Micro-grid real-time load balancing scheduling method and system based on deep reinforcement learning

Abstract

The invention relates to the technical field of micro-grids, and discloses a micro-grid real-time load balancing scheduling method and system based on deep reinforcement learning, and the system comprises a multi-source sensing module, an intelligent decision module, a safety protection module, an execution control module, a digital twin module and an energy efficiency evaluation module. When real-time load balancing scheduling of the micro-grid is carried out, a multi-time-scale scheduling strategy is generated in real time by dynamically coordinating economical efficiency, environmental protection and power supply reliability targets through a deep reinforcement learning agent, so that the problems of high operation cost, standard exceeding of carbon emission and voltage instability caused by a single optimization target in a traditional method are further solved, and the real-time load balancing scheduling of the micro-grid is realized. According to the method, the multi-dimensional collaborative optimization of the micro-grid in a complex operation environment is ensured, and meanwhile, a cloud global optimization module generates a long-cycle strategy, so that the problems of response delay and insufficient expandability of a centralized control architecture are further solved, and the real-time power balance capability in a high-permeability renewable energy source scene is improved.

Description

Technical Field

[0001] This invention relates to the field of microgrid technology, specifically to a real-time load balancing scheduling method and system for microgrids based on deep reinforcement learning. Background Technology

[0002] A microgrid, also translated as micro-network, refers to a small-scale power generation and distribution system consisting of distributed power sources, energy storage devices, energy conversion devices, loads, monitoring and protection devices.

[0003] The microgrid real-time load balancing scheduling method and system based on deep reinforcement learning is a key technology that integrates artificial intelligence and power system optimization. It aims to improve the stability and economy of microgrid operation. The system coordinates distributed power sources, energy storage units and load demand in real time through deep reinforcement learning algorithms to achieve dynamic power balance. This is one of the core functions of the system. It relies on the agent to continuously perceive the grid status and optimize decision-making to ensure the maintenance of power supply quality under the conditions of renewable energy fluctuations and load changes.

[0004] Currently, due to the intermittent nature of renewable energy generation in microgrids, it is difficult to accurately match the output power of distributed power sources with load demand during real-time load dispatch. When the power deviation continues to accumulate, it can cause the bus voltage to exceed the limit and frequency instability, leading to malfunctions of system protection devices or even partial power outages.

[0005] Therefore, a real-time load balancing scheduling method and system for microgrids based on deep reinforcement learning is proposed to solve the above problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a real-time load balancing scheduling method and system for microgrids based on deep reinforcement learning. This solves the problem mentioned in the background technology that continuous accumulation of power deviation can lead to bus voltage exceeding limits and frequency instability, resulting in malfunction of system protection devices or even partial power outages.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a real-time load balancing scheduling method and system for microgrids based on deep reinforcement learning, wherein the method includes the following steps: S1. Real-time acquisition of multi-source heterogeneous data from microgrids to generate dynamic monitoring datasets; S2. Construct the state space of a deep reinforcement learning agent, integrating grid topology parameters, renewable energy output fluctuation characteristics, and load demand time series characteristics; S3. Design a composite reward function that dynamically integrates economic indicators, power supply stability indicators, and environmental protection indicators; S4. The agent is trained using a dual network architecture. A baseline policy model is generated through offline pre-training, and the policy parameters are dynamically updated based on an online learning mechanism. S5. Deploy a real-time decision engine, input the current microgrid operating status into the baseline strategy model, and output multi-time-scale scheduling instructions; S6. Perform security constraint verification. When a scheduling instruction violates the equipment security threshold, activate the elastic correction mechanism to generate an alternative scheduling scheme.

[0008] Preferably, S1 includes: S11. Collect photovoltaic array output curves, wind turbine power fluctuation data, and load node electricity demand through a distributed sensor network. S12, Integrated energy storage system charging and discharging status, diesel generator response delay characteristics and grid interconnection line transmission margin data; S13. Construct a spatiotemporally aligned data matrix to generate a dynamic monitoring dataset with timestamps.

