dynamic association model for the source-grid-load-storage system, including topological connectivity. This model abstracts grid nodes, power generation equipment, load clusters, and energy storage facilities into graph nodes. Based on historical operating data, the interaction weight matrix between nodes is trained to extract spatial features and perform time series prediction of the system's operating status.

[0010] S3: Establish a multi-objective decision-making model based on a hierarchical optimization framework. The upper layer aims to minimize the operating cost of the entire system, combining power market price signals and carbon emission indicators to generate a source-storage power generation plan and load distribution benchmark. The lower layer uses network security constraints as conditions to construct a constrained mixed integer programming model to verify the feasibility and dynamically modify the benchmark plan.

[0011] S4: Constructing a Markov decision process based on the system state space, action space, and reward function. Using a deep Q-network to train a control strategy, the system dynamically adjusts the source and storage output and load response in scenarios where renewable energy output fluctuates and load changes occur.

[0012] S5: Build a hierarchical control architecture with a time-scale matching module. The minute-level scheduling layer performs rolling optimization of the day-ahead plan. The second-level control layer uses a model predictive control algorithm to quickly compensate for power fluctuations. The millisecond-level protection layer triggers emergency control measures based on transient stability criteria.

[0013] S6: Establish a closed-loop feedback correction mechanism to align the actual operation data of the system after regulation with the predicted data in time and space dimensions, update the dynamic correlation model parameters through the Kalman filter algorithm, and perform adaptive iterative optimization of the regulation strategy.

[0014] Furthermore, in step S2, the graph neural network model uses a multi-head attention mechanism to perform weighted aggregation on different types of node features, and its node state update formula is:

[0015]

[0016] in, represents the feature vector of node v in layer l, is the feature set of node v’s neighbor nodes, is the weight coefficient of the i-th attention head, is the weight matrix of the i-th attention head, K is the number of attention heads, and σ is the activation function. The model captures the time evolution characteristics of the system state through the time series graph convolution layer, and the convolution kernel parameters are dynamically adjusted according to the grid topology.

[0017] Furthermore, in step S3, the upper model in the multi-objective decision-making model of the hierarchical optimization framework introduces the carbon transaction cost function Where ρ is the carbon price, T is the dispatch period, G is the set of power generation equipment, S is the set of energy storage equipment, e i，t With e j，t are the carbon emissions of power generation equipment i and the carbon absorption of energy storage equipment j at time t respectively; the lower model uses a robust optimization method to deal with the renewable energy forecast error and constructs a constraint condition containing the uncertainty set Γ Where ΔP g is the power fluctuation range considering the prediction error.

[0018] Furthermore, in step S4, the state space of the reinforcement learning model includes the real-time output of the source-side power generation equipment, the grid-side line flow distribution, the load classification status of the load side, the storage-side charge status and the system frequency deviation; the action space is defined as a discretized set of source-storage power regulation amounts and load interruptibility amounts; the reward function is designed as R = -λ1ΔC -λ2ΔP -λ3Δf, where ΔC is the control cost increment, ΔP is the power imbalance amount, Δf is the frequency deviation, and λ1, λ2, and λ3 are weight coefficients.

[0019] Furthermore, in step S5, the time scale matching module of the hierarchical control architecture uses wavelet transform to decompose and reconstruct multi-time scale data, and establishes a bidirectional data channel including prediction error compensation between the minute-level scheduling layer and the second-level control layer; the model predictive control algorithm adopts a rolling time domain optimization strategy to match the predicted trajectory of the next N control cycles with the current state, and the optimization objective function includes a tracking error term and a control quantity change term.

[0020] Furthermore, in step S6, when the closed-loop feedback correction mechanism aligns the spatiotemporal dimensions of the data, a dynamic time warping algorithm is used to process the time series matching problem of data with different sampling frequencies; the state transfer matrix of the Kalman filter algorithm is dynamically adjusted according to changes in the system operation mode, and the observation matrix contains the accuracy parameters and installation location information of the measuring equipment to achieve adaptive update of the control strategy parameters.

[0021] Furthermore, in step S1, the adaptive filtering algorithm adopts a variable step size minimum mean square error LMS algorithm, and the step size factor is dynamically adjusted according to the data variance; the standardized data sequence adopts a feature extraction method based on principal component analysis PCA to map high-dimensional data to a low-dimensional space, thereby reducing data redundancy and computational complexity.

[0022] Furthermore, in step S2, the dynamic association model introduces a topological constraint term to regularize the graph neural network training, and the constraint term expression is Ω=∑ (u，v)∈E |θ u -θ v |, where (u, v) is the set of edges in the graph, θ u and θ v are the phase differences between nodes u and v respectively; by penalizing the situation where the characteristics of adjacent nodes are too different, the model's ability to express the power grid topology is improved.

[0023] Furthermore, in step S4, the deep Q network adopts a dual network architecture, including an evaluation network and a target network, and reduces the value function estimation deviation during training by regularly updating the target network parameters; and adopts an experience replay mechanism to store the historical state-action-reward-next state quadruple to alleviate data correlation problems and improve training stability.

