Source-grid-load-storage integrated micro-grid cluster multi-energy collaborative optimization scheduling method based on AI

Abstract

The invention discloses an AI-based source-grid-load-storage integrated micro-grid cluster multi-energy collaborative optimization scheduling method, and relates to the technical field of power system automation, artificial intelligence and distributed optimization scheduling, and the method comprises the steps: collecting source-grid-load-storage operation data, and constructing a multi-dimensional input feature set; inputting the multi-dimensional input feature set into a time sequence Transform model to carry out wind and light output and load ultra-short-term rolling prediction; inputting the predicted value into a digital twinborn model for simulation and load flow calculation, and generating a rigid safety constraint; inputting the rigid security constraint and the predicted value into a scaling dual descent alternating direction multiplier method model to execute distributed collaborative optimization solution, and generating and executing a power distribution and mutual aid instruction; and correcting the parameters of the time sequence Transform model and the scaling dual descent alternating direction multiplier method model according to the execution result. Therefore, efficient coordination and stable operation of the source, the network, the load and the storage in the micro-grid cluster are realized.

Description

Technical Field

[0001] The embodiments of the present invention relate to the fields of power system automation, artificial intelligence and distributed optimization scheduling technology, and in particular to an AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters with source, grid, load and storage. Background Technology

[0002] With the in-depth advancement of the construction of new power systems, integrated microgrid clusters with generation, grid, load and storage have become the core grid connection form for achieving high-proportion renewable energy consumption and safe and low-carbon power supply in mainstream scenarios such as industrial parks, county distribution networks and industrial and commercial parks. Their operating characteristics, such as independent operation by multiple property owners, nonlinear coupling of multiple energy flows and stringent grid connection safety requirements, have put forward clear engineering requirements for cluster-level collaborative scheduling technology.

[0003] Existing microgrid dispatching technologies mostly focus on optimizing the operational characteristics of energy storage units within a single microgrid. Among them, the invention patent with application number CN202610039789.4 discloses a microgrid intelligent dispatching method based on multi-energy collaborative optimization. This scheme constructs a set of basic state parameters by acquiring the operating status information of energy storage sub-units, accurately characterizing the power capacity, energy boundary, and operational limitations of energy storage units, thereby determining their unique dispatching state identifier within a single dispatching cycle, limiting energy storage units to participate in only one type of dispatching behavior, and realizing refined dispatching of individual energy storage units in a single microgrid scenario.

[0004] However, the existing technologies (including the above-mentioned solutions) still have core defects that are seriously out of touch with the actual implementation scenarios of microgrid clusters in the power industry, as follows: (1) Insufficient adaptability of the scheduling architecture: The existing solutions only focus on the scheduling optimization of individual energy storage units within a single microgrid, and have not built a distributed collaborative architecture that adapts to the operation of multiple ownership entities and the hierarchical scheduling management requirements of the distribution network in the microgrid cluster, and cannot meet the core engineering requirements of cross-microgrid power mutual assistance. (2) Inadequate solution convergence: At the same time, conventional optimization algorithms cannot adapt to the non-convex optimization scenario of multi-element nonlinear coupling at the cluster level, and are prone to iterative divergence and solution failure, making it difficult to meet the response requirements of real-time scheduling in power engineering. (3) Lack of a full-process closed-loop control mechanism: The existing solutions only realize the energy storage status determination and static scheduling within a single scheduling cycle, and cannot solve the problem of continuous disconnect between the scheduling model and the actual working conditions on site, making it difficult to ensure the long-term stable grid-connected operation of the microgrid cluster.

[0005] Therefore, developing a fully closed-loop optimization scheduling method that adapts to non-convex constraints, distributed collaboration, deep coupling of simulation and scheduling, and refined control of energy storage has become a core technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters with source, grid, load, and storage. By integrating time-series AI prediction, digital twin simulation, distributed collaborative scheduling using the scaling dual alternating descent multiplier method (ADMM), and hierarchical control of hybrid energy storage, this method achieves efficient collaboration and stable operation of all elements (source, grid, load, and storage) within the microgrid cluster.

[0007] In a first aspect, embodiments of the present invention provide an AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters with source-grid-load-storage architecture, including:

[0008] Real-time collection of operational data of all elements of source, grid, load and storage, and data preprocessing, and construction of multi-dimensional input feature set based on the preprocessed data;

[0009] The multidimensional input feature set is input into a pre-built time series Transformer model to perform ultra-short-term rolling prediction of wind and solar power output and load, and the predicted value and confidence interval are output.

[0010] The predicted values ​​are input into the digital twin model for power grid operation simulation and power flow calculation. Based on the simulation results, safety warnings are issued and rigid safety constraints are generated.

[0011] The rigid safety constraints and the predicted values ​​are input into the scaling dual descent alternating direction multiplier method model, and distributed cooperative optimization is performed to generate power allocation and mutual assistance commands.

[0012] The power allocation and mutual assistance commands are sent to the hybrid energy storage system, controllable loads, and distributed power sources for execution.

[0013] The prediction parameters of the temporal Transformer model and the scheduling parameters of the scaling dual descent alternating direction multiplier method model are adjusted based on the execution results.

[0014] As a preferred embodiment, the method further includes constructing an operational framework that links a three-level distributed architecture with a multi-time-scale scheduling mechanism;

[0015] The three-level distributed architecture includes a regional cluster layer, a single microgrid layer, and a device control layer. The regional cluster layer is used to formulate global power mutual assistance strategies and tie-line power boundary plans among multiple microgrids. The single microgrid layer is used to perform local source-grid-load-storage optimization scheduling and security verification. The device control layer is used to execute device commands at the millisecond to minute level.

[0016] The multi-timescale scheduling mechanism includes day-ahead baseline planning, intraday rolling correction, and ultra-short-term real-time smoothing.

[0017] In a preferred embodiment, the multidimensional input feature set is input into a pre-built time-series Transformer model for ultra-short-term rolling prediction of wind and solar power output and load, and the predicted values ​​and confidence intervals are output, including:

[0018] Based on the self-attention mechanism of the temporal Transformer model, long temporal dependency features are extracted from the multidimensional input feature set to generate a prediction sequence containing multiple time points;

[0019] The time-series Transformer model is pre-trained offline with the goal of minimizing prediction error, and the model parameters are fine-tuned online using a rolling update method.

[0020] Based on the predicted sequence, output the predicted point values ​​and confidence intervals for a predetermined future time scale.

[0021] In a preferred embodiment, the predicted values ​​are input into a digital twin model for power grid operation simulation and power flow calculation. Based on the simulation results, safety warnings are issued and rigid safety constraints are generated, including:

[0022] The real-time collected operational data is mapped to the digital twin model to achieve virtual-real synchronization between the physical entity and the virtual model;

[0023] Based on the real-time collected operational data, multi-timescale power flow calculations are performed according to the preset power flow algorithm in the digital twin model.

[0024] The power grid operating status is assessed based on the power flow calculation results, and abnormal scenarios in the power grid operating status are identified and safety warnings are issued according to the preset anomaly judgment rules.

