A simulation method and system for power distribution network multi-scene power balance

Abstract

The application discloses a kind of simulation method and system for power distribution network multi-scene power balance, this method first constructs the dynamic coupling model including "source-net-load-storage" whole element, accurately depict the randomness of new energy and the interactive characteristics of load, energy storage.Furthermore, based on historical data and probability simulation, automatically generate multi-dimensional operation scene covering typical and extreme cases.The core of simulation is the layered rolling scheduling mechanism of minute, hour, day, respectively responsible for rapid response, prediction scheduling and global balance.In this process, integrate intelligent algorithms such as reinforcement learning, according to system frequency, power deviation and cost, real-time closed-loop optimization is carried out on energy storage, load and power purchase strategy, to improve the accuracy and economy of balance.

Description

Technical Field

[0001] This invention relates to the field of power system simulation and optimal scheduling, and in particular to a simulation method and system for power balance in multiple scenarios of distribution networks. This method can be widely applied to the planning, evaluation, and operation scheduling research of modern distribution networks with rapidly growing diversified loads such as high-proportion renewable energy integration, distributed photovoltaic power, and electric vehicle charging stations. Background Technology

[0002] With the accelerated construction of new power systems, the installed capacity of renewable energy sources such as wind power and photovoltaics continues to rise, and diverse loads such as distributed photovoltaics and charging piles are rapidly increasing, leading to unprecedented complexity and uncertainty in the distribution network system. On the one hand, the output of new energy sources is significantly affected by meteorological conditions, exhibiting random fluctuations and uncertainties. Traditional simulation methods based on typical daily or static load assumptions are difficult to accurately depict the dynamic characteristics of new energy sources in actual operation. On the other hand, the concentrated charging of electric vehicles and the rapid changes in loads in commercial buildings and industrial parks cause the system load to exhibit strong peak-valley fluctuations and sudden load response requirements. In addition, the grid connection of user-side and system-side energy storage devices to participate in regulation provides new possibilities for achieving source-grid-load-storage coordination, but also further increases the complexity of simulation models and the difficulty of algorithm design. Currently, research on distribution network simulation optimization in academia and engineering focuses primarily on single scenarios or a limited number of typical operating conditions, single-time-scale scheduling models, and single-factor regulation strategies. Some studies only conduct daily or hourly simulations under typical load scenarios such as sunny days or holidays, neglecting coverage of extreme conditions such as low wind, high load, and feeder faults. Other studies have constructed linear or quadratic programming models at the daytime or time-period granularity, but lack refined processing for the real-time requirements of minute-level energy storage response and load shifting. Many other results focus on single-item optimization of renewable energy absorption efficiency, load response benefits, or energy storage scheduling benefits, without establishing a dynamic coupling simulation framework integrating source-grid-load-storage, and failing to balance real-time performance, computational efficiency, and simulation accuracy. Existing platforms and methods have significant shortcomings in system adaptability, comprehensive scenario simulation, and intelligent optimization closed-loop capabilities, failing to fully meet the comprehensive needs of distribution networks in high-penetration renewable energy environments for multi-scenario coverage, all-time response, and adaptive strategy evolution in planning, evaluation, and operation scheduling. Summary of the Invention

[0003] Therefore, the technical problem to be solved by the present invention is to overcome the defects of the existing technology, such as single scene coverage, single temporal granularity, static scheduling model and fragmentation of source-grid-load-storage elements, which make it difficult to ensure minute-level rapid response and take into account daily overall energy balance and adaptive optimization under the background of high proportion of new energy access, surge of diversified loads and energy storage participation in regulation.

[0004] To address the aforementioned technical problems, this invention provides a simulation method and system for power balance in multiple scenarios of distribution networks, comprising: S1. Input Module: The module sequentially collects and preprocesses data on the output of new energy sources such as wind power and photovoltaics, the distribution network topology and power flow parameters, diverse loads such as residential, commercial, and industrial loads and electric vehicle charging piles, the status and capacity of user-side and system-side energy storage devices, and the simulation scenario generation conditions.

