Transformer area autonomous optimization scheduling intelligent method and system considering multiple flexible resources, and storage medium

Abstract

The invention relates to the crossing field of intelligent power grids and artificial intelligence, in particular to an optimal scheduling and charging and discharging control method and system for flexible resources such as electric vehicles, distributed energy storage and adjustable building loads, and a storage medium. According to the method, electric vehicles, distributed energy storage and building loads are modeled into intelligent agents, a scheduling strategy is generated under a centralized training-distributed execution framework through multi-agent reinforcement learning, and quadratic programming safety projection is introduced to ensure that actions meet power grid physical constraints. Cooperative optimization of multiple types of flexible resources is realized, the operation cost is effectively reduced, the peak load is reduced, and the power grid safety and the user satisfaction are guaranteed.

Description

Technical Field

[0001] This invention relates to the intersection of smart grids and artificial intelligence, and in particular to an optimized scheduling and charging / discharging control method, system, and storage medium for flexible resources such as electric vehicles, distributed energy storage, and adjustable building loads. Background Technology

[0002] With the advancement of the "dual-carbon" strategic goals and the rapid construction of new power systems, a large number of distributed renewable energy sources (photovoltaics, wind power, etc.) have been connected to the distribution network, profoundly changing the operating environment of the power system. Its main characteristics include:

[0003] 1. Uncertainty and volatility have increased significantly.

[0004] With the large-scale integration of renewable energy sources such as distributed photovoltaic (PV) and wind power, the uncertainty and volatility of power system operation have increased significantly. The output of PV and wind power is directly affected by meteorological conditions, exhibiting strong volatility and intermittency. This leads to a significant widening of the peak-to-valley difference in the system load curve, frequent fluctuations in operating indicators such as voltage and current, and challenges to grid security and stability. Furthermore, renewable energy forecasting models often struggle to achieve high-precision predictions, and the existence of forecast errors further exacerbates the difficulty of dispatching, resulting in a significant decline in the effectiveness of traditional optimization methods based on deterministic planning.

[0005] 2. Flexibility: Diverse resource types, but underutilization.

[0006] Currently, various flexible resources have been integrated into distribution substations, but their overall utilization rate is insufficient. Electric vehicles (EVs), as typical mobile energy storage units, are increasing in number year by year, possessing enormous potential for peak shaving and demand response. However, the highly random nature of EV users' charging needs, the uncertainty of their travel times, and the diversity of their preferences make unified and coordinated scheduling of EVs difficult. Distributed energy storage devices (such as battery energy storage systems, BESS) have rapid response capabilities, but their economic operation remains unresolved due to capacity limitations and cycle life degradation. Meanwhile, building loads (such as HVAC systems, lighting, and interruptible loads) have some regulatory capacity, but the control process needs to balance user comfort and economy, lacking a flexible and efficient coordination mechanism. Therefore, the enormous potential of flexible resources has not yet been fully explored.

[0007] 3. Existing scheduling methods have obvious limitations.

[0008] To address the aforementioned challenges, existing scheduling methods have significant shortcomings in practical applications. While centralized optimization scheduling methods can obtain globally optimal solutions, they require modeling and uniformly solving the entire system. The computational complexity increases exponentially with resource scale, making it difficult to meet the demands of real-time scheduling for large-scale distributed resources. Furthermore, they face issues of communication latency and data privacy protection. Heuristic or rule-based methods, such as simple responses based on electricity price signals or experience-based scheduling rules, while computationally simple, lack global coordination and adaptability, often failing to guarantee overall system optimality. Prediction-driven methods rely on high-precision forecasts of load and renewable energy output, but forecast errors significantly weaken scheduling effectiveness, making them unreliable in high-uncertainty scenarios. Summary of the Invention

[0009] The purpose of this invention is to improve the efficiency of coordinated utilization and operational control of various flexible resources in a distribution area. By introducing reinforcement learning and multi-agent decision-making mechanisms, a dynamic, rolling, and adaptive optimization scheduling method is achieved to solve the technical bottlenecks of traditional methods in large-scale resource access scenarios, such as insufficient real-time performance, low scheduling efficiency, and difficulty in balancing user constraints.

