A microgrid whole life cycle planning method based on physical disentangled agent model

Abstract

This invention discloses a microgrid lifecycle planning method based on a physically decoupled surrogate model in the field of microgrid planning and optimized operation technology. This method includes: constructing a two-layer microgrid planning model covering the entire equipment lifecycle; establishing a physically decoupled surrogate model architecture; decomposing the objective vector into a smooth continuous subspace and a sparse abrupt subspace, and designing corresponding loss functions for each; adaptive active learning sampling during the execution phase, utilizing an uncertainty quantification method based on committee queries, and introducing a dynamic decay weighting mechanism; and employing a hybrid oracle-based online correction integer-encoded multi-objective evolutionary algorithm to solve for the optimal configuration scheme. This invention can eliminate the feature masking effect between heterogeneous objectives, improve the accuracy of identifying physical hard constraint boundaries, and achieve efficient and high-fidelity planning scheme solutions under limited computing resources.

Description

Technical Field

[0001] This invention relates to the field of microgrid planning and optimized operation technology, and more specifically, to a microgrid full life cycle planning method based on a physical decoupling agent model. Background Technology

[0002] Microgrid planning aims to achieve a balance between economic efficiency and low carbon emissions while meeting load demand by optimizing the capacity configuration of distributed power sources (such as solar and wind power), energy storage systems, and diesel generators. Traditional microgrid planning methods primarily focus on energy dispatch during system operation, aiming to minimize fuel consumption and grid purchase costs. However, with the increasing penetration rate of renewable energy equipment, the proportion of implicit carbon emissions generated during the manufacturing, transportation, installation, and decommissioning processes of these devices has increased significantly. Ignoring these implicit carbon emissions can lead to a "carbon inversion" phenomenon in the planning results, where increasing energy-intensive equipment to reduce operational emissions actually increases the total carbon emissions over the entire life cycle.

[0003] Integrating life-cycle assessment (LCA) into microgrid planning requires establishing a two-layer optimization model that couples long-term capacity configuration decisions with short-term multi-scenario energy dispatch. This model is a large-scale mixed-integer linear programming (MILP) problem, exhibiting NP-hard characteristics. Solving it directly using traditional branch-and-bound methods or commercial solvers incurs enormous computational loads and is extremely time-consuming; a single optimization cycle can be extremely time-consuming and even lead to memory overflow, failing to meet the demands of rapid iteration for planning schemes in engineering applications.

[0004] To address the aforementioned computational efficiency issues, existing technologies have introduced surrogate model-assisted optimization methods. These methods utilize machine learning algorithms to establish a mapping relationship between device configuration variables and system performance indicators, replacing time-consuming time-series simulation models. However, existing surrogate models suffer from the following technical drawbacks when applied to microgrid lifecycle planning: First, there is the feature masking problem caused by heterogeneous objectives. Microgrid planning objectives include two types of indicators with drastically different physical properties: one is lifecycle cost and carbon emissions, which are large in magnitude and continuous and smooth within the decision space; the other is reliability penalty indicators such as load shedding and power curtailment, which are small in magnitude and exhibit a highly sparse, zero-inflation distribution. Existing technologies typically use a uniform regression model for end-to-end modeling. During training, the gradients generated by the large-scale economic indicators dominate parameter updates, thus masking the sparse penalty signals that determine the physical feasibility of the system. This prevents the surrogate model from accurately learning the physical constraint boundaries of the system.

[0005] Second, sampling strategies lacking physical awareness are inefficient. Existing active learning sampling strategies typically employ the objective-neutral principle, uniformly distributing computational resources across the global decision space. This results in a large number of samples falling into safe regions, while samples in the boundary regions that truly determine system reliability are scarce. Consequently, the surrogate model's prediction accuracy under extreme conditions is insufficient, failing to guarantee the robustness of the planning results.

[0006] In summary, existing technologies cannot eliminate the influence of feature masking when dealing with heterogeneous physical targets, and lack a sampling mechanism that can automatically focus on physical constraint boundaries, making it difficult to ensure the physical feasibility of the planning scheme while guaranteeing computational efficiency. Summary of the Invention