[0005] S2. Multi-scenario generation module: Based on the scenario generation conditions, Monte Carlo sampling and typical scenario combination technology are used to generate multiple operating scenarios such as sunny, cloudy, low wind, high load, centralized access of charging piles, and feeder fault.

[0006] S3. Layered Rolling Simulation Scheduling Module: The simulation scheduling is performed in a layered nested manner at three time scales: minute, hour, and day. These are used for rapid energy storage response and load shifting, short-term new energy output prediction and reserve capacity scheduling, and overall daily energy balance and power purchase decision-making, respectively.

[0007] S4. Feedback Optimization Control Module: During the simulation process, it integrates intelligent optimization algorithms such as reinforcement learning, dynamic programming, or genetic algorithms to perform closed-loop dynamic adjustment of energy storage output, load response, and grid power purchase based on real-time monitored system frequency deviation, power error, and operating cost function.

[0008] S5. Output Evaluation and Visualization Module: After the simulation is completed, it automatically calculates key performance indicators such as peak-valley difference, load balance error, curtailment rate, energy storage utilization rate, system operating cost, and flexible adjustment capability, and generates visual reports and graphs for distribution network planning, evaluation and operation decision-making.

[0009] The multi-scene generation module combines the Monte Carlo method with typical scene generation formulas:

[0010] in, , , They represent the first The scene's meteorological parameters, load parameters, and fault event parameters. Generate functions for complex scenes.

[0011] The multi-scene generation module also includes hierarchical clustering of the generated scenes and combining the k-means algorithm to select cluster center scenes to optimize scene representativeness.

[0012] The minute-level stage of the hierarchical rolling simulation scheduling module adopts the following energy storage charging and discharging control model:

[0013] in, For energy storage at any time The charging and discharging power, , These are the power output and load demand of the same period, respectively.

[0014] Hourly-level phases are based on short-term forecasts and risk reserve capacity models:

[0015] in, For the first Hourly new energy power output forecast Standard deviation, , This is the safety scheduling coefficient.

[0016] The following energy balance and electricity purchase optimization model is adopted at the daily level:

[0017]

[0018] in, For the first day Electricity purchase during specific time periods To correspond to the electricity price, This represents the net change in energy storage charging and discharging.

[0019] The feedback optimization control module employs a policy gradient reinforcement learning algorithm, and its gradient update formula is:

[0020] in, For the policy function, For instant rewards.

[0021] The load response control further includes a price incentive model:

[0022] And with upper and lower limits of load dispatchability constraints:

[0023] The input module updates data in real time by accessing meteorological API, SCADA and PMU systems, and establishes a local time series database to support long-cycle simulation.

[0024] The typical scenarios described in the multi-scenario generation module must cover at least seven operating conditions, including sunny days, cloudy days, high winds, low winds, high loads, holidays, and faulty feeders.

[0025] The hierarchical rolling simulation scheduling module schedules data every minute in the time-level phase. It runs once every few minutes, and every hourly interval. It runs once an hour, and in the daily phase it runs once at midnight every day.

[0026] The feedback optimization control module adopts an alternating optimization strategy, alternately updating energy storage output, load shifting strategy and power purchase decision until each sub-problem converges.

[0027] The output evaluation and visualization module automatically calculates and visualizes peak-valley ratio, load balancing error, curtailment rate, energy storage cycle efficiency, grid purchase cost, and system flexibility index, and supports exporting PDF, Excel reports, and HTML dashboards.