[0024] A second aspect of the present invention provides a system for implementing the above-mentioned source-grid-load-storage system coordinated control method, the system comprising:

[0025] The global data perception unit is equipped with a distributed sensor array and a multi-protocol communication module to collect and pre-process multi-dimensional physical quantities of the source, grid, load and storage system;

[0026] The system dynamic modeling unit includes a graph neural network calculation module and a topological feature extraction module, which is connected to the global data perception unit to build a system dynamic association model;

[0027] The hierarchical optimization decision-making unit integrates the multi-objective decision-making module and the constraint verification module, receives the output data of the system dynamic modeling unit, and performs hierarchical optimization calculations;

[0028] The real-time control execution unit is equipped with a reinforcement learning strategy generation module and a multi-time scale control module, which is connected to the hierarchical optimization decision-making unit to generate and execute the control strategy;

[0029] The data feedback processing unit, which includes a spatiotemporal alignment module and a parameter update module, receives the actual system operation data and feeds it back to the system dynamic modeling unit and the hierarchical optimization decision unit;

[0030] The network communication interaction unit is equipped with a high-speed data transmission protocol and a security encryption module to realize data interaction and instruction transmission between units;

[0031] The output end of the global data perception unit is connected to the input end of the system dynamic modeling unit, the output end of the system dynamic modeling unit is connected to the input end of the hierarchical optimization decision unit, the output end of the hierarchical optimization decision unit is connected to the input end of the real-time control execution unit, the output end of the real-time control execution unit acts on the source grid load storage system, and the system dynamic operation data is fed back to the data feedback processing unit through the network communication interaction unit. The output end of the data feedback processing unit is respectively connected to the system dynamic modeling unit and the hierarchical optimization decision unit.

[0032] Beneficial effects:

[0033] The present invention proposes a method and system for collaborative control of a source-grid-load-storage system. In terms of data processing, a global data acquisition system is constructed through a distributed sensor array and a multi-protocol communication network. Adaptive filtering and standardized processing are combined to ensure accurate and reliable data, providing a solid foundation for subsequent control. When modeling the system, a graph neural network is used in combination with topological constraints and a multi-head attention mechanism to accurately capture the complex relationships between system nodes, achieving efficient prediction and feature extraction of operating status. Hierarchical optimization decision-making introduces strategies such as carbon trading costs and robust optimization, taking into account both economy and safety, and improving the operating efficiency of the system. Real-time control uses reinforcement learning and multi-time scale control to quickly respond to fluctuations in renewable energy and sudden load changes to ensure system stability. The closed-loop feedback mechanism uses Kalman filtering and dynamic time warping to continuously optimize the control strategy and enhance system adaptability. In the corresponding control system, the six major units of global data perception, dynamic modeling, and hierarchical decision-making work closely together to achieve a full process of data collection, analysis, decision-making, and execution. This collaborative control method and system not only significantly improves the economic efficiency of the source-grid-load-storage system operation, reduces costs by optimizing resource allocation, but also greatly enhances the system's security and stability, effectively responding to complex operating scenarios; at the same time, its adaptive iterative capability enables the system to continuously evolve, in line with the development trend of intelligent and low-carbon energy systems, providing strong support for the construction of new power systems, and has broad application prospects and significant technical advantages. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is an operational flow chart of a source-grid-load-storage system coordinated control method provided by the present invention;

[0035] Figure 2 This is a unit composition diagram of a source-grid-load-storage system collaborative control system provided by the present invention. DETAILED DESCRIPTION

[0036] It should be noted that, unless there is a conflict, the embodiments in this application and the features described in the embodiments can be combined with each other. The application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] like Figure 1As shown, a method for coordinated control of a source-grid-load-storage system includes the following steps:

[0038] S1. Based on a distributed sensor array and multi-protocol communication network, a global data acquisition system covering source-side power generation units, grid-side transmission networks, load-side power terminals, and storage-side energy storage devices is constructed. This system synchronously samples multi-dimensional physical quantities such as real-time power, voltage, frequency, temperature, and state of charge, and uses an adaptive filtering algorithm to eliminate data noise and form a standardized data sequence.

[0039] Specifically, this step aims to build a data-aware network covering the entire power generation, grid, load, and storage system. On the power generation side, high-precision power sensors must be deployed. For example, for wind turbines, current and voltage sensors with a measurement accuracy of ±0.5% of rated power are installed to collect real-time generated power. On the grid side, phasor measurement units (PMUs) are installed at key nodes on the transmission lines to acquire voltage phasors, frequency, and other data at a sampling rate of 50 times per second. On the load side, smart meters are deployed in large commercial buildings, industrial plants, and other locations to monitor real-time power consumption data for different types of loads. On the storage side, a dedicated battery management system (BMS) collects parameters such as state of charge (SOC) and charge and discharge current for battery energy storage systems.

[0040] This collected data exists in a variety of communication protocols, such as Modbus and IEC61850. The system utilizes a protocol conversion gateway to standardize the data format. The collected data is often mixed with noise, which is processed using an adaptive filtering algorithm. For example, in the case of photovoltaic power plant output power data, the algorithm dynamically adjusts filtering parameters to effectively remove random fluctuations caused by factors such as cloud cover, ensuring data accuracy.