[0007] The purpose of this invention is to provide a microgrid lifecycle planning method based on a physical decoupling proxy model, which solves the technical problems of carbon inversion, high computational complexity, feature masking, and low efficiency in exploring constraint boundaries. It can eliminate the feature masking effect between heterogeneous targets, improve the identification accuracy of physical hard constraint boundaries, and achieve efficient and high-fidelity planning solutions under limited computing resources.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: A microgrid lifecycle planning method based on a physical decoupling agent model includes the following steps: S1. Construct a two-layer planning model for the entire life cycle of a microgrid. The upper layer of the two-layer planning model is the investment planning layer, which uses the number of installed equipment as the decision variable and aims to minimize the total life cycle cost and carbon emissions. The lower layer of the two-layer planning model is the operation and scheduling layer, which uses hourly power under typical scenarios as the decision variable and aims to minimize the overall operating cost, while being subject to power balance, equipment operation logic, and physical constraints. S2. Construct a physically decoupled surrogate model architecture. Perform physical orthogonal decomposition on the output vector of the bi-level programming model in step S1, dividing it into a smooth continuous subspace and a sparse mutation subspace. Design corresponding loss functions for different subspaces and train independent base regressors to construct a decoupled surrogate model. S3. In the execution phase, adaptive active learning sampling is used. An uncertainty quantification method based on committee query is adopted and a dynamic decay weighting mechanism is introduced. The sampling weight of the sparse mutation subspace target is dynamically adjusted according to the iteration round. High-value samples are selected from the candidate pool, the bi-level programming model of step S1 is called for labeling, and the surrogate model is updated. S4. Online correction and optimization based on hybrid oracle: A multi-objective evolutionary algorithm with integer encoding is used. The surrogate model trained in step S3 is used for fitness evaluation to quickly search for the Pareto optimal solution set. During the search process, key solutions are periodically extracted and the bi-level programming model in step S1 is called for high-fidelity verification. If the prediction error exceeds the limit, the true solution is added to the training set for incremental update. Finally, the verified optimal microgrid configuration scheme is output.

[0009] Furthermore, the evaluation boundary of the microgrid's entire life cycle in S1 is expanded from a single operation phase to the entire life cycle, covering four stages: raw material acquisition and equipment manufacturing, transportation and construction, operation and maintenance, and end of life and recycling.

[0010] Furthermore, in step S1, the total life cycle cost includes initial investment cost, equipment replacement cost, fixed operation and maintenance cost, total life cycle operating cost, and asset residual value recovery; the total life cycle carbon emissions include implicit carbon emissions in the equipment manufacturing, transportation, construction, and recycling stages, as well as direct carbon emissions in the operation stage.

[0011] Furthermore, the physical orthogonal decomposition in S2 specifically involves decomposing the output vector of the bi-level programming model into four independent physical components: basic operating cost, annualized operating carbon emissions, annualized load shedding, and annualized abandoned electricity.

[0012] Furthermore, in S2, the smooth continuous subspace includes basic operating costs and annualized operating carbon emissions, using mean squared error (MSE) as the loss function; the sparse abrupt subspace includes annualized load shedding and annualized power curtailment, trained using the Tweedie distribution loss function, where the variance power exponent of the Tweedie distribution loss function is... Values ​​in .

[0013] Furthermore, the base regressor in S2 is a lightweight gradient booster, and a prediction committee containing multiple independently trained lightweight gradient booster regressors is constructed for each physical component to quantify the uncertainty of the prediction; the lightweight gradient booster adopts a leaf-growing strategy and a histogram-based algorithm structure, and the uncertainty of the prediction committee is the standard deviation of the prediction values ​​of each base regressor.

[0014] Furthermore, the dynamic decay weighting mechanism in S3 is specifically as follows: a comprehensive scoring function is defined, which is a weighted sum of the standardized uncertainties of each sub-objective; the weights of the sparse mutation subspace are dynamically assigned using an inverse decay function, and are given high weights in the early stage of training to prioritize the exploration of physical constraint boundaries, and the weights gradually decay to balance with the smooth continuous subspace as the number of iterations increases.

[0015] Furthermore, the process of selecting high-value samples from the candidate pool in S3 specifically involves: adopting a hybrid query strategy, selecting a preset proportion of samples with the highest comprehensive score as utilization samples in each round of active learning iteration, randomly selecting the remaining proportion of samples from the top 50% of candidate samples as exploration samples, and inputting the selected samples into a bi-layer programming model for high-fidelity labeling.

[0016] Furthermore, the online correction of the hybrid oracle in S4 specifically involves: during the evolution process, periodically extracting the cost minimum point, carbon emission minimum point, and compromise knee point from the current non-dominated solution set as key solutions; inputting the key solutions into the bi-level programming model in step S1 for verification to obtain the true performance indicators; if the relative error between the surrogate model's predicted value and the true value exceeds a preset threshold, then adding the key solutions and their true values ​​to the training set to incrementally update the surrogate model.

[0017] Furthermore, the multi-objective evolutionary algorithm with integer encoding in S4 directly maps the number of installed devices to integer genes. During fitness evaluation, if the load shedding or abandoned power predicted by the surrogate model violates the physical constraints, the individual is marked as an infeasible solution or a high penalty is imposed; otherwise, the physical components are recombined into the total life cycle cost and total life cycle carbon emissions.

