CN-122026530-A - AI-based source network charge storage integrated micro-grid cluster multi-energy collaborative optimization scheduling method

CN122026530ACN 122026530 ACN122026530 ACN 122026530ACN-122026530-A

Abstract

The invention discloses an AI-based source network charge storage integrated micro-grid cluster multi-energy collaborative optimization scheduling method, which relates to the technical field of power system automation and artificial intelligence and distributed optimization scheduling, and comprises the steps of collecting source network charge storage operation data and constructing a multi-dimensional input feature set; the method comprises the steps of inputting a multi-dimensional input characteristic set into a time sequence transducer model to conduct wind-light output and load ultra-short-term rolling prediction, inputting a predicted value into a digital twin model to conduct simulation and power flow calculation to generate rigid safety constraint, inputting the rigid safety constraint and the predicted value into a scaling pair descent alternating direction multiplier method model to conduct distributed collaborative optimization solution to generate power distribution and mutual aid instructions, and conducting execution, and correcting parameters of the time sequence transducer model and the scaling pair descent alternating direction multiplier method model according to an execution result. Therefore, the efficient synergy and stable operation of the micro-grid cluster source, network, load and storage are realized.

Inventors

HU XINYU
TANG HUA
WANG MIN
Shang Jiaojuan
GUO YU
PU CHENG

Assignees

国网江苏省电力有限公司
盐城电力设计院有限公司
国网江苏省电力有限公司盐城供电分公司

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. The AI-based source network load storage integrated micro-grid cluster multi-energy collaborative optimization scheduling method is characterized by comprising the following steps: Collecting source network load storage full-element operation data in real time, preprocessing the data, and constructing a multidimensional input feature set based on the preprocessed data; Inputting the multidimensional input feature set into a pre-built time sequence transducer model to perform wind-light output and load ultra-short-term rolling prediction, and outputting a predicted value and a confidence interval; Inputting the predicted value into a digital twin model to perform power grid operation simulation and tide calculation, performing safety early warning based on simulation results, and generating rigid safety constraint; Inputting the rigid safety constraint and the predicted value into a scaling dual descent alternating direction multiplier method model, and executing distributed collaborative optimization solution to generate power allocation and mutual aid instructions; Issuing the power distribution and mutual aid instruction to a hybrid energy storage system, a controllable load and a distributed power supply for execution; And correcting the prediction parameters of the time sequence transducer model and the scheduling parameters of the scaling dual-reduction alternating direction multiplier method model according to the execution result.
2. The method of claim 1, further comprising constructing a three-level distributed architecture in conjunction with a multi-time-scale scheduling mechanism; The three-level distributed architecture comprises an area cluster layer, a single micro-grid layer and an equipment control layer, wherein the area cluster layer is used for making a global power mutual-aid strategy and a tie line power boundary plan among multiple micro-grids, the single micro-grid layer is used for executing optimal scheduling and safety check of local source network load storage, and the equipment control layer is used for executing equipment instructions of millisecond to minute level; the multi-time scale scheduling mechanism comprises a day-ahead benchmark plan, intra-day rolling correction and ultra-short-term real-time stabilization.
3. The method of claim 1, wherein inputting the multi-dimensional input feature set into a pre-built time sequence transducer model for wind-solar power output and load ultra-short-term rolling prediction, outputting a predicted value and a confidence interval, comprises: extracting long time sequence dependent features in the multi-dimensional input feature set according to a self-attention mechanism of the time sequence transducer model, and generating a prediction sequence comprising a plurality of time points; Performing offline pre-training on the time sequence transducer model with the aim of minimizing the prediction error, and performing online fine tuning on model parameters in a rolling updating mode; And outputting a point predicted value and a confidence interval of a future preset time scale according to the predicted sequence.
4. The method of claim 1, wherein inputting the predicted value into a digital twin model for grid operation simulation and power flow calculation, performing safety precaution based on simulation results and generating rigid safety constraints, comprises: mapping operation data acquired in real time to the digital twin model to realize virtual-real synchronization of a physical entity and a virtual model; Based on the operation data acquired in real time, carrying out multi-time scale tide calculation according to a tide algorithm preset in the digital twin model; Evaluating the running state of the power grid according to the load flow calculation result, identifying an abnormal scene in the running state of the power grid according to a preset abnormal judgment rule, and sending out a safety early warning; and generating operation safety boundary constraint according to the load flow calculation result, and taking the operation safety boundary constraint as the rigid safety constraint of the scaled dual descent alternating direction multiplier method model.