[0009] Preferably, S2 includes: S21. Extract the bus voltage deviation rate and branch power flow over-limit risk index from the power grid topology. S22. Use an LSTM network to generate a characteristic vector of renewable energy power output fluctuation; S23. Extract the spatiotemporal correlation features of load demand using a convolutional neural network; S24. The bus voltage deviation rate, over-limit risk index, fluctuation feature vector, and spatiotemporal correlation feature are fused into a state space.

[0010] Preferably, S3 includes: S31. Calculate power generation costs and peak shaving benefits to generate economic indicators; S32. Monitor voltage and frequency deviations to generate power supply stability indicators; S33, Quantify carbon emission intensity to generate environmental protection indicators; S34. Dynamically adjust the weighting coefficients of the three indicators according to the microgrid operation mode, and insert the formula: ; Character meaning: R represents the reward value, the optimization goal of the reinforcement learning agent. For economic indicators, the cost-effectiveness of power generation. For stability indicators, voltage-frequency deviation, For environmental indicators, carbon emission intensity, α, β, γ are dynamic weighting coefficients; Threshold setting: Weight coefficient threshold: α + β + γ = 1; Metric trigger threshold: When any When this occurs, it indicates performance degradation, activating the dynamic weight adjustment mechanism.

[0011] Preferably, S4 includes: S41. The Actor-Critic network architecture is trained using a historical data pool during the offline pre-training phase. S42. During the online learning phase, real-time interactive data is stored in an experience playback pool; S43. Update the policy network parameters based on the priority sampling mechanism.

[0012] Preferably, S5 includes: S51. Input the second-level data into the short-term scheduling module to generate energy storage charging and discharging commands; S52. Input minute-level data into the mid-term scheduling module to generate diesel generator start / stop commands; S53. Input hourly data into the long-term scheduling module to generate a tie-line power exchange plan.

[0013] Preferably, S6 includes: S61. Verify whether the dispatch command exceeds the energy storage SOC safety range and the generator ramp rate limit; S62. When an out-of-limit condition is detected, a correction strategy based on fuzzy rules is invoked to generate an alternative solution. S63. Verify the safety margin of the alternative solution using Monte Carlo simulation, inserting the formula: ; Character meaning: In energy storage state, As the minimum safety threshold, The maximum safety threshold; Threshold setting: Safe operation threshold: Based on lithium battery aging tests, to prevent overcharging / over-discharging; Over-limit trigger threshold: when and At that time, the S62 correction strategy is activated.

[0014] Preferred options also include: S7. Construct a digital twin verification platform to map the scheduling instructions to a virtual microgrid model for pre-verification; S8. Once the pre-verification is successful, an execution command is sent to the physical device.

[0015] Preferred, including: The multi-source sensing module collects dynamic data of the microgrid through distributed sensing units, generates a standardized monitoring dataset using a spatiotemporal alignment unit, and outputs a clean data stream through a quality verification unit. The intelligent decision-making module receives the clean data stream, generates a reinforcement learning state space through the state construction unit, dynamically generates optimization objectives through the reward calculation unit, and outputs a scheduling instruction set through the policy network unit. The security protection module receives the scheduling instruction set, identifies the risk of exceeding limits through the constraint verification unit, generates alternative solutions through the elastic correction unit, and outputs a safe scheduling instruction through the pre-verification unit. The execution control module receives the safety scheduling instructions, decomposes the control tasks through the multi-timescale parsing unit, and sends instructions to the generator, energy storage and switching equipment through the device communication unit. The digital twin module receives the clean data stream and the security scheduling instructions, constructs a dynamic model of the microgrid through the virtual mapping unit, pre-simulates the scheduling results through the simulation and deduction unit, and adjusts the strategy parameters through the feedback optimization unit. The energy efficiency assessment module integrates the scheduling instruction set and execution results, calculates economic efficiency, stability and environmental protection scores through a multi-indicator fusion unit, and generates a scheduling efficiency heatmap using a visualization unit.

[0016] Preferably, the digital twin module has a built-in equipment aging prediction unit that predicts the lifespan decay curve of key equipment based on historical scheduling data; The energy efficiency assessment module includes a carbon emission tracking unit that generates a carbon footprint traceability report in real time.