[0018] By adopting the above technical solution, the present invention has the following advantages: 1. This invention provides a microgrid full life cycle planning method based on a physically decoupled proxy model. By constructing a physically orthogonal decoupled architecture, it separates and models continuous and smooth economic indicators from sparse and abrupt physical penalty indicators, and introduces the Tweedie distribution loss function to process zero-inflation data. This significantly improves the identification accuracy of physical hard constraint boundaries (such as load shedding events), avoids the risk of power outages in actual operation of the planning scheme, and eliminates the "feature masking" effect between heterogeneous targets.

[0019] 2. This invention provides a microgrid full life cycle planning method based on a physical decoupling surrogate model. By using a trained surrogate model to replace time-consuming MILP time series simulation, the optimization time of the entire process can be compressed from several hours to minutes while ensuring that the global prediction error is less than 5%, achieving a speedup of hundreds of times. This meets the needs of rapid engineering iteration and achieves an order-of-magnitude improvement in computational efficiency at high fidelity.

[0020] 3. This invention provides a microgrid lifecycle planning method based on a physically decoupled surrogate model. Through a phase-adaptive active learning sampling strategy, it dynamically adjusts the sampling focus, prioritizing the exploration of physical constraint boundaries before shifting to global accuracy development. Under small sample conditions, it significantly improves prediction accuracy and reliability recall, enhances data utilization efficiency, and greatly reduces reliance on expensive simulation data.

[0021] 4. This invention provides a microgrid full life cycle planning method based on a physical decoupling agent model. By introducing full life cycle evaluation, the implicit carbon emissions of equipment are explicitly included in the planning model, which can accurately identify and eliminate ineffective emission reduction investment schemes, providing a scientific decision-making basis for the deep decarbonization of microgrids and effectively avoiding the risk of "carbon inversion".

[0022] 5. This invention provides a microgrid lifecycle planning method based on a physically decoupled surrogate model. By embedding a hybrid oracle online correction mechanism into the evolutionary algorithm, the key solutions of the Pareto front are verified and incrementally updated with high fidelity. This eliminates the prediction bias of the surrogate model in the extrapolation region, ensuring the physical strict feasibility of the final scheme and the engineering practical value of the final output scheme. Attached Figure Description

[0023] Figure 1 This is a flowchart of the overall process for a microgrid lifecycle planning method based on a physical decoupling agent model. Figure 2 This is a schematic diagram of the microgrid system topology and life cycle evaluation boundary of the present invention; Figure 3 This is a logical architecture diagram of the microgrid full life cycle two-layer planning model of the present invention; Figure 4 This is a schematic diagram of the physical decoupling proxy model of the present invention. Detailed Implementation

[0024] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings. It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0025] The microgrid lifecycle planning method proposed in this invention, based on physical decoupling agent-assisted optimization, is as follows: Figure 1 As shown, it includes the following steps: S1. Construct a two-layer planning model for the entire life cycle of a microgrid. The upper layer of the two-layer planning model is the investment planning layer, which uses the number of installed equipment as the decision variable and aims to minimize the total life cycle cost and total life cycle carbon emissions. The lower layer of the two-layer planning model is the operation and scheduling layer, which uses hourly power under typical scenarios as the decision variable and aims to minimize the overall operating cost, while being subject to power balance, equipment operation logic and physical constraints.

[0026] The purpose of this step is to establish a high-fidelity mathematical model to describe the economic costs, carbon emissions, and physical operating characteristics of a microgrid throughout its entire lifecycle. This model serves not only as the objective function for the final optimization but also as the ground truth for generating training data and constructing surrogate models.

[0027] The specific implementation process includes the following sub-steps: Define system boundaries and physical topology: First, determine the physical structure and evaluation boundaries of the microgrid system.

[0028] Physical topology: specifically as follows Figure 2 As shown, the system includes a photovoltaic (PV) array, wind turbine (WT) generators, and diesel generator (DG) generators as the power supply side; a battery energy storage system (ESS) as the flexibility adjustment unit; and load side including conventional loads and interruptible loads. The system is connected to the external main grid through a point of common coupling (PCC), allowing bidirectional power exchange.

[0029] Life Cycle Assessment (LCA) Boundaries: To avoid the "carbon inversion" phenomenon, this invention extends the assessment boundaries from a single operational phase to the entire life cycle, covering four core phases: 1. Raw material acquisition and equipment manufacturing stage (generating initial occult carbon); 2. Transportation and construction phase (generating construction carbon emissions and initial investment); 3. Operation and maintenance phase (generating carbon emissions from fuel consumption, carbon emissions from electricity purchases, and operation and maintenance costs). 4. End-of-life and recycling phase (involving the negative carbon benefits of residual value recovery and material regeneration).

[0030] Secondly, constructing such Figure 3 The microgrid lifecycle bi-level programming model shown has an upper level for investment planning, where the decision variables are the number of each type of equipment installed in the system (an integer vector). The goal of the upper level is to minimize the total lifecycle cost (LCC) and total lifecycle carbon emissions (LCE).