[0025] Based on the power flow calculation results, operational safety boundary constraints are generated as rigid safety constraints for the scaling dual descent alternating direction multiplier method model.

[0026] In a preferred embodiment, the rigid safety constraints and the predicted values ​​are input into a scaling dual descent alternating direction multiplier method model, and a distributed cooperative optimization solution is performed to generate power allocation and mutual assistance instructions, including:

[0027] Taking each sub-microgrid as an independent optimization block, the power allocation problem of the microgrid cluster is modeled as a distributed non-convex optimization problem with nonlinear coupling constraints. The rigid safety constraints serve as local operating constraints of the distributed non-convex optimization problem, and the predicted values ​​serve as input parameters of the distributed non-convex optimization problem.

[0028] Construct a regularized augmented Lagrange function by introducing a regularization term for the dual variable into the standard augmented Lagrange function;

[0029] Each sub-microgrid solves its local subproblem synchronously and in parallel based on the power information exchanged with neighboring microgrids on the current dual variables;

[0030] Each sub-microgrid only interacts with the power residuals of adjacent tie lines to perform coordinated updates of consistency variables;

[0031] The dual variable is updated using the scaling dual descent formula;

[0032] The iterative process is controlled by adaptive parameter adjustment and early stop mechanism until the preset convergence condition is met or the preset maximum number of iterations is reached. Then, the energy storage charging and discharging power, controllable load adjustment amount and microgrid interconnection power exchange command are output.

[0033] As a preferred embodiment, the adaptive parameter adjustment and residual early stopping mechanism specifically includes:

[0034] Calculate the original residual and dual residual for each iteration, and dynamically adjust the penalty parameter based on the comparison results of the original residual and dual residual, and adjust the scaling factor accordingly based on the adjustment of the penalty parameter.

[0035] The iteration is terminated early when both the original residual and the dual residual are less than a preset convergence threshold, or when the number of iterations reaches a preset maximum number of iterations.

[0036] In a preferred embodiment, the power allocation and mutual assistance commands are sent to the hybrid energy storage system for execution, including:

[0037] The hybrid energy storage system adopts a composite energy storage architecture consisting of lithium batteries, supercapacitors, and cascaded batteries.

[0038] A filtering algorithm is used to decompose the scheduling power in the power allocation and mutual assistance command into high-frequency components, medium-frequency components, and low-frequency components, and allocate them to the supercapacitor, the lithium battery, and the cascade battery for execution, respectively.

[0039] The power allocation priority is adjusted based on the health status ranking of each energy storage unit, and the adjusted power command is issued for execution.

[0040] In a preferred embodiment, adjusting the power allocation priority based on the health status ranking results of each energy storage unit specifically includes:

[0041] The state of charge and health status of each energy storage unit are estimated in real time, and the energy storage units are sorted according to their health status.

[0042] Based on the sorting results, the energy storage units are divided into priority scheduling units and non-priority scheduling units. The priority scheduling units are responsible for high-power charging and discharging, while the non-priority scheduling units limit their charging and discharging depth.

[0043] The number of charge-discharge cycles per day for each energy storage unit is limited to a preset threshold.

[0044] In a preferred embodiment, the prediction parameters of the temporal Transformer model and the scheduling parameters of the scaling dual descent alternating direction multiplier method model are corrected based on the execution results, including:

[0045] Based on the execution result data, the time-series Transformer prediction model is incrementally fine-tuned online, and the model prediction parameters are updated at preset time intervals.

[0046] The penalty parameters and scaling factors of the scaling dual descent alternating direction multiplier method model are optimized based on the execution result data for subsequent scheduling cycle optimization calculations.

[0047] In a preferred embodiment, after the adjusted power command is issued and executed, the method further includes:

[0048] The actual output power of each energy storage converter is collected in real time. The actual output power is compared with the adjusted power command to calculate the execution error, and the execution error is eliminated by using a PID correction algorithm.

[0049] Secondly, embodiments of the present invention also provide an AI-based integrated microgrid cluster multi-energy collaborative optimization scheduling system, comprising:

[0050] The data acquisition and preprocessing module is used to collect real-time operational data of all elements of source, grid, load and storage and perform data preprocessing, and construct a multi-dimensional input feature set based on the preprocessed data;

[0051] The time series prediction module is used to input the multi-dimensional input feature set into a pre-built time series Transformer model to perform ultra-short-term rolling prediction of wind and solar power output and load, and output the predicted value and confidence interval.

[0052] The digital twin simulation module is used to input the predicted values ​​into the digital twin model for power grid operation simulation and power flow calculation, and to generate rigid safety constraints based on the simulation results;

[0053] The distributed cooperative optimization module is used to scale the rigid safety constraints and the predicted values ​​into a dual descent alternating direction multiplier method model, perform distributed cooperative optimization solutions, and generate power allocation and mutual assistance instructions.

[0054] The instruction execution module is used to send the power allocation and mutual assistance instructions to the hybrid energy storage system, controllable load and distributed power source for execution.

[0055] The parameter correction module is used to correct the prediction parameters of the temporal Transformer model and the scheduling parameters of the scaling dual descent alternating direction multiplier method model based on the execution results.

[0056] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising:

[0057] One or more processors;

[0058] Storage device for storing one or more programs;

[0059] When the one or more programs are executed by the one or more processors, the one or more processors implement the AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters based on source, grid, load and storage as described in any embodiment of the present invention.

[0060] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters with source, grid, load and storage as described in any embodiment of the present invention.

[0061] Compared with existing technologies, the present invention achieves the following beneficial effects:

[0062] (1) This invention constructs a three-level distributed architecture consisting of a regional cluster layer, a single microgrid layer, and an equipment control layer. Combined with a multi-timescale scheduling mechanism of day-ahead baseline planning, intraday rolling correction, and ultra-short-term real-time smoothing, it achieves power coordination of multiple microgrid clusters without a central node and with lightweight communication. Compared with existing centralized scheduling, this scheme eliminates the dependence on full data from the central node, significantly improves the scalability and robustness of the system, and can adapt to the operation scenarios of multi-entity, cross-regional microgrid clusters.

[0063] (2) This invention applies the Scaling Dual Descent Alternating Direction Multiplier Method (ADMM) to the field of microgrid cluster scheduling. By introducing a regularization term for the dual variable into the standard augmented Lagrangian function, the dual variable in non-convex scenarios is constrained within a bounded range. Furthermore, the Scaling Dual Descent formula replaces the traditional Dual Ascending update, effectively solving the iterative oscillation and dual variable divergence problems of traditional ADMM under non-convex problems such as nonlinear constraints of microgrid energy storage and power coupling. Theoretical analysis shows that, under the conditions of bounded decision variables, smooth objective function, and continuously differentiable coupling constraints, this algorithm can achieve stable convergence, filling the technical gap of non-convex distributed optimization in the field of microgrid scheduling.