[0017] Compared with existing technologies, this invention provides a real-time load balancing scheduling method and system for microgrids based on deep reinforcement learning, which has the following beneficial effects: 1. In this invention, by setting up a multi-objective dynamic optimization mechanism, when performing real-time load balancing scheduling of microgrids, the deep reinforcement learning agent dynamically coordinates the objectives of economy, environmental protection and power supply reliability, and generates multi-time-scale scheduling strategies in real time. This further solves the problems of high operating costs, excessive carbon emissions and voltage instability caused by the single optimization objective in traditional methods, and ensures that microgrids achieve multi-dimensional collaborative optimization in complex operating environments.

[0018] 2. In this invention, by setting up an edge-cloud collaborative decision-making architecture, when responding to load fluctuations at the second level, the edge node executes a lightweight reinforcement learning model to achieve local fast scheduling, while the cloud global optimization module generates long-cycle strategies, further solving the problems of response latency and insufficient scalability of centralized control architecture, and improving the real-time power balance capability in high-penetration renewable energy scenarios.

[0019] 3. In this invention, by setting up an uncertainty adaptive module, when scheduling renewable energy output and load demand, the spatiotemporal feature decomposition technology is used to dynamically identify fluctuation patterns. Combined with the online exploration mechanism of the reinforcement learning agent, the strategy parameters are autonomously adjusted to further solve the supply and demand matching deviation problem caused by the intermittency of photovoltaic / wind power and the randomness of load, and reduce the risk of frequency exceeding the limit caused by power balance deviation.