[0041] After processing, the data is organized into standardized sequences, such as assigning unified timestamps to various data types and sampling them at minute-level intervals. For example, in a regional power grid, this step yields a complete operational data set covering 10 power sources, 50 transmission lines, 300 load nodes, and five energy storage plants, providing a reliable data foundation for subsequent system analysis and regulation.

[0042] S2. Build a dynamic association model for the source-grid-load-storage system with topological connectivity using a graph neural network. This model abstracts grid nodes, power generation equipment, load clusters, and energy storage facilities into graph nodes. Based on historical operating data, it trains an interaction weight matrix between nodes to extract spatial features and conduct time series prediction of the system's operating status.

[0043] Specifically, this step uses a graph neural network to construct a system model, abstracting each entity in the source-grid-load-storage system into graph nodes. For example, thermal power plants, wind farms, and photovoltaic power stations are considered source nodes, substations and transmission lines are considered network nodes, residential communities and factories are considered load nodes, and battery energy storage stations are considered storage nodes. The edges connecting the nodes represent power transmission relationships or energy interaction relationships.

[0044] The model training utilizes historical system operation data from the past year, such as hourly power generation, load demand, and line flow data, to optimize the interaction weight matrix between nodes through a backpropagation algorithm. In practice, when a transmission line in the system fails, the model can quickly analyze the impact on surrounding nodes based on the trained weight relationships. For example, it can predict which power points require output adjustments and which loads may be affected.

[0045] The model can also capture the time series characteristics of the system's operating status. Taking wind power prediction as an example, by analyzing the relationship between wind speed, wind direction and power generation in historical data and combining it with the structure of the graph neural network, it not only considers the historical power changes of a single wind farm, but also integrates information from surrounding wind farms, grid loads and other related nodes to more accurately predict wind power output in the future, providing forward-looking information support for system regulation.

[0046] S3: Establish a multi-objective decision-making model based on a hierarchical optimization framework. The upper layer aims to minimize the operating cost of the entire system, combining power market price signals and carbon emission indicators to generate a source-storage power generation plan and load allocation benchmark. The lower layer uses network security constraints as a condition to construct a mixed integer programming model that includes constraints such as line capacity, voltage deviation, and frequency fluctuation. The model verifies the feasibility of the benchmark plan and dynamically modifies it.

[0047] Specifically, the upper-level model in the multi-objective decision-making model of the hierarchical optimization framework focuses on the economic optimization of the entire system. Taking the carbon trading cost as an example, in a certain provincial power system, the carbon price is set at 50 yuan / ton, the scheduling cycle is one day (24 hours), and the total carbon trading cost is calculated based on the carbon emission intensity and power generation plan of each power generation equipment. At the same time, combined with the real-time electricity price in the power market, such as 1 yuan / kWh during peak hours and 0.3 yuan / kWh during off-peak hours, a source-storage power generation plan and load distribution benchmark are formulated to minimize power generation costs and carbon trading costs.

[0048] The lower-level models in the multi-objective decision-making model of the hierarchical optimization framework focus on ensuring the safety of system operation. A robust optimization model is constructed to address the uncertainty of renewable energy output. For example, for a photovoltaic power plant, assuming a power forecast error range of ±20%, by setting uncertainty sets, the power generation plan is designed to ensure that, even under the worst-case error scenario, the system still meets safety constraints such as line capacity limits (e.g., a line's maximum transmission power is 500 megawatts) and voltage deviation limits (e.g., node voltage must remain within ±5% of the rated voltage). If the lower-level model identifies safety risks in the upper-level baseline plan, it will dynamically modify it to ensure the plan's feasibility.

[0049] S4: Design a real-time control strategy generation mechanism based on reinforcement learning. This involves constructing a Markov decision process using the system state space, action space, and reward function. This strategy is trained using a deep Q-network to dynamically adjust source and storage output and load response in scenarios where renewable energy output fluctuates and load changes occur.

[0050] Specifically, the reinforcement learning mechanism generates a control strategy by defining a state space, an action space, and a reward function. The state space encompasses key operating parameters of the source-grid-load-storage system, such as the real-time output of each power generation device on the source side, the power flow distribution of each line on the grid side, the real-time power consumption of various loads on the load side, the charge state of each energy storage station on the storage side, and system frequency deviation. The action space discretizes the power regulation of the source and storage devices and the interruptibility of the loads, for example, dividing the power regulation into multiple levels such as ±10% and ±20%.

[0051] The reward function is used to evaluate the quality of control actions. In actual operation, when the system experiences a sudden drop in wind power output, the control strategy may choose to increase the discharge power of the energy storage power station and appropriately reduce some interruptible loads. If this action reduces the system power imbalance, restores frequency stability, and keeps the control cost within an acceptable range, a positive reward will be given; conversely, if it causes system instability or a significant increase in costs, a negative reward will be given. Through continuous training, the deep Q network can learn the optimal control strategy in various complex scenarios. For example, during peak load periods on holidays, it can quickly make decisions such as adjusting source and storage output and guiding load shifting.

[0052] S5: Build a hierarchical control architecture with a time-scale matching module. The minute-level scheduling layer performs rolling optimization of the day-ahead plan. The second-level control layer uses a model predictive control algorithm to quickly compensate for power fluctuations. The millisecond-level protection layer triggers emergency control measures based on transient stability criteria.