[0064] (3) This invention designs a programmable adaptive parameter adjustment and early stopping mechanism. By dynamically adjusting the penalty parameter and scaling factor through real-time calculation of the comparison results between the original residual and the dual residual, and presets a convergence threshold directly related to the engineering scheduling accuracy. This mechanism enables the algorithm to terminate the iteration early while meeting the scheduling accuracy requirements, with the maximum number of iterations controlled within 15, meeting the millisecond-level real-time scheduling requirements, and overcoming the technical difficulty of existing distributed optimization algorithms in balancing solution accuracy and engineering real-time performance.

[0065] (4) This invention adopts a synchronous parallel iteration and lightweight communication strategy. During the iteration process, each sub-microgrid only exchanges the power residuals of adjacent tie lines, without exchanging local decision variables or full operational data. Compared with the existing centralized scheduling and distributed scheduling with full data exchange, this scheme significantly reduces the communication bandwidth requirements and computational burden, and significantly improves the scheduling efficiency and communication robustness of the microgrid cluster. It is suitable for practical engineering scenarios with limited communication conditions.

[0066] (5) This invention adopts a composite energy storage architecture composed of lithium batteries, supercapacitors, and cascaded batteries. Based on frequency domain decomposition, the dispatch power is distributed in layers. High-frequency fluctuation components are smoothed by supercapacitors, mid-frequency continuous components are supported by lithium batteries, and low-frequency time-shifted components are undertaken by cascaded batteries, giving full play to the electrical characteristics advantages of different types of energy storage. At the same time, by monitoring the health status of energy storage units in real time, units with higher health status are prioritized to undertake high-power charging and discharging, while limiting the charging and discharging depth of units with lower health status, and constraining the number of charging and discharging cycles per day, the overall lifespan of the energy storage system is extended by more than 30%, effectively reducing equipment wear and maintenance costs.

[0067] (6) This invention collects execution result data and performs online incremental fine-tuning of the temporal Transformer prediction model. At the same time, it dynamically optimizes the penalty parameters and scaling factors of the scaling dual descent ADMM model, realizing deep coupling and dynamic adaptation of prediction, simulation, scheduling, and execution. This closed-loop mechanism effectively solves the problems of prediction, simulation, scheduling, and execution being independent of each other and the model being out of sync with actual operating conditions in the prior art. It ensures that the scheduling accuracy remains stable in long-term operation and avoids the continuous decline in scheduling performance caused by model drift. Attached Figure Description

[0068] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0069] Figure 1This is a flowchart of the AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters based on source, grid, load and storage provided in this embodiment of the invention;

[0070] Figure 2 This is a schematic diagram of the three-level distributed architecture and multi-time-scale scheduling mechanism provided in the embodiments of the present invention;

[0071] Figure 3 This is the scaling dual descent ADMM distributed collaborative optimization iterative logic diagram provided in the embodiments of the present invention;

[0072] Figure 4 This is a schematic diagram of the hybrid energy storage stratified control and lifetime balancing strategy provided in an embodiment of the present invention;

[0073] Figure 5 This is a schematic diagram of the structure of the AI-based integrated microgrid cluster multi-energy collaborative optimization scheduling system provided in an embodiment of the present invention;

[0074] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0075] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0076] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as being processed sequentially, many of these operations (or steps) may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operation is completed, but may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0077] Example 1

[0078] like Figure 1 The diagram shows a flowchart of an AI-based multi-energy collaborative optimization scheduling method 100 for integrated microgrid clusters with source, grid, load, and storage provided in Embodiment 1 of the present invention. The method 100 specifically includes the following steps:

[0079] S110: Real-time collection of operational data of all elements of source, grid, load and storage, and data preprocessing, and construction of multi-dimensional input feature set based on the preprocessed data.

[0080] The aforementioned full-element operational data includes: photovoltaic power generation, wind power generation, load power consumption, energy storage system charge status and health status, grid connection point voltage, tie line power, and upper-level grid dispatch instructions.

[0081] As a preferred embodiment, the collected full-element operational data is preprocessed, specifically including the following steps:

[0082] Outlier removal: The 3σ criterion is used to remove extreme outliers. Data points that exceed the mean ± 3 standard deviations are removed.

[0083] Missing value imputation: Linear interpolation is used to imput missing values ​​to ensure the continuity of the time series;

[0084] Time alignment: unifies sampling times to eliminate time asynchrony issues caused by sampling delay deviations between different devices;

[0085] Normalization: Min-max normalization is used to map all feature data to the [0,1] interval, eliminating the influence of units and facilitating model training.

[0086] As a preferred embodiment, a multi-dimensional input feature set is constructed based on the preprocessed data, covering meteorological features, time-series features, environmental features, calendar features, and power grid correlation features, wherein...

[0087] Meteorological characteristics include solar radiation intensity, wind speed, wind direction, ambient temperature, humidity, air pressure, and precipitation probability; temporal characteristics include historical 15s / 1min / 5min / 15min wind and solar power output values, load power values, and temporal differential characteristics; calendar characteristics include hour, minute, weekday / rest day, season, month, and holiday labels; and power grid characteristics include grid connection point voltage, tie line power, and upper-level power grid dispatch instructions.

[0088] In a preferred embodiment, the present invention employs a three-level hierarchical distributed architecture and a multi-time-scale scheduling dual-system linkage to achieve full-scenario, full-process collaborative scheduling. Specifically, the three-level hierarchical distributed architecture includes:

[0089] Regional cluster layer: responsible for power mutual assistance among multiple microgrids, tie-line power constraints, global power balance, and issuing total cluster power regulation requirements and boundary power plans;

[0090] Single microgrid layer: Executes local source-grid-load-storage power allocation, economic dispatch, and security constraint verification; receives instructions from the cluster layer and completes local optimization solutions.

[0091] Equipment control layer: Targeting photovoltaic, wind turbine, conventional energy units, hybrid energy storage systems, and flexible controllable loads, it enables millisecond to minute-level instruction execution;

[0092] Among them, the multi-timescale scheduling mechanism covers the 24-hour baseline plan, the 1-hour rolling correction within the day, and the 5-minute real-time smoothing in the ultra-short term, matching the accuracy and response requirements of different scheduling cycles.

[0093] The aforementioned three-level distributed architecture, in conjunction with the multi-time-scale scheduling mechanism, enables different scheduling levels to perform their respective functions at different time scales, jointly ensuring the stable, efficient, and economical operation of the microgrid cluster.

[0094] S120: Input the multidimensional input feature set into the pre-built time series Transformer model to perform ultra-short-term rolling prediction of wind and solar power output and load, and output the predicted value and confidence interval.

[0095] As a preferred embodiment, the above-described temporal Transformer model employs an encoder-decoder structure, with its core components being a multi-head self-attention mechanism and a feedforward neural network.

[0096] Encoder: It consists of 6 stacked encoder layers, each containing a multi-head self-attention layer, layer normalization and feedforward network, used to extract long temporal dependency features from the input sequence;

[0097] Decoder: Composed of 4 stacked decoders, it combines encoder output with historical time series data to generate future power prediction sequences;

[0098] Position coding: Sine position coding is introduced to embed the temporal sequence of time-series data into the model, enabling the model to perceive the relative position information at different time points.