[0020] 4. In this invention, by setting up a multi-level security protection mechanism, a digital twin platform is used for pre-verification before executing scheduling instructions, and an elastic correction strategy is embedded to dynamically generate alternative solutions, thereby further solving the problem of power outages caused by equipment safety threshold exceeding limits and sudden faults, and ensuring that the microgrid maintains the ability to supply power to critical loads under abnormal operating conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart of the microgrid real-time load balancing scheduling method based on deep reinforcement learning according to the present invention. Figure 2 This is a diagram illustrating the architecture of the microgrid real-time load balancing scheduling system based on deep reinforcement learning, as described in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 The method and system for real-time load balancing scheduling of microgrids based on deep reinforcement learning include the following steps: S1. Real-time acquisition of multi-source heterogeneous data from microgrids to generate dynamic monitoring datasets; S2. Construct the state space of a deep reinforcement learning agent, integrating grid topology parameters, renewable energy output fluctuation characteristics, and load demand time series characteristics; S3. Design a composite reward function that dynamically integrates economic indicators, power supply stability indicators, and environmental protection indicators; S4. The agent is trained using a dual network architecture. A baseline policy model is generated through offline pre-training, and the policy parameters are dynamically updated based on an online learning mechanism. S5. Deploy a real-time decision engine, input the current microgrid operating status into the baseline strategy model, and output multi-time-scale scheduling instructions; S6. Perform security constraint verification. When a scheduling instruction violates the equipment security threshold, activate the elastic correction mechanism to generate an alternative scheduling scheme. S1 includes: S11. Collect photovoltaic array output curves, wind turbine power fluctuation data, and load node electricity demand through a distributed sensor network. S12, Integrated energy storage system charging and discharging status, diesel generator response delay characteristics and grid interconnection line transmission margin data; S13. Construct a spatiotemporally aligned data matrix to generate a dynamic monitoring dataset with timestamps. Add the following formula: ; Character meaning: S is the normalized state feature matrix; This is a raw data matrix of photovoltaic / wind turbine power and load demand. This is the time decay weight matrix. Use the Sigmoid activation function; Threshold setting: Data timeliness threshold: when The weight decays to below 0.1; Feature fusion threshold: when any component (Feature failure) triggers the data re-acquisition mechanism; S2 includes: S21. Extract the bus voltage deviation rate and branch power flow over-limit risk index from the power grid topology. S22. Use an LSTM network to generate a characteristic vector of renewable energy power output fluctuation; S23. Extract the spatiotemporal correlation features of load demand using a convolutional neural network; S24. The bus voltage deviation rate, over-limit risk index, fluctuation characteristic vector, and spatiotemporal correlation characteristics are fused into a state space, and a new formula is added: ; Character meaning: For normalized action vectors, The raw values ​​are for energy storage charging and discharging power, diesel generator power, and grid interconnection power. Assign weighting coefficients to power allocation; Threshold setting: Weight constraints: Energy storage is the priority. Diesel engine as backup; Action exceeding threshold: when When charging and discharging are overloaded, the S6 elastic correction is activated; S3 includes: S31. Calculate power generation costs and peak shaving benefits to generate economic indicators; S32. Monitor voltage and frequency deviations to generate power supply stability indicators; S33, Quantify carbon emission intensity to generate environmental protection indicators; S34. Dynamically adjust the weighting coefficients of the three indicators according to the microgrid operation mode, and insert the formula: ; Character meaning: R represents the reward value, the optimization goal of the reinforcement learning agent. For economic indicators, the cost-effectiveness of power generation. For stability indicators, voltage-frequency deviation, For environmental indicators, carbon emission intensity, α, β, γ are dynamic weighting coefficients; Threshold setting: Weight coefficient threshold: α + β + γ = 1; Metric trigger threshold: When any When this occurs, it indicates performance degradation and activates the dynamic weight adjustment mechanism; S4 includes: S41. The Actor-Critic network architecture is trained using a historical data pool during the offline pre-training phase. S42. During the online learning phase, real-time interactive data is stored in an experience playback pool; S43. Update policy network parameters based on priority sampling mechanism, with the following new formula: ; Character meaning: For the optimization objective of the policy network, The time decay factor, For the future The instant reward for each step, D is the data distribution of the experience replay pool, E is the expected value, (s, a) is the state-action pair (s, a) sampled from the experience replay pool D, and T is the upper limit of the number of time steps; Threshold setting: Attenuation factor threshold: Long-term optimization is the dominant approach; Reward sparsity threshold: when If the reward signal is weak, activate the course learning enhancement mechanism. S5 includes: S51. Input the second-level data into the short-term scheduling module to generate energy storage charging and discharging commands; S52. Input minute-level data into the mid-term scheduling module to generate diesel generator start / stop commands; S53. Input hourly data into the long-term scheduling module to generate a tie-line power exchange plan; S6 includes: S61. Verify whether the dispatch command exceeds the energy storage SOC safety range and the generator ramp rate limit; S62. When an out-of-limit condition is detected, a correction strategy based on fuzzy rules is invoked to generate an alternative solution. S63. Verify the safety margin of the alternative solution using Monte Carlo simulation, inserting the formula: ; Character meaning: In energy storage state, As the minimum safety threshold, The maximum safety threshold; Threshold setting: Safe operation threshold: Based on lithium battery aging tests, to prevent overcharging / over-discharging; Over-limit trigger threshold: when and At that time, the S62 correction strategy is activated; Also includes: S7. Construct a digital twin verification platform to map the scheduling instructions to a virtual microgrid model for pre-verification; S8. After pre-verification passes, an execution command is sent to the physical device, with the following formula added: ; Meaning of the characters: K represents the digital twin consistency index. The power of the i-th node in the virtual / physical system. This is the lower limit of power measurement error; Threshold setting: Validation pass threshold: ; Model mismatch threshold: when When the model fails, it triggers an online update of the twin model; include: The multi-source sensing module collects dynamic data of the microgrid through distributed sensing units, generates a standardized monitoring dataset using a spatiotemporal alignment unit, and outputs a clean data stream through a quality verification unit. The intelligent decision-making module receives the clean data stream, generates a reinforcement learning state space through the state construction unit, dynamically generates optimization objectives through the reward calculation unit, and outputs a scheduling instruction set through the policy network unit. The security protection module receives the scheduling instruction set, identifies the risk of exceeding limits through the constraint verification unit, generates alternative solutions through the elastic correction unit, and outputs a safe scheduling instruction through the pre-verification unit. The execution control module receives the safety scheduling instructions, decomposes the control tasks through the multi-timescale parsing unit, and sends instructions to the generator, energy storage and switching equipment through the device communication unit. The digital twin module receives the clean data stream and the security scheduling instructions, constructs a dynamic model of the microgrid through the virtual mapping unit, pre-simulates the scheduling results through the simulation and deduction unit, and adjusts the strategy parameters through the feedback optimization unit. The energy efficiency assessment module integrates the scheduling instruction set and execution results, calculates economic efficiency, stability and environmental protection scores through a multi-index fusion unit, and generates a scheduling efficiency heatmap using a visualization unit. The digital twin module has a built-in equipment aging prediction unit that predicts the lifespan degradation curve of key equipment based on historical scheduling data. The energy efficiency assessment module includes a carbon emissions tracking unit, which generates real-time carbon footprint traceability reports and adds a new formula: ; Character meaning: The overall energy efficiency index of the system. Scoring is based on economic efficiency, stability, and environmental friendliness. These are dynamic weighting coefficients; Threshold setting: Operating mode threshold: Economic model: ; Safe Mode: ; Energy efficiency alarm threshold: when When performance deteriorates, the system initiates a self-test.