[0010] To achieve the above objectives, the first aspect of this invention provides an intelligent method for autonomous optimization scheduling of transformer substations, considering various flexible resources. This method integrates a multi-agent framework with deep reinforcement learning (DRL) techniques. Starting from the operational status of resources such as electric vehicles, distributed energy storage, and building loads, it achieves adaptive optimization scheduling of multiple types of flexible resources through environment modeling, action space design, reward function construction, and a centralized training-distributed execution mechanism. This improves the system's peak shaving and valley filling capabilities, operational economy, and user satisfaction. The technical process mainly includes the following four steps:

[0011] Step S1: Environment Modeling and State Space Construction: Electric vehicles, distributed energy storage, and adjustable building loads are modeled as intelligent agents. Their operating status (such as SoC, electricity price, travel demand, load level, etc.) and grid-side information (electricity price curve, load forecast, system constraints) are collected to construct a unified environmental state space representation, providing input for reinforcement learning training.

[0012] Step S2: Action Space Definition and Candidate Policy Generation: Design a feasible action space for each agent, including charging / discharging power levels, load adjustment amplitude, etc. Based on this, a set of feasible policies that meet the grid operation constraints (voltage, current, and capacity limits) is generated using reinforcement learning algorithms, significantly reducing the computational complexity of global optimization.

[0013] Step S3: Reinforcement Learning Training and Rolling Optimization Scheduling: Construct a reward function with operating cost, peak load, and user satisfaction as core indicators. Employ algorithms such as Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO) to jointly train multiple agents within a "centralized training—distributed execution" framework. Continuously update the policy through rolling temporal optimization, enabling the agents to generate real-time optimal scheduling schemes in dynamic environments.

[0014] Step S4: Online Execution and Closed-Loop Feedback Control: In actual operation, each agent selects the optimal action (such as charging / discharging power, load adjustment) based on its local state and grid signals, and feeds back the execution results to the environment and central coordination module, forming a closed-loop scheduling and control process of "perception—decision—execution—feedback". This mechanism ensures that the system can maintain efficient scheduling and stable operation even under uncertain situations such as electricity price fluctuations, sudden travel disruptions, or equipment malfunctions.

[0015] Through the above technical solution, the present invention realizes the coordinated scheduling of multiple types of flexible resources, taking into account system operation safety, economy and user experience, and can be widely applied to new power distribution network operation optimization and smart energy management scenarios.

[0016] A second aspect of the present invention is to provide an autonomous optimization scheduling system for implementing the above-described method, the system comprising:

[0017] The environmental perception and state construction module is used to execute step S1 of claim 1, model electric vehicles, distributed energy storage, and adjustable building loads as intelligent agents, collect their operating status and grid-side information, and construct a unified environmental state space representation.

[0018] The action space definition and strategy generation module is used to execute step S2 of claim 1, design a feasible action space for each agent, including charging / discharging power level and load adjustment range, and generate a set of feasible strategies that meet the grid operation constraints using reinforcement learning algorithms;

[0019] The reinforcement learning training and rolling optimization module is used to execute step S3 of claim 1, construct a reward function with operating cost, peak load and user satisfaction as core indicators, adopt a deep deterministic policy gradient or proximal policy optimization algorithm, jointly train multiple agents under a centralized training-distributed execution framework, and continuously update the policy through rolling time domain optimization.

[0020] The online execution and closed-loop control module is used to execute step S4 of claim 1. In actual operation, each intelligent agent selects the optimal action according to the local state and power grid signal, and feeds back the execution result to the environment and central coordination module to form a closed-loop scheduling and control process.

[0021] The safety projection module, located in the online execution and closed-loop control module, is used to minimize and correct the original actions output by the agent using a quadratic programming method, so as to ensure that the corrected actions meet physical and operational constraints, including node voltage, line current, resource SoC, and user comfort.