[0099] As a preferred embodiment, the model training employs a strategy that combines offline pre-training with online fine-tuning:

[0100] Offline pre-training: The initial training of the model was completed using source-grid-load-storage operation data from the past 6 months. The optimization objective was to minimize the root mean square error, so that the model would have basic predictive capabilities.

[0101] Online fine-tuning: The latest running data is collected every 15 minutes, and the model parameters are incrementally updated using mini-batch gradient descent, so that the model can adapt to the dynamic changes in the operating conditions and avoid model drift.

[0102] Error correction: By integrating historical prediction error sequences, an error correction model is constructed to correct the final prediction results and further improve prediction accuracy.

[0103] As a preferred embodiment, the time-series Transformer model outputs ultra-short-term predictions for the next 0 to 4 hours, specifically including:

[0104] Predicted values: Accurate predicted values ​​of wind and solar power generation and load power consumption;

[0105] Interval predictions: provided as upper and lower bounds with 95% confidence, for subsequent robust scheduling.

[0106] The prediction results are pushed to the digital twin simulation module and the scaling dual descent ADMM scheduling module in 15-minute increments, with the prediction error controlled within 5%.

[0107] For example, taking a microgrid cluster containing three sub-microgrids as an example, meteorological, wind and solar power output, and load data of the cluster over the past month are collected, with a sampling granularity of 15 minutes, constructing a 96-step input sequence (corresponding to a 24-hour historical data window). The model predicts the following output results:

[0108] Forecasted photovoltaic output of microgrid 1: 1.2MW (0.3MW shortfall);

[0109] Forecast wind power output for microgrid 2: 0.9MW (0.5MW shortfall);

[0110] Forecasted photovoltaic output of microgrid 3: 0.8MW (0.4MW shortfall);

[0111] Cluster total load forecast: 6.5MW, with a total power deficit of 2MW, matching the scheduling demand.

[0112] S130: Input the predicted value into the digital twin model to perform power grid operation simulation and power flow calculation, and generate a safety warning and rigid safety constraints based on the simulation results.

[0113] As a preferred embodiment, this step constructs a 1:1 digital twin virtual entity of the microgrid cluster, realizing real-time synchronization between the physical entity and the virtual model, full-element simulation of the operating status, and dynamic generation of rigid safety constraints, providing rigid operating boundaries for scaling dual descent ADMM scheduling. The specific implementation steps are as follows:

[0114] Step S131, Full-element digital twin modeling:

[0115] Constructing a 1:1 digital twin virtual entity of a microgrid cluster, fully mapping all elements of the physical entity, including source, grid, load, and storage, specifically including:

[0116] Source-side models: power output characteristic models, efficiency curves, and start-stop constraints for photovoltaic, wind turbine, gas turbine, and diesel generators;

[0117] Grid-side model: topological connections of transmission lines, transformers, switches, and tie lines; resistance / reactance parameters; power flow constraints.

[0118] Load-side models: electricity consumption characteristics and regulation response characteristics of industrial load, commercial load, residential load, and flexible controllable load;

[0119] Storage-side model: charging and discharging characteristics of lithium batteries, supercapacitors, and cascade batteries; SOC / SOH constraints; power response delay; topology model adopts IEC 61850 standard to ensure accurate electrical connection relationships.

[0120] Step S132, Real-time data interaction and synchronization mechanism:

[0121] Data acquisition: Real-time data collection of source-grid-load-storage operation data is acquired via IoT terminals at a granular level of 1 second;

[0122] Virtual-real synchronization: A data-driven synchronization mechanism is adopted to map the measured data to the twin model in real time, with a synchronization latency of no more than 100ms;

[0123] Data interface: Data transmission is completed using the MQTT protocol to ensure high concurrency and low latency communication.

[0124] Step S133: Run the multi-timescale simulation engine (adapting to the scheduling timescale and adopting a hierarchical simulation mode):

[0125] Ultra-short-term simulation (5-minute granularity): Steady-state power flow calculation, analysis of power balance and voltage distribution;

[0126] Intraday simulation (1-hour granularity): Electromechanical transient simulation to analyze power grid stability under wind and solar fluctuations;

[0127] Day-to-day simulation (24-hour granularity): Mid- to long-term operation simulation to predict global power supply and demand.

[0128] Step S134, Power Flow Calculation and Operational Status Assessment:

[0129] The Newton-Raphson method is used for power flow calculation, and the following electrical indicators are calculated in real time: node voltage amplitude / phase angle, line transmission power, power loss, and grid frequency. Based on the calculation results, the operating status is divided into three categories: normal, warning, and fault. The judgment threshold follows the GB / T 12325-2008 power quality standard.

[0130] In a preferred embodiment, early warning is achieved based on preset anomaly detection rules:

[0131] Sudden increase / decrease in wind and solar power output: Output fluctuation exceeds 30% of rated power within 15 minutes; Sudden increase in load: Load increases by more than 25% within 10 minutes; Line overload: Transmission power exceeds 90% of the line's rated capacity; Voltage exceeding limits: Node voltage exceeds the rated voltage range of ±7%.

[0132] Step S135, Generation and scheduling mapping of rigid safety constraints:

[0133] The simulation results are transformed into rigid constraints for the scaled dual descent ADMM scheduling, and these constraints are pushed to the scheduling module in real time as boundary conditions for the optimization problem.

[0134] Node voltage constraints: ;

[0135] Power flow constraints: ;

[0136] Frequency constraints: ;

[0137] in, Node voltage lower limit, in kilovolts (kV), is the lowest voltage at which a microgrid node is allowed to operate. : No. The actual operating voltage value of each node, in kilovolts (kV); The node voltage upper limit, in kilovolts (kV), is the highest voltage that a microgrid node is allowed to operate at. : No. The actual transmission power of the line, in megawatts (MW); : No. The maximum allowable transmission power limit for each line, in megawatts (MW). The actual operating frequency of the microgrid system, measured in Hertz (Hz).

[0138] For example, taking a microgrid cluster containing three sub-microgrids as an example, a 1:1 twin model is built, importing the topologies, line parameters (two 10kV tie lines), and equipment parameters of the three microgrids. Real-world data is accessed at a 1-second granularity to achieve virtual-real synchronization, with a synchronization delay of 80 milliseconds. Power flow simulation and security constraint generation are performed, and the results are as follows:

[0139] Node voltage: 10.1~10.3kV, meeting the rated voltage ±7% constraint;

[0140] The power flow limit of the connecting line is 3MW, with no risk of overload.

[0141] Safety constraints for output: energy storage power ±1.5MW, tie line switching power ≤2.5MW, voltage constraints [9.3kV, 10.7kV].

[0142] No abnormalities such as wind and solar fluctuations or sudden load increases were detected during the simulation, and the operation was normal.

[0143] S140: Input the rigid safety constraint and the predicted value into the scaling dual descent alternating direction multiplier method model, perform distributed cooperative optimization solution, and generate power allocation and mutual assistance instructions.