[0031] The lower layer is the operation and scheduling layer: it receives device capacity configurations from the upper layer as hard constraints. For typical daily scenarios, it optimizes hourly power flow and energy state. The goal of this lower layer is to minimize system operating costs (including carbon tax and reliability penalties).

[0032] The upper and lower layers have a two-layer coupling relationship: the mathematical form is shown in formula (1), where the upper layer decision variables are... The feasible domain of the lower layer is defined, and the optimal operating indicators fed back from the lower layer are... Used to calculate the objective function value of the upper layer.

[0033] (1)

[0034] Establish a lower-level operation optimization model (based on mixed-integer linear programming MILP): The lower-level model simulates the hourly energy flow of the microgrid under a given configuration, with specific constraints including: Power balance constraint: ensures that at any given time The sum of renewable energy output, diesel engine output, energy storage discharge power, and grid power purchase is equal to the sum of load demand, energy storage charging power, and grid power sales.

[0035] To ensure mathematical feasibility, a non-negative relaxation variable, load shedding power, is introduced. and abandoned power See the formula below for details.

[0036] (2)

[0037] 1. Diesel generator operating constraints: These include generator start / stop logic, minimum load rate limits (to prevent inefficient operation), and ramp constraints. See the following formula for details: (3) (4) (5) 2. Dynamic constraints of energy storage systems: describing the battery's state of charge (SOC) as a function of charge and discharge power and efficiency. The evolution across time periods includes charging and discharging mutual exclusion constraints. See the following formula for details.

[0038] (6) (7) (8) Lower-level objective function: Minimize the overall daily operating cost It consists of fuel costs, start-up costs, net grid interaction costs, carbon tax costs, and reliability penalty costs (weighted penalties for load shedding and power curtailment). See the formula below for details.

[0039] (9) (10) (11) (12) (13) (14) Establish an upper-level full lifecycle evaluation model: Based on the operational results fed back from the lower level, the upper-level model calculates two core objective functions: 1. Objective Function 1: Life Cycle Cost (LCC) (15) Lifecycle cost (LCC) includes initial investment cost. Equipment replacement costs Fixed operation and maintenance costs Total lifecycle operating costs and asset residual value recovery It consists of five parts.

[0040] The detailed calculations for each cost item are shown in the formulas below. In particular, the replacement cost takes into account the equipment's lifespan. Shorter than the project cycle Multiple changes in time and the time value of money (discount rate) ).

[0041] (16) (17) (18) (19) (20) 2. Objective Function Two: Life Cycle Carbon Emissions (LCE) (twenty one) Lifecycle carbon emissions include embodied carbon emissions from equipment manufacturing, transportation, construction, and recycling. Direct carbon emissions during operation It consists of two parts.

[0042] Hidden carbon emissions Calculation: Based on the equipment lifecycle assessment factors and the number of replacements during the project cycle calculate, (twenty two) direct carbon emissions Calculation: Annualized carbon emissions calculated based on the lower-level model The summation is performed, as detailed in the formula below.

[0043] (twenty three)

[0044] Data generation and scenario reduction: To improve computational efficiency while preserving characteristics of extreme operating conditions, a K-Means clustering algorithm combined with an extreme value screening mechanism was used to extract data from historical meteorological and load data. A typical scenario and Each extreme scenario is assigned a corresponding weight. Using the aforementioned two-level programming model for the entire lifecycle of a microgrid, through the decision space... Internal sampling and invoking the MILP solver generate a dataset containing decision variables. Corresponding physical performance indicators A high-fidelity dataset (basic cost, carbon emissions, load shedding, and abandoned electricity) is used for training subsequent proxy models.

[0045] S2. Construct a physically decoupled surrogate model architecture. Perform physical orthogonal decomposition on the output vector of the bi-level programming model in step S1, dividing it into a smooth continuous subspace and a sparse mutation subspace. Design corresponding loss functions for different subspaces and train independent basis regressors to construct a decoupled surrogate model. Specifically, the physical orthogonal decomposition is to decompose the output vector of the bi-level programming model into four independent physical components: basic operating cost, annualized operating carbon emissions, annualized load shedding, and annualized abandoned electricity.

[0046] This step aims to establish a surrogate model architecture based on physics-disentangled decoupling to replace the computationally intensive MILP solution process. Addressing the feature masking problem caused by the strong heterogeneity of the objective function in microgrid lifecycle planning (i.e., huge and continuous economic indicators, and sparse and zero-expansion distribution of reliability indicators), this step employs orthogonal decoupling and adaptive distribution mapping strategies to ensure that the surrogate model can simultaneously and accurately capture both smooth economic surfaces and abrupt physical constraint boundaries.