[0028] The power balance simulation method for distribution networks in multiple scenarios and all time periods provided in this application has the following advantages: The power balance simulation method for distribution networks across multiple scenarios and all time periods provided in this application has the following advantages: By unifying the modeling of four key elements—"sources," renewable energy sources such as wind and solar power; "grid," distribution network power flow and topology; "load," diverse loads including residential, commercial, industrial, and electric vehicle charging piles; and "storage," user-side and system-side energy storage devices—it achieves full-element coupling and coordinated scheduling, overcoming the shortcomings of traditional single-element simulations that cannot accurately depict system dynamics. Based on a multi-scenario generation mechanism combining Monte Carlo sampling and hierarchical clustering of typical scenarios, it can systematically cover typical and extreme operating conditions such as sunny, cloudy, low wind, high load, concentrated electric vehicle charging, and feeder faults, greatly enhancing the robustness and wide applicability of the simulation results. It innovatively constructs three time-level hierarchical rolling simulation architectures: minute-level focusing on rapid energy storage response and load peak shifting, hour-level focusing on... The system considers short-term forecasting of new energy sources and reserve capacity scheduling, while daily-level optimization of energy balance and power purchase strategies ensures rapid peak-shaving and valley filling for sudden power deviations, while also taking into account accurate planning for medium- and long-term operating trends. Intelligent optimization strategies such as reinforcement learning, dynamic programming, and genetic algorithms are incorporated into the simulation process. Using system frequency deviation, power error, and operating cost functions as evaluation indicators, real-time adaptive closed-loop regulation of energy storage output, load response, and power purchase decisions is achieved, significantly improving the economy and self-optimization capabilities of the scheduling scheme. Furthermore, after the simulation, multiple performance indicators are automatically generated, including peak-valley difference, load balance error, curtailment rate, energy storage utilization rate, power purchase cost, and system flexibility, and output in intuitive report and graph formats. This provides systematic and engineering-oriented decision support for distribution network planning, evaluation, and operation scheduling, significantly reducing operational risks and costs, and demonstrating good engineering feasibility and application value. Attached Figure Description

[0029] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This application provides a simulation method for power balance in multiple scenarios of power distribution networks, along with a schematic diagram of the overall system structure and data flow. Figure 2 The hierarchical rolling simulation scheduling flowchart provided for this application; Figure 3 Block diagram of the intelligent feedback optimization control module provided in this application; Figure 4 Line charts comparing system balancing capabilities under multiple scenarios provided for this application; Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0031] Reference Figure 1 As shown in the diagram, the overall structure and data flow of a multi-scenario power balance simulation method for distribution networks based on multi-level rolling scheduling provided in this application include: A multi-scenario power balance simulation method for distribution networks based on multi-level rolling scheduling, characterized by comprising: S1. Input Module: Through interfaces such as SCADA / PMU and meteorological API, it collects and preprocesses in real time the "source" power output of new energy sources such as wind power and photovoltaics, the "grid" power flow and topology data of distribution network such as bus voltage level, line impedance and capacity, and topology switch status, the "load" load curves of residential, commercial, industrial and electric vehicle charging piles, and the "storage" initial simulation data of the SOC and capacity C of user-side and system-side energy storage devices.

[0032] S2. Data Acquisition and Preprocessing Module: Cleans, interpolates, and converts the original data of the above elements, constructs a time-series database, and calculates the load forecast and available energy storage capacity according to simulation requirements to ensure the consistency and completeness of subsequent simulation inputs.

[0033] S3. Multi-Scene Generation Module: Based on the time-series data output by S2, it uses Monte Carlo random sampling to generate data containing meteorological parameters. Load parameters and fault parameters By combining the sample set with hierarchical clustering of typical scenarios, a systematic framework is constructed that includes scenarios such as sunny, cloudy, low wind, high load, centralized charging of electric vehicles, and feeder faults. Typical and extreme operating scenarios .

[0034] S4. Layered Rolling Simulation Scheduling Module: Performs rolling scheduling nested across three time scales (minute, hour, and day) within the same simulation platform. Minute-level: For each scenario Every Minute-level call-to-response energy storage rapid response and load peak shifting control strategy; Hourly level: Based on historical output and uncertainty risk assessment, calculated per... Hourly dispatch of new energy forecasting and backup energy storage capacity; Daily Level: Performs full-day energy balancing and electricity purchase optimization at midnight every day.