[0053] Specifically, the hierarchical control architecture consists of a minute-level scheduling layer, a second-level control layer, and a millisecond-level protection layer. The minute-level scheduling layer performs rolling optimization based on the day-ahead plan, combined with the latest weather forecasts, load forecasts, and other information. For example, if a significant increase in photovoltaic output is predicted by noon in the morning, the generation plan of the thermal power units is adjusted in advance, reducing their output to make room for photovoltaic integration. The plan is updated every 15 minutes.

[0054] The second-level control layer primarily addresses rapid power fluctuations. When sudden changes in wind power occur, the model predictive control algorithm calculates the power adjustments for the source and storage devices based on the current system state and the predicted trajectory over the next few minutes. For example, if a wind farm's power suddenly drops by 5 megawatts, the second-level control layer can issue a command within 1 second, prompting nearby energy storage plants to quickly discharge power to replenish the system's power and maintain system power balance.

[0055] The millisecond-level protection layer is the last line of defense for ensuring system transient stability. When a severe disturbance, such as a short-circuit fault, occurs, the millisecond-level protection layer, based on transient stability criteria, triggers emergency control measures, such as rapidly removing non-critical loads or adjusting generator excitation, to prevent system collapse within milliseconds.

[0056] S6: Establish a closed-loop feedback correction mechanism to align the actual operation data of the system after regulation with the predicted data in time and space dimensions, update the dynamic correlation model parameters through the Kalman filter algorithm, and realize adaptive iterative optimization of the regulation strategy.

[0057] Specifically, a closed-loop feedback correction mechanism ensures continuous optimization of control strategies. During data alignment, due to the varying sampling frequencies of different devices—for example, PMUs sample 50 times per second, while smart meters sample once per minute—a dynamic time warping algorithm is employed to precisely align these data across time. For example, if a power grid fault occurs at a specific moment, this algorithm can accurately align the operating data of each device before and after the fault.

[0058] The Kalman filter algorithm is used to update system model parameters. During system operation, the model's state transition matrix and observation matrix are dynamically adjusted based on the deviation between the actual collected data and the model's predicted data. For example, when a discrepancy is detected between the actual charge and discharge efficiency of an energy storage plant and the model's set values, the algorithm adjusts the relevant parameters so that the model more accurately reflects the characteristics of the energy storage plant. Through continuous feedback correction, the system can adapt to conditions such as equipment aging and changes in grid topology, ensuring that the control strategy remains effective. For example, after a new transmission line is added to the grid, after a period of feedback learning, the system can quickly adjust the control strategy to fully utilize the new line.

[0059] In step S2, the graph neural network model uses a multi-head attention mechanism to perform weighted aggregation on different types of node features, and its node state update formula is:

[0060]

[0061] in, represents the feature vector of node v in layer l, is the feature set of node v’s neighbor nodes, is the weight coefficient of the i-th attention head, is the weight matrix of the i-th attention head, K is the number of attention heads, and σ is the activation function. The model captures the time evolution characteristics of the system state through the time series graph convolution layer, and the convolution kernel parameters are dynamically adjusted according to the grid topology.

[0062] Specifically, the model introduces a multi-head attention mechanism, which is like equipping the system with multiple "observation perspectives," each of which focuses on extracting and fusing features from different types of nodes. For example, in a large regional power grid system, some attention heads focus specifically on the output characteristics of wind power and photovoltaic power generation at the source node, while others focus on the load of the transmission lines at the network node. Through weighted aggregation, the model's understanding of the complex relationships between nodes is more accurate. During actual training, the convolution kernel parameters of the time series graph convolution layer are dynamically adjusted according to the power grid topology. For example, when a new transmission line is added to the power grid, the convolution kernel parameters will change adaptively to ensure that the model can quickly adapt to topological changes. Taking a provincial power grid as an example, after adopting this improved model, the accuracy of system failure prediction has increased by 15%, which has gained valuable time for formulating control strategies in advance.

[0063] Among them, in step S3, the upper model of the hierarchical optimization framework introduces the carbon trading cost function Where ρ is the carbon price, T is the dispatch period, G is the set of power generation equipment, S is the set of energy storage equipment, e i，t With e j，t are the carbon emissions of power generation equipment i and the carbon absorption of energy storage equipment j at time t respectively; the lower model uses a robust optimization method to deal with the renewable energy forecast error and constructs a constraint condition containing the uncertainty set Γ Where ΔP g is the power fluctuation range considering the prediction error.

[0064] Specifically, the upper-level model incorporates carbon trading costs into economic optimization objectives, which is highly relevant. For example, in a regional power grid with a high concentration of energy-intensive industries, when the carbon price is set at 80 yuan / ton, the model calculates that thermal power units proactively reduce their generation plans during high-carbon periods and instead increase the use of low-carbon energy sources such as wind power and photovoltaics. This results in a 20% reduction in carbon emissions within a single scheduling cycle. Combined with peak and valley electricity prices, this reduces overall power generation costs. The lower-level model utilizes robust optimization methods to account for renewable energy forecast errors. Assuming a ±18% forecast error for photovoltaic power plants, the model reserves sufficient power margin when formulating power generation plans. This ensures that even in the event of extreme weather conditions that cause a sudden drop in photovoltaic output, the grid can still meet safety constraints such as line capacity (for example, a key line capacity limit of 600 megawatts) and voltage deviation (node voltage must remain within ±4% of the rated value), thus avoiding operational risks associated with inaccurate forecasts.