[0022] A third aspect of the present invention is to provide an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method as claimed in any one of claims 1 to 4.

[0023] A fourth aspect of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described herein.

[0024] Compared with existing technologies, the intelligent method for autonomous optimization scheduling of distribution areas that considers various flexible resources provided by this invention has the following beneficial effects:

[0025] 1. This invention integrates a multi-agent reinforcement learning framework with a distributed execution mechanism. By sensing the real-time operating status of flexible resources such as electric vehicles, distributed energy storage, and building loads, it adaptively generates scheduling strategies that meet grid constraints, achieving efficient coordination and optimization of multiple resource types. This method overcomes the limitations of poor real-time performance in traditional centralized scheduling and insufficient global scope in heuristic methods, significantly improving the system's scheduling flexibility and resource utilization efficiency.

[0026] 2. Starting from the operation level of the distribution area, this invention constructs a full-process scheduling link that includes "state awareness - strategy generation - rolling optimization - closed-loop execution". It comprehensively considers various operational objectives and constraints such as operating costs, electricity price signal response, load peak shaving and valley filling, and user constraints, effectively ensuring the feasibility and economic efficiency of the scheduling scheme.

[0027] 3. The "centralized training-distributed execution" mode proposed in this invention ensures global coordination while enabling various flexible resources to make independent local decisions, significantly reducing the computational burden of central optimization, enhancing the scalability and real-time performance of the method, and making it suitable for flexible scheduling in large-scale resource access and complex scenarios.

[0028] Through simulation verification on a typical system containing 500 electric vehicles, 10 distributed energy storage units, and 20 building loads, the results show that the method proposed in this invention outperforms traditional centralized optimization methods, rule-based heuristic methods, and single-agent reinforcement learning methods in terms of peak load reduction rate, operating cost reduction, and user satisfaction improvement. It can effectively improve the operation optimization level of distribution network areas under the participation of multiple flexible resources, and is suitable for the operation scenario of new distribution systems with high penetration of renewable energy and distributed resource access. It has good engineering adaptability and promotion value. Attached Figure Description

[0029] Figure 1 This is a flowchart of the flexible resource optimization scheduling method based on reinforcement learning of the present invention. Detailed Implementation

[0030] The following, in conjunction with the accompanying drawings and implementation steps, provides a more detailed description of the reinforcement learning-based optimization scheduling method for flexible resources (electric vehicles, distributed energy storage systems (BESS), adjustable building loads, etc.) proposed in this invention. The overall process follows a cyclical mechanism of "state awareness—policy generation (centralized training)—distributed execution—safe projection—closed-loop feedback," achieving rolling and adaptive optimization scheduling at the operational level.

[0031] Step 1: Modeling System Participants and the Graph Environment

[0032] To unify the power grid topology and flexibility resource status into a single representation space, this invention employs a graph structure to model the distribution network: Let the graph be... ,in It is a set of nodes (bus, load nodes, and nodes containing resources). For edge collection (line, tie line, switch). It is an adjacency matrix.

[0033] Node feature vectors:

[0034] (1)

[0035] Symbol explanation:

[0036] :node time Voltage amplitude, in pu;

[0037] , :node time Unadjustable base reactive / non-reactive load (MW / Mvar);

[0038] :node time EV combined charge and discharge power (MW; discharge can be written as negative);

[0039] :node time The average or polymeric state of charge (pu or %) of the EV.

[0040] , :node time BESS has active power and SoC;

[0041] :node time Indoor temperature (°C);

[0042] :node time The set temperature (°C);

[0043] :time Electricity price (¥ / MWh);

[0044] Time step encoding (e.g., 15-minute slot index);

[0045] Edge feature vectors:

[0046] (2)

[0047] :line Resistance / Reactance (pu or Ω / Ω);

[0048] :line Apparent power limit (MVA);

[0049] :line time Switch status (0 / 1);

[0050] Adjacency matrix definition:

[0051] (3)

[0052] Step 2: Multi-agent (EV / BESS / HVAC) state and action modeling

[0053] Flexibility resources are divided into three types of agent sets. Electric vehicles (EVs), Building Energy Storage System (BESS), and Building Adjustable Load (HVAC) – Total Intelligent System Integration .