[0035] S5. Feedback Optimization Control Module: During the simulation iteration process in S4, intelligent optimization algorithms such as reinforcement learning, dynamic programming, or genetic algorithms are introduced to perform closed-loop adaptive adjustment of energy storage output, load response, and power purchase strategy. The system frequency deviation, power deviation, and operating cost are used as feedback basis to achieve "simulation is optimization".

[0036] S6 Output Evaluation and Visualization Module: After the simulation is completed, it automatically summarizes multi-dimensional performance indicators such as peak-valley difference, load balance error, curtailment rate, energy storage cycle efficiency, electricity purchase cost and system flexibility adjustment capability under various scenarios, and generates interactive reports and visualization results such as line charts and bar charts. It supports exporting to PDF and Excel for planning evaluation and operation decision-making.

[0037] The multi-scene generation module uses the following formula to generate long letters based on composite scenes:

[0038] in,

[0039] In the formula , , ,

[0040] This formula is derived from meteorological parameters. Load parameters Sampling is performed using a normal distribution, and a fault indication vector is introduced. This enables a unified portrayal of both typical and random scenarios.

[0041] The multi-scene generation module further includes a hierarchical clustering step based on k-means clustering:

[0042]

[0043] in, For the first Scene clusters, As the cluster center, the cluster center scenario is selected as the final typical scenario.

[0044] The minute-level stage of the hierarchical rolling simulation scheduling module adopts the following energy storage charging and discharging optimization model:

[0045]

[0046]

[0047] in, , These are the weighting coefficients. The time window length is in minutes. It is in energy storage mode.

[0048] The hourly phase employs the following multi-objective reserve capacity scheduling model:

[0049]

[0050] in, and The first Expected and standard deviation of hourly new energy power output , , This is the scheduling coefficient.

[0051] The daily-level stage adopts the following multi-time period electricity purchase and energy balance joint optimization model:

[0052]

[0053]

[0054] in, To balance the error penalty coefficient, This represents the maximum amount of electricity that can be purchased.

[0055] The feedback optimization control module is based on a reinforcement learning algorithm using policy gradients, and its policy gradient update formula is as follows:

[0056] in, To accumulate rewards, For the baseline function, This is a parameterized strategy.

[0057] The reward function is defined as follows:

[0058] in, , These are the weighting coefficients. and These are time-of-use electricity prices and the amount of electricity purchased, respectively.

[0059] The multidimensional visualization performance indicators include peak-to-valley ratio, load balancing error, curtailment rate, energy storage cycle efficiency, electricity purchase cost, and system flexibility adjustment capability, and support output in the form of line graphs, heat maps, and interactive dashboards.

[0060] The simulation platform is deployed in a Kubernetes cloud environment, adopts a microservice architecture, and its modules communicate with each other through RESTful APIs. It also supports real-time data push via WebSocket and visualization front-end display.

[0061] Reference Figure 2 As shown, the hierarchical rolling simulation scheduling flowchart provided in this application includes: During the simulation preparation phase, a typical operating scenario is considered. The system obtains the input dataset for this scene from the multi-scene generation module, including: New energy instantaneous output sequence ; Load demand ; Initial State of Charge (SOC) and upper and lower capacity limits for energy storage; This scenario corresponds to time division parameters at the day, hour, and minute levels.

[0062] Entering a daily loop, the outer index uses the simulated number of days. This is a loop variable. For each day, the system pre-sets the upper and lower limits of the total amount of electricity that can be purchased that day, as well as the electricity price curve. And clear the cumulative energy deviation for the day.

[0063] Within the daily timeframe, an hourly sub-loop is triggered to... The index is updated hourly. At the top of each hour, the system first invokes an optimization based on the "uncertainty reserve capacity" model:

[0064]

[0065] in and These represent the predicted average output and fluctuation range coefficient for that hour, respectively. , , This is determined by the user's risk appetite and system reliability requirements. Once this step is completed, the energy storage reserve capacity for subsequent minute-level use is obtained. .