[0047] In a specific embodiment, the implementation process is as follows: Figure 4 As shown, it includes the following sub-steps: Physical orthogonal decoupling and subspace partitioning: the output vector of the bi-level programming model (Including multi-dimensional physical performance indicators) physical orthogonal decomposition is performed instead of using the traditional end-to-end single model for prediction.

[0048] Decomposition strategy: Decompose the output vector Decomposed into four independent physical components: basic operating cost Annualized carbon emissions Annualized load shedding and annualized abandoned electricity .

[0049] Subspace definition: Based on the statistical distribution of each physical component in the decision space, it is divided into two orthogonal characteristic subspaces: 1. Smooth Continuous Subspace: Includes basic operating costs. and annualized carbon emissions These two indicators reflect the flow of funds and energy in a system under normal power supply conditions. They exhibit a globally continuous and gradient-smooth distribution across the entire decision space, and their numerical magnitudes are typically large (e.g., costs within a certain range). level).

[0050] 2. Sparse Zero-Inflated Subspace: Contains annualized shear load. and annualized abandoned electricity These two indicators characterize the physical reliability boundary of the system (i.e., whether power outages or curtailment of solar and wind power occur). Within the vast majority of decision regions where power balance is satisfied, their values ​​are strictly zero; only when the configuration scheme touches the physical hard constraint boundary will their values ​​exhibit a nonlinear positive step. Their distribution characteristics are typical of a zero-inflated distribution.

[0051] Design of adaptive loss function for distribution: For the statistical differences between the two subspaces mentioned above, adaptive loss functions are configured for each to eliminate the gradient masking effect during backpropagation. For smooth continuous subspaces: mean squared error (MSE) is used as the loss function to minimize the Euclidean distance between the predicted and true values, ensuring high-precision regression of economic and carbon emission surfaces.

[0052] For sparse mutation subspaces (key technical point): The Tweedie distribution is introduced as the core assumption of the loss function, and the Tweedie distribution loss function is used for training. The variance power exponent of the Tweedie distribution loss function is... Values ​​in .

[0053] Explanation of principle: The Tweedie distribution belongs to the family of exponential dispersion models. When its variance power exponent takes the value of... When the interval is defined, it corresponds to the Compound Poisson-Gamma Process. This mathematical property allows for the joint modeling of both the probability quality concentrated at zeros (corresponding to the system's safe state) and the continuous distribution over the positive real number field (corresponding to the degree of violation).

[0054] Technical effect: By configuring the Tweedie loss function, the surrogate model can adaptively suppress small prediction noise within the safe region (forced regression to zero) while keenly capturing abrupt gradients at the constraint boundaries, thereby avoiding the gradients of high-level economic indicators from drowning out sparse physical violation signals.

[0055] Construction and Integration of Base Learners: The base regressors are Lightweight Gradient Boosting Machines (LightGBM), which are used as the base regressors for the four decoupled sub-models mentioned above. A prediction committee containing multiple independently trained LightGBM regressors is constructed for each physical component to quantify the uncertainty of the predictions. The Lightweight Gradient Boosting Machine employs a leaf-growing strategy and a histogram-based algorithm structure. The uncertainty of the prediction committee is the standard deviation of the predictions from each base regressor.

[0056] Algorithm selection criteria: Lightweight Gradient Boosting Machine (LightGBM) employs a leaf-wise growth strategy, which reduces loss compared to a level-wise growth strategy, thus achieving higher prediction accuracy in physical boundary regions with significant gradient changes. Furthermore, its histogram-based algorithm structure effectively handles high-dimensional mixed-integer features.

[0057] Uncertainty Quantification Preparation: To support subsequent active learning, a bootstrap aggregating (Bagging) strategy is employed for each physical sub-objective to construct a system containing... A prediction committee of independently trained LightGBM regressors. For any input... The model not only outputs the predicted mean It also outputs the predicted standard deviation. The formula for quantifying cognitive uncertainty is as follows.

[0058] (twenty four).

[0059] S3. In the execution phase, adaptive active learning sampling is used. An uncertainty quantification method based on committee query is adopted and a dynamic decay weighting mechanism is introduced. The sampling weight of the sparse mutation subspace target is dynamically adjusted according to the iteration round. High-value samples are selected from the candidate pool, the bi-level programming model of step S1 is called for labeling, and the surrogate model is updated. This step aims to drive the training focus of the surrogate model smoothly from the initial "Feasibility Boundary Discovery" to the later "Global Accuracy Refinement" through a dynamic sampling strategy. Given the extreme scarcity of physical constraint boundary (e.g., load shedding events) samples in microgrid planning, this step employs a hybrid strategy based on Query-By-Committee (QBC).