[0144] As a preferred embodiment, this step, based on the rigid safety constraints issued by the digital twin simulation module (which are transformed into the equipment safety operation boundary in the local operation constraints), the output / load data issued by the time series prediction module, and the total power regulation requirements of the cluster and the boundary power plans of each sub-microgrid issued by the regional cluster layer, uses the scaling dual descent alternating direction multiplier method to coordinately optimize the allocation of energy storage power and mutual assistance power between sub-microgrids, specifically including the following steps:

[0145] Step S141, Decomposition and Modeling of Multi-block Non-convex Distributed Optimization Problem:

[0146] The power allocation problem of microgrid clusters is modeled as a distributed nonconvex optimization problem with multiple sub-microgrids and nonlinear coupling constraints. The model satisfies the basic convergence conditions of bounded decision variables, continuous differentiability of the objective function, and smooth coupling constraints.

[0147] Optimization Unit Division: Taking each sub-microgrid as an independent optimization entity, it is divided into a total of 10 sub-microgrids. Each optimization block relies solely on local data for computation;

[0148] Local decision variables: Define the first Individual microgrid decision variables This includes energy storage charging and discharging power, controllable load regulation, and power exchanged via interconnection lines between adjacent sub-microgrids.

[0149] Local operating constraints: upper and lower limits of source-load-storage capacity, energy storage SOC constraints, controllable load adjustment range, and equipment safety operation boundaries issued by digital twin;

[0150] Global coupling constraints: The sum of the net power regulation of all sub-microgrids matches the total regulation demand of the cluster; the power exchanged by the interconnection lines of adjacent sub-microgrids is consistent in both directions; and the nonlinear equality coupling form is satisfied.

[0151] ;

[0152] in, The total number of sub-microgrids within a microgrid cluster, which is a positive integer; The sub-microgrid's designation, with a value of [value missing]. ; : No. Local decision variables for each sub-microgrid include energy storage charging and discharging power, controllable load regulation, and power exchanged via interconnection lines between adjacent sub-microgrids. : No. The nonlinear coupling constraint function corresponding to each sub-microgrid characterizes the contribution of that sub-microgrid to the global power balance; The sum of the nonlinear coupling constraint functions of all sub-microgrids is 0, indicating that the nonlinear equality coupling condition that the global power balance and tie-line power exchange consistency are satisfied is met.

[0153] Optimization objectives: To achieve multiple objectives, including minimizing cluster operating costs, maximizing renewable energy consumption, and minimizing tie-line power fluctuations, a local non-convex optimization objective function is constructed.

[0154] Step S142: Construction of regularized augmented Lagrangian function and distributed iterative solution of scaling dual descent ADMM:

[0155] To address the issues of unbounded dual variables and iterative divergence in traditional ADMM in non-convex scenarios, an augmented Lagrangian function with a dual variable regularization term is constructed. Scaling dual descent is used instead of traditional dual ascending updates, forming a stable distributed iterative mechanism.

[0156] Step S1421, Construction of regularized augmented Lagrangian function:

[0157] Based on the standard augmented Lagrangian function, a quadratic penalty term for the dual variable is introduced to constrain the dual variable to a bounded range, thus preventing iterative divergence. The function form is as follows:

[0158]

[0159] in: For regularized augmented Lagrangian functions; For standard augmented Lagrange terms; As dual variables; For penalty parameters; The constraint strength constant; For dual variables, the quadratic regularization term;

[0160] Step S1422, Scaling Dual Descent ADMM Distributed Iterative Core Steps:

[0161] The iterative process employs a local subproblem synchronous parallel solution, consistency coordination, and scaling duality descent update procedure:

[0162] Local subproblems are solved synchronously and in parallel: each sub-microgrid is based on the dual variables of the current iteration. The system exchanges power information with neighboring microgrids in the previous round, solves local subproblems independently using the near-end gradient method, and updates local decision variables according to the Gauss-Seidel model. After all sub-microgrids have been calculated, the system synchronously enters the next iteration.

[0163] Coordinated update of consistency variables: Each sub-microgrid only interacts with the power residuals of adjacent tie lines, and completes the power consistency correction through boundary information exchange;

[0164] Scaling Dual Descent Update: The scaling dual descent formula is used to update the dual variable, achieving synchronous optimization of the original and dual variables. The formula is as follows:

[0165] ;

[0166] in: This is the scaling factor; For coupling constraint residuals; This represents the number of iterations. : No. The dual variable updated in the next iteration is used to penalize the deviation from the global coupling constraint; : Scaling factor, used to control the step size ratio of dual descent to ensure stable convergence of the iteration; : No. The dual variable of the next iteration; The penalty parameter in the augmented Lagrangian function is used to adjust the strength of the constraint penalty. : Constraint strength constant, used to regularize the dual variable and ensure bounded iteration; : No. The global coupling constraint residual of the next iteration represents the deviation between the net power adjustment of all sub-microgrids and the total regulation demand of the cluster.

[0167] Step S143, Adaptive parameter adjustment and real-time early stopping mechanism:

[0168] This embodiment designs a dual-dimensional adaptive parameter adjustment rule plus a residual threshold early stopping mechanism to balance convergence speed, numerical stability, and real-time scheduling requirements.

[0169] Penalty parameters Adaptive adjustment rule: Let the first... The original residual of the second iteration Dual residual The adjustment rules are as follows:

[0170] like The violation of the constraints was serious, and ;

[0171] like Oscillation of dual variables, let ;

[0172] All other cases remain unchanged. ;

[0173] in, : No. The set of all local decision variables of the sub-microgrids after the next iteration update; : No. The original residual of the next iteration represents the degree of violation of the global coupling constraints; : No. The global coupling constraint function value for the next iteration; The norm of the global coupling constraint function value, used to quantify the degree of violation of the original constraint; : No. The dual residual of the next iteration represents the degree of fluctuation in the update of the dual variable; : No. The set of all local decision variables of the sub-microgrids in the next iteration; : No. The set of all local decision variables of the sub-microgrids in the next iteration; The norm of the difference between decision variables in two consecutive iterations, used to measure the magnitude of change during iteration; : No. The penalty parameter is updated in the next iteration; : No. The penalty parameter for the next iteration.

[0174] scaling factor Adaptive adjustment: Adjusts inversely to the penalty parameter to balance convergence speed and stability.

[0175] ;

[0176] in, : No. The scaling factor updated in the next iteration; : The initial value of the scaling factor; : The initial value of the penalty parameter.

[0177] Early stopping mechanism: Preset convergence thresholds for the original residual and the dual residual, as well as a maximum number of iterations. The iteration is terminated early when both the original residual and the dual residual are less than the aforementioned convergence thresholds, or when the number of iterations reaches the preset maximum number of iterations.

[0178] For example, taking a microgrid cluster containing three sub-microgrids as an example, the initial value of the penalty parameter is set. =100, constraint strength constant Initial value of scaling factor The maximum number of iterations is 15, and the convergence threshold is: original residual ≤ 0.05, dual residual ≤ 1. The changes in residuals and parameters during the iteration process are shown in Table 1 below:

[0179] Table 1

[0180]

[0181] As shown in the table above: First iteration: Each sub-microgrid is based on... Solve the local subproblem with a total increased power output of 1.7 MW and residuals. Updated Iterations 2-5: Residuals gradually decrease, penalty parameters are adaptively fine-tuned, and dual variables are updated smoothly; Iteration 6: Original residuals Dual residual If all values ​​are less than the preset threshold, the early stopping mechanism is triggered, and the algorithm converges.