[0066] After hourly optimization is completed, minute-level rolling scheduling begins, dividing the current hour into... Each minute is a fragmented time slice. For each minute, perform the following steps: (1) Read real-time data: Call the simulation platform interface to obtain the real-time output and load value of new energy in this minute; (2) Calculate the energy storage action: Calculate the increment of charging and discharging power for this minute according to the following formula.

[0067] And comply with the power of the energy storage device and Range constraints; (3) Update status: according to Update the remaining battery power and add the load balancing error of this minute to the daily statistics; (4) Determine if the loop ends: Then it increments. And return to step 1; otherwise, exit the current hour-minute loop and return to the hour-level index.

[0068] When a certain hour After all the minute-by-minute segments have been executed, the system will display the final value for that hour. Record the total load error and the utilization of reserve capacity, and then make a judgment. To decide whether to proceed to the next hour's scheduling.

[0069] After the one-minute and hourly cycles are completed, the scheduling of all hours for the day is finished, corresponding to... Time-triggered daily summary and electricity purchase optimization:

[0070]

[0071] in For the energy storage to discharge (positively) or charge (negatively), a penalty coefficient λ is used to balance the energy gap and the cost of purchasing electricity. Furthermore, throughout the three-layer loop, this embodiment employs an online reinforcement learning strategy, using "system real-time frequency deviation" and "daily cost" as environmental feedback signals, and updates the action policy using the following policy gradient algorithm:

[0072] in This is the cumulative reward from now until the end of the day. This mechanism achieves a deep integration of simulation scheduling and adaptive learning.

[0073] when When that happens, the entire process jumps back to the next round of the daily cycle; otherwise, when all Once the daily average is completed, the system enters the final state and outputs the multi-day balance index trend, energy storage utilization rate change, and cost comparison curve through the S6 module.

[0074] Reference Figure 3 As shown in the diagram, the intelligent feedback optimization control module block diagram provided in this application includes: Input layer: at each simulation time step The input layer collects and outputs the following real-time information: Frequency deviation This reflects the impact of system scheduling errors on operational safety; Power balance error To measure the current matching degree of source, load and storage; Current operating costs This includes electricity purchase costs and energy storage loss costs; State vector This is used for subsequent decision-making.

[0075] This layer ensures data synchronization through high-speed caching and parallel bus technology to meet real-time call requirements within minutes or shorter periods.

[0076] Policy candidate generation unit: The policy candidate generation unit is called in parallel. Reinforcement learning strategies Based on current parameters and state Output a set of action probability distributions; The model predicts candidate solutions using dynamic programming or genetic algorithms in the alternative action space. The local optimal solution set is given in the text.

[0077] Candidate set generation:

[0078] in The value evaluation function is based on DP / GA, with parameters... Learned from historical data.

[0079] Experience Library / Historical Data Unit: This peer module maintains a dedicated experience library, storing validated high-quality control actions and their performance curves (such as cost-benefit curves) from previous simulations. Before action evaluation, it can quickly retrieve the historical records closest to the current state for candidate generation or rollback recovery.

[0080] Similarity retrieval formula:

[0081] Using similarity threshold Selecting historical optimal solutions improves the reliability of decision-making.

[0082] Action Evaluation and Selection Unit: This unit is based on a reward function. For candidate actions Sort the actions and select the optimal ones.

[0083] in,

[0084] It is after combining frequency offset correction and cost Composite evaluation index, , It is set by engineering requirements.

[0085] Action Execution and Model Update Unit Selected actions This includes energy storage charging and discharging, and load response. and purchased electricity The changes were sent to the simulation model, and the following updates were completed: Energy storage status:

[0086] Load response: , Electricity purchase records: .

[0087] Parallel scheduling submodule To support large-scale or regionally segmented distribution networks, this implementation introduces parallel sub-modules during the action execution phase to handle the following separately: Multi-region load adjustment involves parallel computation of load clustering and peak shifting strategies for each node. Subnet node power allocation utilizes a linear power flow solver to update node voltage and flow direction in a time-sharing manner within a parallel thread.