[0065] Among them, in step S4, the state space of the reinforcement learning model includes the real-time output of the source-side power generation equipment, the grid-side line flow distribution, the load classification status of the load side, the storage-side charge status and the system frequency deviation; the action space is defined as a discretized set of source and storage power regulation quantities and load interruptibility quantities; the reward function is designed as R = -λ1ΔC -λ2ΔP -λ3Δf, where ΔC is the control cost increment, ΔP is the power imbalance, Δf is the frequency deviation, and λ1, λ2, and λ3 are weight coefficients.

[0066] Specifically, the state space comprehensively covers key operational parameters of the source-grid-load-storage system, creating a "panoramic monitoring room" for the model, enabling real-time monitoring of system dynamics. For example, in a city power grid with a high proportion of renewable energy, the state space collects real-time data such as wind speed at each wind farm, solar intensity at photovoltaic power plants, industrial load power consumption profiles, and the SOC of energy storage power plants. The action space discretizes power regulation and load control into different levels, facilitating rapid decision-making by the model. The reward function provides a quantitative evaluation criterion for the effectiveness of control strategies. For example, in winter scenarios with peak heating loads and insufficient wind power output, the model will receive a higher reward if it appropriately increases energy storage discharge and directs some commercial loads to stagger peak power consumption, thereby stabilizing the system frequency at 50±0.2Hz while remaining within budget. Through continuous training, the model's control response speed in similar complex scenarios has increased by 30%, effectively ensuring stable grid operation.

[0067] Among them, in step S5, the time scale matching module of the hierarchical control architecture uses wavelet transform to decompose and reconstruct multi-time scale data, and establishes a bidirectional data channel including prediction error compensation between the minute-level scheduling layer and the second-level control layer; the model predictive control algorithm adopts a rolling time domain optimization strategy to match the predicted trajectory of the next N control cycles with the current state, and the optimization objective function includes the tracking error term and the control quantity change term.

[0068] Specifically, the time scale matching module uses wavelet transforms to build a "translation bridge" for data at different time scales, enabling efficient data exchange between the minute-level scheduling layer and the second-level control layer. For example, in a certain industrial park power grid, the minute-level scheduling layer predicts that the photovoltaic output will increase significantly in the afternoon based on the weather forecast. After processing through wavelet transforms, the key information is quickly transmitted to the second-level control layer. The second-level control layer adopts a rolling time domain optimization strategy. When it detects that the wind power power suddenly drops by 3 megawatts, based on the predicted trajectory for the next 5 minutes, it accurately calculates the power that the energy storage power station needs to discharge immediately within 1 second to maintain the power balance of the system. At the same time, this optimization improves the collaborative efficiency of the hierarchical control architecture between different time scales by 25%, significantly enhancing the system's ability to cope with power fluctuations.

[0069] Among them, in step S6, when the closed-loop feedback correction mechanism aligns the spatiotemporal dimensions of the data, a dynamic time warping algorithm is used to handle the time series matching problem of data with different sampling frequencies; the state transfer matrix of the Kalman filter algorithm is dynamically adjusted according to changes in the system operation mode, and the observation matrix contains the accuracy parameters and installation location information of the measuring equipment to achieve adaptive update of the control strategy parameters.

[0070] Specifically, the dynamic time warping algorithm plays a key role in data alignment. For data with large differences in sampling frequency, such as PMU data with a sampling frequency of 50 times / second and smart meter data with a sampling frequency of 1 time / minute, the algorithm can act like a "time tailor," precisely aligning data from different devices at the time of the same event. For example, in the case of a short-circuit fault in a power grid, the algorithm quickly aligned data from various devices before and after the fault, providing an accurate basis for subsequent analysis. The Kalman filter algorithm dynamically adjusts system model parameters based on the deviation between actual data and model predictions. During the operation of a certain energy storage power station, when a 5% difference in actual charge and discharge efficiency was detected compared to the model's set value, the Kalman filter algorithm quickly adjusted the relevant parameters, making the model's prediction of the energy storage power station's operating status more accurate. After multiple feedback corrections, the effectiveness of the system's control strategy increased by 20%.

[0071] Among them, the adaptive filtering algorithm in step S1 adopts a variable step size minimum mean square error (LMS) algorithm, and the step size factor is dynamically adjusted according to the data variance; the standardized data sequence adopts a feature extraction method based on principal component analysis (PCA) to map high-dimensional data to a low-dimensional space, reducing data redundancy and computational complexity.

[0072] Specifically, the adaptive filtering algorithm uses a variable-step-size minimum mean square error (LMS) algorithm, which automatically adjusts the step size based on data variance during data processing. For example, in the processing of photovoltaic power station data, when weather changes cause power data to fluctuate dramatically (increase variance), the algorithm automatically reduces the step size, improving filtering accuracy and effectively removing noise caused by cloud cover. When the data is stable, the step size is increased to speed up processing. Principal component analysis (PCA) is used to map standardized data sequences. In the processing of data from a large urban power grid, data containing hundreds of dimensions such as voltage, current, power, and frequency are mapped to a low-dimensional space, removing redundant information and increasing data processing efficiency by 40%. This also reduces the complexity of subsequent model calculations, laying the foundation for rapid system analysis and control.