[0060] The specific implementation process includes the following sub-steps: Committee Construction and Uncertainty Quantification: Using Bootstrap Aggregating for Each Decoupled Physical Objective in Step S2 Build includes Prediction Committee of Individual Base Regressors For any unlabeled sample in the candidate pool The uncertainty of the model's predictions is measured by the degree of disagreement among committee members. Mathematically, it is defined as the standard deviation of the predicted values. For the calculation formula, please refer to formula (24).

[0061] Standardization: Due to the vast differences in magnitude between different physical targets (e.g., the basic cost is approximately...), The order of magnitude, while the load shedding amount may be... (The magnitude may be zero), directly accumulating the original standard deviation will lead to targets with large magnitudes dominating the sampling weights. Therefore, this step utilizes the statistical standard deviation of each target value in the currently labeled sample pool to... By performing division normalization, we obtain the standardized uncertainty index. .

[0062] The dynamic decay weighting mechanism is as follows: a comprehensive scoring function is defined, which is the weighted sum of the standardized uncertainties of each sub-objective; the weights of the sparse mutation subspace are dynamically assigned using an inverse decay function. In the early stage of training, they are given high weights to prioritize the exploration of physical constraint boundaries, and the weights gradually decay to balance with the smooth continuous subspace as the number of iterations increases.

[0063] To shift the training focus from "finding the feasible region boundary" to "optimizing the Pareto front," this step is designed to adapt to different iteration rounds. Change in the comprehensive scoring function .

[0064] Comprehensive score calculation: defined as the weighted sum of the standardized uncertainties of each sub-objective, see formula (25) for details.

[0065] (25)

[0066] Dynamic weight adjustment: targeting objectives in sparse mutation subspaces (such as load shedding). ), its weighting coefficient Dynamic values ​​are assigned using an inverse decay function, as detailed in the following formula: (26) in, The lower bound of the basic weight, For dynamic gain amplitude, This is the decay rate coefficient.

[0067] Physical meaning: In the early stages of training ( (smaller) The significantly higher weights compared to other smoothing objectives force the algorithm to prioritize sampling those samples that cause disagreement within the committee regarding "whether to cut the load," thus quickly locking in the physical hard constraint boundaries; as iterations increase, The sampling weights are reduced to a balance with the smoothing target weights, and the sampling focus is regressed to a fine regression of overall economic efficiency and carbon emissions.

[0068] Specifically, the selection of high-value samples from the candidate pool involves: adopting a hybrid query strategy, selecting a preset proportion of samples with the highest comprehensive score as utilization samples in each round of active learning iteration, randomly selecting the remaining proportion of samples from the top 50% of candidate samples as exploration samples, and inputting the selected samples into a bi-layer programming model for high-fidelity labeling.

[0069] Hybrid Query Strategy: To prevent the surrogate model from getting stuck in locally high variance regions and to maintain the diversity of the feature space, this step selects from the candidate pool in each round of active learning iteration. High-fidelity annotation is performed on each new sample (using the MILP model from step S1). The selection process employs a hybrid strategy combining "exploitation sampling" and "exploration sampling": 1. Exploitation sampling: Selecting a comprehensive score highest A sample of samples. This sample is used to eliminate the largest cognitive blind spot in the current model.

[0070] 2. Exploration Sampling: From the remaining candidate samples, first select the top 50% subset by rating, and then randomly select from it. A sample size. This sample size is used to prevent the sampling distribution from becoming too concentrated and to ensure coverage of the entire decision space.

[0071] 3. Closed-loop iterative update: The above-selected... A new sample is input into the bi-level programming model in step S1 to solve the problem and obtain the actual physical performance indicators. This data is then added to the training dataset. All LightGBM base regressors are then updated, and the process continues until the preset maximum number of iterations is reached or the global prediction error converges.

[0072] S4. Online correction and optimization based on hybrid oracle: A multi-objective evolutionary algorithm with integer encoding is used. The surrogate model trained in step S3 is used for fitness evaluation to quickly search for the Pareto optimal solution set. During the search process, key solutions are periodically extracted and the bi-level programming model in step S1 is called for high-fidelity verification. If the prediction error exceeds the limit, the true solution is added to the training set for incremental update. Finally, the verified optimal microgrid configuration scheme is output.

[0073] This step aims to use the high-precision surrogate model trained in steps S2 and S3 to replace the computationally intensive MILP simulation, enabling a rapid search for the Pareto optimal solution set within a large integer decision space. To prevent prediction bias caused by sample sparsity in the later stages of the search (especially misjudgment of the shear load boundary), this step introduces an online correction closed-loop mechanism.

[0074] Integer-coded multi-objective evolutionary search: An integer-coded multi-objective evolutionary algorithm is used as the global search algorithm for upper-level planning. The integer-coded multi-objective evolutionary algorithm directly maps the number of installed devices to integer genes. During fitness evaluation, if the load shedding or abandoned power predicted by the surrogate model violates the physical constraints, the individual is marked as an infeasible solution or a high penalty is imposed; otherwise, the physical components are recombined into the total life cycle cost and total life cycle carbon emissions.