[0182] Dispatch command output: Microgrid 1 energy storage discharge 0.6 MW, Microgrid 2 supercapacitor discharge 0.7 MW, Microgrid 3 diesel generator additional generation 0.7 MW, total additional generation 2 MW.

[0183] S150: The power allocation and mutual assistance commands are sent to the hybrid energy storage system, controllable load and distributed power source for execution.

[0184] In a preferred embodiment, this module adopts a multi-type hybrid energy storage architecture, based on power timing decomposition and health state balancing strategies, to achieve rapid power response, fluctuation smoothing, and extended lifespan of energy storage, and to receive and execute scaling duality descent ADMM scheduling commands. Wherein:

[0185] The hybrid energy storage system adopts a triple-layer hybrid energy storage architecture consisting of lithium batteries, supercapacitors, and secondary batteries, with functions divided according to electrical characteristics:

[0186] Supercapacitors: power density ≥5kW / kg, response delay ≤10ms, capable of smoothing high-frequency power fluctuations (≥1Hz); Lithium batteries: energy density ≥150Wh / kg, cycle life ≥3000 cycles, capable of supporting medium-frequency continuous power (0.1–1Hz); Cascaded batteries: reorganized retired power batteries, reducing costs by 40%, capable of handling low-frequency long-term power (≤0.1Hz) time shifts; Energy storage systems are equipped with DC / DC converters to achieve independent control and rapid response.

[0187] Power hierarchical allocation and fluctuation decomposition employ a second-order Butterworth high-pass / low-pass filtering algorithm to decompose the total scheduled power into three components:

[0188] High-frequency components (fluctuation components): extracted through high-pass filtering and distributed to supercapacitors to smooth out instantaneous fluctuations in wind and solar power; mid-frequency components (support components): extracted through band-pass filtering and distributed to lithium batteries to meet continuous power regulation; low-frequency components (time-shift components): extracted through low-pass filtering and distributed to secondary batteries to achieve peak charging and peak discharging; the decomposed time constant is adaptively adjusted according to the response speed of scheduling instructions.

[0189] Real-time energy storage status sensing and estimation: State of charge (SOC) is estimated using ampere-hour integration and open-circuit voltage correction, and State of harm (SOH) is estimated using impedance spectroscopy.

[0190] SOC estimation: ,error ;

[0191] SOH estimation: Real-time monitoring of battery health degradation status;

[0192] Sampling period: Voltage / current sampling period 10 ms, SOC / SOH update period 1s.

[0193] in, :time The energy storage state of charge; Initial moment The energy storage state of charge; Rated capacity of the energy storage system; :time The charging and discharging current is positive for charging and negative for discharging. From the initial moment At the time Accumulated ampere-hours; SOH: State of Health of the energy storage battery; The current actual usable capacity of the energy storage battery; : Rated capacity of the energy storage battery at the time of manufacture.

[0194] As a preferred embodiment, the adaptive charge / discharge strategy is generated by: scaling the dual decreasing ADMM scheduling power as the objective, and combining energy storage constraints to generate charge / discharge commands.

[0195] Power constraints: To avoid operating at excessive power;

[0196] Capacity constraints: Lithium battery SOC control at The SOC of the supercapacitor is controlled between 10% and 90%.

[0197] Response constraint: Charge / discharge switching delay ≤ 50ms, ensuring rapid execution of scheduling instructions.

[0198] in, The minimum allowable charge and discharge power of the energy storage system; Actual charging and discharging power of the energy storage system; The maximum allowable charge and discharge power of the energy storage system; The minimum state of charge allowed for an energy storage system.

[0199] As a preferred embodiment, the lifetime balancing scheduling rule achieves lifetime balancing of energy storage units based on the real-time SOH value:

[0200] Health ranking: Real-time calculation of SOH for each energy storage unit, and ranking by SOH from high to low;

[0201] Power allocation: Prioritize cells with higher State of Harmony (SOH) to perform high-power charging and discharging, and reduce the depth of charge and discharge of cells with lower SOH;

[0202] Cycle balancing: Limit the maximum number of charge and discharge cycles per day, ≤60 times / day for lithium batteries and ≤5000 times / day for supercapacitors; avoid overcharging and over-discharging, and frequent charging and discharging, extending the overall energy storage system life by more than 30%.

[0203] The power command after hierarchical allocation is sent to each energy storage converter to complete the power output; the actual output power is collected in real time, compared with the dispatch command, and the execution error is calculated.

[0204] A PID correction algorithm is used to eliminate execution errors and ensure that the output power accuracy is ≤2%.

[0205] S160: Based on the execution results, correct the prediction parameters of the temporal Transformer model and the scheduling parameters of the scaling dual descent alternating direction multiplier method model.

[0206] In a preferred embodiment, based on the execution result of step S150, the prediction parameters of the temporal Transformer model and the scheduling parameters of the scaling dual descent alternating direction multiplier method model are corrected for optimization calculations in subsequent scheduling cycles, specifically including:

[0207] S161, Correct data acquisition:

[0208] Key data collected during step S150 is used as the basis for parameter correction, specifically including:

[0209] Iterative residual data: Records the original and dual residual sequences during the scaling dual descent ADMM iteration process;

[0210] Command execution deviation: Record the deviation between the actual output power of each executing entity (hybrid energy storage system, controllable load, distributed power source) and the dispatch command;

[0211] Status update data: Records the updated values ​​of SOC and SOH for each energy storage unit, as well as the actual values ​​of wind and solar power output and load power.

[0212] S162. Time-series Transformer prediction model parameter correction:

[0213] Online incremental fine-tuning of the time-series Transformer prediction model using execution result data includes:

[0214] Online fine-tuning: Incrementally update model parameters using the latest collected execution result data at preset time intervals (e.g., every 15 minutes);

[0215] Error correction: By fusing historical prediction error sequences, an error correction model is constructed to correct subsequent prediction results;

[0216] Model update: The corrected prediction parameters are updated to the time series Transformer model for wind and solar power output and load prediction in the next scheduling cycle.

[0217] S163, Scaling Dual Descent ADMM Model Scheduling Parameter Correction:

[0218] The scheduling parameters of the dual descent alternating direction multiplier method model are optimized and scaled based on the execution results data, specifically including:

[0219] Penalty parameter optimization: Dynamically adjust the initial value of the penalty parameter or an adaptive adjustment rule based on the iterative residual sequence and execution bias data;

[0220] Scaling factor optimization: Optimize the initial value of the scaling factor according to the adjustment of the penalty parameter;

[0221] Convergence threshold optimization: Based on the actual scheduling accuracy requirements, the convergence thresholds of the original residual and the dual residual are adaptively adjusted.

[0222] S164, Parameter Feedback and Closed-Loop Formation:

[0223] The corrected parameters are then fed back to the corresponding module:

[0224] The corrected prediction parameters are fed back into the time series Transformer model for calculation of subsequent prediction periods;

[0225] The corrected scheduling parameters are fed back to the scaling dual descent ADMM model for calculation of subsequent optimization cycles.