[0073] The dynamic association model in step S2 introduces a topological constraint term to regularize the graph neural network training. The constraint term expression is Ω = ∑ (u，v)∈E |θ u -θ v |, where (u, v) is the set of edges in the graph, θ u and θ v are the phase differences between nodes u and v respectively; by penalizing the situation where the characteristics of adjacent nodes are too different, the model's ability to express the power grid topology is improved.

[0074] Specifically, the introduced topology constraints act as "structural guidelines" for the model. During training, they penalize excessive differences in the characteristics of adjacent nodes, forcing the model to learn information about the grid's topology. For example, in a complex ring grid, during model training, the topology constraints ensure that characteristics such as the phase difference between nodes conform to the grid's actual structure, enhancing the model's ability to express the grid's topology. In practical applications, when the grid undergoes line modifications or the addition of new substations, the model can adapt to topological changes more quickly, reducing the prediction error for system power flow distribution by 12% compared to the unmodified model, providing more reliable support for grid planning and operational control.

[0075] Among them, the deep Q network in step S4 adopts a dual-network architecture, including an evaluation network and a target network. By regularly updating the target network parameters, the value function estimation deviation during training is reduced; an experience replay mechanism is used to store the historical state-action-reward-next state quadruple to alleviate data correlation problems and improve training stability.

[0076] Specifically, the dual-network architecture, with its evaluation network and target network, acts like a "double think tank" for the model. The evaluation network evaluates control strategies in real time, while the target network regularly updates parameters, effectively reducing bias in value function estimation. In a certain renewable energy distribution network, when faced with dramatic fluctuations in wind power, the evaluation network quickly proposes multiple control solutions, while the target network evaluates these solutions using stable parameters, avoiding erroneous decisions caused by evaluation errors. The experience replay mechanism acts as a "treasure trove of experience," storing historical state-action-reward-next-state quadruples. During training, randomly sampled data for learning, addressing data correlation issues. For example, after the system repeatedly handled peak load scenarios, the experience replay mechanism enabled the model to learn a more optimal load regulation strategy, improving training stability by 35% and significantly enhancing the reliability of the control strategy.

[0077] like Figure 2 As shown, a system for implementing the above-mentioned source-grid-load-storage system coordinated control method includes:

[0078] The global data perception unit is equipped with a distributed sensor array and a multi-protocol communication module to collect and pre-process multi-dimensional physical quantities of the source, grid, load and storage system;

[0079] The system dynamic modeling unit includes a graph neural network calculation module and a topological feature extraction module, which is connected to the global data perception unit to build a system dynamic association model;

[0080] The hierarchical optimization decision-making unit integrates the multi-objective decision-making module and the constraint verification module, receives the output data of the system dynamic modeling unit, and performs hierarchical optimization calculations;

[0081] The real-time control execution unit is equipped with a reinforcement learning strategy generation module and a multi-time scale control module, which is connected to the hierarchical optimization decision-making unit to generate and execute the control strategy;

[0082] The data feedback processing unit, which includes a spatiotemporal alignment module and a parameter update module, receives the actual system operation data and feeds it back to the system dynamic modeling unit and the hierarchical optimization decision unit;

[0083] The network communication interaction unit is equipped with a high-speed data transmission protocol and a security encryption module to realize data interaction and instruction transmission between units;

[0084] The global data perception unit deploys high-precision sensors, such as power sensors with an accuracy of ±0.2% installed on thermal power units on the source side. Phasor measurement units with a sampling frequency of 100 times per second are installed at key nodes on the grid side. Combined with a multi-protocol communication module, it enables conversion between protocols such as Modbus and IEC61850, providing a reliable data foundation for the system. The system dynamic modeling unit, based on a graph neural network, processes massive amounts of data in real time. For example, in a provincial power grid, it can respond to a network topology consisting of 500 nodes in milliseconds, building an accurate system dynamic model. The hierarchical optimization decision-making unit coordinates upper-level economic optimization with lower-level safety verification. For example, in a power grid in a concentrated industrial area, the upper level formulates a power generation plan based on a carbon price of 80 yuan / ton and peak and off-peak electricity prices, while the lower level performs safety verification based on a ±15% forecast error for photovoltaic power generation to ensure the plan's feasibility. The real-time control execution unit utilizes reinforcement learning and multi-timescale control. In a grid where renewable energy accounts for 60%, even in the event of a sudden drop of 10 megawatts in wind power, it can execute energy storage discharge and load adjustment commands within one second. The data feedback processing unit uses a dynamic time warping algorithm to align data of different frequencies, such as accurately matching PMU data with a sampling frequency of 50 times per second with smart meter data at a sampling frequency of 1 time per minute. The unit then optimizes model parameters through Kalman filtering. The network communication interaction unit uses a hybrid communication method of high-speed optical fiber and 5G to ensure data transmission latency is less than 10 milliseconds. This unit enables data exchange and command transmission between units. For example, in the event of a power grid fault, the fault data collected by the global data perception unit is rapidly transmitted to other units via the network communication interaction unit. The system dynamic modeling unit updates the model, the hierarchical optimization decision unit generates new strategies, the real-time control execution unit executes operations, and the data feedback processing unit conducts effect evaluation and parameter correction, thereby achieving efficient coordinated control of the source, grid, load and storage system.