[0024] The real-time load balancing scheduling method for microgrids based on deep reinforcement learning and the system operation steps are as follows: I. Multi-source data acquisition and preprocessing 1. Deployment of Distributed Sensor Networks High-precision sensors are deployed at each node of the microgrid to collect real-time data on photovoltaic array output power curves, wind turbine power fluctuations, load node electricity demand, energy storage system state of charge, diesel generator response delay characteristics, and grid interconnection line transmission margin. Multi-source heterogeneous data is timestamped and quality verified by a spatiotemporal alignment unit to eliminate noise and outliers and generate a standardized monitoring dataset.

[0025] 2. Renewable Energy Output Forecast Based on the collection of solar irradiance, wind speed and temperature data by a distributed meteorological sensing network, a meteorological environment field is generated by a locally weighted Gaussian process regression algorithm. The field is then input into a multi-head self-attention wind power prediction network and an adaptive photovoltaic temperature model, and the real-time power output prediction values ​​of photovoltaic and wind power are output.

[0026] II. State Space Construction and Feature Fusion 1. Power Grid Topology Feature Extraction By analyzing the microgrid topology, key parameters such as bus voltage deviation rate and branch power flow over-limit risk index are extracted to quantify the power grid operation safety boundary.

[0027] 2. Temporal Feature Coding An LSTM network is used to generate a feature vector of renewable energy output fluctuations to capture the intermittent patterns of wind and solar resources.

[0028] By extracting the spatiotemporal correlation features of load demand through convolutional neural networks, the distribution patterns of industrial adjustable load and residential rigid load can be identified.

[0029] 3. State-space integration The grid topology parameters, renewable energy feature vectors, and load time-series features are fused into the state space of the deep reinforcement learning agent, which can be formally represented as: ; in Voltage deviation rate, This is an index indicating the risk of exceeding limits. and These are the characteristic vectors of renewable energy and load, respectively.

[0030] III. Design of Deep Reinforcement Learning Agents 1. Design of Composite Reward Function Dynamically integrating economic, stability, and environmental objectives: Economic indicators: calculate the peak shaving benefits of power generation costs and the cost of purchasing electricity from the external power grid.

[0031] Stability metrics: Monitor voltage frequency deviation rate and penalize out-of-limit events.

[0032] Environmental indicators: quantify carbon emission intensity and promote the consumption of clean energy.

[0033] Reward function normalization: ; The weighting coefficients α, β, and γ are dynamically adjusted according to the microgrid's operating mode.

[0034] 2. Dual Network Architecture Training Offline pre-training: The Actor-Critic network is trained using a historical data pool to generate a baseline policy model. The Actor network outputs scheduled actions, and the Critic network evaluates the value of the actions.

[0035] Online learning: Real-time interactive data is stored in an experience replay pool, and policy network parameters are updated based on a priority sampling mechanism to adapt to source load fluctuations.

[0036] IV. Generation of Dual-Time-Scale Scheduling Strategies 1. Day-ahead scheduling layer The system uses an LSTM network to analyze the load, renewable energy output, and electricity price data for the previous 7 days, and outputs the power dispatch strategy, energy storage target status, and power purchase plan for the next 24 hours through a radial basis function.