[0075] In one specific embodiment, the encoding strategy is: decision variables. The number of photovoltaic, wind turbine, diesel engine and energy storage installations is directly mapped to integer genes, avoiding solution space drift caused by relaxation of continuous variables.

[0076] Fitness assessment: In the evolutionary process, the fitness of each individual in the population (i.e., life-cycle cost (LCC) and life-cycle carbon emission (LCE)) is no longer calculated by calling time-consuming MILP models, but rather by incorporating decision variables... Input the physical decoupling agent model constructed in step S2.

[0077] Calculation process: The proxy model outputs four physical components in parallel (basic operating cost, annualized operating carbon emissions, annualized load shedding, and annualized abandoned electricity).

[0078] If the predicted annualized load shedding... If an individual violates other physical constraints, it is marked as an infeasible solution or subject to a heavy penalty.

[0079] If the constraints are met, then according to formulas (15)-(21), the physical components are reorganized into two target values: life cycle cost (LCC) and life cycle carbon emission (LCE).

[0080] Hybrid Oracle Online Correction: To address the "extrapolation inference" error that may occur when the population tends towards the Pareto front (usually located at the critical boundary of physical constraints) in the later stages of evolutionary algorithms, this step designs an online closed-loop correction strategy: The online correction of the hybrid oracle is as follows: During the evolution process, the cost minimum point, carbon emission minimum point, and compromise knee point are periodically extracted from the current non-dominated solution set as key solutions; the key solutions are input into the bi-level programming model in step S1 for verification to obtain the true performance indicators; if the relative error between the surrogate model's predicted value and the true value exceeds a preset threshold, the key solution and its true value are added to the training set to incrementally update the surrogate model.

[0081] In a specific embodiment, 1. Key Solution Extraction: In each stage of the evolutionary process Every generation (e.g., every 10 generations), evolution is paused, and key candidate solutions with topological representativeness are extracted from the current non-dominated set.

[0082] Extraction strategy: Cover the three key points of the Pareto frontier: the cost minimum, the carbon emission minimum, and the knee point.

[0083] 2. High-fidelity verification (Oracle Labeling): This involves applying the extracted key solutions... The data is then fed back to the underlying high-fidelity MILP model described in step one for rigorous timing simulation to obtain accurate physical performance indicators. .

[0084] Incremental Update: Calculates the predicted values ​​of the surrogate model. Compared with the true value The relative error between them.

[0085] If the relative error exceeds a preset threshold (e.g., 5%), then the prediction in that area is determined to be in failure.

[0086] The system will provide the key solution and its truth label. As high-value samples, they are added to the training dataset.

[0087] The incremental training interface of LightGBM (Booster.update) is called to update the surrogate model parameters only for that local region. This mechanism ensures that the surrogate model can dynamically calibrate its prediction ability for extreme physical boundaries with extremely low computational overhead.

[0088] Output the final planning scheme: After a preset number of algebraic evolutions and online corrections, the algorithm outputs the final Pareto optimal solution set.

[0089] Result format: Includes a series of non-dominated microgrid configuration options (e.g., Option A: Economic priority; Option B: Compromise recommendation; Option C: Low carbon priority).

[0090] Decision support: Each proposal comes with detailed life cycle evaluation indicators (life cycle cost LCC, life cycle carbon emissions LCE) and technical indicators (renewable energy penetration rate, average annual load shedding rate) that have been finally verified by MILP, allowing planners to select based on actual budgets and emission reduction targets.

[0091] Finally, it should be noted that although the present invention has been described with reference to specific embodiments, those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Various equivalent changes or substitutions can be made without departing from the concept of the present invention. Therefore, any changes or modifications to the above embodiments within the essential spirit of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A microgrid lifecycle planning method based on a physical decoupling surrogate model, characterized in that, Includes the following steps: S1. Construct a two-layer planning model for the entire life cycle of a microgrid. The upper layer of the two-layer planning model is the investment planning layer, which uses the number of installed equipment as the decision variable and aims to minimize the total life cycle cost and carbon emissions. The lower layer of the two-layer planning model is the operation and scheduling layer, which uses hourly power under typical scenarios as the decision variable and aims to minimize the overall operating cost, while being subject to power balance, equipment operation logic, and physical constraints. S2. Construct a physically decoupled surrogate model architecture. Perform physical orthogonal decomposition on the output vector of the bi-level programming model in step S1, dividing it into a smooth continuous subspace and a sparse mutation subspace. Design corresponding loss functions for different subspaces and train independent base regressors to construct a decoupled surrogate model. S3. In the execution phase, adaptive active learning sampling is used. An uncertainty quantification method based on committee query is adopted and a dynamic decay weighting mechanism is introduced. The sampling weight of the sparse mutation subspace target is dynamically adjusted according to the iteration round. High-value samples are selected from the candidate pool, the bi-level programming model of step S1 is called for labeling, and the surrogate model is updated. S4. Online correction and optimization based on hybrid oracle: A multi-objective evolutionary algorithm with integer encoding is used. The surrogate model trained in step S3 is used for fitness evaluation to quickly search for the Pareto optimal solution set. During the search process, key solutions are periodically extracted and the bi-level programming model in step S1 is called for high-fidelity verification. If the prediction error exceeds the limit, the true solution is added to the training set for incremental update. Finally, the verified optimal microgrid configuration scheme is output.