[0054] The state and actions of the EV agent:

[0055] The state vector is:

[0056] (4)

[0057] : No. EVs in time SoC ([0,1]);

[0058] : No. EV charging / discharging efficiency ((0,1]);

[0059] : No. Arrival / departure time of EVs;

[0060] No. EVs in time action The unit is MW (discharge can be negative). For the first EVs in time The charging and discharging power, The first The lower and upper limits of EV charging and discharging.

[0061] SoC constraints:

[0062] (5)

[0063] in:

[0064] : No. EVs in time The remaining battery power;

[0065] Battery capacity (MWh); Step size (h);

[0066] ;

[0067] : No. The lower and upper limits of the remaining battery power of an EV.

[0068] BESS agent:

[0069] action , For the first b BESS at time The charging and discharging power, The first b The lower and upper bounds of the BESS charge and discharge power; the SoC constraint is the same as the EV's SoC constraint (5) (efficiency and capacity parameters correspond to BESS). Function: rapid power smoothing, backup support.

[0070] HVAC agent:

[0071] action Unit MW (cooling / heating converted to equivalent electrical power). For the first h HVAC in time The charging and discharging power, For the first h The upper limit of HVAC charging and discharging power.

[0072] First-order RC dynamics at room temperature

[0073] (6)

[0074] in, For the first h HVAC in time The internal temperature, For the time change, For the first h The heat capacity of an HVAC unit For the first h The thermal resistance of an HVAC unit; For time external temperature; For the first h Equivalent cooling / heating efficiency of each HVAC unit; For the first h The comfort constraints of an HVAC system.

[0075] Net power injection at nodes:

[0076] (7)

[0077] in, :node i In time t Renewable energy output (MW).

[0078] Step 3: Reward Function (Cost - Peak - User - Grid)

[0079] System operating costs

[0080] (8)

[0081] in, :node i In time t Power purchased from the grid (MW, with purchase counted as positive, transmission counted as negative, or one-sided); Battery aging cost factor (¥ / MWh).

[0082] Peak suppression index

[0083] (9)

[0084] Function: To suppress the instantaneous peak value of total power purchase in the system.

[0085] User-side losses and comfort

[0086] EV range guarantee:

[0087] (10)

[0088] in, : Insufficient penalty weight (¥ / pu);

[0089] : No. k The initial charge of each EV.

[0090] HVAC comfort:

[0091] (11)

[0092] Comfort weight (¥ / ℃) 2 );

[0093] : No. h HVAC in time t The set temperature.

[0094] Soft constraint penalty for power grid

[0095] (12)

[0096] in, Voltage upper and lower limits (pu); Weighting of penalties for exceeding limits; Branch apparent power (MVA); Penalty weight for exceeding capacity limits.

[0097] Global Rewards

[0098] (13)

[0099] in, : Reward weight (dimensionless); parameter tuning is used to balance economy, peak, user, and power grid.

[0100] Step 4: Centralized Training – Distributed Execution (CTDE) and Policy Learning (PPO)

[0101] PPO actor goals

[0102] (14)

[0103] in, Actor parameters; Cut off half the width; Entropy regularization weights.

[0104] : probability ratio Limited to the range If the value exceeds the specified range, it will be truncated.

[0105] Entropy: The probability distribution of actions, representing the uncertainty of the strategy.