[0088] This parallelization design significantly improves the efficiency of large-scale system simulation.

[0089] Feedback evaluation unit: After the simulation model is updated, the feedback evaluation unit is recalculated.

[0090] Instant rewards: , Cumulative Returns: , And and Stored in the experience base for subsequent strategy updates and candidate retrieval.

[0091] Parameter / policy update unit: Online optimization based on policy gradient or evolutionary algorithm.

[0092] Alternatively, the DP / GA population and fitness weights can be iteratively adjusted to ensure that the controller can continuously learn and improve in a constantly changing operating environment.

[0093] Reference Figure 4 As shown, the line chart comparing the system balancing capabilities under multiple scenarios provided in this application includes: Under five typical operating scenarios—sunny weather, low wind, high load, concentrated charging of electric vehicles, and feeder fault—the proposed multi-scenario, all-time simulation method for power generation, grid, load, and storage demonstrated superior power balancing capabilities. Specifically, in the feeder fault scenario, due to limited external transmission channels, traditional simulations often show significant load gaps. However, this method, by rapidly mobilizing energy storage response and load shifting, reduces the maximum load deviation to 0kW, controls the curtailment rate at 0.8%, achieves an energy storage utilization rate as high as 91%, and reduces grid power purchase costs by 18.2%. In sunny weather scenarios, thanks to the significant fluctuations in photovoltaic output, this method, through minute-level rolling scheduling, controls the maximum load deviation to 320kW, with a curtailment rate of 4.6%, an energy storage utilization rate of 82%, and a 12.5% ​​reduction in electricity purchase costs. In low-wind scenarios, despite insufficient wind power output, this method still ensures that the curtailment rate does not exceed 7.2%, the energy storage utilization rate reaches 75%, and the electricity purchase cost decreases by 9.8%. In high-load scenarios and centralized electric vehicle charging scenarios, the maximum deviations are 500kW and 180kW, respectively, with curtailment rates of 5.5% and 2.1%, energy storage utilization rates of 70% and 69%, and cost reductions of 10.2% and 9.3%. The above data fully demonstrate that the method of this invention can maintain system balance under diverse and extreme operating conditions, balancing reliability and economy, and verifying its robustness and applicability in environments with high penetration of new energy sources and diverse loads.

[0094] Referring to Table 1, the comparison table of full-time rolling simulation and static scheduling provided in this application includes:

[0095] In a continuous 48-hour simulation comparison, the three-level hierarchical rolling simulation of the present invention (minute-hour-day) significantly improves key performance indicators compared with traditional static daily scheduling. First, the hierarchical rolling simulation reduces the total operating cost over the entire cycle from 32,000 yuan to 27,400 yuan, a reduction of 14.4%, mainly due to the dynamic compensation for new energy prediction errors and the real-time optimization of energy storage strategies by hourly and daily scheduling. Second, regarding peak-valley difference regulation, static scheduling struggles to respond promptly to load fluctuations, resulting in a peak-valley difference as high as 560kW. The present method, through minute-level rapid energy storage response and load shifting, compresses the peak-valley difference to 180kW, a reduction of 67.9%, effectively improving system balance accuracy. Furthermore, the energy storage cycle efficiency increases from the traditional 58.4% to 87.2%, an improvement of 49.2%, demonstrating the significant advantages of the present invention in reducing energy storage losses and extending equipment lifespan. In summary, this method not only significantly reduces operating costs but also substantially improves peak-shaving performance and energy storage utilization, demonstrating the engineering value of the "simulation as optimization" framework in all-time, multi-scenario distribution network balance scheduling.