[0037] 2. Real-time scheduling layer Based on a deep deterministic policy gradient algorithm, the day-ahead scheduling prediction results are integrated with real-time state data: Second-level instructions: Generate energy storage charging and discharging power instructions to smooth out wind and solar power fluctuations.

[0038] Minute-level instructions: Adjust the start-stop status of the diesel generator set to cope with sudden load changes.

[0039] Hourly directive: Optimize tie-line power exchange plans to reduce electricity purchase costs.

[0040] V. Safety Constraint Verification and Flexible Correction 1. Monitoring of risks of exceeding limits Verify whether the scheduling instructions violate the device safety threshold.

[0041] 2. Flexible correction mechanism When an out-of-limit situation is detected, a correction strategy based on fuzzy rules is invoked to generate an alternative, and its safety margin is verified through Monte Carlo simulation. The correction logic is as follows: ; .

[0042] VI. Digital Twin Pre-validation 1. Virtual mapping Dispatch commands are mapped to a high-fidelity microgrid model on a digital twin platform to simulate the system state after the commands are executed.

[0043] 2. Consistency Verification Calculate the power deviation consistency index between the virtual model and the physical system: ; when If the verification is successful, the model will be updated online; otherwise, it will be considered valid.

[0044] VII. Control Command Parsing and Issuance 1. Instruction Decoding The action space features are parsed into device-level control commands using a fully connected neural network. Renewable energy control: Photovoltaic maximum power tracking signal, wind turbine pitch angle adjustment signal.

[0045] Energy storage control: charging and discharging power commands, grid connection voltage setpoints.

[0046] Grid interaction: power exchange signals, real-time electricity price response strategies.

[0047] 2. Execution across multiple time scales Edge nodes perform second-level energy storage control, while the cloud center coordinates minute-level unit scheduling to ensure that the delay between commands and equipment response matches.

[0048] VIII. Online Learning and Strategy Optimization 1. Dynamic feedback mechanism Collect the operational status data of each unit after executing control commands and feed it back to the state space construction module.

[0049] 2. Strategy Iteration and Optimization The Critic network parameters are updated based on temporal difference errors, and the Actor network is optimized using the policy gradient ascent method to maximize long-term cumulative rewards. ; in This is a future reward decay factor.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A microgrid real-time load balancing scheduling method based on deep reinforcement learning, characterized in that, The method comprises the following steps: S1, real-time acquisition of micro-grid multi-source heterogeneous data to generate dynamic monitoring data set; S2, constructing a state space of a deep reinforcement learning agent, integrating grid topology parameters, renewable energy output fluctuation characteristics and load demand time series characteristics; S3, designing a composite reward function, dynamically integrating economic indicators, power supply stability indicators and environmental protection indicators; S4, training the agent using a dual network architecture, generating a benchmark strategy model through offline pre-training, and dynamically updating the strategy parameters based on an online learning mechanism; S5, deploying a real-time decision engine, inputting the current micro-grid operating state into the benchmark strategy model, and outputting multi-time scale scheduling instructions; S6, performing safety constraint checking, and when the scheduling instructions violate the device safety threshold, activating an elastic correction mechanism to generate an alternative scheduling scheme.