[0106] Other symbols are shown below:

[0107] probability ratio

[0108] (15)

[0109] in: Indicates the current strategy parameters Under these circumstances, the agent is in a state Select Action The probability of; Indicates the old strategy parameters Under the same state Select Action The probability of. Measure the relative probability of the old and new strategies being in the same state-action pair. When, it indicates that the new strategy is more inclined to choose that action than the old strategy; when If the ratio is less than the ratio of the old policy to the new policy, it indicates that the new policy is less likely to choose that action. This ratio is used in the PPO objective function to constrain the magnitude of policy updates, thereby avoiding excessive differences between the old and new policies and ensuring training stability.

[0110] Critics Network and Generalized Advantage Estimation

[0111] In the PPO algorithm, the critic network is used to estimate the state-value function. Its loss function is defined as: .in, For critics' parameters, Indicates in the parameter The following value estimate, For target value.

[0112] To reduce variance while taking bias into account, we introduce Generalized Advantage Estimation (GAE), which takes the following form:

[0113] (16)

[0114] in,

[0115] Generalized advantage estimator (GAE) measures the advantage in a given state. Take action Then, the degree of advantage compared to the benchmark value function; Time-series difference error (TD error) reflects the difference between the one-step return and the value function estimate; :time t Instant rewards; :GAE parameters; Discount factor; : Backtrack length.

[0116] Step 5: Online Distributed Execution and Safety Projection (QP Safety Filter)

[0117] To ensure strict feasibility of online execution, a minimum modification projection QP is introduced to transform the policy action. Mapping to the feasible region yields .

[0118] (17)

[0119] in, / Original / projected motion vector (MW); Action upper and lower bounds (physical / operational limits).

[0120] Step 6: Closed-loop execution and special event handling

[0121] Closed-loop process

[0122] In each time slot State acquisition → Local decision-making by each agent →Safe projection →Execution Control→Collecting New Status and Rewards →Next time slot.

[0123] Early departure guarantee

[0124] (18)

[0125] in, Minimum SoC required for user off-site operation; : Maximum charging power; when this cannot be met, priority yielding is triggered (sacrificing economy → adjusting comfort → reducing output of other resources) to ensure safety.

[0126] This invention provides a specific embodiment:

[0127] 1. Scene and Data Settings

[0128] Network and Time Domain

[0129] The IEEE 123-node test system was used, with a node count of... , the number of branches Node voltage permissible range ; Rolling intraday scheduling 24 hours, time resolution Total time slots The network uses images. Representation, adjacency matrix Used for power flow and constraint checks.

[0130] Electricity price, load and weather

[0131] Time-of-use electricity pricing It exhibits distinct peaks and valleys, with evening peaks; the basic load at each node Based on actual measurement scaling; outdoor temperature From a typical summer day curve.

[0132] Resource scale and distribution

[0133] Number of electric vehicles (EVs) Distributed across 20 nodes with piles Number of distributed energy storage BESS Adjustable number of HVAC units in buildings Renewable energy penetration is approximately 35%, PV output It exhibits a peak curve at midday.

[0134] Node / Edge Attributes

[0135] Node feature vectors With edge features The definition is the same as step 1 in "Detailed Implementation", including voltage. ,load Resource power and SoC, indoor / set temperature wait.

[0136] 2. Resource Model and Key Parameters

[0137] EV model

[0138] Maximum power per vehicle Battery capacity ,efficiency ; SoC recursion can be seen in formula (5), off-site requirements .

[0139] BESS model

[0140] Capacity per unit ), power limit , SoC boundary [0.1, 0.95].

[0141] HVAC Model

[0142] The heat capacity and thermal resistance are respectively: Temperature constraints The room temperature kinetics are shown in equation (6).

[0143] Node Injection Power

[0144] (19)

[0145] 3. Algorithm Framework and Training Deployment

[0146] Method: Multi-agent reinforcement learning (MARL) + centralized training and distributed execution (CTDE) + secure projection QP, with the global reward as Equation (13).

[0147] PPO Training

[0148] Actor's Goal probability ratio Commentators estimate using GAE .

[0149] Key hyperparameter: learning rate , , , Entropy weight .

[0150] Online execution and secure projection

[0151] Strategy Output The QP minimum correction projection is QP is shown in equation (17), ensuring With SoC / comfort constraints.