[0226] Through the above mechanism, the present invention achieves dynamic adaptation of each link of prediction, simulation, scheduling and execution, forming a closed-loop control of the entire process from data acquisition to execution feedback.

[0227] For example, taking a microgrid cluster containing 3 sub-microgrids as an example, the parameters are corrected based on the execution result of step S150:

[0228] Command execution deviation: Energy storage system execution error 1.2%, controllable load adjustment deviation 0.8%, distributed power generation output deviation 1.0%;

[0229] Iteration residual record: The original residual of the last iteration was 0.02, and the dual residual was 0.48;

[0230] Status update: The SOC of lithium battery in microgrid 1 has decreased from 65% to 58%, the SOC of supercapacitor in microgrid 2 has decreased from 75% to 70%, and the SOH of secondary battery in microgrid 3 has decreased from 78% to 77.5%.

[0231] Prediction model correction: The time series Transformer model was fine-tuned online using execution bias data, and the model weights were adjusted to make subsequent predictions closer to actual operating conditions; the prediction error was reduced from 2.8% to 2.3%.

[0232] Model optimization correction: Based on the iteration residual records, the initial value of the penalty parameter was optimized (from 100 to 90); based on the execution deviation data, the initial value of the scaling factor was optimized (from 1 to 0.95); after the adjustment, the number of iterations in subsequent scheduling cycles was reduced from 6 to 5, improving the solution speed.

[0233] Closed-loop effect: The corrected prediction parameters and scheduling parameters are fed back to the corresponding model for calculation in the next scheduling cycle; forming a complete closed loop of prediction → simulation → optimization → execution → correction, the scheduling accuracy remains stable in long-term operation, and the scheduling performance degradation caused by model drift is avoided.

[0234] Based on the above embodiments, the core beneficial effects of the present invention are as follows:

[0235] (1) This invention constructs a three-level distributed architecture consisting of a regional cluster layer, a single microgrid layer, and an equipment control layer. Combined with a multi-timescale scheduling mechanism of day-ahead baseline planning, intraday rolling correction, and ultra-short-term real-time smoothing, it achieves power coordination of multiple microgrid clusters without a central node and with lightweight communication. Compared with existing centralized scheduling, this scheme eliminates the dependence on full data from the central node, significantly improves the scalability and robustness of the system, and can adapt to the operation scenarios of multi-entity, cross-regional microgrid clusters.

[0236] (2) This invention applies the Scaling Dual Descent Alternating Direction Multiplier Method (ADMM) to the field of microgrid cluster scheduling. By introducing a regularization term for the dual variable into the standard augmented Lagrangian function, the dual variable in non-convex scenarios is constrained within a bounded range. Furthermore, the Scaling Dual Descent formula replaces the traditional Dual Ascending update, effectively solving the iterative oscillation and dual variable divergence problems of traditional ADMM under non-convex problems such as nonlinear constraints of microgrid energy storage and power coupling. Theoretical analysis shows that, under the conditions of bounded decision variables, smooth objective function, and continuously differentiable coupling constraints, this algorithm can achieve stable convergence, filling the technical gap of non-convex distributed optimization in the field of microgrid scheduling.

[0237] (3) This invention designs a programmable adaptive parameter adjustment and early stopping mechanism. By dynamically adjusting the penalty parameter and scaling factor through real-time calculation of the comparison results between the original residual and the dual residual, and presets a convergence threshold directly related to the engineering scheduling accuracy. This mechanism enables the algorithm to terminate the iteration early while meeting the scheduling accuracy requirements, with the maximum number of iterations controlled within 15, meeting the millisecond-level real-time scheduling requirements, and overcoming the technical difficulty of existing distributed optimization algorithms in balancing solution accuracy and engineering real-time performance.

[0238] (4) This invention adopts a synchronous parallel iteration and lightweight communication strategy. During the iteration process, each sub-microgrid only exchanges the power residuals of adjacent tie lines, without exchanging local decision variables or full operational data. Compared with the existing centralized scheduling and distributed scheduling with full data exchange, this scheme significantly reduces the communication bandwidth requirements and computational burden, and significantly improves the scheduling efficiency and communication robustness of the microgrid cluster. It is suitable for practical engineering scenarios with limited communication conditions.

[0239] (5) This invention adopts a composite energy storage architecture composed of lithium batteries, supercapacitors, and cascaded batteries. Based on frequency domain decomposition, the dispatch power is distributed in layers. High-frequency fluctuation components are smoothed by supercapacitors, mid-frequency continuous components are supported by lithium batteries, and low-frequency time-shifted components are undertaken by cascaded batteries, giving full play to the electrical characteristics advantages of different types of energy storage. At the same time, by monitoring the health status of energy storage units in real time, units with higher health status are prioritized to undertake high-power charging and discharging, while limiting the charging and discharging depth of units with lower health status, and constraining the number of charging and discharging cycles per day, the overall lifespan of the energy storage system is extended by more than 30%, effectively reducing equipment wear and maintenance costs.

[0240] (6) This invention collects execution result data and performs online incremental fine-tuning of the temporal Transformer prediction model. At the same time, it dynamically optimizes the penalty parameters and scaling factors of the scaling dual descent ADMM model, realizing deep coupling and dynamic adaptation of prediction, simulation, scheduling, and execution. This closed-loop mechanism effectively solves the problems of prediction, simulation, scheduling, and execution being independent of each other and the model being out of sync with actual operating conditions in the prior art. It ensures that the scheduling accuracy remains stable in long-term operation and avoids the continuous decline in scheduling performance caused by model drift.

[0241] Example 2

[0242] Figure 5 This is a schematic diagram of the structure of the AI-based integrated microgrid cluster multi-energy collaborative optimization scheduling system 500 provided in Embodiment 2 of the present invention, as shown below. Figure 5 As shown, the system includes:

[0243] The data acquisition and preprocessing module 510 is used to collect real-time operational data of all elements of source, grid, load and storage and perform data preprocessing, and construct a multi-dimensional input feature set based on the preprocessed data.

[0244] The time series prediction module 520 is used to input the multi-dimensional input feature set into a pre-built time series Transformer model to perform ultra-short-term rolling prediction of wind and solar power output and load, and output the predicted value and confidence interval.

[0245] The digital twin simulation module 530 is used to input the predicted values ​​into the digital twin model for power grid operation simulation and power flow calculation, and to generate rigid safety constraints based on the simulation results;

[0246] The distributed cooperative optimization module 540 is used to input the rigid safety constraints and the predicted values ​​into the scaling dual descent alternating direction multiplier method model, perform distributed cooperative optimization solution, and generate power allocation and mutual assistance instructions.

[0247] The instruction execution module 550 is used to send the power allocation and mutual assistance instructions to the hybrid energy storage system, controllable load and distributed power source for execution.

[0248] The parameter correction module 560 is used to correct the prediction parameters of the temporal Transformer model and the scheduling parameters of the scaling dual descent alternating direction multiplier method model based on the execution results.