2. The method of claim 1, wherein, The S1 comprises: S11, collecting photovoltaic array output curve, wind turbine power fluctuation data and load node electricity demand through a distributed sensing network; S12, integrating energy storage system charge and discharge state, diesel generator response delay characteristics and grid tie line transmission margin data; S13, constructing a time-space aligned data matrix to generate a dynamic monitoring data set with time stamp. 3.The deep reinforcement learning based microgrid real-time load balancing scheduling method of claim 1, wherein, The S2 comprises: S21, extracting bus voltage deviation rate and branch flow out-of-limit risk index in the grid topology; S22, generating a renewable energy output fluctuation feature vector using an LSTM network; S23, extracting load demand space-time correlation characteristics through a convolutional neural network; S24, fusing the bus voltage deviation rate, out-of-limit risk index, fluctuation feature vector and space-time correlation characteristics into a state space. 4.The deep reinforcement learning based microgrid real-time load balancing scheduling method of claim 1, wherein, The S3 comprises: S31, calculating generation cost peak shaving benefit to generate economic indicators; S32, monitoring voltage frequency deviation to generate power supply stability indicators; S33, quantifying carbon emission intensity to generate environmental protection indicators; S34, dynamically adjusting the weight coefficients of the three indicators according to the micro-grid operating mode, and inserting the formula: ； Character meaning: R is the reward value, the optimization goal of reinforcement learning agent, is the economic index, the cost benefit of power generation, is the stability index, the voltage frequency deviation, is the environmental protection index, the carbon emission intensity, and α, β, and γ are dynamic weight coefficients. Threshold setting: weight coefficient threshold: α + β + γ = 1; Indicator trigger threshold: when any of performance degradation, the weight dynamic adjustment mechanism is activated.

5. The method of claim 1, wherein, The S4 comprises: S41, using historical data pool to train Actor-Critic network architecture in offline pre-training stage; S42, storing real-time interaction data through experience replay pool in online learning stage; S43, updating strategy network parameters based on priority sampling mechanism.

6. The method of claim 1, wherein, The S5 comprises: S51, inputting second-level data into short-term scheduling module to generate energy storage charge and discharge instructions; S52, inputting minute-level data into medium-term scheduling module to generate diesel engine start-stop instructions; S53, inputting hour-level data into long-term scheduling module to generate tie line power exchange plan. 7.The deep reinforcement learning based microgrid real-time load balancing scheduling method of claim 1, wherein, The S6 comprises: S61, checking whether the scheduling instructions exceed the energy storage SOC safety interval and the generator climbing rate limit; S62, when the out-of-limit is detected, calling a correction strategy based on fuzzy rules to generate an alternative scheme; S63, verifying the safety margin of the alternative scheme through Monte Carlo simulation, and inserting the formula: ； Character meaning: is an energy storage state, is a minimum safety threshold, is a maximum safety threshold; Threshold setting: safe operation threshold: , based on lithium battery aging test, prevent overcharge / over discharge; Exceeded threshold: when and S62 the correction strategy is activated. 8.The deep reinforcement learning based microgrid real-time load balancing scheduling method of claim 1, wherein, Further comprising: S7, constructing a digital twin verification platform, mapping the scheduling instructions to a virtual micro-grid model for pre-verification; S8, issuing execution instructions to physical devices when pre-verification is passed.

9. The microgrid real-time load balancing scheduling system based on deep reinforcement learning, realizing the method of claim 1-8, characterized in that, Comprising: A multi-source perception module collects microgrid dynamic data through distributed sensing units, generates a standardized monitoring data set using a time-space alignment unit, and outputs a clean data stream through a quality verification unit; An intelligent decision-making module receives the clean data stream, generates a reinforcement learning state space through a state construction unit, dynamically generates an optimization target using a reward calculation unit, and outputs a scheduling instruction set through a policy network unit; A security protection module receives the scheduling instruction set, identifies the risk of exceeding the limit through a constraint verification unit, generates an alternative solution using an elastic correction unit, and outputs a safe scheduling instruction through a pre-validation unit; An execution control module receives the safe scheduling instruction, decomposes the control task through a multi-time scale analysis unit, and issues instructions to generators, energy storage, and switching devices through a device communication unit; A digital twin module receives the clean data stream and safe scheduling instruction, constructs a microgrid dynamic model through a virtual mapping unit, preplays the scheduling results using a simulation deduction unit, and adjusts the policy parameters through a feedback optimization unit; An energy efficiency evaluation module integrates the scheduling instruction set and execution results, calculates the economic, stability, and environmental scores through a multi-index fusion unit, and generates a scheduling efficiency heat map using a visualization unit. 10.The deep reinforcement learning based microgrid real-time load balancing scheduling method and system of claim 1, wherein: The digital twin module has a built-in device aging prediction unit that predicts the life attenuation curve of key devices based on historical scheduling data; The energy efficiency evaluation module includes a carbon emission tracking unit that generates a carbon footprint traceability report in real time.

Images