2. The microgrid lifecycle planning method based on a physical decoupling agent model according to claim 1, characterized in that, The evaluation boundary of the microgrid's entire life cycle in S1 is expanded from a single operation phase to the entire life cycle, including four stages: raw material acquisition and equipment manufacturing, transportation and construction, operation and maintenance, and end of life and recycling.

3. The microgrid lifecycle planning method based on a physical decoupling agent model according to claim 1, characterized in that, In step S1, the total life cycle cost includes initial investment cost, equipment replacement cost, fixed operation and maintenance cost, total life cycle operating cost and asset residual value recovery; the total life cycle carbon emissions include implicit carbon emissions in the equipment manufacturing, transportation, construction and recycling stages, as well as direct carbon emissions in the operation stage.

4. The microgrid lifecycle planning method based on a physical decoupling agent model according to claim 1, characterized in that, The physical orthogonal decomposition in S2 specifically involves decomposing the output vector of the bi-level programming model into four independent physical components: basic operating cost, annualized operating carbon emissions, annualized load shedding, and annualized abandoned electricity.

5. A microgrid lifecycle planning method based on a physical decoupling surrogate model according to claim 1, characterized in that, In S2, the smooth continuous subspace includes basic operating costs and annualized operating carbon emissions, using mean squared error (MSE) as the loss function; the sparse abrupt change subspace includes annualized load shedding and annualized power curtailment, trained using the Tweedie distribution loss function, where the variance power exponent of the Tweedie distribution loss function is... Values ​​in .

6. The microgrid lifecycle planning method based on a physical decoupling agent model according to claim 1, characterized in that, The base regressor in S2 is a lightweight gradient booster, and a prediction committee containing multiple independently trained lightweight gradient booster regressors is constructed for each physical component to quantify the uncertainty of the prediction. The lightweight gradient booster adopts a leaf-growing strategy and a histogram-based algorithm structure, and the uncertainty of the prediction committee is the standard deviation of the prediction values ​​of each base regressor.

7. The microgrid lifecycle planning method based on a physical decoupling agent model according to claim 1, characterized in that, The dynamic decay weighting mechanism in S3 is as follows: a comprehensive scoring function is defined, which is the weighted sum of the standardized uncertainties of each sub-objective; the weights of the sparse mutation subspace are dynamically assigned using an inverse decay function, and are given high weights in the early stage of training to prioritize the exploration of physical constraint boundaries, and the weights gradually decay to balance with the smooth continuous subspace as the number of iterations increases.

8. The microgrid lifecycle planning method based on a physical decoupling agent model according to claim 1, characterized in that, The specific steps of selecting high-value samples from the candidate pool in S3 are as follows: a hybrid query strategy is adopted. In each round of active learning iteration, a preset proportion of samples with the highest comprehensive score are selected as utilization samples, and the remaining proportion of samples are randomly extracted from the top 50% of candidate samples with the highest comprehensive score as exploration samples. The selected samples are then input into the bi-layer programming model for high-fidelity labeling.

9. A microgrid lifecycle planning method based on a physical decoupling surrogate model according to claim 1, characterized in that, The online correction of the hybrid oracle in S4 specifically involves periodically extracting the cost minimum point, carbon emission minimum point, and compromise knee point from the current non-dominated solution set as key solutions during the evolution process. The key solution is input into the bi-level programming model in step S1 for verification to obtain the true performance index. If the relative error between the predicted value and the true value of the surrogate model exceeds the preset threshold, the key solution and its true value are added to the training set to incrementally update the surrogate model.

10. A microgrid lifecycle planning method based on a physical decoupling surrogate model according to claim 1, characterized in that, The multi-objective evolutionary algorithm with integer encoding in S4 directly maps the number of installed devices to integer genes. During fitness evaluation, if the load shedding or abandoned power predicted by the surrogate model violates the physical constraints, the individual is marked as an infeasible solution or a high penalty is imposed. Otherwise, the physical components are recombined into the total life cycle cost and total life cycle carbon emissions.

Images