[0152] 4. Comparison with other methods

[0153] Rule: Price and threshold-based rule control (nighttime charging, peak time limit / stop).

[0154] Opt: Centralized convex optimization (prediction-driven, rolling solution, ignoring some discrete / nonlinear terms).

[0155] Single-RL: Single agent RL (which aggregates resources as a single entity for control).

[0156] Proposed: This invention (MARL+CTDE+QP secure projection).

[0157] Evaluation indicators

[0158] Peak Reduction Rate (PR):

[0159] (20)

[0160] : The system at time under no control (baseline case) t Total active power;

[0161] After adopting the scheduling strategy (control scheme), the system at time... t Total active power;

[0162] The system's maximum peak load under baseline conditions;

[0163] : The maximum peak load of the system after control.

[0164] Cost reduction rate (CS):

[0165] (twenty one)

[0166] EV Satisfaction Rate (EV-Sat):

[0167] (twenty two)

[0168] Mean square of comfort violation (HVAC-MSE):

[0169] (twenty three)

[0170] Voltage over-limit duration (V-viol, min):

[0171] (twenty four)

[0172] result

[0173] Table 1 Key Indicators for the Whole Day

[0174]

[0175] Conclusion: This invention significantly outperforms other inventions in both PR and CS metrics, while also achieving higher EV-Sat and lower HVAC-MSE, indicating that the system benefits were not achieved at the expense of user experience.

[0176] Table 2 Peak Segment Performance (18:00–21:00)

[0177]

[0178] The contributions from the three categories of BESS discharge, EV reverse transmission (V2G), and HVAC load shifting are all highest in this invention, thus achieving the best performance in both peak value and peak quantity.

[0179] Table 3 Constraints Compliance and Real-Time Performance

[0180]

[0181] QP safety projection guarantees zero limit exceedance / zero overload, and a trigger rate of 2.7% indicates that most of the original policy actions are close to the feasible region; P99=118 ms satisfies the requirement. .

[0182] Table 4 Ablation Experiment

[0183]

[0184] Conclusion: Secure projection and CTDE are key to performance and feasibility; HVAC peak shifting capability contributes significantly to PR / CS.

[0185] Sensitivity and robustness analysis:

[0186] PV penetration rate: from 20% to 50%, the PR of this invention increased from 16.0% to 20.4%, CS increased from 10.8% to 14.2%, and HVAC-MSE and EV-Sat remained basically unchanged.

[0187] EV Scale: PR increased from 15.1% to 21.7%, CS increased from 10.2% to 15.3%, and EV-Sat remained above 98%.

[0188] Price / Travel Disturbances: ARR / DEP noise ±20%, PR / CS ​​decrease not exceeding 2.5%; V-viol is always 0 (thanks to QP).

[0189] In the IEEE 123-node system, this invention outperforms heuristic, centralized optimization, and single-agent RL methods in terms of public relations (PR) and service response (CS), outperforms user-side metrics, achieves zero network constraint violations, and meets real-time performance targets. Peak segment energy decomposition shows the strongest synergistic contribution from the three resource types, demonstrating clear engineering interpretability and promotional value.

[0190] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the inventive concept, and these all fall within the protection scope of the present invention.