[0085] The output end of the global data perception unit is connected to the input end of the system dynamic modeling unit, the output end of the system dynamic modeling unit is connected to the input end of the hierarchical optimization decision unit, the output end of the hierarchical optimization decision unit is connected to the input end of the real-time control execution unit, the output end of the real-time control execution unit acts on the source grid load storage system, and the system dynamic operation data is fed back to the data feedback processing unit through the network communication interaction unit. The output end of the data feedback processing unit is respectively connected to the system dynamic modeling unit and the hierarchical optimization decision unit.

[0086] Currently, the coordinated regulation of power generation, grid, load and storage faces many challenges. This method and system innovates in data, models, regulation and other aspects, effectively overcoming the shortcomings of existing technologies.

[0087] In terms of data processing and model building, existing technologies have problems such as incomplete data collection and low model accuracy. This collaborative control method and system uses a global data collection system, high-precision sensors, and a multi-protocol communication network to accurately collect multi-dimensional physical quantities of the entire source, grid, load, and storage system, solving the problems of incomplete data coverage and inconsistent communication protocols. At the same time, the use of graph neural networks combined with a multi-head attention mechanism and topological constraints can accurately capture the complex dynamic relationships between system nodes. Compared with traditional static models, the prediction accuracy of system faults is improved by 15%, effectively overcoming the shortcomings of existing models that cannot adapt to changes in grid topology and cannot accurately predict.

[0088] Traditional approaches to optimizing control strategies and architectures often focus on single objectives or local optimization, resulting in slow control responses and insufficient flexibility. This collaborative control system utilizes a hierarchical, multi-objective decision-making model. The upper layer incorporates carbon trading costs for economic optimization, while the lower layer utilizes robust optimization to ensure system security. For example, in a power grid in a high-energy-consuming region, carbon emissions were reduced by 20% within a single dispatch cycle, while also lowering power generation costs. The real-time control execution unit leverages reinforcement learning and multi-timescale control to respond to power fluctuations within one second in power grids with a high proportion of renewable energy. This improves control response speed by 30%, significantly enhancing the system's ability to cope with complex scenarios.

[0089] In terms of system closure and adaptability, existing systems lack effective feedback mechanisms, making them difficult to adapt to system changes. This collaborative control method establishes a closed-loop feedback correction mechanism. Through a dynamic time warping algorithm and a Kalman filter algorithm, it achieves precise alignment of data with different sampling frequencies and dynamic updates of model parameters. In actual operation of energy storage power stations, it can quickly adjust model deviations caused by differences in charging and discharging efficiencies, improving the effectiveness of the control strategy by 20%. This ensures that the system can continue to operate efficiently despite equipment aging and changes in grid topology, thus compensating for the poor adaptability of existing technologies.

[0090] While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for coordinated control of a source-grid-load-storage system, characterized in that: The following steps are involved: S1: Build a global data acquisition system covering source-side power generation units, grid-side transmission networks, load-side power terminals, and storage-side energy storage devices. This system synchronously samples real-time power, voltage, frequency, temperature, and state of charge, and eliminates data noise through adaptive filtering algorithms. S2: Utilizes graph neural networks to construct a dynamic association model for the source-grid-load-storage system, including topological connectivity. This model abstracts grid nodes, power generation equipment, load clusters, and energy storage facilities into graph nodes. Based on historical operating data, the interaction weight matrix between nodes is trained to extract spatial features and perform time series prediction of the system's operating status. S3: Establish a multi-objective decision-making model based on a hierarchical optimization framework. The upper layer aims to minimize the operating cost of the entire system, combining power market price signals and carbon emission indicators to generate source storage power generation plans and load distribution benchmarks. The lower layer uses network security constraints as conditions to build a mixed integer programming model with constraints, and performs feasibility verification and dynamic correction on the benchmark plan; S4: A Markov decision process is constructed using the system state space, action space, and reward function of the reinforcement learning model. A deep Q-network is used to train a control strategy to dynamically adjust the source and storage output and load response in scenarios with renewable energy output fluctuations and load mutations. S5: Build a hierarchical control architecture with a time-scale matching module. The minute-level scheduling layer performs rolling optimization of the day-ahead plan. The second-level control layer uses a model predictive control algorithm to quickly compensate for power fluctuations. The millisecond-level protection layer triggers emergency control measures based on transient stability criteria. S6: Establish a closed-loop feedback correction mechanism to align the actual operation data of the system after regulation with the predicted data in time and space dimensions, update the dynamic correlation model parameters through the Kalman filter algorithm, and perform adaptive iterative optimization of the regulation strategy.

2. The coordinated control method of the source-grid-load-storage system according to claim 1 is characterized in that: In step S2, the graph neural network model uses a multi-head attention mechanism to perform weighted aggregation on different types of node features, and its node state update formula is: in, represents the feature vector of node v in layer l, is the feature set of node v’s neighbor nodes, is the weight coefficient of the i-th attention head, is the weight matrix of the i-th attention head, K is the number of attention heads, and σ is the activation function. The model captures the time evolution characteristics of the system state through the time series graph convolution layer, and the convolution kernel parameters are dynamically adjusted according to the grid topology.