[0249] The AI-based multi-energy collaborative optimization scheduling system for integrated microgrid clusters with source, grid, load, and storage provided in this embodiment of the invention can execute the AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters with source, grid, load, and storage provided in any of the embodiments of the invention. It has the corresponding functions and beneficial effects of executing the AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters with source, grid, load, and storage. For detailed process, please refer to the relevant operations of the AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters with source, grid, load, and storage in the foregoing embodiments.

[0250] Example 3

[0251] Figure 6 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0252] like Figure 6As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0253] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0254] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 executes the AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters described above.

[0255] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0256] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An AI-based multi-energy collaborative optimization scheduling method for integrated microgrid clusters with source-grid-load-storage architecture, characterized in that: include: Real-time collection of operational data of all elements of source, grid, load and storage, and data preprocessing, and construction of multi-dimensional input feature set based on the preprocessed data; The multidimensional input feature set is input into a pre-built time series Transformer model to perform ultra-short-term rolling prediction of wind and solar power output and load, and the predicted value and confidence interval are output. The predicted values ​​are input into the digital twin model for power grid operation simulation and power flow calculation. Based on the simulation results, safety warnings are issued and rigid safety constraints are generated. The rigid safety constraints and the predicted values ​​are input into the scaling dual descent alternating direction multiplier method model, and distributed cooperative optimization is performed to generate power allocation and mutual assistance commands. The power allocation and mutual assistance commands are sent to the hybrid energy storage system, controllable loads, and distributed power sources for execution. The prediction parameters of the temporal Transformer model and the scheduling parameters of the scaling dual descent alternating direction multiplier method model are adjusted based on the execution results.

2. The method according to claim 1, characterized in that, The method also includes constructing a runtime framework that links a three-level distributed architecture with a multi-time-scale scheduling mechanism; The three-level distributed architecture includes a regional cluster layer, a single microgrid layer, and a device control layer. The regional cluster layer is used to formulate global power mutual assistance strategies and tie-line power boundary plans among multiple microgrids. The single microgrid layer is used to perform local source-grid-load-storage optimization scheduling and security verification. The device control layer is used to execute device commands at the millisecond to minute level. The multi-timescale scheduling mechanism includes day-ahead baseline planning, intraday rolling correction, and ultra-short-term real-time smoothing.

3. The method according to claim 1, characterized in that, The multidimensional input feature set is input into a pre-built time-series Transformer model for ultra-short-term rolling prediction of wind and solar power output and load, and the predicted values ​​and confidence intervals are output, including: Based on the self-attention mechanism of the temporal Transformer model, long temporal dependency features are extracted from the multidimensional input feature set to generate a prediction sequence containing multiple time points; The time-series Transformer model is pre-trained offline with the goal of minimizing prediction error, and the model parameters are fine-tuned online using a rolling update method. Based on the predicted sequence, output the predicted point values ​​and confidence intervals for a predetermined future time scale.

4. The method according to claim 1, characterized in that, The predicted values ​​are input into a digital twin model for power grid operation simulation and power flow calculation. Based on the simulation results, safety warnings are issued and rigid safety constraints are generated, including: The real-time collected operational data is mapped to the digital twin model to achieve virtual-real synchronization between the physical entity and the virtual model; Based on the real-time collected operational data, multi-timescale power flow calculations are performed according to the preset power flow algorithm in the digital twin model. The power grid operating status is assessed based on the power flow calculation results, and abnormal scenarios in the power grid operating status are identified and safety warnings are issued according to the preset anomaly judgment rules. Based on the power flow calculation results, operational safety boundary constraints are generated as rigid safety constraints for the scaling dual descent alternating direction multiplier method model.

5. The method according to claim 1, characterized in that, The rigid safety constraints and the predicted values ​​are input into a scaling dual descent alternating direction multiplier method model, and a distributed cooperative optimization solution is performed to generate power allocation and mutual assistance instructions, including: Taking each sub-microgrid as an independent optimization block, the power allocation problem of the microgrid cluster is modeled as a distributed non-convex optimization problem with nonlinear coupling constraints. The rigid safety constraints serve as local operating constraints of the distributed non-convex optimization problem, and the predicted values ​​serve as input parameters of the distributed non-convex optimization problem. Construct a regularized augmented Lagrange function by introducing a regularization term for the dual variable into the standard augmented Lagrange function; Each sub-microgrid solves its local subproblem synchronously and in parallel based on the power information exchanged with neighboring microgrids on the current dual variables; Each sub-microgrid only interacts with the power residuals of adjacent tie lines to perform coordinated updates of consistency variables; The dual variable is updated using the scaling dual descent formula; The iterative process is controlled by adaptive parameter adjustment and early stop mechanism until the preset convergence condition is met or the preset maximum number of iterations is reached. Then, the energy storage charging and discharging power, controllable load adjustment amount and microgrid interconnection power exchange command are output.

6. The method according to claim 5, characterized in that, The adaptive parameter adjustment and residual early stopping mechanism specifically includes: Calculate the original residual and dual residual for each iteration, and dynamically adjust the penalty parameter based on the comparison results of the original residual and dual residual, and adjust the scaling factor accordingly based on the adjustment of the penalty parameter. The iteration is terminated early when both the original residual and the dual residual are less than a preset convergence threshold, or when the number of iterations reaches a preset maximum number of iterations.

7. The method according to claim 1, characterized in that, The power allocation and mutual assistance commands are sent to the hybrid energy storage system for execution, including: The hybrid energy storage system adopts a composite energy storage architecture consisting of lithium batteries, supercapacitors, and cascaded batteries. A filtering algorithm is used to decompose the scheduling power in the power allocation and mutual assistance command into high-frequency components, medium-frequency components, and low-frequency components, and allocate them to the supercapacitor, the lithium battery, and the cascade battery for execution, respectively. The power allocation priority is adjusted based on the health status ranking of each energy storage unit, and the adjusted power command is issued for execution.

8. The method according to claim 7, characterized in that, The adjustment of power allocation priority based on the health status ranking of each energy storage unit specifically includes: The state of charge and health status of each energy storage unit are estimated in real time, and the energy storage units are sorted according to their health status. Based on the sorting results, the energy storage units are divided into priority scheduling units and non-priority scheduling units. The priority scheduling units are responsible for high-power charging and discharging, while the non-priority scheduling units limit their charging and discharging depth. The number of charge-discharge cycles per day for each energy storage unit is limited to a preset threshold.

9. The method according to claim 1, characterized in that, The prediction parameters of the temporal Transformer model and the scheduling parameters of the scaling dual descent alternating direction multiplier method model are corrected based on the execution results, including: Based on the execution result data, the time-series Transformer prediction model is incrementally fine-tuned online, and the model prediction parameters are updated at preset time intervals. The penalty parameters and scaling factors of the scaling dual descent alternating direction multiplier method model are optimized based on the execution result data for subsequent scheduling cycle optimization calculations.

10. The method according to claim 7, characterized in that, After the adjusted power command is issued and executed, the method further includes: The actual output power of each energy storage converter is collected in real time. The actual output power is compared with the adjusted power command to calculate the execution error, and the execution error is eliminated by using a PID correction algorithm.

Images