Claims

1. A smart method for autonomous optimization scheduling of distribution transformer areas considering multiple flexible resources, characterized in that, The method includes the following steps: Step S1. Environment Modeling and State Space Construction: Electric vehicles, distributed energy storage, and adjustable building loads are modeled as intelligent agents, their operating status and grid-side information are collected, and a unified environmental state space representation is constructed to provide input for reinforcement learning training; Step S2. Action Space Definition and Candidate Policy Generation: Design a feasible action space for each agent, including charging / discharging power level and load adjustment range. Based on this, use reinforcement learning algorithm to generate a set of feasible policies that meet the power grid operation constraints. Step S3. Reinforcement learning training and rolling optimization scheduling: Construct a reward function with operating cost, peak load and user satisfaction as core indicators, adopt deep deterministic policy gradient or proximal policy optimization algorithm, and jointly train multiple agents under the centralized training-distributed execution framework. The policy is continuously updated through rolling time domain optimization, so that the agents can generate real-time optimal scheduling schemes in dynamic environments. Step S4. Online execution and closed-loop feedback control: In actual operation, each agent selects the optimal action based on the local state and power grid signal, and feeds back the execution result to the environment and central coordination module, forming a closed-loop scheduling and control process of perception-decision-execution-feedback; In step S4, before executing the action, a safety projection module is introduced, and a quadratic programming method is used to minimize and correct the original action output by the agent, so as to ensure that the corrected action meets physical and operational constraints including node voltage, line current, resource SoC and user comfort.

2. The intelligent method for autonomous optimization scheduling of distribution areas considering multiple flexible resources as described in claim 1, characterized in that, The reward function is as follows: (13) in, : Reward weight (dimensionless); System operating costs (8) in, :node i In time t Power purchased from the grid (MW, with purchase counted as positive, transmission counted as negative, or one-sided); Battery aging cost factor (¥ / MWh); Peak suppression index (9) User-side losses and comfort EV range guarantee: (10) in, : Insufficient penalty weight (¥ / pu); : No. k The initial charge of each EV; HVAC comfort: (11) Comfort weight (¥ / ℃) 2 ); : No. h HVAC in time t The set temperature; Soft constraint penalty for power grid (12) in, Voltage upper and lower limits (pu); Weighting of penalties for exceeding limits; Branch apparent power (MVA); Penalty weight for exceeding capacity limits.

3. The intelligent method for autonomous optimization scheduling of distribution areas considering multiple flexible resources as described in claim 1, characterized in that, The multi-agent reinforcement learning algorithm employs a proximate policy optimization (PPO) algorithm, where the objective function of the actor network in the PPO algorithm is: (14) in, Actor parameters; Cut off half the width; Entropy regularization weights; : probability ratio Limited to the range If the value exceeds the specified range, it will be truncated. Entropy: The probability distribution of actions, representing the uncertainty of the strategy.

4. The intelligent method for autonomous optimization scheduling of distribution areas considering multiple flexible resources according to claim 1, characterized in that, The secure projection module in step S4 is implemented by solving the following quadratic programming problem: Introducing Minimal Modification Projection (QP) to transform policy actions Mapping to the feasible region yields , (17) in, / Original / projected motion vector (MW); Action upper and lower bounds (physical / operational limits).

5. A regional autonomous optimization scheduling system for implementing the method of any one of claims 1 to 4, characterized in that, The system includes: The environmental perception and state construction module is used to execute step S1 of claim 1, model electric vehicles, distributed energy storage, and adjustable building loads as intelligent agents, collect their operating status and grid-side information, and construct a unified environmental state space representation. The action space definition and strategy generation module is used to execute step S2 of claim 1, design a feasible action space for each agent, including charging / discharging power level and load adjustment range, and generate a set of feasible strategies that meet the grid operation constraints using reinforcement learning algorithms; The reinforcement learning training and rolling optimization module is used to execute step S3 of claim 1, construct a reward function with operating cost, peak load and user satisfaction as core indicators, adopt a deep deterministic policy gradient or proximal policy optimization algorithm, jointly train multiple agents under a centralized training-distributed execution framework, and continuously update the policy through rolling time domain optimization. The online execution and closed-loop control module is used to execute step S4 of claim 1. In actual operation, each intelligent agent selects the optimal action according to the local state and power grid signal, and feeds back the execution result to the environment and central coordination module to form a closed-loop scheduling and control process. The safety projection module, located in the online execution and closed-loop control module, is used to minimize and correct the original actions output by the agent using a quadratic programming method, so as to ensure that the corrected actions meet physical and operational constraints, including node voltage, line current, resource SoC, and user comfort.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 4.

Images