5. The method of claim 1, wherein scaling the rigid safety constraint with the predicted value input into a dual descent alternating direction multiplier model, performing a distributed collaborative optimization solution, generating power allocation and mutual-match instructions, comprising: modeling a power distribution problem of a micro-grid cluster into a distributed non-convex optimization problem containing nonlinear coupling constraint by taking each sub-micro-grid as an independent optimization block, wherein the rigid safety constraint is used as a local operation constraint of the distributed non-convex optimization problem, and the predicted value is used as an input parameter of the distributed non-convex optimization problem; constructing a regularized augmented Lagrangian function, and introducing a regularized term of the dual variable into the standard augmented Lagrangian function; each sub-micro-grid synchronously and parallelly solves respective local sub-problems based on power information exchanged between the current dual variable and the adjacent micro-grid; Each sub-micro grid only interacts adjacent tie line power residual errors, and performs consistency variable collaborative updating; updating the dual variables by adopting a scaling dual descent formula; and controlling the iterative process through self-adaptive parameter adjustment and early shutdown until a preset convergence condition is met or a preset maximum iteration number is reached, and outputting the energy storage charging and discharging power, the controllable load adjustment quantity and the micro-grid inter-connection line exchange power instruction.
6. The method of claim 5, wherein the adaptive parameter tuning and residual early-shutdown system specifically comprises: calculating an original residual error and a dual residual error of each iteration, dynamically adjusting punishment parameters according to a comparison result of the original residual error and the dual residual error, and correspondingly adjusting scaling factors according to adjustment conditions of the punishment parameters; and stopping iteration in advance when the original residual error and the dual residual error are smaller than a preset convergence threshold value, or stopping iteration when the iteration number reaches a preset maximum iteration number.
7. The method of claim 1, wherein issuing the power allocation and mutual-match instruction to a hybrid energy storage system for execution comprises: The hybrid energy storage system adopts a composite energy storage framework consisting of a lithium battery, a super capacitor and a echelon battery; The power distribution and dispatch power in the mutual aid instruction is decomposed into a high-frequency component, a medium-frequency component and a low-frequency component by adopting a filtering algorithm, and is respectively distributed to the super capacitor, the lithium battery and the echelon battery for execution; And adjusting the power distribution priority according to the health status sequencing result of each energy storage unit, and issuing and executing the adjusted power instruction.
8. The method of claim 7, wherein the adjusting the power allocation priority according to the health sequencing result of each energy storage unit specifically comprises: estimating the state of charge and the health state of each energy storage unit in real time, and sequencing each energy storage unit according to the health state; Dividing the energy storage unit into a priority scheduling unit and a non-priority scheduling unit according to the sequencing result, wherein the priority scheduling unit bears high-power charge and discharge, and the non-priority scheduling unit limits the charge and discharge depth; and limiting the single-day charge and discharge cycle times of each energy storage unit not to exceed a preset cycle time threshold.
9. The method of claim 1, wherein modifying the predicted parameters of the time series fransformer model and the scaled dual reduced alternating direction multiplier model scheduling parameters based on the execution result comprises: performing online incremental fine adjustment on the time sequence transducer prediction model according to the execution result data, and updating model prediction parameters at preset time intervals; And optimizing the penalty parameter and the scaling factor of the scaling dual-descent alternating direction multiplier method model according to the execution result data, and using the penalty parameter and the scaling factor for optimizing calculation of a subsequent scheduling period.
10. The method of claim 7, wherein after issuing the adjusted power instruction, the method further comprises: and acquiring the actual output power of each energy storage converter in real time, comparing the actual output power with the adjusted power instruction to calculate an execution error, and eliminating the execution error by adopting a PID correction algorithm.