3. The coordinated control method of the source-grid-load-storage system according to claim 1, characterized in that: Step S3: the upper model in the multi-objective decision-making model of the hierarchical optimization framework introduces the carbon transaction cost function Where ρ is the carbon price, T is the dispatch period, G is the set of power generation equipment, S is the set of energy storage equipment, e i，t With e j，t are the carbon emissions of power generation equipment i and the carbon absorption of energy storage equipment j at time t respectively; the lower model uses a robust optimization method to deal with the renewable energy forecast error and constructs a constraint condition containing the uncertainty set Γ Where ΔP g is the power fluctuation range considering the prediction error.

4. The coordinated control method of the source-grid-load-storage system according to claim 1, characterized in that: In step S4, the system state space of the reinforcement learning model includes the real-time output of the source-side power generation equipment, the grid-side line flow distribution, the load classification status of the load side, the storage-side charge status and the system frequency deviation; the action space is defined as a discretized set of source-storage power regulation amounts and load interruptibility amounts; the reward function is designed as R = -λ1ΔC -λ2ΔP -λ3Δf, where ΔC is the control cost increment, ΔP is the power imbalance amount, Δf is the frequency deviation, and λ1, λ2, and λ3 are weight coefficients.

5. The coordinated control method of the source-grid-load-storage system according to claim 1, characterized in that: In step S5, the time scale matching module of the hierarchical control architecture uses wavelet transform to decompose and reconstruct multi-time scale data, and establishes a bidirectional data channel including prediction error compensation between the minute-level scheduling layer and the second-level control layer; the model predictive control algorithm adopts a rolling horizon optimization strategy to match the predicted trajectory of the next N control cycles with the current state, and the optimization objective function includes a tracking error term and a control quantity change term.

6. The coordinated control method of the source-grid-load-storage system according to claim 1, characterized in that: In step S6, when the closed-loop feedback correction mechanism aligns the spatiotemporal dimensions of the data, a dynamic time warping algorithm is used to process the time series matching problem of data with different sampling frequencies; the state transfer matrix of the Kalman filter algorithm is dynamically adjusted according to changes in the system operation mode, and the observation matrix contains the accuracy parameters and installation location information of the measuring equipment to perform adaptive updates of the control strategy parameters.

7. The coordinated control method of the source-grid-load-storage system according to claim 1, characterized in that: In step S1, the adaptive filtering algorithm adopts a variable step-size minimum mean square error (LMS) algorithm, and the step-size factor is dynamically adjusted according to the data variance; the standardized data sequence adopts a feature extraction method based on principal component analysis (PCA) to map high-dimensional data to a low-dimensional space, thereby reducing data redundancy and computational complexity.

8. The coordinated control method of the source-grid-load-storage system according to claim 1, characterized in that: Step S2: The dynamic association model introduces a topological constraint term to regularize the graph neural network training. The constraint term expression is Ω=∑ (u，v)∈E |θ u -θ v |, where (u, v) is the set of edges in the graph, θ u and θ v are the phase differences between nodes u and v respectively.

9. The coordinated control method of the source-grid-load-storage system according to claim 1, characterized in that: In step S4, the deep Q network adopts a dual network architecture, including an evaluation network and a target network, and reduces the value function estimation deviation during training by regularly updating the target network parameters; an experience replay mechanism is used to store the historical state-action-reward-next state quadruple.

10. A coordinated control system for source, grid, load and storage systems, characterized in that: The system comprises: The global data perception unit is equipped with a distributed sensor array and a multi-protocol communication module to collect and pre-process multi-dimensional physical quantities of the source, grid, load and storage system; The system dynamic modeling unit includes a graph neural network calculation module and a topological feature extraction module, which is connected to the global data perception unit to build a system dynamic association model; The hierarchical optimization decision-making unit integrates the multi-objective decision-making module and the constraint verification module, receives the output data of the system dynamic modeling unit, and performs hierarchical optimization calculations; The real-time control execution unit is equipped with a reinforcement learning strategy generation module and a multi-time scale control module, which is connected to the hierarchical optimization decision-making unit to generate and execute the control strategy; The data feedback processing unit, which includes a spatiotemporal alignment module and a parameter update module, receives the actual system operation data and feeds it back to the system dynamic modeling unit and the hierarchical optimization decision unit; The network communication interaction unit is equipped with a high-speed data transmission protocol and a security encryption module to realize data interaction and instruction transmission between units; The output end of the global data perception unit is connected to the input end of the system dynamic modeling unit, the output end of the system dynamic modeling unit is connected to the input end of the hierarchical optimization decision unit, the output end of the hierarchical optimization decision unit is connected to the input end of the real-time control execution unit, the output end of the real-time control execution unit acts on the source grid load storage system, and the system dynamic operation data is fed back to the data feedback processing unit through the network communication interaction unit, and the output end of the data feedback processing unit is respectively connected to the system dynamic modeling unit and the hierarchical optimization decision unit.

Images