Claims

1. A simulation method and system for power balance in multiple scenarios of distribution networks, characterized in that, include: S1. Input Module: The module collects and preprocesses data on the output of new energy sources such as wind power and photovoltaics, the distribution network topology and power flow parameters, the diverse loads such as residential, commercial, and industrial loads and electric vehicle charging piles, the status and capacity of energy storage devices on the user side and system side, and the conditions for generating simulation scenarios. S2. Multi-scenario generation module: Based on the scenario generation conditions, Monte Carlo sampling and typical scenario combination technology are used to generate multiple operating scenarios such as sunny, cloudy, low wind, high load, centralized access of charging piles, and feeder fault; S3. Layered Rolling Simulation Scheduling Module: The simulation scheduling is performed in a layered nested manner at three time scales: minute, hour, and day. These are used for rapid energy storage response and load shifting, short-term new energy output prediction and reserve capacity scheduling, and overall daily energy balance and power purchase decision-making, respectively. S4. Feedback Optimization Control Module: During the simulation process, it integrates intelligent optimization algorithms such as reinforcement learning, dynamic programming or genetic algorithms, and performs closed-loop dynamic adjustment of energy storage output, load response and grid power purchase based on real-time monitoring of system frequency deviation, power error and operating cost function; S5. Output Evaluation and Visualization Module: After the simulation is completed, it automatically calculates key performance indicators such as peak-valley difference, load balance error, curtailment rate, energy storage utilization rate, system operating cost, and flexible adjustment capability, and generates visual reports and graphs for distribution network planning, evaluation and operation decision-making.

2. The method according to claim 1, characterized in that, The multi-scene generation module combines the Monte Carlo method with typical scene generation formulas: ； in, , , They represent the first The scene's meteorological parameters, load parameters, and fault event parameters. Generate functions for complex scenes.

3. The method according to claim 2, characterized in that, The multi-scene generation module also includes hierarchical clustering of the generated scenes and combining the k-means algorithm to select cluster center scenes to optimize scene representativeness.

4. The method according to claim 3, characterized in that, The minute-level stage of the hierarchical rolling simulation scheduling module adopts the following energy storage charging and discharging control model: ； in, For energy storage at any time The charging and discharging power, , These are the power output and load demand of the same period, respectively.

5. The method according to claim 4, characterized in that, Hourly-level phases are based on short-term forecasts and risk reserve capacity models: ； in, For the first Hourly new energy power output forecast Standard deviation, , This is the safety scheduling coefficient.

6. The method according to claim 5, characterized in that, The following energy balance and electricity purchase optimization model is adopted at the daily level: ； ； in, For the first day Electricity purchase during specific time periods To correspond to the electricity price, This represents the net change in energy storage charging and discharging.

7. The method according to claim 6, characterized in that, The feedback optimization control module employs a policy gradient reinforcement learning algorithm, and its gradient update formula is: ； in, For the policy function, For instant rewards.

8. The method according to claim 7, characterized in that, The load response control further includes a price incentive model: ； And with upper and lower limits of load dispatchability constraints: 。 9. The method according to claim 1, characterized in that, The input module updates data in real time by accessing meteorological API, SCADA and PMU systems, and establishes a local time series database to support long-cycle simulation.

10. The method according to claim 1, characterized in that, The typical scenarios described in the multi-scenario generation module must cover at least seven operating conditions, including sunny days, cloudy days, high winds, low winds, high loads, holidays, and faulty feeders.

11. The method according to claim 1, characterized in that, The hierarchical rolling simulation scheduling module schedules data every minute in the time-level phase. It runs once every few minutes, and every hourly interval. It runs once an hour, and in the daily phase it runs once at midnight every day.

12. The method according to claim 1, characterized in that, The feedback optimization control module adopts an alternating optimization strategy, alternately updating energy storage output, load shifting strategy and power purchase decision until each sub-problem converges.

13. The method according to claim 1, characterized in that, The output evaluation and visualization module automatically calculates and visualizes peak-valley ratio, load balancing error, curtailment rate, energy storage cycle efficiency, grid purchase cost, and system flexibility index, and supports exporting PDF, Excel reports, and HTML dashboards.

Images