Description

AI-based source network charge storage integrated micro-grid cluster multi-energy collaborative optimization scheduling method Technical Field The embodiment of the invention relates to the technical field of power system automation, artificial intelligence and distributed optimal scheduling, in particular to a source network charge storage integrated micro-grid cluster multi-energy collaborative optimal scheduling method based on AI. Background With the deep advancement of the construction of a novel power system, a source network charge storage integrated micro-grid cluster becomes a core grid-connected form for realizing high-proportion new energy consumption and safe low-carbon power supply in main flow scenes such as industrial parks, county distribution networks, industrial and commercial parks and the like, and the multi-production-right main body of the micro-grid cluster has the operation characteristics of independent operation, nonlinear coupling of multi-energy flows and severe grid-connected safety requirements and provides clear engineering requirements for a cluster-level collaborative scheduling technology. The scheme constructs a basic state parameter set by acquiring operation state information of an energy storage subunit, accurately characterizes power capacity, energy boundary and operation limit of the energy storage unit, judges a unique dispatching state identification of the energy storage unit in a single dispatching cycle, limits the energy storage unit to only participate in one dispatching behavior type, and realizes fine dispatching of the energy storage unit in a single micro-grid scene. However, the prior art (comprising the scheme) still has the core defect of serious disconnection from the actual landing scene of the micro-grid cluster in the power industry, and particularly has the following steps that (1) the scene suitability of a dispatching architecture is insufficient, the prior scheme only focuses on dispatching optimization of energy storage monomers in a single micro-grid, a distributed cooperative architecture which is suitable for multi-property main operation of the micro-grid cluster and is required by hierarchical dispatching management of a distribution network is not constructed, and the core engineering requirement of cross-micro-grid power mutual economy cannot be met. (2) Meanwhile, the conventional optimization algorithm cannot adapt to a non-convex optimization scene of cluster-level multi-element nonlinear coupling, iteration divergence and solving failure problems are easy to occur, and response requirements of real-time scheduling of power engineering are difficult to meet. (3) The full-flow closed loop control mechanism is lacking, the existing scheme only realizes the energy storage state judgment and static scheduling in a single dispatching cycle, the problem that a dispatching model is continuously disjointed from the actual working condition on site cannot be solved, and the long-term stable grid-connected operation of the micro-grid cluster is difficult to ensure. Therefore, developing a full closed loop optimized scheduling method which is adaptive to non-convex constraint, distributed collaboration, simulation-scheduling deep coupling and energy storage refined control becomes a core technical problem to be solved in the field. Disclosure of Invention Aiming at the defects of the prior art, the embodiment of the invention provides an AI-based source network charge storage integrated micro-grid cluster multi-energy collaborative optimization scheduling method, which realizes the efficient collaborative and stable operation of all elements of the micro-grid cluster endogenous, network, charge and storage through full-flow fusion of sequential AI prediction, digital twin simulation, scaling dual descent Alternating Direction Multiplier Method (ADMM) distributed collaborative scheduling and hybrid energy storage hierarchical control. In a first aspect, an embodiment of the present invention provides an AI-based source network load storage integrated micro-grid cluster multi-energy collaborative optimization scheduling method, including: Collecting source network load storage full-element operation data in real time, preprocessing the data, and constructing a multidimensional input feature set based on the preprocessed data; Inputting the multidimensional input feature set into a pre-built time sequence transducer model to perform wind-light output and load ultra-short-term rolling prediction, and outputting a predicted value and a confidence interval; Inputting the predicted value into a digital twin model to perform power grid operation simulation and tide calculation, performing safety early warning based on simulation results, and generating rigid safety constraint; Inputting the rigid safety constraint and the predicted value into a scaling dual descent alternating direction multiplier method model, and executing distributed collaborative op